POLARISATION: APPLICATIONS IN REMOTE SENSING

This page intentionally left blank

Polarisation
Applications in Remote Sensing

S. R. CLOUDE

1
3
Great Clarendon Street, Oxford OX2 6DP
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York
© S. R. Cloude 2010
The moral rights of the author have been asserted
Database right Oxford University Press (maker)
First Edition 2010
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose the same condition on any acquirer
British Library Cataloguing in Publication Data
Data available
Library of Congress Cataloging in Publication Data
Cloude, Shane.
Polarisation : applications in remote sensing / S.R. Cloude.
p. cm.
ISBN 978–0–19–956973–1 (hardback)
1. Electromagnetic waves—Scattering. 2. Polarimetric remote sensing.
3. Interferometry. I. Title.
QC665.S3C56 2009
539.2—dc22
2009026998
Typeset by Newgen Imaging Systems (P) Ltd., Chennai, India
Printed in Great Britain
on acid-free paper by
CPI Antony Rowe, Chippenham, Wiltshire

ISBN: 978–0–19–956973–1 (Hbk.)

10 9 8 7 6 5 4 3 2 1
Preface

An alternative title considered for this book was Which Way is Up? Questions
and Answers in Polarisation Algebra. On advice it was rejected in favour of a
more conventional approach. Still, it is a good question. Which way is up? A
question with a literal scientific interpretation—namely, how to define vertical
in a free reference frame for electromagnetic waves, but also one with a col-
loquial interpretation about the best route to progress. At a technical level this
book is concerned with the answer to the former, but hopefully will serve to
promote in the reader some idea of the latter. It arises from over twenty years’
personal experience of research in the topic, but also through the privilege of
having met and collaborated with many of those who made fundamental con-
tributions to the subject. Much of this original work remains, unfortunately,
scattered in the research literature over different years and journals. This book,
then, is an attempt to bring it all together in a didactic and coherent form suitable
for a wider readership.
The book aims to combine—I believe for the first time—the topics of wave
polarisation and radar interferometry, and to highlight important developments
in their fusion: polarimetric interferometry. Here indeed we shall see that the
whole is greater than the sum of the parts, and that by combining the two we
open up new possibilities for remote sensing applications.
It is intended as a graduate level text suitable for a two-semester course for
those working with radar remote sensing in whatever context, but is also aimed
at working scientists and engineers in the broad church that is remote sensing.
Hopefully it will also appeal to those working in optical physics—especially
polarimetry and light scattering—and to mathematicians interested in aspects
of polarisation algebra.
Before reviewing the structure of the book, certain spelling requires clarifi-
cation. Polarisation or Polarization? The usual response is that British English
uses ‘s’, and American ‘z’. However, in this text we reserve spelling with ‘s’ for
the property of a transverse wave, while we use ‘z’ for the effect of electromag-
netic fields on matter. Hence waves remain polarised while matter is polarized.
In this way we take advantage of both forms.
Chapter 1 first provides an introduction to the physical properties of polarised
waves using the formal machinery of electromagnetic wave theory. The idea is
to provide motivation and a foundation for many concepts used in later chapters.
For example, the concepts of matrix decomposition, the use of the Pauli matrices
in wave propagation and scattering and, most importantly of all, the idea of
using unitary matrices to form a bridge between mathematical descriptions of
polarisation in terms of complex and real numbers, are all introduced in this
chapter. This is in addition to the more prosaic elements of polarisation theory,
such as the polarisation ellipse, the Stokes vector, and the Poincaré sphere, all
of which are covered. The chapter is organized around three main themes: how
vi Preface

to generate polarised waves and describe them in various coordinate systems,

how to represent the propagation of such waves between two points A and
B, and finally how to describe their interaction with particles via the process
of scattering. The idea throughout is to develop the concept of the ‘memory’
imprinted on a wave of its original polarisation and how this may be lost through
the complexities of propagation and scattering.
This idea of ‘loss of memory’ is developed further in Chapter 2, where
stochastic effects are treated in more detail. We start by considering the
coherency matrix of a wave and show how it leads to the wave dichotomy;
namely, two different ways in which to model the loss of polarisation informa-
tion to noise. This then opens up a new approach to describing the effects of
noise, not just on a freely propagating wave but also on a scattering system as a
whole via the concept of scattering entropy. Entropy is an important concept in
this book and here we show how entropy from a generalized coherency matrix
description can be formally linked to the classical Mueller/Stokes formulation.
This leads, for example, to a formal test for isolating the set of physical Mueller
matrices from the much wider set of 4 × 4 real matrices—something which
is quite difficult to do from the Mueller calculus itself. We also show how
the entropy concept can be applied to multiple dimensions, including general
bistatic or forward scattering, so freeing it from the important but special case
of backscatter widely used in radar.
Chapter 3 was in many ways one of the most difficult to write. Here we
attempt to apply the ideas of entropy to electromagnetic models of surface
and volume scattering (where polarization becomes important). What makes it
difficult is the sheer scope of the problem. There are so many such models that
they perhaps deserve a whole book to themselves. Instead we concentrate on
a few simple models to convey the key ideas, and also link to developments
in later chapters on decomposition theory and interferometry. Given that the
main application of this book is to microwave scattering, we further concentrate
on low-frequency models, whereby the wavelength is quite large compared to
the size of the scattering feature, which has the further advantage that closed-
form analytic formulae are available to calculate, for example, the scattering
entropy. Having discussed this, we provide some treatment of high-frequency
models and how they differ in polarisation properties from the low-frequency
approach.
Chapter 4 deals with the important new topic of decomposition theorems.
These now have widespread application in microwave remote sensing, and
basically seek to isolate or separate various contributions in a mixture of scat-
tering processes. The most important such idea is to separate surface from
volume scattering. Microwaves have the ability to penetrate vegetation and
other land cover (snow, ice, and so on) and thus generally incorporate a
complicated mixture of processes in the scattered signal. Decomposition the-
orems are an attempt to separate these and hence improve interpretation and
parameter retrieval in quantitative remote sensing applications. There are two
basic classes of decomposition–coherent and incoherent–and within each class
several authors have proposed different models. Here we provide a unified
survey of all such methods and illustrate their various strengths and weak-
nesses by linking their physical structure to the ideas developed in earlier
chapters.
 Preface vii

One key conclusion we will see from the first four chapters is that entropy or
‘loss of memory’ about polarisation is often linked directly to the randomness
of the scattering medium, and that the remote sensing ‘observer’ has little con-
trol over this. This is problematic for applications, for example, in vegetation
remote sensing, where randomness in the volume leads to loss of polarisation
information. A key idea for the second part of the book is therefore how to
achieve some kind of entropy control in remote sensing of random media. One
way to do this is to employ interferometry. Radar interferometry is a mature
established topic, so in Chapter 5 we provide only a brief introduction for those
not familiar with the key concepts. However, the chapter also contains one or
two novel developments required in later chapters. In particular we develop a
Fourier–Legendre series approach to a description of coherent volume scatter-
ing in interferometry. This then provides a bridge between the two halves of the
book, and allows us to consider, in Chapter 6, the combination of polarisation
diversity with interferometry.
The combination of polarisation diversity with radar interferometry has been
a key development over the past decade. It was first made possible from an
experimental point of view by late additions to the NASA Shuttle imaging radar
mission SIR-C in 1994, and since then has evolved through a combination of
theoretical studies and airborne radar experiments. In Chapter 6 we outline
the basic theory of the topic, showing how to form interferograms in different
polarisation channels before considering mathematically the idea of coherence
optimization, whereby we seek the polarisation that maximizes the coherence
(or minimizes the entropy). In this way we provide a link with earlier chapters by
showing how polarimetric interferometry leads to a form of ‘entropy control’,
even in random media applications.
In Chapter 7 we therefore revisit the ideas of surface and volume scattering
first introduced in Chapter 4, but this time we investigate their properties in both
interferometry and polarimetry. This is built around the idea of a coherence loci,
a geometrical construct to bound the variation of interferometric coherence with
polarisation, and closely related to the coherence region, the latter taking into
account spread due to statistical estimation of coherence from data. Given the
importance of surface/volume decompositions in microwave remote sensing,
we treat in some detail the two-layer scattering problem of a volume layer on
top of a surface and use it to review several model variations that are found in
the literature.
In Chapter 8 we use these ideas to investigate the inverse problem: the esti-
mation of model parameters from observed scattering data. We concentrate on
the two-layer geometry and investigate four classes of problem. We start with
the simplest: estimation of the lower bounding surface position, which is a
basic extension of conventional interferometry and allows us, for example, to
locate surface position beneath vegetation and hence remove a problem called
vegetation bias in digital elevation models (DEMs). We then look at estimating
the top of the layer, which corresponds in vegetation terms to finding forest
height. This is an important parameter for estimating forest biomass, for exam-
ple, and in assessing the amount of carbon stored in above-ground vegetation.
We then look at the possibility of imaging a hidden layer using polarimetric
interferometry. In this case we wish to filter out the scattering from a volume
layer to image a surface beneath. The next logical step is to image the vertical
viii Preface

variation of scattering through the layer itself, and this we treat as the topic
of polarisation coherence tomography or PCT, which combines the Fourier–
Legendre expansion of coherent volume scattering with decomposition theory
in an interesting example of what can happen when two of the major themes of
this book—polarisation and interferometry—are fused.
Finally, in Chapter 9 we turn attention to illustrative examples of these theo-
retical concepts. By far the most important current application area is in radar
imaging or synthetic aperture radar (SAR), and so we begin by reviewing
the basic concepts behind this technology, always highlighting those issues
of particular importance to polarisation. We treat a hierarchy of such imaging
systems, from SAR to POLSAR and POLInSAR, and then consider illustra-
tive current applications in surface, volume, and combined surface and volume
scattering.
We then present supportive material in three Appendices. In the first we
provide a basic introduction to matrix algebra. This is used extensively in
descriptions of polarised wave scattering, and is provided here to help those
not familiar with the terminology and notation employed.
As mentioned earlier, one key idea in this book is the role played by uni-
tary matrix transformations in linking (or mapping) different representations
of polarisation algebra. For this reason, in Appendix 2 we provide a detailed
mathematical treatment of the algebra behind such relationships, introducing
concepts from Lie algebra, group theory, and matrix transformations to illus-
trate the fundamental relationships between complex and real representations
of polarised wave scattering.
Finally, in Appendix 3 we provide a short treatment of stochastic signal
theory as it relates to polarisation and interferometry. Here we treat aspects of
speckle noise in coherent imaging, and show how estimation errors impact on
estimation of scattered field parameters in remote sensing.
This book is the culmination of many years of study and research, and
acknowledgement must be given to those many colleagues and students who
provided the impetus and curiosity to study and develop these topics. Acknowl-
edgements and thanks are extended to the European Microwave Scattering
Laboratory (EMSL) at Ispra, Italy, for their permission to use data from their
large anechoic chamber facility; to the German Aerospace Centre (DLR) in
Oberpfaffenhofen, Germany, for provision of airborne radar data from their
E-SAR system; and to Michael Mishchenko of NASA Goddard Space Center,
USA, for provision of his latest numerical simulations of multiple scattering
from particle clouds. Thanks also to the Japanese Space Agency (JAXA) for
provision of the PALSAR satellite data used in Chapter 9. All these datasets
play a vital role in illustrating the theory outlined in this book, and, I believe,
help enormously in clarifying what would otherwise remain abstract concepts.
Key personal thanks go to five colleagues in particular. Firstly, to Profes-
sor Wolfgang Boerner of the University of Illinois, Chicago, USA. His early
vision and boundless energy have inspired several generations of researchers
in these topics, including my own early studies as a PhD student. Secondly,
thanks to Professor Eric Potter of the University of Rennes, France. Our early
collaboration on radar polarimetry, and particularly on decomposition theory,
was inspiring, and has lead, I am pleased to say, to a lifelong friendship and
collaboration. Thanks also to Drs Irena Hajnsek and Kostas Papathanassiou of
the German Aerospace Centre, DLR. Their support and their contributions to
 Preface ix

the development of polarimetric interferometry have been key in the maturation

of the subject. Finally, however, I would like to acknowledge the late Dr. Ernst
Luneburg of DLR, Germany. His combination of scholarship and passion for
the application of mathematics to remote sensing was the true inspiration for
me to write this book, and I feel I can now finally answer his oft-posed question:
‘Wo ist das Buch?’
Shane Cloude
January 2009
This page intentionally left blank
Contents

1 Polarised electromagnetic waves 1

1.1 The generation of polarised waves 3
1.2 The propagation of polarised waves 9
1.3 The geometry of polarised waves 34
1.4 The scattering of polarised waves 46
1.5 Geometry of the scattering matrix 60
1.6 The scattering vector formulation 67

2 Depolarisation and scattering entropy 71

2.1 The wave coherency matrix 72
2.2 The Mueller matrix 78
2.3 The scattering coherency matrix formulation 85
2.4 General theory of scattering entropy 91
2.5 Characterization of depolarising systems 110
2.6 Relating the Stokes/Mueller and coherency matrix
formulations 113

3 Depolarisation in surface and volume scattering 115

3.1 Introduction to surface scattering 116
3.2 Surface depolarisation 133
3.3 Introduction to volume scattering 142
3.4 Depolarisation in volume scattering 155
3.5 Simple physical models for volume scattering and
propagation 167

4 Decomposition theorems 178

4.1 Coherent decomposition theorems 178
4.2 Incoherent decomposition theorems 189

5 Introduction to radar interferometry 208

5.1 Radar interferometry 208
5.2 Sources of interferometric decorrelation 219

6 Polarimetric interferometry 234

6.1 Vector formulation of radar interferometry 234
6.2 Coherence optimization 240
xii Contents

7 The coherence of surface and volume scattering 252

7.1 Coherence loci for surface scattering 253
7.2 Coherence loci for random volume scattering 255
7.3 The coherence loci for a two-layer
scattering model 265
7.4 Important special cases: RVOG, IWCM
and OVOG 270

8 Parameter estimation using polarimetric

interferometry 284
8.1 Surface topography estimation 284
8.2 Estimation of height hv 295
8.3 Hidden surface/target imaging 315
8.4 Structure estimation: extinction and Legendre
parameters 322

9 Applications of polarimetry and interferometry 340

9.1 Radar imaging 340
9.2 Imaging interferometry: InSAR 345
9.3 Polarimetric synthetic aperture radar (POLSAR) 347
9.4 Polarimetric SAR interferometry (POLInSAR) 360
9.5 Applications of polarimetry and interferometry 363

Appendix 1 Introduction to matrix algebra 401

Appendix 2 Unitary and rotation groups 411

Appendix 3 Coherent stochastic signal analysis 425

Bibliography 435

Index 451
Polarised
electromagnetic waves 1
The term ‘wave polarisation’ is relatively recent in the history of optics. It was
first used by Étienne Malus (1775–1812) in 1809, although the ‘orientabil-
ity’ of optical waves was certainly known by Isaac Newton (1643–1727) and
Christiaan Huygens (1629–1695). They were concerned with a description of
the strange phenomenon of double refraction in Iceland spar (calcite), first pre-
sented by Rasmus Bartholin (1625–1698) in 1670, and an explanation was set
to challenge the best minds in optics for the ensuing 150 years. (For an introduc-
tion to the historical importance of polarisation in optics and its role in nature,
see Collet, 1993; Iniesta, 2003; Konnen, 1985.)
It was, however, Thomas Young (1773–1829) who first suggested, in 1817,
that polarisation may arise due to a transverse wave component of light—a
controversial suggestion at the time, but an idea that was further developed
and quantified by Augustin-Jean Fresnel (1788–1827) in 1821, with the devel-
opment of the Fresnel equations for polarisation by surface reflection. This
was followed in 1852 by the development of a mathematical theory of par-
tially polarised waves by George Gabriel Stokes (1819–1903), based on the
concept of a four-element Stokes vector (Stokes, 1852). However, it was
only with the development of the electromagnetic wave theory of James Clerk
Maxwell (1831–1879) in 1861 that light and indeed all electromagnetic waves
were formally shown to be transverse and thus ‘carry a memory of orien-
tation’ in propagating from source to observer (Jones, 1989). The reader
should note, however, that Maxwell’s theory caused some controversy at the
time, and an interesting and readable account of the (sometimes turbulent)
evolution of what we now call Maxwell’s equations can be found in Hunt
(1991).
In this book we concentrate on this orientation ‘memory effect’ and inves-
tigate ways in which it can be used for remote sensing. In a more general
sense, this can be considered a subset of the wider, more formal topic of vector
electromagnetic inverse problems (Boerner, 1981, 1992; Hopcraft, 1992).
In the post-Maxwell era there were four main developments of historical
interest in the description of polarised waves. Firstly we mention the work
of Henri Poincaré (Poincaré, 1892), who formalized many useful concepts in
polarisation optics using a strongly geometrical approach. This was followed
in 1941 by the first use of formal matrix algebra to describe the propagation of
vector waves, by R. Clark Jones of the Polaroid Corporation and Harvard Uni-
versity. At about the same time, Hans Mueller, at the Massachussetts Institute
of Technology, developed a matrix calculus for dealing with partially polarised
2 Polarised electromagnetic waves

Polarised Vector The

Radiation Wave Scattering
Fig. 1.1 A tripartite decomposition of active
remote sensing systems Propagation Matrix

waves. In the radar community, early application of matrix algebra to scatter-

ing was carried out by Edward Kennaugh at Ohio State University. Finally,
the concept of coherency matrices, first developed by Norbert Wiener in 1930,
were first applied to polarisation algebra by Emil Wolf in 1954, and in 1960
Parrant and Roman formally linked polarisation algebra to the density matrix
of statistical quantum mechanics. However, the coherency matrix formulation
has much wider applicability to polarisation algebra than was originally fore-
seen, and in this book we explore this relationship in more detail and provide
an updated treatment of these concepts.
Before treating these advanced topics, however, in this first chapter we use the
machinery of electromagnetic wave theory to consider the basic mechanisms
behind generation, propagation, and scattering of polarised electromagnetic
waves. The formalism so developed will allow us to propagate a wave from the
source to a scattering object and back again, so forming a basic template for the
treatment of active remote sensing systems. Figure 1.1 shows a schematic rep-
resentation of this tripartite decomposition of wave problems. We shall follow
the logical progression of the diagram and begin with a description of the gen-
eration of polarised waves. We start with a general, coordinate free description
based on the vector form of Maxwell’s equations (Chen, 1985) before quickly
focusing on three important coordinate systems, first classified in antenna theory
in Ludwig (1973), and now widely used in analysis, engineering measurements
and physical modelling. From these we can then define the concept of co- and
crosspolarised fields. and ask the basic question as to whether the perfectly
polarised source exists, even theoretically. (For the answer, see equation (1.16)
and subsequent discussion.)
Having described how to generate polarised waves we then introduce vector
wave propagation. This is a major topic in itself, and so in order to quickly
bring forward the main ideas we require later in this book, we proceed by
considering three specific examples. We start with the simplest—wave prop-
agation in homogenous isotropic media—before examining two more exotic
cases, where we will see how the concept of wave orthogonality can be for-
mally defined and the family of polarisation types extended to include elliptical
and circular polarisations.
Finally we consider the complex process of wave scattering, whereby sec-
ondary currents are induced in an object by the incident field and act as new
sources of radiation to transfer information about the scatterer back to the
observer. This process is characterized in the far field (that is, for large sepa-
rations of source and object) for all polarisation states by a complex scattering
amplitude matrix [S], the measurement and analysis of which forms a central
theme of this book.
 1.1 The generation of polarised waves 3

1.1 The generation of polarised waves

1.1.1 Maxwell’s equations and vector plane waves
Electromagnetic waves are generated by accelerating charges (Jackson, 1999;
Cloude, 1995a). The time and space variation of electric and magnetic fields
are governed by a set of four partial differential equations, called Maxwell’s
equations, which can be written succinctly in the form shown on the left in
equation (1.1) (Chen, 1985; Born and Wolf, 1989; Ishimaru 1991):
∂B 
∇ ×E =− 
∂t 


∂D 

D = ε0 E wave equation ∂ 2E ∂J
∇ ×H =J + −−−− − − −→ ∇ × ∇ × E + ε µ = −µ0
∂t  B = µ0 H 0 0
∂t 2 ∂t


∇.B = 0 




∇.D = ρ
(1.1)

It is an interesting consequence of Maxwell’s equations that even vacuum or

free space is characterized by a pair of important constants: the permeability
µ0 and permittivity ε0 , which have values derived from experiment, as shown
in equation (1.2):

µ0 = 4π × 10−7 H /m ε0 = 8.854 × 10−12 F/m (1.2)

Equation (1.1) then relates the radiated vector fields E, B, D, and H to source
vector currents J and scalar charge density ρ. The explicit differential equations
relating these quantities can then be derived from equation (1.1) by treating the
∇ operator as a vector of partial derivative operations and using the following
results from linear algebra:
 
  i j k 
∂ ∂ ∂ 
∇= , , a × b = ax ay az  (1.3)
∂x ∂y ∂z bx by bz 

The cross-product of ∇ with a vector is called the ‘curl’ (or sometimes ‘rot’ for
rotor) operator, and the dot product the divergence or ‘div’. In this book we
shall be primarily concerned with the ‘memory’ these fields have for the vector
nature of their source (that is, its orientation in space and structure in time), and
how this may be used for remote sensing purposes.
The vector currents J and scalar charges ρ are sources of the fields in equation
(1.1). To demonstrate these as equivalent to a time derivative of current, we
generate a vector wave equation by first forming a secondary vector product
as ∇ × ∇ × E and then using the ∇ × H Maxwell equation plus constitutive
relations to eliminate B. The result is shown on the right-hand side of equation
(1.1). Note that on the left of this equation we have mixed second time and space
derivatives of the electric field vector, while on the right we have the source of
these fields, localized as the time derivative of vector currents. As current itself
is caused by the time derivative of charge, it follows that radiation is caused by
the second time derivative or charge acceleration. This acceleration is a vector
quantity, the orientation of which is transferred into the radiated fields in the
4 Polarised electromagnetic waves

form of propagating waves. While the form of these waves can be very general
(see Cloude, 1995a), it is useful to start with a special type of solution: namely,
vector plane waves.
For plane wave solutions we postulate electric field and driving current vec-
tors of the form shown in equation (1.4). By adopting these simple plane wave
solutions the space and time derivatives take on the simplified form shown on
the right-hand side of this equation:

 
i ωt−β.r  ∂
≡ iω
E = E0e ∂t
⇒ (1.4)
J = J 0 eiωt  ∇ ≡ −iβ

Where ω = 2πf and β = 2π/λ, f is the frequency of the √ wave in Hertz, λ its
wavelength in metres, and throughout this text we set i = −1. Note that our
notation, with a positive sign for the time derivative, is chosen by convention
and leads to a complex refractive index for lossy material with a negative
imaginary part (see Section 3.1.1.1). Be aware, however, that other notations
exist in the literature, with some authors choosing E* for the plane wave, which
changes the sign of the time derivative and leads to a complex refractive index
with positive imaginary part (and which also impacts on the sense of circular
polarisations, as we shall see). By direct substitution
  we then find that the vector
 
wave equation in (1.1) has a solution when β  = β = ω2 ε0 µ0 , and we obtain
a vector Helmholtz equation of the form shown in equation (1.5), where we
have now eliminated explicit time dependence.

∇ × ∇ × E − β 2 E = iωµo J (1.5)

The importance of such simple vector plane waves follows from the linearity of
Maxwell’s equations since, by superposition, the field at any location x can then
be obtained as a sum of contributions from all the source currents at locations
y. Hence we can express the solution of equation (1.5) as an integral or sum of
the form shown in equation (1.6) (Chen, 1985).

E(x) = iωµ0 G(x, y).J (y)dV (1.6)

V

The propagator of vectors from y to x is termed the dyadic Green’s function,

and by formally solving the vector Helmholtz equation for a Dirac delta source
(a point source in space) it can be shown to have the following general form
(Chen, 1985):

 i 1
G(x, y) = I − r r g + (I − 3r r)g − 2 2 (I − 3r r)g (1.7)
kR k R

Here I is the 3 × 3 unit dyad and has the form of a unit matrix, R = |x − y|,
r = (x − y)/|x − y|, and the scalar Green’s function g accounts for causality
 1.1 The generation of polarised waves 5

and energy conservation as shown in equation (1.8):

 
 
−iβ x−y
e
g(x, y) =   (1.8)
 
4π x − y

As we move further away from the source currents, then R → ∞ and the
first term of G dominates. Hence in the far-field, the dyadic Green’s function
simplifies by definition to the following form:

 e−iβR
G∞ = I − r r (1.9)
4πR

The first part of this expression shows that only the components of J transverse
to the direction of propagation r contribute to the radiated field. It follows from
this that the radiated fields are transverse to the direction of propagation of the
wave (called transverse electromagnetic or TEM waves). The electric field is
defined from an integral sum over all currents, but the resultant must always
lie in a plane perpendicular to r. This is called the plane of polarisation and
the resultant time locus of the electric field in this plane, the polarisation of the
radiated wave.
To illustrate this, consider the fields radiated by elementary dipoles. In
electromagnetic theory there are two types to consider: electric and magnetic
(Jackson, 1999). For the electric dipole, current is localized at the origin, and
an electric dipole moment p0 generates an effective current distribution of the
form shown in equation (1.10). Now evaluating the integral using the far field
Green’s dyadic (equations (1.6) and (1.9)), we obtain the fields radiated by
the dipole as shown in equation (1.10). In the far field, all components have
the structure of transverse electromagnetic (TEM) waves for which the electric
and magnetic field amplitudes are related by the free space wave impedance
Zo ≈ 377  as shown. Note, for example, that the radiation in the direction
r = p 0 is zero (the cross-product is zero), producing the characteristic dumb-
bell radiation pattern. The radiated magnetic field vector can always be derived
from the electric field as shown.

J (r) = iωp0 δ(r)

β 2 e−iβR  β 2 e−iβR 
⇒ E(r) = I − r r .p0 = r × r × p0
po 4π ε0 R 4π ε0 R
µo
TEM waves r × E(r) = H (r) = Zo H (r)
εo
βωe−iβR 
⇒ H (r) = r × p0
4πR
(1.10)

A magnetic dipole, on the other hand, can be generated by a small loop carrying
a uniform current I. The magnetic dipole moment m is then defined from the
product of current and loop area, and is a vector normal to the plane of the loop,
6 Polarised electromagnetic waves

as shown in equation (1.11).

β 2 e−iβR
m H (r) = r × (r × m)
4π µ0 R (1.11)
βωe−iβR
E(r) = − (r × m)
4π R

By treating the time variation of loop current I as an equivalent magnetic cur-

rent source in a symmetrized version of Maxwell’s equations, the radiated fields
can be obtained directly from those of the electric dipole using a duality trans-
formation (Baum, 1995). This symmetry in the equations of (1.1) is useful,
as it permits solution of a completely different ‘dual’ problem to the original
without the need for recalculation. Radiation by electric and magnetic dipoles
is an example of such dual problems.
The corresponding fields radiated by a magnetic dipole are shown in equation
(1.11), where we see that the E and H fields have been interchanged by the
duality transformation, but that the structure of the fields is again due to vector
cross and triple products. We shall use these results to formulate scattering by
small chiral or handed particles (like a helix), where currents flow in both linear
and circular components, in Section 3.3.

1.1.2 Polarised wave coordinate systems

So far our treatment has avoided reference to any specific coordinate system,
but in practice the radiation and scattering of waves is projected onto coordi-
nates relevant to the problem at hand. Hence one is faced with the problem of
choosing the best coordinate system. One reason why this choice is so important
is because we very often want to set up currents J on an antenna system so that
the radiated wave in the far field has a well-defined orientation or polarisation.
However, depending on the coordinates chosen we may find that in some direc-
tions the polarisation has components orthogonal to that desired. This is termed
crosspolarisation, and in radiation problems is normally undesirable (Collin,
1985). In scattering, on the other hand, it can be useful for identification of
the orientation of the induced currents on the scatterer. To illustrate the prob-
lems involved in defining crosspolarisation, we outline three commonly used
coordinate systems first derived in Ludwig (1973).

1.1.2.1 System I: Cartesian coordinates

This coordinate system is commonly used to describe wave propagation in a
paraxial approximation or where there is one well-defined direction. It is defined
Plane of in terms of a right-handed triplet of unit vectors i, j, and k such that the direction
j Polarisation of propagation k = i × j, as shown in Figure 1.2. This system can be related to
spherical polar coordinates (System II) by transformation equations, as shown
i in equation (1.12):

Plane Wave i = sin θ cos φr + cos θ cos φθ − sin φφ

k Propagation Direction
i× j = k j = sin θ sin φr + cos θ sin φθ + cos φφ (1.12)
Fig. 1.2 Ludwig system I: Cartesian k = cos θr − sin θ θ
coordinates
 1.1 The generation of polarised waves 7

φ
k

θ θ = cosθ cosφ i +θ sinφ j − sinθk

θ r = sinθ cosφ i + sinθ sinφ j + cosθk

j
φ = −sinφ i − cosφ j
φ
Fig. 1.3 Ludwig system II: spherical polar
i coordinates

For example, consider radiation by an elementary horizontal dipole antenna

with dipole moment p 0 = pi. The radiated electric field Cartesian components
can then be obtained from equation (1.10) as shown in equation (1.13):

pe−iβz  −pβ 2 e−iβz

E= β × β ×i = i (1.13)
4π ε0 z 4π ε0 z

This is polarised in the same direction as the antenna current vector. In this way
we can consider the EM wave as transferring a ‘memory’ of the orientation of
the dipole source into the far field, with zero crosspolarisation. Such a con-
venient result does not, however, apply in all coordinate systems, as we now
demonstrate.

1.1.2.2 System II: spherical polar coordinates

In theoretical considerations of the radiation and scattering of waves in three-
dimensional space, spherical polar coordinates are widely used. Here we can
locate a source or scatterer at the origin and consider the fields in the surround-
ing three-dimensional space, as shown in Figure 1.3. The wave propagation
direction is then associated with the r unit vector, and the transverse plane
formed by the θ and φ unit vectors generates the plane of polarisation of the
wave. Figure 1.3 shows how these two unit vectors can be specified by two
angles and related to a local Cartesian system. Again considering an elemental
x-directed dipole at the origin, we now obtain the radiated field components as
shown in equation (1.14):

pβ 2 e−iβR  pβ 2 e−iβR 
E= r× r×i = cos θ cos φθ − sin φφ (1.14)
4π ε0 R 4π ε0 R

Here we see that although our source has a well-defined orientation, the radi-
ated field has components that vary with direction and hence are not so neatly
constrained as in the Cartesian case. Although providing a convenient general
format for three-dimensional radiated fields, the spherical polar system is not the
only choice for describing general polarised systems. An alternative, favoured
in the antenna measurement community, is based on a hybrid combination of
Cartesian and polar concepts, considered as follows.

1.1.2.3 System III: hybrid measurement system

Although the Cartesian and spherical polar systems are convenient for the-
oretical analyses, in practice antenna patterns and scattering diagrams are
referenced to a third coordinate system formed as a hybrid of these two. The
key idea here is to define the polarisation unit vectors as Cartesian components i
8 Polarised electromagnetic waves

z
a y = sinφ θ + cosφ φ
θ a x = cosφ θ − sinφ φ

y
φ
Fig. 1.4 Ludwig system III: hybrid measure-
ments coordinates x

and j, but then to permit three-dimensional field structures by allowing parallel

transport of these unit vectors according to spherical polar angles (with the
source antenna or scatterer located at the origin). Figure 1.4 shows a schematic
of this system. It is clear from the geometry of this transport process that the
unit vectors ax and ay are generated by the spherical angle φ, as shown in
Figure 1.4.
Returning to our example of radiation by an x-directed dipole, we can now
establish a systematic method for calculating the level of crosspolarisation
radiated by projecting the field in spherical polar coordinates onto the ax ay
system. The desired copolarised field is then by definition the ax component,
while ay is the crosspolarised field. By direct calculation we have the following
results:

pβ 2 e−iβR
copolarfield = E.ax = (cos θ cos2 φ + sin2 φ)
4π ε0 R
(1.15)
pβ 2 e−iβR
crosspolarfield = E.ay = sin φ cos φ(cos θ − 1)
4π ε0 R

Note that in the principal planes (when θ and φ are zero) there is zero crosspolar-
isation. However, for radiation in other directions the ratio of cross- to copolar
fields (the XPOL ratio) is given by equation (1.16), which can rise to a maximum
of −15 dB when φ = θ = π/4.
 
 sin φ cos φ(cos θ − 1) 
XPOL = 20 log10   (1.16)
cos θ cos2 φ + sin2 φ 

This level is often too high for radar and communication applications, and hence
more sophisticated antennas with even lower crosspolarisation have been devel-
oped. To illustrate how such a low crosspolar antenna might be constructed,
consider the case of radiation by a Huygens source (Collin, 1985). This can be
considered a ’patch’ of a plane wave. According to Huygens’ principle, such
a patch radiates elementary secondary wavelets, the superposition of which
marks the advance of the wave front. Figure 1.5 shows such a patch of plane
wave of square dimension 2a, where the fields are constant across the aperture
and zero elsewhere.
The field radiated by such a structure can be obtained from Maxwell’s equa-
tions by employing equivalent electric and magnetic currents Jes , Jms in the
aperture (Collin, 1985; Cloude, 1995a). These are defined from the transverse
components of the field, as shown in equation (1.5). The radiation is then defined
by the expression shown. Note that with a distributed current source such as
 1.2 The propagation of polarised waves 9

J es = n × H J ms = n × E
Ex
Es = (1 + cosθ ) f (cosφθ − sin φ φ )
e –ibR
R
Hy sin(b sin θ cos φ a) sin( bsin θ sin φ a)
f=
bsin θ cos φ a b sin θ sin φ a
Fig. 1.5 Radiation by a Huygens patch: the
Huygens Source ideal zero cross-polarised source

this, the radiation integral can be explicitly evaluated and produces a Fourier
Transform relation between the aperture distribution and the far field. In this
case the rectangular distribution gives rise to a SINC function. However, for
our purposes, interest centres more on the polarisation properties of the radiated
field.
From the polarisation point of view we observe a very interesting result. The
radiation from this ‘aperture antenna’has zero crosspolarisation in all directions.
This shows that in theory, low crosspolarisation can be obtained, although
in practice securing the right kind of symmetric aperture distribution can be
difficult to engineer, especially over a broad band of frequencies (Collin, 1985;
Mott, 1992).
Having established the influence of coordinate systems on the definition
of co- and crosspolarised waves in free space, we now turn to consider the
propagation of waves in more complex environments. In particular, we consider
constraints posed by the presence of the medium on the allowed polarisation
states of the propagating field, and thus establish a calculus for dealing with
the distortion of the ‘memory’ effect in the transfer of orientation information
from source to far field.

1.2 The propagation of polarised waves

In the absence of sources, waves propagate according to an homogeneous form
of Maxwell’s equations, as shown on the left in equation (1.17). Complexity
now arises in the way in which the presence of material matter influences the
way in which the wave can propagate. In this section we consider unbounded
wave propagation in each of three special cases: isotropic, when ε and µ are
scalar quantities; anisotropic, when ε becomes a tensor or matrix; and chiral
materials, where electric and magnetic effects are coupled in the material by
helical current flow (Kong, 1985). Without loss of generality we employ the
Cartesian coordinate system I with propagation in the +z-direction.

Case I Case II Case III


∂B
∇ ×E =−  

D = εr ε0 E D = εE D = εE + ηB
∂t
⇒ B = µ0 H B = µ0 H H = γ E + µ−1
0 B
∂D 

∇ ×H = Isotropic material Anisotropic material Chiral material
∂t
(1.17)

As when considering the radiated field, we first generate a set of vector homoge-
neous wave equations from the Maxwell curl equations. The resulting systems
10 Polarised electromagnetic waves

for each of the three classes of material are shown in equation (1.18).

∂ 2E
∇ × ∇ × E + εµ0=0 Case I
∂t 2
∂ 2E
∇ × ∇ × E + εµ0 2 = 0 Case II (1.18)
∂t
∂ ∂ 2E
∇ × ∇ × E − µ0 (η + γ ) ∇ × E + εµ0 2 = 0 Case III
∂t ∂t

We set µ = µ0 for applications of interest in remote sensing (That is, we ignore

variations in magnetic properties of materials). Also, we postulate vector plane
wave solutions propagating in the +z direction of the general form shown in
equation (1.19):

E = ei(ωt−βz) (ex i + ey j + ez k) = E 0 ei(ωt−βz) (1.19)

With these two assumptions we can simplify the homogeneous wave equations
to the general form shown in equation (1.20):

∇ × ∇ × E − ω2 εµ0 E = 0 Case I
∇ × ∇ × E − ω2 εµ0 E = 0 Case II (1.20)
∇ × ∇ × E − iωµ0 (η + γ ) ∇ × E − ω2 εµ0 E = 0 Case III

We now seek conditions on the three complex coefficients ex , ey and ez such

that the plane wave satisfies the vector wave equations in equation (1.20). To
do this we shall make use of the following spatial derivatives of the plane wave
solution in our search for a match:

∇ × E = ei(ωt−βz) iβ(ey i − ex j)
(1.21)
∇ × ∇ × E = ei(ωt−βz) β 2 (ex i + ey j)

1.2.1 Case I: wave propagation in isotropic media and

C2 symmetry
It is one of the unfortunate ambiguities of scientific notation that the word
‘polarisation’ is used to describe both the electric field orientation of plane
waves and also the effect of electric fields on matter. In an attempt to avoid
this ambiguity we establish a notation to spell polarisation with ‘s’ to describe
wave properties and with ‘z’to describe material interactions. Material therefore
becomes polarized, while a wave is polarised.
In the simplest case, material becomes polarized by the scalar amplitude of
an electric field, and the influence of the material on the wave is then determined
by the dielectric constant ε, which in general is a complex scalar. Under these
circumstances the wave equation also becomes simplified, and has the form of
 1.2 The propagation of polarised waves 11

a vector Helmholtz equation, as shown in equation (1.22):

D = εE → ∇ × ∇ × E − ω2 µεE = 0 (1.22)

In order for the plane wave to be a solution, its components must then satisfy
the following equation set obtained by explicit evaluation of equation (1.22):
     
ex ex 0
β 2 ey  − ω2 εµ0 ey  = 0 ⇒ β 2 = ω2 εµ0 = β0 n ⇒
0 ez 0
1 c c
f =√ √ =√ = (1.23)
εr ε0 µ0 λ εr λ nλ

Note that ez = 0; that is, these plane wave solutions represent transverse elec-
tromagnetic (TEM) waves. These waves are also non-dispersive; that is, they
all propagate with the same phase velocity, itself determined from the free
space velocity c = 2.997 × 108 m/s, and the refractive index n of the medium,
which is related to the square root of the dielectric constant εr , as shown in
equation (1.23).
These constraints do not specify ex and ey . In fact, any complex pair will
satisfy the wave equation. This we call a C2 symmetry, in that any element of
a two-dimensional complex space is a solution. Note, however, that the pair
(ex , ey ) are independent of time and space, and therefore represent a spatio-
temporal invariant of the wave. They define the polarisation of the plane wave.
Since their resultant always lies in the xy plane transverse to the propagation
direction, this is now called the plane of polarisation of the wave.
Without loss of generality we can write the pair as a column vector in C2—
the space of two-dimensional complex numbers—as shown in equation (1.24),
where m is the amplitude of the wave, and the trigonometric factors arise directly
from the requirement that w itself has unit amplitude, or is unitary.
      
ex  2 cos αw eiφx cos αw eiφx
E0 = = |ex | + ey
2   =m = mw
ey sin αw eiφy sin αw eiφy
(1.24)

We often wish to compare waves with the same amplitude, and therefore set
m = 1. In this case the column vector is unitary and has three free parameters.
Importantly, each choice of unitary vector w then defines a new class of vectors
w⊥ , being orthogonal to the first. As is conventional for complex vector spaces,
orthogonality is based on the Hermitian inner product of column vectors, as
shown in equation (1.25):
   
cos αw eiφx ∗T iχ − sin αw e
iφx π
w= ⇒ w ⊥ · w = 0 ⇒ w ⊥ = e ⇒ α⊥ = αw +
sin αw eiφy cos αw eiφy 2
(1.25)

We see that the orthogonal state is not uniquely defined. There is a phase angle
χ left undetermined from w by the combined Hermitian and unitary constraints.
This problem can be resolved by considering how the pair w and w⊥ are to be
combined to provide a coordinate system or polarisation basis or frame for the
representation of arbitrary wave states.
12 Polarised electromagnetic waves

To find the components of an arbitrary vector E in terms of the unitary states

w and w⊥ we form a 2 × 2 transformation matrix through projections, with the
unitary vectors as columns, as shown in equation (1.26):
 
  cos αw eiφx − sin αw ei(φx +χ )
E = w w⊥ ·E = · E = [U ] · E (1.26)
sin αw eiφy cos αw ei(φy +χ )

We must still deal with the free parameter χ . One way to resolve this issue is
to force the matrix U to be special unitary; that is, to have unit determinant.
This not only establishes a consistent method for change of base but, as shown
in Appendix 2, links directly via group theory to the geometry of the real space
of the Poincaré sphere and Stokes vector. With this added condition we obtain
the following constraint equation for χ :

Det(U ) = 1 ⇒ φx + φy + χ = 0 ⇒ χ = −(φx + φy ) (1.27)

Consequently the general special unitary change of base matrix can be written
as shown in equation (1.28):
 
cos αw eiφx − sin αw e−iφy
[U2 ] = (1.28)
sin αw eiφy cos αw e−iφx

Hence we can summarize by saying that if we find a solution to the wave

equation E in isotropic material, then there is an infinite set of other solu-
tions generated by the relation [U2 ]E. This is a formal representation of the
C2 freedom we spoke of in equation (1.23). We see that the properties of spe-
cial unitary matrices are central to the development of polarimetry theory, and
a general description of the properties of such complex matrices is given in
Appendix 2. We shall make extensive use of this 2 × 2 change of base matrix,
and also higher-dimensional unitary forms, in analytical manipulations involv-
ing polarised waves. To develop [U2 ] we involved the idea of orthogonality
of complex vectors. In this case it was a mathematical convenience in order
to develop a frame or coordinate system. However, orthogonality also arises
naturally in many physical systems, as we now consider.

1.2.2 Case II: wave propagation in anisotropic media

In this more complicated case the orientation of the induced polarization vector
inside the material is no longer parallel to the orientation of the field excitation,
and ε therefore becomes a tensor or matrix. In this case the vector wave equation
assumes a tensor form shown in equation (1.29):

D = ε.E
→ ∇ × ∇ × E − ω2 µ0 .ε · E = 0 (1.29)
B = µ0 H

From energy conservation, ε must be a positive definite (PD) Hermitian tensor

(see Appendix 1), which means that it is always possible to find a coordinate
system inside the material for which the matrix is diagonal (Kong, 1985) and
 1.2 The propagation of polarised waves 13

of the form shown in equation (1.30):

 
εa 0 0
ε = 0 εb 0 0 < εc ≤ εb ≤ εa (1.30)
0 0 εc

Mathematically this is an example of an eigenvalue decomposition, which as

we shall see throughout this book often simplifies the treatment of propagation
and scattering of polarised waves. As the permittivity tensor is PD Hermitian
it has positive real eigenvalues (εa , εb , εc ) and orthogonal eigenvectors, which
define the abc axes of the material. If two of the eigenvalues are equal then
the material is uniaxial, while if all three are distinct then it is biaxial. Such
degeneracy can arise through symmetry, as for example in crystal optics, in
which cubic symmetry gives rise to triple degeneracy and isotropic propagation.
Double degeneracy is found in three crystal groups (tetragonal, hexagonal, and
rhombohedral) which are consequently uniaxial. Again we shall see this theme
arise in more general scattering problems, whereby symmetry in the medium
controls the distribution of eigenvalues of a polarisation matrix.
The abc coordinate system forms what are called the principal axes of the
material, and in general these will not coincide with the xyz of our plane wave
propagation system. However, when they do, analysis of propagation greatly
simplifies, as we now show. In order for our plane wave to be a solution of the
wave equation, the coefficients ex and ey must now satisfy the following matrix
equation:
       
ex ε11 ε12 ε13 ex 0
β 2 ey  − ω2 µ0 ε21 ε22 ε23  . ey  = 0 (1.31)
0 ε31 ε32 ε33 ez 0

This is generally made complicated because the ε tensor is full. In this case
it is more convenient to rewrite equation (1.31) in terms of the electric dis-
placement vector D rather than E. We then obtain the modified form shown in
equation (1.32):

 −1      
ε11 ε12 ε13 dx dx 0
∇ × ∇ × ε21 ε22 ε23  . dy  − ω2 µ0 . dy  = 0 (1.32)
ε31 ε32 ε33 dz dz 0

For plane wave solutions, the vector on the far left of this expression has only x
and y components, from which it follows that dz = 0; that is, that the D vector
(not the E vector) is always transverse to the direction of propagation. For this
reason D is often preferred to the electric field E when describing the polar-
isation of waves in anisotropic media. Now assuming that our external wave
system xyz corresponds to abc we obtain the following simplified dispersion
relation:
      
1
εa 0 dx dx 0
β 2
1 − ω 2 µ0 . = (1.33)
0 εb
dy dy 0
14 Polarised electromagnetic waves

We see that in this case we no longer have the C2 freedom of isotropic mate-
rial, and that for a wave to propagate it must be polarised along the a (x) or
b (y) directions. Furthermore, the velocity of propagation is different for the
two waves—a phenomenon that leads to differential phase shifts between com-
ponents of the wave, and is known as birefringence. Any general polarisation
state can be expressed as a linear mixture of a and b through the basis projec-
tion matrix of equation (1.28). Thus, when a polarisation state is launched at
z = 0 then its a and b components will propagate at different velocities (and
also in general with different extinction rates), and hence as it progresses into
the material it will change its polarisation state. The only exceptions to this
are the states a and b themselves. If they are launched into the material then
they progress without distortion. If we represent the effect of propagation up
to a plane z = z0 as a 2 × 2 complex matrix [Mz0 ] we can write the following
eigenvalue problem:

    
E (z0 ) = Mz0 .E (0) = λE (0) ⇒ Mz0 − λ [I2 ] E(0) = 0 (1.34)

We then see that the states that remain unchanged due to propagation are eigen-
vectors of the matrix [Mz0 ]. Consequently we refer to these as eigenpropagation
states, or simply eigenstates, of the material. We now show how [Mz0 ] can be
related to the electric field wave equation.
Returning to equation (1.31) for the electric field, and now imposing the
constraint that dz = 0, we can remove the ez dependence and obtain a pair of
equations for ex and ey only. The following equation is then obtained for an
arbitrary polarisation state, where in the last step we have expressed the spatial
term as an ordinary derivative with respect to z, itself obtained from integration
of the second derivative appearing from the vector wave equation (assuming
[Kz ] does not depend on z).

 
  |ε13 |2 ε13 ε23    
2 ex
ε11 − ε ε12 −
ε  ex 0
β − ω µ0  
2 · e =
2 33 33
ε ∗ ε∗
ey |ε | y 0
ε12∗ − 13 23
23
ε22 −
ε33 ε33
 
0
⇒ β 2 E − ω2 µ0 [Kz ] .E =
0
d 2E
⇒ = −ω2 µ0 [Kz ] · E
dz 2
dE
⇒ = [N ] · E = −iω µ0 [Kz ] · E (1.35)
dz

The most important part of the above analysis is the derivation of a simple matrix
differential equation governing the propagation of the C2 column vector E in
terms of a differential matrix [N ], which may be easily integrated to obtain the
[M ] matrix at distance z0 , as shown in equation (1.36).

 
dE   z0
= [N ] E ⇒ Mz0 = [M0 ] exp [N ]dz (1.36)
dz 0
 1.2 The propagation of polarised waves 15

If [N ] is constant and we assume [M0 ] = [I2 ], then this simplifies to equation

(1.37):

[Mz ] = e[N ]z (1.37)

where the matrix exponential function can be conveniently defined in terms of

its infinite series expansion as shown in equation (1.38), which is defined under
matrix multiplication for all square matrices [A] (see Appendix 1).

[A]2 z 2 [A]n z n
exp([A]z) = I + [A]z + + ··· + ··· (1.38)
2! n!

We shall now make use of the following six important properties of the
matrix exponential function, where the matrix commutator bracket is defined
as [A, B] = AB − BA. We see from property II that the eigenvectors of [Mz ]
and [N ] are identical, and that the eigenpolarisation states are determined by
the eigenvectors of the reduced dielectric tensor [Kz ] in equation (1.35).

I exp(A) · exp(B) = exp(C)

1 1
⇒C =A+B+ [A, B] + ([A, [A, B]] + [B, [B, A]]) + · · ·
2 12
II exp(SAS −1 ) = S exp(A)S −1
III det(exp(A)) = exp(Tr(A))
IV exp(A)−1 = exp(−A)
d
V exp(Az) = A exp(Az)
dz
d
VI exp(−Az) = − exp(−Az)A (1.39)
dz

In the special case of zero absorption by the material, [Mz ] must be unitary
(norm-preserving). If this is the case then its inverse is just its conjugate trans-
pose, and from property IV it follows that [N ] = i[H ] where [H ] is Hermitian.
If the matrix [Mz ] is special unitary (that is, with unit determinant) then from
property III it follows that the matrix [N ] must also be traceless. Note that we
can always factor a determinant phase term from a unitary propagation matrix
[Mz0 ] to leave a special unitary form, as shown in equation (1.40):
 
e−iβa z0 0
[Mz0 ] = [U2 ] · . [U2 ]∗T
0 e−iβb z0
 (βa −βb ) 
−i(βa +βb ) e−i 2 z0 0
=e 2
z0
[U2 ] · (βa −βb ) z · [U2 ]∗T (1.40)
0 ei 2 0

The determinant phase represents the ‘mean’ propagation constant in the

medium, and the differential terms are all placed inside the special unitary
16 Polarised electromagnetic waves

component. We have already encountered special unitary matrices for change

of base in C2, and now we see that we can also use them to represent propa-
gation in lossless materials using the matrix exponential function as shown in
equation (1.41):

 
cos αw eiφx − sin αw e−iφy
[M2 ] = [U2 ] = = exp(iH )
sin αw eiφy cos αw e−iφx
 
h1 h2 − ih3
⇒ [H ] = (1.41)
h2 + ih3 −h1

where the three coefficients h1 , h2 and h3 are all real. This last result introduces
the idea of matrix decomposition to polarimetry. In principle, we take a com-
plex matrix (such as [M2 ]) and express it as the sum of component matrices,
each of which has some simpler physical interpretation. In this way we can
‘model’ the processes giving rise to the observed matrix in terms of a combina-
tion of elementary physical mechanisms. To see this, note that the matrix [H ]
can be formally expressed as a linear combination of elementary matrices as
follows:

!
3      
1 0 0 1 0 −i
[H ] = hl [σl ] = h1 + h2 + h3 (1.42)
0 −1 1 0 i 0
l=1

The triplet of matrices σ =[σ 1 , σ 2 , σ 3 ] are called the Pauli spin matrices, as
they were first applied to problems of spin in quantum mechanics by Wolfgang
Pauli (1900–1958). More generally, as we shall see, they are useful for decom-
posing classical vector wave scattering problems involving complex matrix
transformations.
Considering each elementary Pauli matrix at a time, we can use the series
expansion of equation (1.38) to derive the corresponding unitary matrices. The
key stage is to derive the square of the elementary matrix, and we note that for
all three Pauli matrices we have the following relation:

 
1 0
σi2 = = σ0 (1.43)
0 1

where we have defined a new element σ 0 as the 2 × 2 matrix identity. Hence

we can generate the mappings shown in equation (1.44) and give each matrix
a simple physical interpretation as follows:

σ1 : represents birefringence between the eigenstates a and b.

σ2 : represents birefringence between eigenstates at ±45◦ to the basis states;
that is, a ± b.
σ3 : represents birefringence between quadrature combinations; that is, a ± ib,
which corresponds, as we can see, to a plane rotation—a result we shall
use extensively in this book.
 1.2 The propagation of polarised waves 17

θ 2 σ12 θ n σ1n
exp(iθ σ1 ) = σ0 + iθ σ1 − + · · · (i)n + ···
2! n!
   
θ2 θ3
= σ0 1 − + · · · + iσ1 θ − + ···
2! 3!
 
cos θ + i sin θ 0
= cos θ σ0 + i sin θ σ1 =
0 cos θ − i sin θ
   iθ   
1 0 e 0 1 0
= . .
0 1 0 e−iθ 0 1

θ 2 σ22 θ n σ2n
exp(iθ σ2 ) = σ0 + iθ σ2 − + · · · (i)n + ···
2! n!
   
θ2 θ3
= σ0 1 − + · · · + iσ2 θ − + ···
2! 3!
  (1.44)
cos θ i sin θ
= cos θ σ0 + i sin θ σ2 =
i sin θ cos θ
   iθ   
1 1 −1 e 0 1 1
= . .
2 1 1 0 e−iθ −1 1

θ 2 σ32 θ n σ3n
exp(iθ σ3 ) = σ0 + iθ σ3 − + · · · (i)n + ···
2! n!
   
θ2 θ3
= σ0 1 − + · · · + iσ3 θ − + ···
2! 3!
 
cos θ sin θ
= cos θ σ0 + i sin θ σ3 =
− sin θ cos θ
   iθ   
1 1 i e 0 1 −i
= · ·
2 i 1 0 e−iθ −i 1

In order to generalize this procedure we need to repeat the series expansion

using the most general [H ] matrix, itself decomposed as a linear combination
of the Pauli matrices. This again requires evaluation of the square of the matrix,
which can now be written as shown in equation (1.45):

[H ]2 = (h1 σ1 + h2 σ2 + h3 σ3 ) . (h1 σ1 + h2 σ2 + h3 σ3 )

= (h21 + h22 + h23 )σ0 = θ 2 σ0 (1.45)

from which we see it is convenient to define the scalar amplitude of the matrix
[H ] as θ and to normalize the vector of coefficients h = θn where n · n = 1.
With this modification the series again simplifies into elementary trigonometric
18 Polarised electromagnetic waves

functions as follows:

θ 2 (n.σ )2 θ n (n.σ )n
exp(iθn.σ ) = σ0 + iθn.σ − + · · · (i)n + ···
2! n!
   
θ2 θ3
= σ0 1 − · · · + in.σ θ − + ···
2! 3!
= cos θ σ0 + i sin θn.σ
 
cos θ + i sin θn1 i sin θ (n2 − in3 )
=
i sin θ (n2 + in3 ) cos θ − i sin θn1
 
eiθ 0
= [U2 ] [U2 ]∗T (1.46)
0 e−iθ

This represents the most general special unitary matrix and an alternative
parameterization to that used in equation (1.28). We shall see in Section 1.3
that there is a simple geometrical interpretation of both sets of parameters in
terms of spherical trigonometry on the Poincaré sphere. From the form of the
eigenvalue decomposition we can see that the general unitary matrix repre-
sents birefringence between a pair of orthogonal elliptical polarisations. Such
a propagation channel is called a retarder, and θ is called the retardence of the
channel.

1.2.2.1 Radio wave propagation through the ionosphere

As an important use of the [N ] matrix formalism, we now consider the propaga-
tion of waves through a gyrotropic or handed medium. An important example
of this type is radio wave propagation through a part of the atmosphere called
the ionosphere (located at an approximate altitude between 50 and 400 km)
(Collin, 1985). Due to ionization by the Sun’s radiation, this thin part of the
atmosphere can be modelled as a cold plasma in the presence of the Earth’s
magnetic field. In the absence of a DC magnetic field the dielectric constant
of an ionized gas at frequency ω can be written (in the absence of collision
damping) in terms of the plasma frequency ωp as shown in equation (1.47)
(Ishimaru, 1991, Chapter 8):
"
ωp2 Ne e 2
εr = 1 − ωP = (1.47)
ω2 mε0

where Ne is the electron number density in the material (between 1010 and
1012 m−3 for the ionosphere) and e/m is the charge-to-mass ratio for an elec-
tron. Such a material, although frequency-dispersive, is isotropic, and therefore
does not distort the polarisation of the propagating wave. However, in the pres-
ence of an applied DC magnetic field the situation changes. Here we restrict
attention to the case where the DC field is applied along the z-direction (along
the direction of propagation for our plane wave). In this case the effect of an
electric field depends on its polarisation, and the medium becomes gyrotropic
with a dielectric tensor of the form shown in equation (1.48) (Ishimaru 1991,
 1.2 The propagation of polarised waves 19

Chapter 8):

  
 ωp2
εa iεb 0 
 εa = 1 −
ω2 − ωc2
ε = ε0 −iεb εa 0 ⇒ (1.48)

 −ωc ωp2
0 0 εr 
 εb =
ω(ω2 − ωc2 )

where ωc is the cyclotron frequency defined in terms of the applied mag-

netic field strength and the charge to mass ratio for the electron, as shown
in equation (1.49):

eB0
ωc = (1.49)
m
To give a typical example, the Earth’s magnetic field strength is around 5 ×
10−4 Tesla, which leads to a cyclotron frequency of 1.4 MHz. Considering
propagation in the z direction, we can now use equation (1.48) to generate the
2 × 2 [Kz ] matrix directly from this tensor, as shown in equation (1.50):
 
εa iεb
[Kz ] = ε0
−iεb εa
     
ε0 1 i ε − εB 0 1 −i
= . a . (1.50)
2 i 1 0 εa + εB −i 1

where we have also shown the corresponding eigenvector decomposition

of [Kz ]. This decomposition immediately exposes the physical structure of
the propagation problem. The eigenpolarisations are identified as left and
right circular polarisation. However, these two states propagate with different
propagation constants, determined by the eigenvalues of [Kz ].

1.2.2.2 Defining the sense of circular polarisation

Before proceeding, we first establish some notation concerning the handedness
of circular polarisation. In common with IEEE engineering standards we define
the sense of polarisation from the time variation of the electric field vector
in a fixed spatial plane. (Note that spatial variation for a fixed time would
be equally valid, but confusingly leads to the opposite definitions.) Again by
convention, we define the sense by looking in the −z direction; that is, against
the direction of propagation. With this established, left-hand circular is defined
as clockwise rotation, and right-hand anticlockwise. These give rise to the
polarisation vectors shown in Figure 1.6.
Returning to the gyrotropic medium, we see that left-hand circular polar-
isation is associated with an eigenvalue εa − εb while right-hand circular
polarisation is associated with εa + εb . We can now calculate the [N ] matrix
for this medium, as shown in equation (1.51):
√
[N ] = −iω µ0 [Kz ]
  √   
√ 1 1 i εa − εb 0 1 −i
= −iω ε0 µ0 √ . (1.51)
2 i 1 0 εa + εb −i 1
20 Polarised electromagnetic waves

Left-Hand Circular Polarisation

 π 1 1
x e L = cosω t i + cos ω t +  j ⇒ E L = 
 2 2  i
 π 1  1
y e R = cosω t i + cos ω t −  j ⇒ E R =  
 2 2 − i
Right-Hand Circular Polarisation

x
Fig. 1.6 Definition of left- and right-hand
circular polarisations

and finally we obtain the propagation matrix [Mz ] using the exponential function
as shown in equation (1.52):

[Mz ] = exp([N ]z)

     
1 1 i exp(−iβl z) 0 1 −i
= . . (1.52)
2 i 1 0 exp(−iβr z) −i 1

where the two propagation constants are defined in equation (1.53):

&) *
'
√ ' ω 2
β l = β εa − ε b = β ( 1 −
p
ω(ω + ωc )
&) *
(1.53)
'
√ ' ωp2
βr = β εa + εb = β ( 1−
ω(ω − ωc )

We can see that the right circular wave has a resonance when ω = ωc . This
wave is forcing the electrons to move in their ‘natural’ direction about the
magnetic field (according to the Lorentz force equation F = q(E + v × B)). For
this reason it is called the extraordinary wave. The left circular wave, on the
other hand, forces the electrons in the opposite direction and therefore shows
no resonance. It is termed the ordinary wave. Note that when ω is less than
some critical frequency ω1 then βL becomes imaginary and the ordinary wave
does not propagate. The cut-off frequency can be easily obtained from equation
(1.53), as shown in equation (1.54):
"
ω2 ωc
ω1 = ωp2 + c − (1.54)
4 2
Importantly, the extraordinary wave can propagate at low frequencies when the
ordinary wave is below cut-off. Hence low-frequency waves can penetrate the
ionosphere along lines of the Earth’s magnetic field. This is the main mechanism
behind the low-frequency whistler mode of atmospheric propagation. These
results are summarized in Figure 1.7. Here we show typical dielectric constant
variation with frequency and polarisation.
We see that the ordinary wave has a relatively simple behaviour with a cut-off
frequency of ω1 . The extraordinary wave shows more complex behaviour, with
 1.2 The propagation of polarised waves 21

4
Ordinary wave (left handed)
Extraordinary wave (right handed)

3
Dielectric constant

vp 
1+4 
2


vc

c/2
1

vc vc
v1 = v2p + –
4 2
0

1
2 v2 vc
v2 = v2 + +
4 2
–1
0.5 1 1.5 2 2.5 3
Normalized frequency v
vc
Fig. 1.7 Vector propagation modes in
gyrotropic media

two branches to its propagation behaviour, one at low frequencies, and one
at high. Note that at high frequencies (compared to ωc ) the medium becomes
isotropic and transparent with εr = 1. There is, however, a second impor-
tant polarisation phenomenon arising from this result: the distortion of linear
polarisations as they propagate via Faraday rotation, as we now discuss.

1.2.2.3 Faraday rotation

We now consider a Pauli matrix expansion of [N ] and show how it leads natu-
rally to a description of Faraday rotation in gyrotropic media. We first rewrite
equation (1.52) for [Mz ] as shown in equation (1.55):
     
1 βl +βr 1 i exp(−iβz) 0 1 −i
[Mz ] = e−i 2 z · ·
2 i 1 0 exp(iβz) −i 1
 
βl +βr cos βz − sin βz
= e−i 2 z
sin βz cos βz
 
βl +βr cos θF − sin θF
= e−i 2 z
sin θF cos θF
βl +βr
= e−i 2
z −iθF σ3
e (1.55)

Here we have factored the average propagation constant as indicated in equation

(1.40), and defined a differential wavenumber between the left- and right-
handed waves as β = (βl − βr )/2. By expanding the matrix product we
obtain a unitary plane rotation matrix as shown. This in turn may be expressed
as the matrix exponential of a single Pauli matrix, σ3 . The result is that incident
linear polarisations are rotated through an angle θf = β z. This is called Fara-
day rotation, and arises as a consequence of the circular polarised eigenstates
for gyrotropic media (Ishimaru, 1991; Collin, 1985; Bickel and Bates, 1965).
Physically we can consider a linearly polarised wave as decomposed into two
22 Polarised electromagnetic waves

counter-propagating circular waves, and as the two circular components prop-

agate with different velocities so they accrue a phase difference. This phase
difference yields a rotation of the linear polarisation state. The connection
between phase shifts of circular polarisation and rotations of linear polarisation
is of fundamental importance in radar polarimetry, and we shall encounter it
several times in our analysis. Again we note that the Pauli matrix decomposi-
tion provides a natural formalism for identifying the physical consequences of
wave propagation in such media.
One interesting property of Faraday rotation is its invariance to the direction
of wave propagation. If we now consider a plane wave propagating in the
–z direction as a first step, the above formulae remain the same but with –z
replacing z. In this case the rotation matrix is apparently transposed, as the
Faraday angle changes sign since θF = βz. However, the DC magnetic field
has a fixed polarity (+z direction), and hence the matrix [Kz ] is conjugated
for –z propagation (since the off-diagonal elements change sign with B0 ; see
equation (1.48)):

[Kz ] = [K−z ]∗ (1.56)

Consequently the left and right circular polarisations exchange eigenvalues,

and hence both k and z change sign. This leaves the sign of the Faraday angle
unchanged, as a consequence of which the matrix for –z propagation [M−z ] can
be written as follows:
 
i l 2 r z cos θF − sin θF
β +β βl +βr
[M−z ] = e = ei 2 z e−iθF σ3 (1.57)
sin θF cos θF

Surprisingly, the rotation is in the same direction as for +z; that is, if the wave
first propagates through the medium and is then returned to its starting point
then the Faraday rotation is not cancelled but doubled, since
 2
cos θF − sin θF
[Mz ][M−z ] = = e−iθF σ3 e−iθF σ3
sin θF cos θF
 
cos 2θF − sin 2θF
= e−i2θF σ3 = (1.58)
sin 2θF cos 2θF

This can be traced to the presence of the DC magnetic field, which has a
polarity of its own and causes this lack of reciprocity. This is in contrast to
a second type of circularly polarised wave propagation that occurs in many
natural media, such as sugar solutions (optical activity) and in manmade chiral
materials such as helical microwave dielectric composites. Here again, circular
eigenpolarisations are generated, but this time the effect does not double with
space reversal and has a fundamentally different physical origin, as we now
consider.

1.2.3 Case III: Propagation in chiral media

Returning to the vector wave equations (1.18), we now consider the allowed
propagation states in media with coupled electric and magnetic field effects.
The simplest example to consider of such a material is a cloud of small helical
 1.2 The propagation of polarised waves 23

particles embedded in a host material. The application of an electric field will

then cause polarization of the particles but also magnetization through circu-
lating induced currents, which in turn will generate a magnetic field. Hence
the constitutive material equations require a coupling of electric and magnetic
field effects. In the general case all coupling terms can be tensors, as an exten-
sion of that described in case II (for a fuller treatment see (Lakhtakia, 1989)).
However, an important class of systems can be characterized by scalar coupling
terms. These chiral media are characterized by the usual scalar permittivity ε
and permeability µ, but also by chiral admittance parameters γ and η such that
the constitutive equations have the form shown in equation (1.59) (Ablitt, 1999,
2000):

D = εE + ηB
(1.59)
H = γ E + µ−1
0 B

For simplicity we here consider the case of lossless chiral media where η = γ
and both are purely imaginary, so the constitutive equations have the special
form shown in equation (1.60):

D = εE − iγ B
(1.60)
H = −iγ E + µ−1
0 B

We now consider the properties of polarised plane wave propagation in such

materials. Before proceeding to the wave equation, we note that by using
Maxwell’s curl equations we can rewrite these relations in the form shown
in equation (1.61):

D = ε E + ∇ × E
 (1.61)
B = µ0 H + ∇ × H

where  is related to the chiral admittance γ as shown in 1.62:

γ
= (1.62)
ωε

These show that D not only depends on the local value of E at a point in the
material, but also on neighbouring values through the local spatial derivative of
E. This is termed spatial dispersion, and is characteristic of this type of material.
With this notation established we now return to the vector wave equation for
plane waves in such media, and obtain

∇ × ∇ × E − 2ωµ0 γ ∇ × E − ω2 εµ0 E = 0 (1.63)

Performing the spatial derivatives for the plane wave we obtain the following
matrix equation for the electric field components:
       
ex ey ex 0
β 2 ey  − 2iβωµ0 γ −ex  − ω2 εµ0 ey  = 0 (1.64)
0 0 ez 0
24 Polarised electromagnetic waves

from which see that, unlike the case for anisotropic material, ez = 0 is always
true and so these are TEM waves. We can now obtain the [Kz ] matrix by
inspection, as shown in equation (1.65):
   
2iβγ
1 εω 1 iεb
[Kz ] = ε =ε
− 2iβγ
εω 1 −iεb 1
     
ε 1 i 1 − εb 0 1 −i
= · . (1.65)
2 i 1 0 1 + εb −i 1

The eigenvector decomposition again yields left and right circular eigenpo-
larisations and differential propagation phase (circular birefringence) due to a
splitting of the eigenvalues. Note that because of spatial dispersion this matrix
is itself a function of the desired unknown wavenumber β. The [N ] matrix can
be obtained by taking the square root of [Kz ], as shown in equation (1.66):
√
[N ] = −iω µ0 [Kz ]
√   √   
iω µ0 ε 1 i 1 − εb −i
=− . √ 0 .
1
(1.66)
2 i 1 0 1 + εb −i 1

Finally, by using the exponential function we obtain the [Mz ] matrix for
propagation to z, as shown in equation (1.67):

[Mz ] = exp([N ]z)

     
1 1 i exp(−iβl z) 0 1 −i
= · · (1.67)
2 i 1 0 exp(−iβr z) −i 1
√
where the two propagation constants are defined from β0 = ω µ0 ε as

βL = β0 1 − 2βL ) ⇒ βL = β0 (−β0  + 1 + β02 2 )
 (1.68)
βR = β0 1 + 2βR ) ⇒ βR = β0 (β0  + 1 + β02 2 )

The sign of  determines the handedness of the medium as follows:

Clockwise or d-rotatory material:

>0 βR > βL and the phase velocity for RHC is slower than LHC.

Anticlockwise or l-rotatory material:

<0 βR < βL and the phase velocity for LHC is slower than RHC.

The parameter β0  is typically small, lying in the range 10−6 < |β0 | < 10−4
for natural materials. In practice we can therefore expand the square root and
obtain the following simplified relationships:
 
βL ≈ β0 (1 − β0 ) n ≈ (1 − β0 ) nR − nL
⇒ L ⇒ β0  ≈ (1.69)
βR ≈ β0 (1 + β0 ) nR ≈ (1 + β0 ) nR + nL
 1.2 The propagation of polarised waves 25

where nr is the refractive index of the material for right-hand circular polarisa-
tion. Thus  can be obtained experimentally by measuring the refractive index
of the material for right and left circularly polarised wave transmission. Alter-
natively, the chiral parameter can be estimated by measuring the rotation of a
linearly polarised wave. This rotation effect is similar in form to that for Fara-
day rotation, and can be made explicit by using the Pauli matrix representation
of [N ] which leads to an [Mz ] matrix of the form
 
−i l 2 r z cos θA − sin θA
β +β βl +βr
[Mz ] = e = e−i 2 z e−iθA σ3 (1.70)
sin θA cos θA

where
1 π
θA = (βL − βR )z ≈ −β02 z = z (nR − nL ) (1.71)
2 λ
This rotation is clockwise when  is positive, and anticlockwise when  is
negative (looking into the source of the wave). To compare different materials,
the specific rotatory power is defined as θA /z. This has a value ranging from
around 20 degrees/mm for solids such as quartz, down to 0.4 degrees/mm for
liquids such as turpentine.
Note that a key difference between this and Faraday rotation is the behaviour
under inversion of space coordinate. In this case, when we consider propaga-
tion in the −z direction, θA changes sign, and hence this type of rotation is
cancelled with propagation back through the medium. We now turn to consider
a generalization of these ideas to enable us to model propagation in arbitrary
media.

1.2.4 The Jones calculus: homogeneous and

inhomogeneous propagation channels
In the last section we established an important matrix differential equation
relating the effect of a propagation channel on the polarisation of a plane wave,
as shown in equation (1. 72):
 z0 
dE  
= [N ] E ⇒ Mz0 = [M0 ] exp [N ]dz (1.72)
dz 0

We further saw that the matrix [N ] can be conveniently expanded in terms of

the Pauli spin matrices so that it has the following general form:

!
3  
h0 + h1 h2 − ih3
[N ] = i[H ] = i hj σj = i (1.73)
h2 + ih3 h0 − h1
j=0

The study of general solutions of this equation was first carried out in optics in
Jones (1941, 1948), and has subsequently been termed the Jones calculus. In this
section we outline the general structure of this method and detail a classification
of different propagation channels based on the eigenvector decomposition of
the propagation matrix (Azzam, 1987; Lu, 1994).
To complete our study of wave propagation, we first consider generalization
to the case where the propagation channel involves losses and the matrix [Mz ]
26 Polarised electromagnetic waves

is no longer unitary. The main extension required is to permit the presence

of complex Pauli coefficients hj . The argument of the exponential function
then also becomes complex in general. Nonetheless, the series still splits into
odd and even components, and so the limits are modified to hyperbolic rather
than trigonometric functions. Wehave then, in the general case, the following
mapping for complex θ = h0 + h21 + h22 + h23 :

[M2 ] = |M2 |.eiH = |M2 |. exp(θ n.σ )

θ 2 (n.σ )2 θ n (n.σ )n
exp(θn.σ ) = σ0 + θn.σ + + ··· + ···
2! n!
   
θ2 θ3
= σ0 1 + · · · + n.σ θ + + ···
2! 3!
⇒ [M2 ] = |M2 |(cosh θ σ0 + sinh θn.σ )
 
cosh θ + sinh θn1 sinh θ (n2 − in3 )
= |M2 |. (1.74)
sinh θ (n2 + in3 ) cosh θ − sinh θn1

where we have factored the determinant of [M2 ] as a complex scalar. We will

later treat the special case when [M2 ] has zero determinant. This matrix is no
longer unitary, but represents a combination of differential wave attenuation
and phase shifts. Interesting special cases arise when the hj coefficients are
either all real or all imaginary. The [M2 ] matrices for these two special cases
can be written as shown in equation (1.75).

 
cosh θ + sinh θn1 sinh θ (n2 − in3 )
|M2 | · exp(θn · σ ) = |M2 | ·
sinh θ (n2 + in3 ) cosh θ − sinh θn1
 θ 
−φ e 0
= e . [U2 ] . · [U2 ]∗T
0 e−θ
 
cos θ + i sin θn1 i sin θ (n2 − in3 )
|M2 | · exp(iθn.σ ) = |M2 | ·
i sin θ (n2 + in3 ) cos θ − i sin θn1
 iθ 
−iφ e 0
= e . [U2 ] . · [U2 ]∗T
0 e−iθ
(1.75)

These two forms are of special importance in practical applications, as they

generate two common classes of propagation channel. We first summarize
their general structure, and then show an important result that any propagation
channel can be decomposed into a cascade of such effects.

1.2.4.1 Diattenuator channels

For h pure imaginary, [M2 ] has the general form of a 2 × 2 Hermitian matrix
(the upper case in equation (1.75)). Such a channel introduces differential atten-
uation between the (orthogonal) eigenstates defined by the eigenvector matrix
[U2 ]. It is termed a diattenuator channel (Lu, 1996), and is characterized by its
 1.2 The propagation of polarised waves 27

diattenuation D, defined in equation (1.76):

D = tanh 2θ 0 ≤ |D| ≤ 1 (1.76)

As z tends to infinity, D tends to 1, and such a channel will completely extinguish

the eigenvector corresponding to the smallest eigenvalue of [N ]. An example
of such a case is a polarising filter in optics (such as sheet Polaroid), whereby
a propagation matrix of the form shown in equation (1.77) is used to polarise a
light source into a preferred state:
 
λ 0
[M2 ] = 1 (1.77)
0 0

In microwave remote sensing such a situation can also arise for propaga-
tion through an oriented grid or volume of aligned scatterers, such as occurs
in the growth of many aligned agricultural crops such as wheat stalks (see
Section 3.5.2).

1.2.4.2 Retarder channels

For h real, [M2 ] has the form of a 2 × 2 unitary matrix (lower case in equation
(1.75)). Such a channel shows no attenuation, but introduces differential phase
shift between orthogonal eigenstates. It is termed a retarder channel, and is
characterized by its retardence R, defined in equation (1.78):

R = tan 2θ 0≤R≤π (1.78)

1.2.5 The polar decomposition

Very often a propagation channel is composed of a cascade of several composite
channels, as shown schematically in Figure 1.8. One important property of the
Jones calculus is that the [M2 ] matrix for the overall channel can be decomposed
into a product of [Mz ] matrices for each channel so that

[M2 ] = [MzN ] . [MzN −1 ] · · · · · · [Mz1 ] (1.79)

Hence, in general, a propagation channel will show both retardence and diatten-
uation and the [M2 ] matrix will be neither pure Hermitian nor unitary. However,
by employing a polar decomposition of the [M2 ] matrix we can always express
a general channel as an equivalent cascade of just two elements: a pure retarder
[U ] followed by a pure diattenuator [H ] (or vice versa) (Lu, 1996). In matrix

z2 zN

E in E out
[M1] [M2] [M3] [MN]

Fig. 1.8 General propagation channel as a

z1 z3 cascade of composite channels
28 Polarised electromagnetic waves

form we can then always write the following decomposition:

[M2 ] = [U ] · [HR ] = [HL ] · [U ] (1.80)

where the Hermitian diattenuator matrices are uniquely defined from [M2 ], as
shown in equation (1.81):

[HR ]2 = [M2 ]∗T [M2 ]

(1.81)
[HL ]2 = [M2 ] · [M2 ]∗T

If [M2 ] is non-singular, then likewise the retarder [U ] can be determined as

shown in equation (1.82):

[U ] = [M2 ] · [HR ]−1 = [HL ]−1 [M2 ] (1.82)

If [M2 ] is singular then [U ] is not uniquely defined, but this case can be accom-
modated by employing the singular value decomposition, as shown below. In
the most general case we can write [M2 ] in terms of a singular value decom-
position (see Appendix 1) as a product of unitary matrices [V ] and [W ] and a
diagonal matrix of singular values [D], as shown in equation (1.83):

[HR ] = [W ] · [D] · [W ]∗T
[M2 ] = [V ] · [D] · [W ]∗T ⇒ [HL ] = [V ] · [D] · [V ]∗T (1.83)

[U ] = [V ] · [W ]∗T

This decomposition emphasizes the importance of retarder and diattenuator

channels, but also leads to an important classification according to the eigenvec-
tors of [M2 ]. We can see from this result that if [M2 ] has orthogonal eigenstates,
then [V ] = [W ] and [HR ] = [HL ]; that is, the diattenuator is uniquely defined
for [M2 ] matrices with orthogonal eigenvectors. These are called homogeneous
propagation channels, as the diattenuation and retardence D and R depend only
on the eigenvalues and not on the eigenvectors (Lu, 1994).
It is clear that the most general form of homogeneous channel is generated
by the product of retarder and diattenuator channel matrices, where the only
constraint is that the [M2 ] matrix has orthogonal eigenvectors. This can be
secured by employing a product of matrix exponentials with the same n vector.
We then have the following explicit form for the general lossy homogeneous
channel:

[M2 ] = |M2 | · exp(φn · σ ) exp(iχ n · σ ) = exp[(φ + iχ )n.σ ]

 
cosh(φ + iχ ) + sinh(φ + iχ )n1 sinh(φ + iχ )(n2 − in3 )
= |M2 |.
sinh(φ + iχ )(n2 + in3 ) cosh(φ + iχ ) − sinh(φ + iχ )n1
(1.84)

which is the same as the general lossy matrix shown in equation (1.74) with
the substitution θ = φ + χ. However, this does not exhaust the possibilities in
terms of possible types of propagation channel, as we now show.
If [M2 ] has non-orthogonal eigenvectors, then [V ] and [W ] are no longer
equal, and [HR ] and [HL ] are therefore distinct diattenuators with the same
 1.2 The propagation of polarised waves 29

eigenvalues but different eigenvectors. These are called inhomogeneous

propagation channels.
To illustrate how inhomogeneous channels may arise from a simple cas-
cade of homogeneous elements, consider the example of a two-layer channel
composed of a rotator (such as arises with Faraday rotation or optical activ-
ity) followed by a polariser. The composite [M2 ] matrix is then represented as
shown in equation (1.85):
     
1 0 cos θ − sin θ cos θ − sin θ
[M2 ] = · = (1.85)
0 0 sin θ cos θ 0 0

The eigenvectors of [M2 ] are not orthogonal and are easily calculated from the
eigenvalues, as shown in equation 1.86:


sin θ 
λ1 = 0 ⇒ e 1 =
cos θ 
  ⇒ e∗T
1 e2 = 0 (1.86)
1 
λ2 = cos θ ⇒ e2 = 
0

These are propagation states that remain unchanged after transmission through
the channel. However, as we shall now show using the SVD, they do not have
the optimum transmittance. The singular value decomposition is easily obtained
by inspection, as follows:
     
∗T 1 0 1 0 cos θ − sin θ
[M2 ] = [V ] . [D] . [W ] = · · (1.87)
0 1 0 0 sin θ cos θ

from which it follows that the polar decomposition yields an equivalent retarder
and pair of diattenuators of the form shown in equation (1.88):
     
cos θ − sin θ 1 0 cos2 θ − sin θ cos θ
[U ] = [HL ] = [HR ] =
sin θ cos θ 0 0 − sin θ cos θ sin2 θ
(1.88)

The transmittance or gain of the propagation channel is defined as a function

of polarisation E, as shown in equation (1.89):
 
[M2 ] · E 2 E ∗T [M2 ]∗T [M2 ]E
T=  2 = (1.89)
E  E ∗T E

The extrema of this function define the maximum and minimum transmittance of
the channel. This is a classical Rayleigh quotient (see Appendix 1) for which the
maximum and minimum values are obtained for the eigenvalues of [M2 ]∗T [M2 ].
From the above we see that these are just the square of the singular values of
[M2 ]. We further see that we can decompose the matrix product, as shown in
equation (1.90):

[M2 ]∗T [M2 ] = [W ] · [D] · [W ]∗T (1.90)

30 Polarised electromagnetic waves

and so the maximum (minimum) transmittance is obtained using states (columns

of [W ]) which are not the eigenstates of the channel. In the above example we
see that the maximum transmittance is 1, obtained when the incident state is
   
cos θ max T 1
E in = ⇒ E out = (1.91)
− sin θ 0

In this way we see that the singular value decomposition is more relevant to
channel gain optimization studies than the eigenvalue expansion of [M2 ]. Note,
however, that the price to pay for maximization of transmittance is distor-
tion of polarisation state. For homogeneous systems there is little difference
between expansions, but for inhomogeneous propagation channels the SVD
or eigenvalue decompositions must be selected only after consideration of the
application in mind.

1.2.6 Propagation in stochastic channels

Much of this book is concerned with problems of wave scattering in random
media. It is then of interest to consider the effect of volume scattering on
polarised wave propagation and in particular to examine the way in which
the Jones calculus needs to be modified. In this section we show how the
attenuation of a coherent wave in a scattering medium can be estimated using
Foldy’s approximation (Tsang, 1985). While strictly valid only for low particle
concentrations, this method leads to a simple extension of the Jones calculus
for a range of important scattering problems.
The basic geometry is shown schematically in Figure 1.9, in which a wave
is propagating in the z-direction through a host medium. This medium, we
assume, is isotropic, and thus is characterized by a C2 polarisation degeneracy
with a scalar propagation constant β0 . However, inside this medium are located
discrete scattering particles, which for the sake of simplicity we assume are
ellipsoidal in shape. We also assume that the position and orientation of these
scattering particles are random variables with some underlying probability dis-
tributions. We wish to have a means of calculating the equivalent [M2 ] and
[N ] matrices for this channel using properties of the particles themselves. To
begin, we note that each particle acts to transform the incident plane wave
into a spherical wave (scattering). In addition, each particle can transform the
polarisation of the plane wave due to its own shape, orientation and material
structure. To account for these two effects we define for each particle a 2 × 2
scattering matrix as shown in equation (1.92)
     inc 
Exscat ei(ωt−βr) Sxx Sxy E ei(ωt−βr)
= · xinc = [S] · E inc (1.92)
Eyscat r Syx Syy Ey r

E in Eout = [M] Ein

Fig. 1.9 Schematic representation of general Particle

stochastic propagation channel scattering
 1.2 The propagation of polarised waves 31

∆z
∆z

z axis
r
(x,y)

x2 + y2
f0 f1 r= x2 + y2 + ∆z 2 ≈ ∆ z + Fig. 1.10 Derivation of the propagation
2∆z matrix in scattering media

At present there is some ambiguity with the coordinates defined for this equa-
tion. So far we have dealt only with propagation in the z-direction, but clearly
the scattering particle generates a spherical wave, which propagates in three
dimensions. We delay discussion of the definition of coordinates for the scat-
tering matrix until Section 1.4.2. At the moment we are interested only in the
effect of the particle in the same (z) direction as that of the incident wave. This
is termed forward scattering, and later we shall see it as a special case of the
general scattering problem.
The matrix [S] contains four complex numbers: two diagonal copolar scat-
tering coefficients which change the amplitude and phase of the wave but
maintain its polarisation in the same state as that of the incident plane wave.
The off-diagonal terms represent the possibility for crosspolarisation. Here the
amplitude and phase of the wave is modified by the particle, which now also
scatters radiation into the orthogonal polarisation to that incident. We can use
this scattering matrix to derive a Fresnel approximation for the change in elec-
tric field with z, and so determine the form of the [N ] matrix as follows (van
de Hulst, 1981). Figure 1.10 shows a component of the electric field at two
positions: f0 is the field at z while f1 is the new field at position z + z. We
are interested in establishing the change in field f1 − f0 for small z (but still
large enough so that βz  1, otherwise more rigorous diffraction theory is
required). The starting point is to decompose both fields into incident (fi ) and
scattered (fs ) field components so that we can write

f0 = fi
fi = e−iβz (1.93)
f1 = f1i + fs

where we have allowed f1 to be modified by the presence of scattering particles.

The change in incident field is easily calculated from the exponential function
so that

fi = f1i − fi = −iβ.z.fi (1.94)

It remains only to estimate the scattered field component fs in order to estimate

the total change in the field. This can be obtained from the scattering matrix, as
shown in equation (1.95):
! e−iβr
fs = .sxy .e0 (1.95)
r
where the sum must be taken over all the particles contributing to the field
at z + z. The factor sxy is one element of the scattering amplitude matrix,
32 Polarised electromagnetic waves

depending on our choice of polarisation channel. We now again face the problem
that [S] relates a plane wave to a spherical wave, and so formally we must
consider variations in x, y and z. However, we are concentrating on forward
scattering only, and so any consideration of x,y variations must be constrained
to a paraxial or small angle approximation so that only the forward scattering
amplitude is appropriate. To secure this, we replace r in equation (1.95) with its
approximation shown on the right-hand side of Figure 1.10. We also consider
a large number of particles with density N0 per unit volume. In this case the
summation becomes an integral, and we can rewrite equation (1.95) in the form
shown in equation (1.96), where we have included analytical evaluation of the
resulting Fresnel integral with infinite limits:

 2 2)

+y
−iβz −1 − iβ(x2z
fs = e sxy fi N0 z e dxdy dz

∞  1   (1.96)
βx2
−i 2z 2π z 2
−iβz 2π
e dx = ⇒ fs = e sxy fi N0 −i dz
iβ β
−∞

Finally, keeping terms only to first order in dz we obtain

2π No
fs ≈ −i fi zsxy (1.97)
β

In practice, the particles contributing to this integral will not have the same
orientation and size, and we therefore need one further modification to account
for variation over these parameters. Note that changing the size or orientation
does not alter the Fresnel integral, but requires only that we now replace the
constant factor sxy by a configurational average over distributions of size and
shape. So, for example, for ellipsoidal particles with dimensions a, b and c and
major-axis orientation defined by three Euler angles A, B and C (Goldstein,
1980) we can write:

∞ ∞ ∞ 2π π 2π

S = sxy p(a, b, c; A, B, C) da db dc dA dB dC (1.98)

0 0 0 0 0 0

where p(…) is the probability density function (PDF) of the distribution; and
so finally we have the following expression for the scattered field:

2π No
fs ≈ −i S fi z (1.99)
β

This is the desired result which allows us to write the total change of field for
each of the four scattering channels separately as shown in equation (1.100),
where we note that the crosspolarised channels contain, by definition, no
 1.2 The propagation of polarised waves 33

contribution from the incident field.

 
2πN0
fxx ≈ −iβ − i Sxx  zf0
β
 
2πN0 + ,
fxy ≈ −i Sxy zf0
β
  (1.100)
2πN0 + ,
fyx ≈ −i Syx zf0
β
 
2πN0 + ,
fyy ≈ −iβ − i Syy zf0
β

1.2.7 The Foldy–Lax equations

Collecting these results into matrix form, and using our notation from previous
sections, we can now establish the corresponding [N ] matrix for propagation
in random media, as shown in equation (1.101) (Tsang, 1985):
 + , 
−iβ − i 2πβN0 Sxx  −i 2πβN0 Sxy
[N ] = + , + , ⇒ [Mz ] = exp([N ]z)
−i 2πβN0 Syx −iβ − i 2πβN0 Syy
(1.101)

This provides a formal connection with the Jones calculus developed in the
previous section. Again we have two eigenstates for any such material, and
these may or may not be orthogonal, depending now on the distribution of
particle size, shape and orientation. In general, the medium will act as a mixed
retarder/diattenuator, which changes the polarisation of the incident wave as
it progresses into the material. As an example, consider propagation through
a cloud of dipoles. We assume all particles have the same size and shape and
that all are located in the xy plane, and hence their orientation distribution is
controlled by a single parameter function p(θ ). The forward scattering matrix
for each particle is then a function of just two parameters of the form shown in
equation (1.102):
       
cos θ sin θ ε 0 cos θ − sin θ cos2 θ − sin θ cos θ
S= · . =ε
− sin θ cos θ 0 0 sin θ cos θ − sin θ cos θ sin2 θ
(1.102)

The configurational average can be written as follows:

- - 
- p(θ) cos2 θd θ − -p(θ ) sin θ cos θd θ
S = ε (1.103)
− p(θ ) sin θ cos θd θ p(θ) sin2 θd θ

In the special case that the distribution is azimuthally symmetric in the xy plane
we have
 
1 ε 1 0
p (θ) = ⇒ S = (1.104)
2π 2 0 1
34 Polarised electromagnetic waves

which leads to isotropic propagation with zero diattenuation D and the same
extinction rate for all polarisations depending on the particle density, size and
dielectric constant. This simple example will be very important when we con-
sider coherent volume scattering in Chapter 7, and demonstrates how it is not
only the particle shape but also the orientation distribution that determines the
propagation channel in random media problems.
We have seen that wave polarisation can be transformed by propagation
through a channel. It is therefore of interest to consider the complete set of
possible wave polarisation states with a view to investigating what fraction of
all states are generated by a particular transforming channel or system. Such a
complete set is geometrically represented by the surface of a sphere called the
Poincaré sphere, as we now consider.

1.3 The geometry of polarised waves

Geometry is important for the analysis of polarised wave propagation and scat-
tering problems (Deschamps, 1951; Born and Wolf, 1989; Nye, 1999). It
often yields a simplified pictorial representation of the complicated processes
involved and, more fundamentally, offers a general procedure for the identifica-
tion of invariants under transformation of polarisation base. In this section we
develop a systematic approach to polarisation geometry that emphasizes two
key aspects of the problem: namely, the intimate relationship between complex
and real number representations of wave polarisation, and the way in which
transformation invariants can be derived for general scattering problems. The
former will lead us to consider mappings from complex to real spaces, and
ultimately to the Poincaré sphere and Stokes vector. The latter will lead us to
consider matrix eigenvalue decompositions for the treatment of general wave
scattering and propagation problems.

1.3.1 The polarisation ellipse

Our starting point is the structure of the polarisation ellipse. As the electric
field E(r, t) evolves in three-dimensional space and time it traces out a geo-
metrical structure. If we look at the time variation of this structure in a fixed
plane transverse to the direction of propagation and, without loss of generality
restrict attention to harmonic plane waves, then this locus is always elliptical.
To show this, consider the (real) time components of a general harmonic wave
propagating in the z-direction. These can be written in the plane z = 0, as shown
in equation (1.105):
.
ex = ax cos ωt
(1.105)
ey = ay cos (ω (t − to )) = ay cos (ωt − φ)

where ax and ay are the amplitudes of the components, and φ is the relative phase
shift. We can eliminate time from these equations using standard trigonometric
 1.3 The geometry of polarised waves 35

y
θ π π
− ≤θ <
2 2
τ π π
x − ≤τ ≤
4 4

ωt = π/2
q
b a
Change in time ωt = φ
origin
Fig. 1.11 Geometry of the polarisation
p ωt = 0 ellipse

identities to obtain the equation of an ellipse, as shown in equation (1.106):

"
ex cos φ ey ex2
− = sin ωt sin φ = 1− sin φ
ax ay ax
(1.106)
e2 ey2 ex ey
⇒ x2 + 2 − 2 cos φ − sin2 φ = 0
ax ay ax ay

This equation is written in terms of three parameters: ax , ay and φ. However,

a more convenient geometrical representation of the ellipse is in terms of two
angles θ and τ , defined as shown in Figure 1.11. Note that the ellipse also has
amplitude a = ax2 + ay2 but initially we set this to unity. This simplifies the
geometry, but we shall reintroduce it again later when considering connection to
the geometry of the Lorentz transformation. The angle θ is the inclination of the
major axis of the ellipse, while τ is called the ellipticity angle and is a measure
of the shape of the ellipse. It is zero for the special case of linear polarisation,
and has a maximum absolute value of π/4 for circular polarisation. Note that
we define positive and negative ranges for both angles. The variation of θ is
evident from a consideration of plane rotations, but the sign of τ is related to
the sense of the ellipse.
The ‘sense’ is a consequence of the fact that although the ellipse is a static
object, we are actually dealing with a dynamic process: namely, the time evolu-
tion of the electric field. This implies that for any given ellipse we can generate
a time variation in either the clockwise or anticlockwise direction. Note that
this is only true when we have a well-defined direction of propagation (the z
axis). For three-dimensional plane waves we shall have to be more careful in
the definition of sense. As shown in Figure 1.6, we then define left-hand polar-
isations as a clockwise rotation (corresponding to positive τ ) and right-hand
polarisations as anticlockwise (negative τ ).
We can further develop this geometry by considering a complex representa-
tion of the ellipse. From Figure 1.11, assuming for the moment that θ = 0 and
36 Polarised electromagnetic waves

using the complex exponential function, we can write:

 
cos τ iωt
E= e (1.107)
i sin τ

where it is understood that the time domain components can be obtained by

forming the real part of this expression. We see that by using a complex vector
we obtain a time invariant description of the ellipse. Note that in equation
(1.107) the time t = 0 origin lies at the tip of the major axis. In general this origin
could be located at any point around the ellipse. To allow for this transformation
of time origin in the geometry we rewrite the complex representation in the form
shown in equation (1.108):

E = a + ib eiωt ⇒ a.b = 0 (1.108)

where a and b are vectors in the x–y plane lying along the major and minor axes
of the polarisation ellipse. For a general time origin we then have the following
relation:
/
 iωt iφ  p = a cos φ − b sin φ
E = a + ib e e = p + iq e ⇒ iωt
⇒ p.q = 0
q = a sin φ + b cos φ
(1.109)

where the vectors p and q again lie in the xy plane, but are not orthogonal and
correspond to what are called the conjugate semidiameters of the polarisation
ellipse. The relationship between a, b and p, q is summarized in Figure 1.11.

1.3.1.1 Absolute phase of polarised waves

The absolute phase of the wave can now be defined in terms of the ellipse
geometry, as shown in equation (1.110):

1
φ= arg(p2 − q2 + 2ip.q) + nπ (1.110)
2

The importance of this result is that it relates absolute phase to spatial vectors
p and q. For example, points of circular polarisation are then defined from
singularities of φ, for which we have the following geometrical conditions,
which corresponds in general to a line given by the intersection of two surfaces
in space:

p.q = p2 − q2 = 0 (1.111)

These circular polarisation lines are termed C-lines, and are important in a full
characterization of fields based on their singularities as used, for example, in
the study of catastrophe optics (see Nye (1999) for more details).

1.3.1.2 Polarised waves in three dimensions

A second key advantage of using the conjugate semidiameters is that there
then exists a straightforward extension to three-dimensional plane waves. This
 1.3 The geometry of polarised waves 37

becomes important in applications when we combine waves from multiple

sources and directions. In this more general case we can write the electric field
vector as shown in equation (1.112):

 0 1
E r, t = P(r) + iQ r eiωt (1.112)

where P and Q are now three-element vectors. From this, the wave polarisation
is still an ellipse but now lying in the plane defined by P and Q with a normal
to the plane n = P × Q. Special cases arise when P × Q = 0, in which
case the polarisation is linear and n is not defined. Again this will generally
occur along special lines in space called L-lines. Circular polarisation occurs
for P·Q = P2 −Q2 = 0, and so there exist C-lines also in the three-dimensional
case, although these are in general not the same as those for the two-dimensional
case. Note that the direction of wave propagation is not so clearly defined in
three dimensions, and some care is required in the definition of sense of elliptical
polarisation and absolute phase. Although resort can always be made for general
fields to the Poynting vector S = E × H, which is the direction of energy flow,
this involves a combination of electric and magnetic field components which is
undesirable for polarisation studies. To resolve this we again define the absolute
phase from the definition of major and minor axes, as shown in equation (1.113),
where A · B = 0:

P + iQ = eiφ A + iB (1.113)

The problem is now to relate the phase of an elliptical wave at one point
to the phase at a neighbouring point where the ellipse has a different size,
shape and orientation vector n. One elegant way to resolve this is to use the
Pancharatnam geometric phase (Pancharatnam, 1956), defined from the inter-
ferogram between the two complex signals. In this way the phase difference
between points is defined from maxima and minima in a (real) intensity pattern.
Mathematically it is derived from the Hermitian product, as shown in equation
(1.114):

ψ = arg(E ∗1 E 2 ) (1.114)

so that when considering a spatial continuum of polarisation ellipses, the

infinitesimal phase change can be written as shown in equation (1.115):

2 ∗
3 Im E ∗ · d E 2B · d A
δψ = arg E E + dE =  2 = −  2  2 + d φ (1.115)
E  A + B

where the last term results since A·B = 0 implies that d (A·B) = dA·B+d B·A =
0. This then confirms that the change in phase can be decomposed into the
gradient of the phase plus a contribution from the change in geometry. Note that
there still remains a problem with this definition, as ψ is not formally integrable,
causing problems4when we try to define terms such as δψ/δr, although for the
second term, ∂φ ∂r = ∇φ is acceptable. Hence there is no suitable unique
indicator of propagation direction over the whole field. However, we can define
38 Polarised electromagnetic waves

a phase gradient based on the above decomposition of phase as follows:

2AB ∂ε ∂φ 2AB
t=− + =− 2 t + ∇φ (1.116)
A2+ B ∂r
2 ∂r A + B2 ε

where A and B are the magnitudes of the vectors, and the angle ε is the angle
of rotation of the axes A, B about n, defined as B · dA = ABd ε. The component
of t parallel to n represents the twist of the ellipse about n, while the transverse
component represents the bend of the axial directions of the ellipse in its own
plane. This expression is well behaved at most points in the field, and in partic-
ular it can be made to yield sensible definitions even along C-lines, when grad
φ is singular and lines of linear polarisation when tε is not defined (see Nye
(1999) for more details). This means that we can take t as a general definition
of the propagation direction and then use it to define handedness at each point
in the wave field as the sign of the triple product [A B t].

1.3.2 Polarisation geometry for paraxial waves

In many applications of interest the full machinery of this three-dimensional
wave field formulation is not required, and two-dimensional fields with a well-
defined direction of propagation are considered. This is especially true in remote
sensing applications. Returning to these, we now develop several important
geometrical properties of these simplified fields.
We start by considering the change of absolute phase of a wave as a special
unitary transformation of the complex vector in C2. If we multiply one state by
a scalar phase factor exp(iφ), then in order to secure a special unitary transfor-
mation we must apply a conjugate phase change of exp(−iφ) to the orthogonal
state. In transformation terms, therefore, we cannot change the phase of one
polarisation state vector without also considering the effect on its orthogonal
partner. In matrix terms we can represent such changes of absolute phase as
shown in equation (1.117):
   
eiφ 0 cos τ
E= . φ = ωto (1.117)
0 e−iφ i sin τ

This expression then generalizes in a straightforward way if the ellipse orien-

tation θ is not zero. In this more general case, the unitary complex vector that
represents an ellipse is as shown in equation (1.118):
     
cos θ − sin θ cos τ cos αw
E= . = (1.118)
sin θ cos θ i sin τ sin αw eiδ

where we have also shown the general element of C2 as defined in equation

(1.24). By direct expansion and use of trigonometric identities we can relate
the two sets of parameters, as shown in equation (1.119):

cos 2αw = cos 2θ cos 2τ tan 2θ = tan 2αw cos δ

(1.119)
tan δ = tan 2τ csc 2θ sin 2τ = sin 2αw sin δ
 1.3 The geometry of polarised waves 39

Shortly we shall see a simple geometrical interpretation of this result, but for
the moment consider a change of ellipticity of the wave from τ to τ + τ . This
can be represented by a matrix equation as shown in equation (1.120):
     
cos (τ + τ ) cos τ i sin τ cos τ
= . (1.120)
i sin (τ + τ ) i sin τ cos τ i sin τ

Combining this with the rotation matrix involving θ we can extend the analysis
by representing not only unitary vectors in C2, but also special unitary matrix
transformations, so that we can transform between any pair of ellipses by first
rotating the major axis and then changing the ellipticity. The general compound
unitary transformation can then be written as shown in equation (1.121):
     
e−iφ 0 cos τ −i sin τ cos θ − sin θ
[U2 ] = . . (1.121)
0 eiφ −i sin τ cos τ sin θ cos θ

This should be compared with the alternative forms shown in equations (1.28)
and (1.46). Note that in this case we have obtained a product decomposition
with a strong geometrical interpretation. We can also rewrite equation (1.121)
as a product of matrix exponential functions as defined in equation (1.42). The
result is shown in equation (1.122):

[U2 ] = (cos φσo − i sin φσ1 ) (cos τ σo − i sin τ σ2 ) . (cos θ σo − i sin θ σ3 )

= exp(−iφσ1 ) exp(−iτ σ2 ). exp(−iθ σ3 ) (1.122)

This demonstrates how we can consider the effect of a general transformation

as a compound of three simpler operations, each generated by a single Pauli
matrix as shown.
So far our treatment has considered the use of complex matrix transfor-
mations for representation of the polarisation ellipse. There is, however, an
alternative approach based entirely on real quantities. This has more than the-
oretical significance. By transforming a ‘complex’ problem into an equivalent
‘real’ one, we can dispense with the need for a coherent measurement sys-
tem. This is of great practical benefit, especially at high frequencies where it
is difficult to design detectors capable of following the fast field oscillations
directly. Visible optics is a prime example in which widespread use is made
of this mapping for polarisation studies. A simple example of this complex-to-
real equivalence is in the formation of an interference pattern. Here the spatial
variation of a ‘complex’ field is visualized as a ‘real’ fringe intensity variation
across the field of view. In polarimetry this real-to-complex mapping is more
subtle but nonetheless central to the development of techniques for the inter-
pretation of vector scattering problems. In this context it leads to the concept
of the Poincaré sphere, as follows.

1.3.3 The Poincaré sphere

We start by considering each of the complex matrix factors in equation (1.122)
to show how it can be mapped into an equivalent orthogonal transformation of
40 Polarised electromagnetic waves

a real three-vector r. In the context of polarimetry this is called the Stokes vec-
tor representation, after the nineteenth-century English mathematician George
Gabriel Stokes (1819–1903) (Born and Wolf, 1989; Bickel, 1985; Schmeider,
1969).
The first stage is to define a suitable real vector r. As there are three matri-
ces in the cascade of equation (1.122) we anticipate a three-dimensional real
space
 R3 and so set r = (x, y, z). The next key idea is to recognize that
r  = x2 + y2 + z 2 (amplitude) should be a transformation invariant, so that
we deal with orthogonal transformations in R3. As we are going to deal with
matrix products we need to generate a matrix from r with a metric that is a
function of the desired invariant. One way to do this is to construct a matrix
using the Pauli spin matrices, as shown in equation (1.123):

 
x y − iz
[R] = xσ1 + yσ2 + zσ3 = ⇒ det([R]) = −(x2 + y2 + z 2 )
y + iz −x
(1.123)

[R] Is called a spin matrix, and is constructed so that its determinant is a

function of the desired invariant. Note the negative sign. This will be useful
when we consider extension to Lorentz transformations in Section 1.5. Having
established a 2 × 2 matrix representation of r, we can now generate transforma-
tions using our complex 2 × 2 exponential functions so that in general we can
form a transformation as shown in equation (1.124) (Goldstein, 1980; Misner,
1973):

   
R ϕ, n = exp iϕσ .n . [R] . exp −iϕσ .n (1.124)

This equation preserves the determinant of [R] on both sides, as the matrix
exponentials both have unit determinant and the determinant of a matrix product
is just the product of determinants (which was the reason why we focused on
special unitary matrices in equation (1.26)). We have therefore constructed a
transformation of r with the correct metric.
We must now determine other invariants under this transformation. In par-
ticular we seek the axis of any plane rotations. This will allow us to finally
determine a mapping from the 2 × 2 complex matrix into a real 3 × 3 orthogo-
nal matrix. To do this we need only consider infinitesimal transformations, so
we can conveniently use only the first few terms of the series expansion of the
exponential function (see equation
5 (1.38)). Equation (1.125) shows this expan-
sion. We note that since [R] = 3j=1 rj σj , invariants can then be conveniently
identified from products of the Pauli matrices that commute; that is, for which
σi σj − σj σi = 0.

 
[R (ϕ)] = σ0 + iϕσ · n + · · · · [R] · σ0 − iϕσ · n + · · ·

≈ [R] + iϕ [R] σ · n − σ · n [R] (1.125)

Alternatively we can (in this special case of three-dimensional vectors) proceed

by noting that the product of two spin matrices can be expressed in terms
 1.3 The geometry of polarised waves 41

 
of scalar and vector
 products of the original real vectors as a · σ b · σ =
a · b + i a × b σ . Hence it follows from this and equation (1.125) that the only
other invariant is given by equation (1.126):

[R] σ · n − σ · n [R] = r · n (1.126)

This is a very important result. It shows that the transformation of equation

(1.124) has a very simple interpretation in the space of the vector r. Recall that
n is a unit real vector but lies in the original complex domain (equation (1.46)),
and so this relation demonstrates that under the transformation the component
of r in the direction of n is invariant. This is just another way of defining a
plane rotation about the n axis, but critically it is located in the ‘real’ space of
the vector r. Hence we have derived a mapping from unitary transformations
in C2 into orthogonal transformations in R3—a real three-dimensional space.
We can now generate this mapping for each of the elementary matrix expo-
nentials in equation (1.44). Using the fact that n1 = (1, 0, 0), n2 = (0, 0, 1),
and n3 = (0, 0, 1) we obtain the explicit mappings shown in equation (1.127).
Note that double angles appear in the ‘real’ space. This is a direct consequence
of employing the triple matrix product in equation (1.124). It means that the
mapping from complex to real is unfortunately not 1-to-1 but 2-to-1. For exam-
ple, if we add 2π radians to any of the angles in the real space we obtain an
unchanged rotation matrix. However, in the corresponding complex space we
have two distinct matrices, since the half-angle used in the mapping yields only
a π change in the complex domain. So, for every real rotation matrix there are
two corresponding unitary partners:
 
  cos 2θ sin 2θ 0
cos θ sin θ  
exp(iθ σ3 ) = ⇔ − sin 2θ cos 2θ 0
− sin θ cos θ
0 0 1
 
  cos 2τ 0 sin 2τ
cos τ i sin τ  
exp(iτ σ2 ) = ⇔ o 1 0 
i sin τ cos τ
− sin 2τ 0 cos 2τ
 
  1 0 0
exp(iφ) 0  
exp(iφσ1 ) = ⇔ 0 cos 2φ sin 2φ 
0 exp(−iφ)
0 − sin 2φ cos 2φ

[U2 ] = exp(iφσ1 ) exp(iτ σ2 ) exp(iθ σ3 ) = exp(iχ σ .n) = cos χ σ0 + i sin χ σ .n

     
1 0 0 cos 2τ 0 sin 2τ cos 2θ sin 2θ 0
     
⇔ [O3 ] = 0 cos 2φ sin 2φ  .  o 1 0  . − sin 2θ cos 2θ 0
0 − sin 2φ cos 2φ − sin 2τ 0 cos 2τ 0 0 1
(1.127)

We can formalize this correspondence by defining the groups SU(2) and O+ 3 as,
respectively, the group of all 2 × 2 complex unitary and 3 × 3 real orthogonal
matrices with determinant +1 (see Appendix 2). If two groups X and Y have a
1-to-1 correspondence between elements they are said to be isomorphic, while
if the correspondence is many-to-one they are homomorphic. Hence we have
demonstrated that SU(2) and O+ +
3 are homomorphic. This SU(2)–O3 homo-
morphism is fundamental to the development of polarisation geometry. One of
42 Polarised electromagnetic waves

z
P
2α P 2φ
2τ

δ 2τ
2θ y
2θ

Fig. 1.12 The Poincaré sphere and polarisa-

tion space x

the most important consequences of this mapping is that matrix multiplication

is preserved, and so a cascade of matrices in the complex domain maps into
a cascade of matrices in the real domain. This, combined with the result of
equation (1.124), allows us to map any unitary matrix as a set of Euler angle
transformations in the real space, as shown in equation (1.127). This in turn
leads to a simple geometrical interpretation of a unitary matrix transformation
as rotations over a sphere in R3. Each point P on the surface of this sphere
corresponds to a triplet of angles (θ , τ , φ) as shown in Figure 1.12. Here we can
see on the right how the angles can be used to locate any point P in the space of
the vector r. The two most important angles are θ and τ . These are the longitude
and latitude of the point P on the surface of a sphere. The third angle φ corre-
sponds to a rotation about the radius of the sphere and, as shown in equation
(1.117), relates to the absolute phase of the wave. The set of all polarisations
of the same amplitude then covers the surface and generates what is called the
Poincaré sphere, after the French mathematician Henri Poincaré (1854–1912).
This sphere has all linear polarisations mapped around the equator, with left
circular at the north pole and right circular at the south pole. The upper and
lower hemispheres then correspond to elliptical states with opposite sense.
One very important consequence of the double angle mapping is that orthog-
onal polarisations lie diametrically opposite on the sphere and correspond to
antipodal points. This is in distinction to the conventional geometrical notion of
orthogonality, which involves 90-degree angular separation between vectors.
We can also use this sphere to interpret the relationships derived in equation
(1.119). The angles θ and τ are then related to α, δ as elements of a spherical tri-
angle, as illustrated on the left-hand side of Figure 1.12. In this way we see them
as alternate angular coordinates to locate a polarisation state P in polarisation
space.
To proceed, we must now modify this formalism to include the amplitude
of the wave, a. This we can do by modifying the spin matrix [R], as shown in
equation (1.128) (Misner, 1973; Goldstein, 1980).
 
1 + x y − iz
[R] = a(σ0 + xσ1 + yσ2 + zσ3 ) = a
y + iz 1 − x
⇒ det([R]) = a2 (1 − (x2 + y2 + z 2 )) (1.128)

Here we see that by adding the amplitude as a multiple of the identity matrix
we can secure a new invariant: namely, the difference between the amplitude
squared and the square of the length of the vector r. For a polarised wave this
metric will always be zero. This modifies the geometry to that of a Lorentz
 1.3 The geometry of polarised waves 43

transformation, where equation (1.128) arises naturally through the use of the
Minkowski metric. The Lorentz spin transformation is then obtained as a gen-
eralization of equation (1.124) as shown in equation (1.129), where [L] is called
a Lorentz spin matrix, det([L]) = 1, but γ can be complex, and [L] does not
have to be unitary.
    
R γ , n = [L γ , n ] · [R] · [L γ , n ]∗T [L] = exp iγ σ · n γ = γr + iγi
(1.129)

Physically this allows us to consider differences in scattering amplitude or

absorption across polarisation channels, and therefore represents an important
extension of the geometry.

1.3.4 The Stokes vector

We note that by including the amplitude ‘a’ in a description of polarised waves,
we must now formally consider not just a three-vector r but a four-vector g,
called the Stokes vector of the wave, formed as a composite of the three-vector
r and the wave intensity with elements given by the Pauli matrix expansion
of the spin matrix [R], as shown in equation (1.130) (Bickel, 1985; Born and
Wolf, 1989).
   
g0 1
g1  
2 cos 2θ cos 2τ 

g= 
g2  = a  sin 2θ cos 2τ 
g3 sin 2τ
 
1
 cos 2αw 
= a2 
sin 2αw cos δ 
 g02 − g12 − g22 − g32 = 0 (1.130)
sin 2αw sin δ

Figure 1.13 shows some examples of the Stokes vectors for linearly and cir-
cularly polarised waves. Equation (1.130) defines the Stokes vector in terms
of the geometrical parameters of the ellipse. An alternative definition is based
on the wave coherency matrix [J ]. This is a 2 × 2 Hermitian matrix, and is
defined from the product of the complex electric field vector with its conjugate
transpose. Using α and δ parameters this matrix can be defined as shown in
equation (1.131):
 
cos αw
E=a
sin αw eiδ
 
∗T cos2 αw sin αw cos αw e−iδ
⇒ [J ] = EE =a 2
= [R]
sin αw cos αw eiδ sin2 αw
(1.131)

To make this consistent with equation (1.128) the matrix [J ] must be related
to the Stokes vector g as shown in equation (1.132), where I , Q, U and V are
44 Polarised electromagnetic waves

Horizontal Polarisation
g = (1,1,0,0)

Vertical Polarisation
g = (1,–1,0,0)

+45 degree Polarisation

g = (1,0,1,0)

–45 degree Polarisation

g = (1,0,–1,0)

Left Hand circular

g = (1,0,0,1)

Right Hand circular

Fig. 1.13 Example Stokes vectors for g = (1,0,0,–1)
polarised waves

symbols widely used in the optics literature for the Stokes parameters, elements
of the Stokes vector.
   
    Ex Ex∗ + Ey Ey∗ I
Ex Ex∗ Ex Ey∗ 1 g0 + g1 g2 − ig3  E E∗ − E E∗  Q 
 x x y y   
[J ] = = ⇒g= = 
Ey Ex∗ Ey Ey∗ 2 g2 + ig3 g0 − g1  Ex Ey∗ + Ey Ex∗  U 
∗ ∗
i(Ex Ey − Ey Ex ) V
(1.132)

[J ] has the following two important properties:

• Det([J ]) = 0. This follows from the Minkowski metric, and is true for
all polarised monochromatic waves.
• [J ] = [J ]∗T ; that is, [J ] is complex Hermitian, and therefore has orthog-
onal eigenvectors and real eigenvalues (see Appendix 1). For polarised
waves it follows that the eigenvector decomposition of [J ] can be written
as shown in equation (1.133):

     
cos αeiφ − sin αe−iδ λ 0 cos αe−iφ sin αe−iδ
[J ] = . 1 .
sin αeiδ cos αe−iφ 0 0 − sin αeiδ cos αeiφ
(1.133)

In publications not concerned directly with the geometrical structure of the

Stokes vector, the modified Stokes vector gm is often used (Tsang, 1985). This
again is defined from the wave coherency matrix but in a more straightforward
 1.3 The geometry of polarised waves 45

manner, as shown in equation (1.134):

  


1 1 0 0

  1 −1 0 0
  
 g = [G]g M =  g
|Ex |2 
  0 M
  

0 0 1
 Ey 2   0 0 0 1
gM = 
2Re(Ex E ∗ ) ⇒  1 1
 (1.134)
 0 0
y 
 
2 2
∗
2Im(Ex Ey ) 
 g = [G]−1 g =  2 − 2

1 1
0 0 g

  1 0
 M
 0 0
0 0 0 1

The matrices [G] and [G]−1 may then be used to transform between the two
conventions. The use of matrices to represent changes in Stokes vectors forms
the basis of an important calculus. In equation (1.135) we show two other impor-
tant examples of matrices used in Stokes algebra. The first [M ]conj represents
the physical process of conjugation, or taking the complex conjugate of the
wave. As shown, this has the effect of changing the sign of the fourth Stokes
parameter or changing the sense of the polarisation ellipse. On the right we
show the transformation matrix that converts a Stokes vector into its orthogo-
nal polarisation state. This involves reflection of the point in the origin to obtain
the coordinates of the antipodal point on the Poincaré sphere.
   
1 0 0 0 1 0 0 0
0 1 0 0  0 −1 0 0
[M ]conj = 0 0 1 0 
 [M ]⊥ = 0 0 −1 0  (1.135)


0 0 0 −1 0 0 0 −1

Finally, we note that an extension of this Stokes approach to cover the case of
three-dimensional waves, when the wave coherency matrix [J ] becomes 3 × 3,
has been developed by Roman (1959a) as a set of nine Hermitian matrices anal-
ogous to the Pauli spin matrices, the coefficients of which define a generalized
set of Stokes parameters ri for the representation of waves with arbitrary form
(Roman 1959b). In this case the generalization of equation (1.132) is shown in
equation (1.136):
 
Ex Ex∗ Ex Ey∗ Ex Ez∗
[J3 ] = Ey Ex∗ Ey Ey∗ Ey Ez∗ 
Ez Ex∗ Ez Ey∗ Ez Ez∗
 
r0 + r1 + r4 r2 + ir3 r7 + ir8
=  r2 − ir3 r0 r2 − r5 + i(r3 − r6 ) (1.136)
r7 − ir8 r2 − r5 − i(r3 − r6 ) r0 − r1 + r4

Here we concentrate on the simplified 2×2 case, which is the most commonly
used in remote sensing. However, special applications that require analysis
of dynamic vector near field effects should make use of this more general
formalism.
We have now considered the generation and geometry of polarised waves and
the distortion effects caused by propagation channels. Now we turn to consider
the final stage in Figure 1.1: namely, the scattering of polarised waves and the
interpretation of a scatterer as a transformer of polarisation state.
46 Polarised electromagnetic waves

1.4 The scattering of polarised waves

When a plane wave illuminates an object, metallic or dielectric, vector currents
are induced, the magnitude and direction of which depend on the shape and
material composition of the object. These in turn act as secondary sources that
reradiate into space. In general, this reradiation occurs in directions different to
that of the incident wave. This is termed wave scattering, and forms the basis
for our studies of remote sensing using electromagnetic waves. In particular
we are concerned with the polarisation properties of the scattered wave, and
to what extent they maintain a ‘memory’ of the original vector nature of the
induced currents.
In order to formulate a general scattering problem we first locate the scatterer
in free space at the origin of a spherical polar coordinate system, and then
define an incident plane wave direction A using two angles θ and φ. This
wave has a polarisation with C2 freedom, defined in the transverse xy plane as
shown in Figure 1.14. The wave is then scattered, and interest centres on the
radiation scattered into a new direction B, again defined by two polar angles
θ and φ . The two wave-direction vectors β i and β s define a plane called
the scattering plane, and the bisector between the incident and scattered wave
vectors is called the bisectrix vector. This bisector plays an important role
in a geometrical interpretation of wave scattering, as we show later. It also
provides an important symmetry axis for consideration of the basic polarisation
properties of the scattering system. The angle in this plane between incident and
scattered wave directions is called the scattering angle, and important special
cases arise for forward scattering ψ = 0◦ , lateral scattering ψ = 90◦ , and
backscatter ψ = 180◦ .
However complicated the scattering problem may be, the scattered fields
obey Maxwell’s equations, which form a linear system of equations. Hence
there is no loss of generality in postulating a linear mapping from A to B. Since
both A and B have C2 degeneracy, this mapping can be represented by a 2 × 2
complex matrix [S] called the amplitude or scattering matrix. In this way an
arbitrary polarisation Ei at A is mapped by the scatterer into a state Es at B
by the relationship shown in equation (1.137), where the factor in front of the
matrix accounts for the phase and amplitude variation of a spherical wave of

Scatterer
S

x'
x bi
bs
Α y' Β
y
Incident wave bisectrix b scattered wave
Fig. 1.14 General wave scattering coordi-
nates for the amplitude scattering matrix
 1.4 The scattering of polarised waves 47

radius ‘r’ centred on the scattering point:

   
e−iβr Sxx Sx y e−iβr S11 S12
Es = .E i = .E i (1.137)
r Sy x Syy r S21 S22

Note that sometimes in the literature (van de Hulst, 1981; Hovenier, 2004) a
normalizing factor of 1/iβ is used on the right-hand side in equation (1.137). This
is supported by the discussion leading up to equation (1.99), but readers should
be warned that there exist slight differences of definition for the scattering
amplitude matrix.
Cleary, knowledge of this mapping (that is, knowledge of all four complex
numbers in [S]) enables full polarimetric characterization of the scattering prob-
lem from A to B. We can then determine the amplitude and polarisation of the
scattered wave at B for any incident polarisation state Ei on the Poincaré sphere
by combining the C2 freedom with matrix multiplication arising from system
linearity. This is called polarisation synthesis or virtual polarisation adaptation
(Giuli, 1986; Poelman, 1991). The key idea is that measurement of [S] can
lead to adaptive selection of polarisation to enhance features based on signal
processing rather than on physical antenna changes. The matrix [S] therefore
plays an important role in polarimetry theory (Luneburg, 1996, 1997), and we
now turn to discuss its properties in more detail.

1.4.1 The scattering amplitude matrix [S]

A key benefit of knowledge of the [S] matrix is that it provides independence
from the measurement basis used. Above, we chose xy and x y as a natural
coordinate system, but in principle any pair of C2 orthogonal states will be
equally acceptable. This can be formally represented by allowing independent
unitary transformations of the incident and scattered coordinates (Kostinski,
1986; Luneberg, 1996, 1997) so that the change of frame is represented by
the pair of matrix equations shown in equation (1.140), where in general
U 2A = U 2B :

E i = [U2A ] .E i
(1.140)
E s = [U2B ] .E s

With this flexibility, the scattering matrix in the new base pair defined by U 2A
and U 2B is related to that in xy and x y by a unitary similarity transformation
as shown in equation (1.141):

[U2A ]−1 = [U2A ]∗T ⇒ E s = [U2B ] . [S] . [U2A ]∗T E i (1.141)

The form of this transformation suggests that a singular value decomposition

(SVD) of the matrix [S] may be used to simplify the transformation caused by
[S] itself. According to the SVD (see Appendix 1), any complex matrix [S] may
be written in terms of two unitary matrices [U ] and [V ] and a diagonal matrix
48 Polarised electromagnetic waves

[] as shown in equation (1.142):

  /
γ1 0 [U ] . [U ]∗T = [I ]
[S] = [U ] . [] . [V ]∗T = [U ] . . [V ]∗T
0 γ2 [V ] . [V ]∗T = [I ]
(1.142)
In the context of the scattering matrix, this result means that there always exists
a pair of orthogonal bases for which the scattering matrix is diagonal; that is, has
zero cross-coupling between states. We shall see that these represent the condi-
tion for maximum scattering cross-section of the object, and that for the special
case of backscatter the matrices [U ] and [V ] are simply related so that a single
antenna can be used to obtain this maximum backscatter signal. To see this,
and to establish an important connection between the SVD and optimization
based on Lagrange multipliers, consider solving the problem of maximising the
received scattered power. To do this we first establish a Lagrangian function L,
as shown in equation (1.143):

L = E ∗T ∗T ∗T
2 [S] E 1 + λ1 (E 1 E 1 − 1) + λ2 (E 2 E 2 − 1) (1.143)

Here we see in the first term the received complex voltage when we transmit a
state E 1 and receive with an antenna with polarisation E 2 . In addition we have
two constraint equations (and two corresponding unknown Lagrange multipli-
ers) to ensure that the antenna polarisation vectors E are unitary. To solve for the
maximum received power LL∗ we can sidestep the need to expand the product
directly and instead set the partial derivatives of L and L∗T separately to zero,
as shown in equation (1.144). (We shall use this same strategy when solving
coherence optimization in polarimetric interferometry in Section 6.2.)
∂L 
= [S] E + λ E = 0 

∂E ∗T
1 2 2 

2 [S]∗T [S] E 1 = λ∗1 λ2 E 1
⇒ (1.144)
∂L∗T 
 [S] [S]∗T E 2 = λ∗1 λ2 E 2
∗T ∗
= [S] E 2 + λ1 E 1 = 0 
∂E ∗T
1

Hence we see in the optimum case that E 1 is an eigenvector of the Hermitian

matrix [S]∗T [S], while E2 is an eigenvector of the matrix [S][S]∗T . In addition,
as these matrices are Hermitian they each have a pair of orthogonal eigenvec-
tors. These are just the left and right singular vectors of the [S] matrix, as easily
verified from equation (1.142). In the context of radar polarimetry, the Hermi-
tian matrix formed from products such as [S]∗T [S] is called the Graves power
matrix (Graves, 1956), as it contains information through its eigenvalues and
eigenvectors on the variation of scattered power by the object. Note that only
if [S] is symmetric are the matrices [S]∗T [S] and [S][S]∗T equal and the left
and right singular vectors equal. In this case the maximum scattered power is
formed from a copolarised signal (the same antenna for transmit and receive).
We shall return to this issue in the context of backscatter, but first turn to con-
sider some elementary symmetry properties of the general scattering amplitude
matrix.
One basic issue of importance in polarimetry is the identification of any
transformations of the problem in Figure 1.14, such that if we know the [S]
matrix for the original system then can we predict its form for the new prob-
lem without the need for recalculation. This operation would then constitute a
 1.4 The scattering of polarised waves 49

symmetry operation of the system. We have already seen one such example:
the problem of changing the incident wave polarisation. We have seen that we
can easily predict the polarisation of the scattered wave using C2 symmetry
coupled to matrix multiplication. However, this does not exhaust all possibil-
ities. There are several such symmetry transformations (van der Hulst, 1982;
Cloude, 1995b). As a trivial example we can translate the transmitter A or
receiver B along the vectors β i and β s and effect only a scalar phase change in
[S] → eiϕ [S], where φ is related to the change of range between A and B. This
forms the basis for a discussion of interferometry with polarised waves, to be
discussed at length in Chapter 6. Other interesting transformations are offered
by considering rotations and reflections of the original problem in the bisectrix
and scattering planes as we now consider.
We start by considering the matrix S (dispensing with square-bracket nota-
tion for the moment, for convenience) to be the ‘mother’ configuration, with
complex elements a, b, c, and d . There are then three important ‘daughter’
matrices we can write for completely different scattering problems, as shown
in equation (1.145):
  

 a −c

 Sα =

 −b d
  
  
a b a −b
S= ⇒ Sβ = a, b, c, d ∈ C (1.145)
c d 
 −c d

  



 Sγ =
a c
b d

These matrices arise from the following fundamental geometrical transforma-

tions (see Figure 1.14):

S α – Rotate the original problem by π about the bisectrix.

S β – Mirror the original problem with respect to the scattering plane.
S γ – Mirror the original problem with respect to the bisectrix plane.

The first of these, S α , follows from the vector reciprocity theorem for electro-
magnetic waves. This theorem relates the exchange of transmitter and receiver
positions, and states that if we transmit a polarisation state PA from A, then the
component polarised in the PB direction at B is equal to the PA component of
the scattered radiation when we illuminate the same object from B with polar-
isation PB . Figure 1.15 shows the reciprocal problem to that in Figure 1.14.

Scatterer

x'
x y'
y –bs

–bi
Scattered Β Incident
Α Bisectrix b
wave wave
Fig. 1.15 Definition of the reciprocal scatter-
ing problem
50 Polarised electromagnetic waves

Note how the incident and scattered wave vectors have changed direction. The
scattering matrices for the two cases are then related by the reciprocity theorem
as shown in equation (1.146):
0  1 
E AS = S β i β s · P A 

0  1 P B · E As = P A · E Bs


E S = S −β i − β s · P B Reciprocity Theorem
B

0  1 0  1T
⇒ S βi βs = S −β i − βs (1.146)

Note that the y and y’coordinates have been reversed in direction for the problem
in Figure 1.15 compared to the master problem in Figure 1.14. To account for
this we need to invert the y axis so that equation (1.146) becomes
   
  1 0  T 1 0
S βi βs = S −β i −β s
0 −1 0 −1
   
a b a −c
⇒S= ⇒ Sα = (1.147)
c d −b d

This leads directly to S α in equation (1.145). The other two daughter problems
are more easily derived: S β follows from an inversion of one coordinate, while
S γ follows from successive application of α and β.
The reciprocity relation is especially important for backscatter problems,
when the transmitter and receiver are at the same position. In this case it follows
that S = S α , and from equation (1.147) this is only possible if the original
‘mother’ scattering matrix has the following form:
 
a b
[S]backscatter = (1.148)
−b d

This is called the backscatter theorem, and demonstrates that in this special
case the S matrix has only three independent elements. Note that while there
exists a class of non-reciprocal backscatter problems, these are not common
in the remote sensing of natural land and sea surfaces (an important exception
being propagation through the ionosphere when Faraday rotation occurs; see
Section 1.2.2). The reciprocity theorem is therefore an important basic symme-
try, widely employed in radar backscatter systems for information extraction
and calibration.
It is of interest to now reconsider an SVD of the backscatter matrix of equation
(1.148) and to relate the problem to that of matched antenna illumination. To do
this we must first consider a correction to account for the differences between
transmitter and receiver antenna coordinates—a generalization of that following
from equation (1.147).

1.4.2 Back and forward scattering alignment

(BSA and FSA) systems
In the previous section we considered a description of the scattering matrix in a
coordinate system that in a sense ‘follows’ the wave propagation direction. This
is called the forward scattering alignment or FSA system. In radar and antenna
 1.4 The scattering of polarised waves 51

Incident Field Ei = Eθ θ +Eφφ

Voc = h T*.E i
Fig. 1.16 Complex effective length of an
h = hθ θ + hϕ ϕ complex effective length of the antenna antenna

studies, however, a different system is employed. This is called the backscatter

alignment or BSA system. In this section we motivate the need for such a
change and outline the main differences between FSA and BSA descriptions of
scattering (Kostinski, 1986).
When an antenna receives a signal it converts the incident electric field (with
units of Volts/m) into a circuit measurement in Volts (Collin, 1985; Mott, 1992).
The transfer function from field to circuit therefore has the units of length (m).
Given the C2 freedom of propagating waves, this transfer function must be
a two-component complex vector h, called the complex effective length of
the antenna. This basic antenna relationship is defined in Figure 1.16. If the
antenna and field are orthogonally polarised then zero voltage will be received,
regardless of the amplitude of the incident field. At the opposite extreme, when
field and antenna are matched, we have maximum voltage and optimum cou-
pling between antenna and field. In this section we consider the details of this
coupling process and how it relates to the structure of the scattering matrix.
An obvious starting point, common to vectors in C2, is to assume that the open
circuit voltage is given by the Hermitian product between receiving antenna
complex effective length h and the incident polarisation vector E i . This has
the desired property of being zero for orthogonality and a real maximum value
under the condition h = E i
The first problem we face is that in active systems antennas are used both for
transmission and reception, and we must therefore take care over the definition
of antenna polarisation vector h. In common with engineering standards we
define the antenna polarisation in transmit mode. Hence a +45-degree  linearly
1
polarised antenna has a normalized effective length h = √ 1
, defined
2 1
in a right-hand coordinate system as shown on the top line of Figure 1.17.
Now consider an experiment where two identical such antennas are used as
the basis for a communication system. Figure 1.17 shows a schematic of such
an arrangement, and here we see that, although identical, when used in trans-
mit/receive mode the antennas are orthogonal and the communication channel
is null. Clearly, in this case we cannot assume the matched condition as Ei = h.
The source of this problem can be traced to a mismatch in right-hand coor-
dinate systems for transmit and receive. We see that one component of the
receiver is 180-degree phase-shifted compared to that of the transmitter, due
to the z-coordinate reversal inherent in point-to-point communication. To com-
pensate for this we must introduce a minus sign or differential 180-degree
phase shift between the polarisation channels, as shown in Figure 1.17. With
this in place the Hermitian product between effective length and incident field
produces the correct null result. Since this phase correction is in the form of
a matrix product it can be incorporated into the [S] matrix definition. First,
however, we must also consider the fact that the waves propagate in different
directions, from which a further complication arises when dealing with elliptical
polarisations.
52 Polarised electromagnetic waves

1
1 h1 h2 h2 = [1 1]T
[1 1]
T 2
h1 =
2
Antenna 2
Antenna 1

Receiver
Transmitter

1 0  1 0 
Ei =  h ⇒ Voc = h*T   h1 = 0
*

0 −1 1 2 .E = h 2 
0 −1

1
[1 i]
T
1 T h2 =
h1 = [1 i] h1
h2 2
2

Antenna 1 Antenna 2

Transmitter Receiver

Fig. 1.17 Transmission and reception

1 0 
between +45◦ polarised antennas (top) Ei =
*T
 h1 ⇒ Voc = h 2 . E i = 0
and left-hand circularly polarised antennas 0 −1
(bottom), showing the need for coordinate 1 0  * *T
reversal and conjugation in point-to-point but E i =   h ⇒ Voc = h 2 . E i = 1
communication between antennas 0 −1 1

To illustrate the nature of the problem we again consider a communica-

tion link, established this time between two identical left circularly polarised
antennas. The lower part of Figure 1.17 shows the details of this arrangement.
In this case the 180-degree coordinate phase correction with Hermitian prod-
uct again leads to a predicted mismatch between antennas. However, this is
incorrect. Identical circularly polarised antennas remain matched when used in
point-to-point communication. The reason for this change of behaviour when
switching from linear to circular is that while the sense of circular polarisation
does change with the coordinate inversion, it also changes with time reversal.
This sense reversal is mathematically represented by the complex conjugate
operator applied to the incident field, since if P is a polarisation vector then P*
has the same polarisation ellipse, but with opposite sense.
With these coordinate transformations in place, the voltage received by an
antenna with effective length h2 when the scatterer is illuminated by a wave
originating from an antenna with complex length h1 , when both h1 and h2 are
defined in the transmitting mode of the antennas, can be written as shown in
equation (1.149):

   
∗ 1 0 a b
Voc = h∗T . . .h1 = hT2 [S]sensor h1 (1.149)
2 0 −1 c d

Consequently, it is common practice in radar—where backscatter problems

predominate and so the same antenna is used for transmission and reception—
to write the open circuit voltage in the backscatter alignment or BSA system as a
complex inner rather than Hermitian product of a coordinate reflected scattering
 1.4 The scattering of polarised waves 53

matrix, as shown in equation (1.150).

Voc = hT [S]sensor .h
       
1 0 1 0 a b a b
⇒ [S]BSA = [S]FSA = . =
0 −1 0 −1 c d −c −d
(1.150)

Also shown in equation (1.150) is the relationship between the scattering ampli-
tude matrix in the sensor coordinates (which becomes the backscatter alignment
system or BSA system in backscatter) and the wave or FSA coordinates. Note
the phase change of elements of [S]. Care is therefore required in always speci-
fying which coordinate system is being employed in a description of scattering.
In the engineering literature the sensor or BSA system is almost universally
used, while in optics and physics the conventional FSA or wave system is
preferred.
In equation (1.148) we saw that from the backscatter theorem, b = −c in
backscatter. It follows from equation (1.150) that the BSA matrix becomes
complex symmetric; that is, b = c, and the received backscatter voltage
for reciprocal problems can be written in the most general case, as shown
in equation (1.151):
 
T a b
Voc = h2 [S]BSA h1 = h2
T
h (1.151)
b −d 1

This is the most common form of this equation used in radar backscattering
theory. Some examples will be used to illustrate its application.

1.4.2.1 Example 1 : specular backscatter matrix Et E s = [R] E t

Consider the simple problem of normal incidence reflection from a flat surface,
as shown schematically in Figure 1.18. The reflected field is given by the Fresnel h
Ei
equations at normal incidence, which yield zero crosspolarisation and a 180-
degree phase difference between copolar coefficients (see Section 3.1.1). Hence Fig. 1.18 Specular backscatter from a surface
we can write the reflection matrix [Ref ] in the FSA wave coordinate systems as at normal incidence
shown in equation (1.152):
 
  1 0
Ref = A (1.152)
0 −1

where A is a constant that depends on the dimensions of the surface relative

to a wavelength and on the dielectric constant of the surface material. Note
that the matrix [Ref ] turns the reflection problem into an equivalent point-to-
point communication problem, and so the received voltage for transmission
and reception from an antenna with effective length h can be written as shown
in equation (1.153):
       
1 0 1 0 1 0 1 0
Voc = hT .E i = hT . E s = AhT . · E t = AhT . h
0 −1 0 −1 0 −1 0 1
(1.153)

Hence the scattering matrix in the sensor or BSA coordinates is a multiple of

the identity matrix, and we have the following important examples of antenna
54 Polarised electromagnetic waves

match and mismatch: namely, that circular polarisation yields zero received
voltage, while ±45-degree linear is matched for maximum signal strength (as
incidentally are all linear incident polarisations).
  
 1
  
h = ⇒ Voc = 0
√1
1 0 ±i 2
[S]BSA =A ⇒   (1.154)
0 1 
 1
h = √1 ⇒ Voc = A
2 ±1

B
1.4.2.2 Example 2: dihedral scattering matrix
Es = [R] E t
A second important way in which specular backscattering can occur is in dihe-
dral retroreflection, as shown schematically in Figure 1.19. In this case we have
a 90-degree corner formed from two perfect conductors (PC), and the ray paths
indicate a strong specular return to the transmitter through two 45-degree reflec-
A
tions at each surface. From a polarisation point of view the reflection matrix is
formed from the product of the reflection matrices at A and B. In general, the
Fig. 1.19 Dihedral retro-reflection Fresnel coefficients change with angle of incidence onto the surface (see Section
3.1.1), and away from normal incidence are not given by the simple form of
[Ref ] in equation (1.152). We shall consider the more general case later, but
here concentrate on perfect conductors only. For this special case, [Ref ] applies
for all angles, and using a cascade of matrices of the form of equation (1.152)
we can generate the normalized reflection matrix shown in equation (1.155):
     
      1 0 1 0 1 0
Ref = Ref Ref = . = (1.155)
A B 0 −1 0 −1 0 1

The voltage received for an antenna with effective length h in the BSA system
is then of the following form:
     
1 0 1 0 1 0
Voc = hT .E i = hT . · E t = hT · ·h (1.156)
0 −1 0 1 0 −1

and so the scattering matrix for a dihedral has the form shown in equation
(1.157):
  
1

  
h = ⇒ Voc = A
√1
1 0 ±i 2
[S]BSA = A ⇒   (1.157)
0 −1 
 1
h = √1 ⇒ Voc = 0
2 ±1

We now see that a circularly polarised antenna is matched to the return signal,
while a 45-degree linearly polarised antenna is mismatched, rather like the
communication system of Figure 1.17. Clearly the pattern is now set for higher-
order scattering processes. The BSA scattering matrix for order N scattering
from a set of PC surfaces is denoted as follows:
 2N +1
1 0
[S]BSA = (1.158)
0 −1
 1.4 The scattering of polarised waves 55

Fig. 1.20 Image of a trihedral corner reflec-

tor used for polarimetric calibration

so for N odd (1,3,5 and so on) there is zero phase difference between copolar
elements, while for N even (2,4,6 and so on) there is π phase difference. The
generalization of these ideas to dielectric surfaces is dealt with in Section 3.1,
and leads us eventually to use of the scattering alpha parameter to accommodate
the changes in boundary conditions for general surfaces.
As an important example of equation (1.158), consider the case N = 3, which
occurs for a trihedral corner reflector, widely used for radar backscatter cali-
bration, as it combines high radar cross-section with a broad radiation pattern,
so simplifying alignment with the sensor. (Contrast this with a flat plate N = 1,
which can always be made bright by increasing its size but only at expense of the
radiation pattern, which becomes more localized around the specular direction
in a narrow pencil beam.) Figure 1.20 shows an image of a metallic trihedral N
= 3 reflector. From a polarisation point of view, this scatterer behaves the same
as specular normal surface reflection (N = 1), and has a BSA scattering matrix
equal to the 2 × 2 identity matrix. Hence it too is ‘blind’ to circularly polarised
antennas, regardless of its physical size.

1.4.3 Singular value analysis of the scattering matrix

In the previous section we saw that the boundary conditions at perfectly con-
ducting surfaces lead to a set of ‘blind polarisations’ for even and odd number
of reflections (circular for odd and 45-degree linear for even). In order to extend
these ideas to arbitrary materials and shapes we now turn to consider a gener-
alization of blind polarisations through a singular value decomposition (SVD)
of the backscattering matrix in the sensor or BSA coordinate system. To do
this we note that when we change the polarisation basis of the incident field by
a unitary matrix U , then the open circuit voltage can be written as shown in
equation (1.159), from which we see that the backscattering matrix transforms
not as a similarity transformation but as a unitary congruent transformation
(involving only the transpose and not the transpose conjugate) (Luneburg,
1996).

Voc = hT [U ]T [S]sensor [U ] h ⇒ [S] = [U ]T [S]sensor [U ] (1.159)

56 Polarised electromagnetic waves

On the other hand, for complex symmetric matrices the two unitary components
[U ] and [V ] of the SVD are related, since we have the following identity:

[S] = [U ] . [] . [V ]∗T = [S]T = [V ]∗ . [] . [U ]T

⇒ [U ] = [V ]∗
⇒ [S] = [U ] . [] . [U ]T (1.160)

and so we see from the SVD that a symmetric [S] matrix can always be diagonal-
ized by a congruent unitary transformation, just as we found for the scattering
matrix in the sensor coordinates (equation (1.159)). Some examples will help
illustrate how this formalism can be used in practice. Firstly we consider the
simple case of rotation of the reference plane of polarisation for backscattering
by a horizontal dipole and a dihedral reflector. The scattering amplitude matri-
ces for these two cases can be derived from standard similarity transformations
involving a plane rotation matrix, as shown in equation (1.161):
       
cos θ − sin θ 1 0 cos θ sin θ cos2 θ cos θ sin θ
[Sθ ] = . . =
sin θ cos θ 0 0 − sin θ cos θ cos θ sin θ sin2 θ
       
cos θ − sin θ 1 0 cos θ sin θ cos2 θ − sin2 θ 2 cos θ sin θ
[Sθ ] = . . =
sin θ cos θ 0 −1 − sin θ cos θ 2 cos θ sin θ sin2 θ − cos2 θ
(1.161)

Some care is required, however, when considering change to elliptical polarisa-

tion bases. As an important special case we consider the form of the scattering
matrix in the circular polarisation base, obtained by a congruent unitary trans-
formation of the linear [S] matrix, as shown in equation (1.162). Note how the
left and right side transformation matrices are now identical. The reason for
this is the coordinate manipulations inherent in the BSA system.
     
SLL SLR 1 1 i SHH SHV 1 i
[S]circ = =
SRL SRR 2 i 1 SHV SVV i 1
 
1 SHH − SVV + 2iSHV i(SHH + SVV )
= (1.162)
2 i(SHH + SVV ) SVV − SHH + 2iSHV

For example, consider again the rotated dihedral matrix now expressed in the
circular base (substitute the lower example in equation (1.161) into (1.162)).
The result is shown in equation (1.163). Now we see a much-simplified diagonal
form with rotation influencing only the phase between LL and RR polarisations.
This is an example of diagonalization via the SVD.

 
cos2 θ − sin2 θ + i2 sin θ cos θ 0
[S]circ =
0 sin2 θ − cos2 θ + i2 sin θ cos θ
   i2θ 
cos 2θ + i sin 2θ 0 e 0
= =
0 − cos 2θ + i sin 2θ 0 −e−i2θ
(1.163)
 1.4 The scattering of polarised waves 57

P
XPOL null
COPOL null
P⊥ Fig. 1.21 Schematic representation of the
Scatterer S XPOL and COPOL nulls

The main conclusion is that for any given symmetric complex scattering matrix
we can always find an orthogonal base ‘ab’ such that [S] is diagonal in that
base; that is, there is zero crosspolarisation. Incidentally, this is reminiscent
of eigenpolarisations in wave propagation (Section 1.2.2), except here we
are dealing with the physically distinct case of wave scattering. These states
are sometimes confusingly called eigenstates, but should more accurately be
called crosspolar nulls, or often abbreviated to XPOL nulls of the scattering
matrix, as originated by Kennaugh (1952) and further developed by Huynen
(1987) and Boerner (1981). Note that for symmetric matrices there are always
two orthogonal XPOL nulls. They have two very interesting properties, as
follows:
• The XPOL nulls maximize the backscattered signal (for a fixed antenna
polarisation), and thus maximize the detectability of the object in the
presence of noise.
• The XPOL nulls remain unchanged on scattering from the object, and
thus ‘carry’ information about any symmetry properties of the object.
Figure 1.21 shows a schematic representation of this invariance property of
the XPOL nulls. By definition, when P is an XPOL null then the scattered field
is also P polarised (in the BSA coordinates) as shown. For example, when the
scatterer has an axis of symmetry in the plane of polarisation then the XPOL
nulls are linear polarisations aligned and perpendicular to this symmetry axis. In
this way we can establish the orientation of an unknown symmetry axis by using
an SVD of the measured scattering matrix. But how does this diagonalization
process link with the idea of blind polarisations? To see this, we now seek inci-
dent polarisation states that are not invariant but rather transformed into their
orthogonal state on scattering. These states would then by definition constitute
the blind polarisations of the system. In distinction to the diagonalization pro-
cess, these states are called copolar nulls or simply COPOL nulls, and we shall
now show that there are always two such states for all [S] matrices, although
unlike XPOL nulls they are only orthogonal under special circumstances. The
trick is to solve for COPOL nulls starting from the diagonal form of the [S]
matrix, as follows. Recall that the XPOL nulls are always orthogonal (forming
a base we shall call ‘ab’) and so we can always write the scattering matrix in
this base in diagonal form, as shown in equation (1.164):
 
γ 0
[S] = a (1.164)
0 γb

where γa and γb are the complex singular values of [S]. We can represent
any polarisation state as a linear combination of these XPOL nulls, that is,
E = Ea a + Eb b, and hence for COPOL nulls we seek solutions of the form
58 Polarised electromagnetic waves

shown in equation (1.165):

   
  γ 0 E
Ea Eb . a . a =0
0 γb Eb
⇒ Ea2 γa + Eb2 γb = 0
Ea γb
⇒ =± − (1.165)
Eb γa
This shows that there are two COPOL nulls (with a complex polarisation ratio
given by the plus and minus square root), and that generally they are not orthog-
onal to each other. As simple examples we return to the matrices for specular
and dihedral scattering (equations (1.154) and (1.157)). These are both diagonal
in the given coordinate system, and their singular values are therefore easily
obtained by inspection. The results for the COPOL nulls are then given as
shown in equation (1.166), which we see agrees with our test cases considered
in equations (1.154) and (1.157). Equation (1.165) then generalizes this result
to arbitrary scattering matrices.
 
1 0 copol nulls Ex
[S] = −−−−−−→ = ±i
0 1 Ey
  (1.166)
1 0 copol nulls Ex
[S] = −−−−−−→ = ±1
0 −1 Ey
Finally, we note that there are three important invariants of the scattering ampli-
tude matrix under unitary congruent transformations, as shown in equation
(1.167):

  
 det([S]) = ad − bc I


a b invariants under UT SU
[S] = −−−−−−−−−−−−→ Span([S]) = |a| + |b| + |c| + |d | II
2 2 2 2
c d 


 i(c − b) III
(1.167)
The first follows directly from the use of special unitary matrices in establishing
the change of basis matrix. The second allows us to define the norm or magnitude
of a scattering matrix, and will be used to define scattering amplitude vectors.
The third assures us that if the scattering matrix is symmetric in one base, it
remains so in all bases. To prove this, consider an expansion of the amplitude
matrix in terms of the complete and orthogonal Pauli matrix set, as shown in
equation (1.168), where the square root of two normalization is included to
 2
ensure that k  = Span([S]) from invariant II.
!
3
1
[S] = (ko σ0 + k1 σ1 + k2 σ2 + k3 σ3 ) ⇒ [S] = ki σi ki = √ Trace([S] σi )
i=0
2
(1.168)
It follows from properties of the Pauli matrices that we can form the identity
[U ]T [S] [U ] = [U ]T (ko σ0 + k1 σ1 + k2 σ2 + k3 σ3 ) [U ] ⇒ [U ]T σ3 [U ] = σ3
(1.169)
which leads to invariant III shown in equation (1.167).
 1.4 The scattering of polarised waves 59

Propagation Scattering Propagation

A-B at B B-A
Fig. 1.22 Three-stage decomposition of
Stage I Stage II Stage III combined wave propagation and scattering

1.4.4 Combined wave propagation and

scattering effects
Now we can combine the Jones propagation calculus of Section 1.2.4 with the
above scattering matrix results to obtain a general expression for the scattering
matrix of an object observed in the presence of non-trivial wave propagation
effects. This is important in many practical applications, such as radio wave
propagation through the ionosphere for space-based radar Earth observation
(Bickel, 1965) and scattering from objects embedded in chiral materials such
as occur in light scattering by biological systems (Ablitt, 1999, 2000).
To develop this we consider a three-stage process as shown schematically
in Figure 1.22. The transmitter is located at A, and the first stage involves
propagation of the wave from A to the object at B. This is represented by a 2 × 2
complex propagation matrix [Mz (A,B)] which, as we saw in Section 1.2, needs
be neither unitary nor homogeneous:

Stage I : E B = [Mz (A, B)] E A

The second stage involves scattering of the wave by the target. This can be
represented by a 2 × 2 complex scattering matrix [S], which we express in the
sensor or BSA coordinate system as follows:

Stage II : E BS = [SBSA ] E B = [SBSA ] . [Mz (A, B)] E A

The third stage involves propagation back through the medium from B to A.
As we are remaining in the A sensor coordinate system then the reciprocity
theorem implies that [Mz (A,B)] = [Mz (B,A)]T (equation (1.146)), so that we
can write the following:

Stage III : E As = [Mz (B, A)] .E Bs = [Mz (A, B)]T . [SBSA ] . [Mz (A, B)] E A

Consequently we can write the general combination of propagation and

scattering in the BSA system as shown in equation (1.170):

[S]observed = [Mz ]T [SBSA ] [Mz ] (1.170)

This is our desired result. It shows that the observed scattering matrix for
a combined reciprocal propagation and scattering channel is, in the sensor
BSA coordinate system, again obtained as a congruent transformation of the
target scattering matrix by the propagation channel matrix. This represents a
generalization of the change of basis, which resulted in a unitary congruent
transformation.
Having now obtained these basic algebraic results for the complex amplitude
matrix, its form under change of base and its fusion with propagation, we now
turn to consider a geometrical interpretation of these transformation properties.
60 Polarised electromagnetic waves

1.5 Geometry of the scattering matrix

In Section 1.3 we showed how all states of wave polarisation can be mapped
onto the surface of the Poincaré sphere. We saw that there are three possible
coordinate systems used to represent polarisation state on this sphere—two
angle-based and one Cartesian—leading to the Stokes four-vector represen-
tation. Here we consider the implications of these results for a geometrical
interpretation of the scattering amplitude matrix [S].
In general terms the scattering matrix represents a polarisation state trans-
former. Each possible input state, represented by a point on the sphere, is
mapped into a second point with contracted radius (assuming the target is pas-
sive and has no inherent signal gain). The details of this mapping are of interest
as a stage towards information extraction from the scattering matrix. For exam-
ple, we shall find that some points are fixed under this transformation and that
their relative location on the sphere is an indicator of target symmetry.

1.5.1 Polarisation signatures

The first stage of development is to be able to map variations in scattered
amplitude with movement of points over the Poincaré sphere. This can be
achieved in several ways, all of which require map projections of the spherical
coordinates onto a plane. Three important such methods are summarized in
Figure 1.23 (van Zyl, 1987; Woodhouse, 2003). On the far left we show the
polarimetric signature method, in the centre a polar projection of the sphere,
and on the right a hybrid polar system that combines the advantages of the other
two into a convenient form for visualization.
One of the first such methods developed was simply to map the surface into
a bounded rectangular region, as shown on the left in Figure 1.23. Each point
inside the rectangle is then mapped into a transmit/receive antenna configu-
ration and the resulting scattered power obtained from the voltage formula of
equation (1.151). A common configuration is to choose a copolar representa-
tion, where the same point on the Poincaré sphere represents the transmit and
receive antenna polarisation states (in the BSA convention). A second popular
choice is to map the crosspolar channel variation. Here a point on the sphere
is chosen for transmit and its antipodal partner for reception. There are two
main problems with this rectangular representation. The first is the artificial
discontinuity introduced in the θ coordinate by using a rectangular region. The
second is the spreading of the circular polarisation poles of the sphere into

Left Right
handed handed
θ
2θ

Cos2τ
π π
− ≤ 2τ ≤
2 2

Fig. 1.23 Planar mapping representations of

the surface of the Poincaré sphere −π < 2θ < π θ
 1.5 Geometry of the scattering matrix 61

extended linear regions. This distorts the information for points located away
from the equator. One advantage, however, is the clear separation of left and
right sense polarisations into upper and lower halves of the diagram. This per-
mits easy visualization of asymmetry due to calibration errors or scattering from
asymmetric objects such as helices.
To overcome the distortion limitation, use is often made of a polar projection
of the sphere, as shown in the centre of Figure 1.23. Here the pole is located at the
centre of the image, and the polar θ coordinate is represented in a more natural
way, without the need for any artificial discontinuities. However, one remaining
problem is that the double angle representation of the Poincaré sphere leads to
upper and lower hemispheres being mapped on top of each other. Hence there is
no discrimination in this diagram between left and right sense polarisations. One
compromise solution is to employ the representation shown on the right-hand
side of Figure 1.23. Here we maintain the polar representation with circular
polarisation at the centre, but now map θ in the range −π/2 to π/2. This frees
the second half of the polar plot for the opposite sense of polarisation. In this
way we can map both senses and all linear states onto a simple polar diagram
(Woodhouse, 2003).
Figure 1.24 shows examples of all three representations for a set of sam-
ple matrices. The first three are the backscattering matrices corresponding to

–80
–60
–40
2*Tau (degrees)

1 0 –20

[S] = 0

0 1 20
40
60
80
–150 –100 –50 0 50 100 150
2*Theta (degrees)

–80
–60
–40
1 0
2*Tau (degrees)

–20
[S] = 0

0 –1 20
40
60
80
–150 –100 –50 0 50 100 150
2*Theta (degrees)

–80
–60
–40

0 1
2*Tau (degrees)

–20

[S] = 0

1 0 20
40
60
80
–150 –100 –50 0 50 100 150
2*Theta (degrees)

–80
–60
–40
2*Tau (degrees)

–20
0
1 i
[S] = 20
40
i –1 60
80
–150 –100 –50 0 50 100 150
2*Theta (degrees)

–80
–60
–40
2*Tau (degrees)

–20
0
1 –1
[S] = 20
40
–1 1 60
80
–150 –100 –50 0 50
2*Theta (degrees)
100 150
Fig. 1.24 Copolarised backscatter power
signatures for canonical scattering matrices
62 Polarised electromagnetic waves

the Pauli set. The first of these is just the identity matrix, and we can see the
characteristic copolar nulls at left and right circular polarisations. Maximum
response is obtained for all linear polarisations, and there is no left/right depen-
dency of scattered power. This represents the most symmetric of scattering
systems, showing dependency on neither orientation nor sense of polarisation.
The second and third Pauli matrices demonstrate rotation dependence. Here the
copolar nulls are located at 45-degree and 0-degree linear polarisations respec-
tively. Maximum backscatter is now achieved for a set of elliptical polarisations
lying along a great circle of the sphere, passing through the poles correspond-
ing to left and right circular polarisation. Again, however, we note that there
is no left/right dependence on backscatter. These clearly show how the third
Pauli matrix can be obtained from the second by coordinate rotation through 45
degrees. The next scattering matrix shown is that for a helix. This is chosen to
illustrate an object that scatters one sense of circular polarisation in preference
to the other. The asymmetry in the copolarised plot is clearly seen in the first and
third diagrams. Finally we show the scattering matrix for a dipole oriented at 45
degrees to the horizontal. The orientation of the scatterer is clearly seen directly
in the third representation. These final two examples demonstrate the benefit of
the hybrid representation for visualising both rotation and sense preference in
scattering.
Of particular interest are the extreme values of these scattering functions.
These can be analytically evaluated with the help of a singular value decomposi-
tion of the [S] matrix. We showed in Section 1.4.3 that for an arbitrary scattering
matrix there exist two pairs of orthogonal states for which the scattered power is
maximized (XPOL nulls). We further showed that in backscatter these maxima
always correspond to copolar combinations and will therefore appear in the
polarisation charts of Figure 1.24. Also, there are two (non-orthogonal) zero
minimum points which for backscatter also correspond to zeros in the copolar
function. In backscatter, therefore, copolar and crosspolar plots are useful for
assessing the maximum variation of scattering with changes in polarisation.
However, in the general non-symmetric scattering matrix case, as arises in
bistatic scattering, such functions are not guaranteed to contain global maxima
and minima.

1.5.2 The polarisation fork

S
The extreme points in Figure 1.24 correspond directly to the XPOL and COPOL
nulls. These four special points on the Poincaré sphere then form the basis for
a geometrical interpretation of the general transformation of polarisation state,
called the polarisation fork (Kennaugh, 1952; Huynen, 1987). The two XPOL
2aw nulls are maxima (minima) in the copolar (crosspolar) signature plots and,
P Q always being orthogonal, correspond geometrically to antipodal points lying at
–2g I
opposite ends of a diameter of the Poincaré sphere. These are shown as P and Q
in Figure 1.25. We showed in equation (1.165) that the COPOL nulls are then
obtained in the basis defined by PQ from the square root of the ratio of complex
singular values of the scattering matrix. The square root having two solutions,
R
there are then two of these, displaced symmetrically about the line PQ. These are
Fig. 1.25 Polarisation fork geometry shown as R and S in Figure 1.25.To see this we note that the unitary polarisation
 1.5 Geometry of the scattering matrix 63

vectors for R and S can be written as shown in equation (1.171):

 
cos αw γP
PR,S = ⇒ tan αw eiδ = ± − (1.171)
sin αw eiδ PQ γQ
 
where γ are the complex singular values |γP | ≥ γQ ) and the α w angle is
defined as shown in Figure 1.25. Note that αw ≥ π2 , and so very often the
angle 2γ = π − 2αw is employed. γ is called the fork angle of the scatter-
ing matrix (Kennaugh, 1952; Huynen, 1970, 1987). We can then rewrite the
diagonal scattering matrix as a function of the fork angle, as shown in equation
(1.172):
   χ 
i
γ 0 iφ cot γ e 2 0
[S] = P = me χ (1.172)
0 γQ 0 tan γ e−i 2

The angle between the singular vectors, χ , is called the skip angle. Note that
the matrix on the right-hand side has unit determinant. This will provide an
important link with the geometry of the Lorentz transformation in the next
section.
We note from this analysis that the four points P, Q, R, and S all lie in a
plane in the space of the Poincaré sphere. This structure is called the polari-
sation fork, and was first developed by Kennaugh (1952) and later studied by
several authors, particularly Huynen (1970, 1987). Consequently, the angu-
lar parameters of this plane and fork geometry are often termed the Huynen
parameters. Kennaugh also devised a simple set of geometrical rules for pre-
dicting the transformation of polarisation state by a scattering matrix. The rules
employ the point I shown as the intersection of the chord joining the COPOL
nulls and the diameter PQ in Figure 1.25. For a given scattering matrix, the
transformation of any incident polarisation can then be found by first invert-
ing the point through I onto the sphere. This new point is then rotated by π
about the diameter normal to PQ. The coordinates of the new point then cor-
respond to the polarisation of the scattered wave. Applying these rules to the
points P and Q themselves confirms that they remain unchanged on scatter-
ing, while R and S are both mapped into their antipodal orthogonal points as
expected.
Two special classes of scattering can be defined based on the α angle. When
the singular values are equal in amplitude, then αw = 45 degrees, the COPOL
nulls are orthogonal, and I is located at the origin. Then, due to the symme-
try, all polarisation states lying on a great circle formed by rotating the fork
about the RS axis are also XPOL nulls. This is the case for the three Pauli
matrices shown at the top of Figure 1.24, where the loci of maximum scat-
tered power are seen as bands of white across the diagrams. This arises as the
Pauli matrices have singular values that differ only in phase but are equal in
amplitude.
The second important class occurs when the scattering matrix is singular
(zero determinant) and consequently one of the singular values is zero. In this
case αw = 90 degrees, R,I,S collapse to the point Q, and all polarisation states
are transformed into the state P, regardless of their initial position on the sphere.
64 Polarised electromagnetic waves

Such a device is a polariser, in that it scatters the same polarisation state for all
input states. Important examples are shown in the lower two sections of Figure
1.24. The dipole and the helix both have singular scattering matrices, and in
both cases the fork collapses to a line.
Note that a similar construct can be used for the case of non-symmetric
scattering matrices. The complex singular values are still well defined, and
thus the fork angle is also uniquely defined. The axis PQ now corresponds
to the left singular vectors and is again a diameter of the sphere, as P and Q
remain orthogonal. Hence the fork geometry is still maintained. However, now
the scattered wave must be interpreted in a new P Q orthogonal base defined
by the right singular vectors. Hence we have a fork for transmit and a second
fork for reception, although the fork angle remains invariant.

1.5.3 Lorentz geometry and the scattering matrix

The Poincaré formulation is based on the geometry of the unit sphere, and thus
ignores any changes in scattered amplitude. The copolar signatures of Figure
1.24 are generated essentially by using a form of Malus’s law with a metric
of cos2 φ/2 (Hecht, 1997). Here φ is the angular separation of the initial and
final polarisation states on the surface of the Poincaré sphere. However, this
works only in a relative sense because we have defined reference points on the
surface (the original point for copolar signatures plus the antipodal point for
crosspolar). In other words, we must choose cross- or copolar projections for
the signature. It would be better if we could devise a method of interpretation
that did not require such a choice.
In more general terms we would like to devise a geometrical interpretation
of the change in both amplitude and polarisation state. Such an approach can be
developed based on the geometry of the Lorentz transformation. Just as there
is an homomorphism between the group of special unitary matrices and that
of proper rotations, there is also a homomorphic (2-to-1) mapping from the
group of 2 × 2 complex unimodular matrices [Q] and pure Lorentz transfor-
mations (see Appendix 2 and Goldstein (1980)). In this section we explore the
geometrical implications of this mapping for an interpretation of the scattering
matrix.
We begin by again considering the diagonal matrix of singular values of [S].
We can rewrite this in unimodular form as shown in equation (1.173), where we
also show explicitly the polar decomposition of [Q] into unitary and Hermitian
matrices [U ] and [H ] respectively.
 
γ 0
[S] = P = a[Q] = a · [U ] · [H ]
0 γQ
 χ   
ei 2 0 cot γ 0
=a χ ·
0 e−i 2 0 tan γ
 χ   
ei 2 0 eb 0
=a χ · (1.173)
0 e−i 2 0 e−b

This is important, because it is the Hermitian component that leads directly

to the new type of geometry. We saw in Section 1.3.3 that in order to include
 1.5 Geometry of the scattering matrix 65

amplitude information we need to extend the Pauli set by addition of the identity
matrix and obtain a modified spin matrix. To see how matrices of the form [Q]
change this spin matrix we form the triple matrix product shown in equation
(1.174):
 
1 g0 + g1 g2 − ig3
[R (χ, τ )] = [Q] · · [Q]∗T (1.174)
2 g2 + ig3 g0 − g1

If we set the skip angle χ to zero and determinant to unity in equation (1.174),
this transformation has the following simple canonical form:
       
g0new + g1new g2new − ig3new 1 eb 0 g0 + g1 g2 − ig3 eb 0
= . .
g2new + ig3new g0new − g1new 2 0 e−b g2 + ig3 g0 − g1 0 e−b
(1.175)

By direct expansion it follows that we can relate the Stokes vectors of the
incident and scattered field by a 4×4 matrix [M ], as shown in equation (1.176):
 new     
g0 m00 m01 m02 m03 g0
g new  m10 m11 m12 m13  g1 
 1new  =    
g  m20 m21 m22 m23  . g2 
2
g3new m30 m31 m32 m33 g3
   
cosh(2b) sinh(2b) 0 0 g0
 sinh(2b) cosh(2b) 0 0 g1 
=  .   (1.176)
0 0 1 0 g2 
0 0 0 1 g3

There are two important consequences of this result:

1. The matrix [M ] is always real.
The elements of [M ] are always real, despite the fact that the singular values
of [S] are complex. This we can show in two stages: in the first case equation
(1.175) generalises to transformation by the full Hermitian component of the
polar decomposition [H ], as shown in equation (1.177):
 
1 g + g1 g2 − ig3
[R] = [H ] · 0 · [H ]∗T ⇒ [M ] mij ∈  (1.177)
2 g2 + ig3 g0 − g1

The left-hand side is then always itself Hermitian, and so the Pauli expansion
coefficients are real and hence all the elements of [M ] must be real. Finally, by
adding the unitary matrix component of the polar decomposition [U ] we obtain
a spin transformation of the form shown in equation (1.178):
 
1 g0 + g1 g2 − ig3
[R] = [U ] · [H ] · · [H ]∗T [U ]∗T
2 g2 + ig3 g0 − g1
   
1 0 1 0
⇒ [M ] = T · [M ] · T (1.178)
0 [O3 ] 0 [O3 ]T

However, we have already shown that the effect of a complex special unitary
transformation maps into a real 3 × 3 rotation matrix [O3 ] that acts on the
Cartesian components of the Stokes vector (equation (1.127)). Hence the uni-
tary matrix component transforms [M ], as shown in equation (1.178), where
66 Polarised electromagnetic waves

Q g0 P
g0 = –g1 new
g0 = g1
g0

d
new
g1
Fig. 1.26 Minkowski diagram for canonical
g1
transformation of the Stokes vector

0 = (0, 0, 0) is a null vector. However, [M ] remains real, and hence we have

shown that for an arbitrary complex scattering matrix [S], there exists a corre-
sponding 4 × 4 real matrix [M ] which transforms the Stokes vectors as g new =
[M ] g. This will be developed further in Section 2.2 concerning the Mueller
matrix.
2. The matrix [M ] represents a Lorentz ‘boost’ in the direction of the
XPOL nulls.
Although the scattering amplitude matrix can always be made diagonal
by using SVD, we have shown in equation (1.176) that the [M ] matrix is
never diagonal. We see that even in its most canonical form it has two off-
diagonal elements. We now give a geometrical interpretation of this as a Lorentz
transformation of the Stokes vector.
The transformation in equation (1.176) does not represent a plane rotation,
but a Lorentz ‘boost’ in the subspace spanned by g0 and g1 . This can be made
clear by reference to Figure 1.26. Here we show the g0 ,g1 plane. The lines at ±45
degrees represent the loci of all possible states P and Q—the left quadrant being
Q polarised and the right P polarised. These two lines form the analogue of the
light cone in special relativity. They represent pure states of wave polarisation,
but with varying amplitude. Note that points which lie below the lines represent
the condition g1 > g0 . As g0 represents the total intensity of the wave it is not
possible to satisfy this condition for plane waves. Hence this region is non-
physical, and corresponds to light speeds greater than c in special relativity.
But what about the region above the lines? Here we have the condition g0 >
g1 . This can be allowed if we choose, for example, some subspace of the
signal to represent g1 . This occurs with wave depolarisation arising from partial
coherence between the states (see Chapter 2).
In the extreme case g1 = 0, and we have noise signals represented by the
vertical axis and showing no preference for P or Q polarisation. We note that
the geometric effect of [M ] on the coordinate system is to distort it by an angle
δ as shown. Hence the set of incident waves defined by a g1 /g0 ratio lying along
the line at δ degrees are transformed on scattering into waves with g1new = 0;
that is, into noise and vice versa. Borrowing terminology from special relativity
this represents a Lorentz ‘boost’ in the direction PQ in polarisation space; that
is, the space of the Poincaré sphere and not physical space. The magnitude of
the ‘boost’ depends only on the ratio of singular values of [S] or fork angle γ
such that the angle δ is given by

1 − tan4 γ
tan δ = tanh 2b = (1.179)
1 + tan4 γ
 1.6 The scattering vector formulation 67

Geometrically, therefore, the pure states P and Q remain unchanged (they are
the XPOL nulls of the scattering matrix and so by definition are invariant). This
corresponds directly to the invariance of the speed of light in special relativity.
However, the coordinates of other points do change, and we shall see in Chapter
2 that this has important physical interpretation in terms of depolarisation.
In summary we have seen that we can extend the idea of transformation of
polarisation basis by a unitary matrix to transformation of polarisation state by
scattering. In the unitary change of base we can write the general change of
base as a matrix exponential function (equation (1.180)), exposing as it does
the elementary geometry of the Poincaré sphere:

E = [U2 ] E = exp(−iθn · σ )E = (cos θ σ0 − i sin θn · σ )E (1.180)

where σ are the set of three Pauli matrices, and n is a real three-vector defining
the axis for rotation through θ in R3—a real three-dimensional space of the
Stokes vector. The most general form of transformation of E that accounts
not only for change of base but also polarisation dependent scattering, can be
written in the form shown in equation (1.181):

E = [S] E = exp((δp − iθn) · σ )E (1.181)

where δ, θ are real, and p and n are three-vectors. For the important case of
symmetric scattering matrices (as occurs for backscatter in the sensor coordi-
nates, for example) it follows that n = (1, 0, 0), in which case the most general
scattering matrix may be written in geometrical form, as shown in equation
(1.182):

[S] = exp((δ cos a − iθ )σx + µ cos bσy + µ cos cσz ) (1.182)

which corresponds to a Lorentz boost of magnitude δ in a direction in Stokes

space given by Euler angles a, b and c, defining the fork axis PQ in Figure 1.25.

1.6 The scattering vector formulation

In the previous section we were concerned with an interpretation of the form
of the scattering matrix based on the extreme points of the amplitude scattering
function. We saw, however, that the scattering matrix can be characterized
using an alternative approach based on its transformation properties. In this
section we start from this idea to develop the scattering vector characterization
(Byrne, 1971; Cloude, 1985). This will lead us to consider several important
new ideas such as orthogonality of matrices and generalized rotation invariance,
and lead us to develop a new method for characterising scattering matrices in
general.
68 Polarised electromagnetic waves

The vectorization of a matrix is the straightforward process of expanding the

matrix using a set of simpler basis elements. The coefficients of this expansion
then form a vector representation. To start with a simple but important example,
we consider a straightforward lexicographic expansion as shown in equation
(1.183):
         
x y 1 0 0 1 0 0 0 0
[S] = =x +y +w +z (1.183)
w z 0 0 0 0 1 0 0 1

One way to consider this expansion is as a representation of the information

in the scattering matrix in terms of a basic set of canonical scattering mecha-
nisms, each represented by the simpler matrices shown on the right-hand side.
In this regard the basis elements for the lexicographic expansion have a par-
ticular physical significance as electric dipole and dihedral scatterers, ‘x’ as
horizontal, ‘z’ as vertical and the combination of y+w as a 45-degree dihedral
(see equation (1.161)). With this idea in mind, we have seen that the set of
Pauli spin matrices have important canonical interpretation in terms of generic
specular and dihedral scattering. This motivates the following so-called Pauli
expansion of the [S] matrix as follows:
 
1 a + b c − id
[S] = √
2 c + id a − b
       
a 1 0 b 1 0 c 0 1 d 0 −i
=√ +√ +√ +√ (1.184)
2 0 1 2 0 −1 2 1 0 2 i 0

where the square-root factor is used to keep the total scattered power constant;
that is, |x|2 + |y|2 + |w|2 + |z|2 = |a|2 + |b|2 + |c|2 + |d |2 . Note that we can
relate the vectors in the two systems—lexicographic and Pauli—by a 4 × 4
unitary matrix transformation as shown in equation (1.185):
    
x 1 1 0 0 a
y 1 0 0 1 −i b
 = √    (1.185)
w 2 0 0 1 i  c
z 1 −1 0 0 d

It is then straightforward to see how we may generalize this idea of expansion

to include any set of four ‘complete’ matrices (‘complete’ in the sense that any
2 × 2 complex matrix must be able to be represented by the set). In general we
can formalize the vectorization of [S] by defining a Pauli scattering vector k as
shown in equation (1.186):
   
  SXX + SYY a
SXX SXY 1  SXX − SYY  
b
 
[S] = ⇒k= √  =
SYX SYY 2 SXY + SYX   c 
i(SXY − SYX ) d
  √
= |k|w , k  = aa∗ + bb∗ + cc∗ + dd ∗ (1.186)

and then representing the effect of using a basis set other than the Pauli
matrices by a 4 × 4 unitary matrix transformation of the Pauli coefficients
 1.6 The scattering vector formulation 69

(a generalization of equation (1.185)), as shown in equation (1.187):

k = [U4 ] k (1.187)

Note, as a special case, that the effect of change of coordinates from wave
(FSA) to sensor (BSA) on the Pauli vector can be represented by such a unitary
matrix, as shown in equation (1.188):
   
x y x y
[S]FSA = ⇒ [S]BSA =
w z −w −z
 
0 1 0 0
 1 0 0 0 
⇒ k BSA =
 0
k (1.188)
0 0 −i  FSA
0 0 i 0

1.6.1 Scattering mechanisms

As shown in equation (1.186), it is always possible to normalize the scattering
vector by its amplitude to generate a complex unitary vector w. This normalized
vector w we call a ‘scattering mechanism’, as it can be used to characterize
differences in polarised wave scattering (see Chapter 3). In particular it is the
transformation properties of this vector that we will now study. Note that the
concept of a scattering mechanism conveniently scales with dimension of the
problem. The only constraint is that w must always have unit amplitude. We
can start with N = 1; that is, scalar fields characterized by a single complex
number. In this case w is a phase term eiφ ; that is, the set of transformations
are those associated with phase shifts of the scattered field. In two dimensions
we can combine phase shifts with amplitude ratios and still keep w unitary, as
shown in equation (1.189). A physical interpretation of the two parameters αw
and φ2 − φ1 in terms of coordinates on the Poincaré sphere was developed for
polarised waves in Section 1.3. Key for our discussion in terms of the scattering
matrix is the extension of this idea to three, four and higher dimensions. These
can be constructed by extension as shown in equation (1.189):
 
cos αw eiφ1
w1 = eiφ w2 =
sin αw eiφ2
 
  cos αeiφ1 (1.189)
cos αeiφ1  sin α cos ψeiφ2 
w3 = sin α cos ψeiφ2  w4 = 
sin α sin ψ cos γ eiφ3 

sin α sin ψeiφ3
sin α sin ψ sin γ eiφ4

As we did for wave states using the Poincaré sphere, we now need to develop a
physical interpretation of the parameters involved in these higher dimensional
vectors. We begin by recognising that in general, any pair of sets of four basis
matrices for the expansion of [S] can be related in a smooth and continuous
way. To move from one set to another we employ unitary transformations (see
Appendix 2 for a discussion of general unitary transformations). Mathemati-
cally we can then relate the vector of coefficients in one basis k to those in
the second basis set k by a unitary matrix transformation, itself expressed as a
70 Polarised electromagnetic waves

matrix exponential function, so forming a natural multidimensional extension

of the simple scalar case, as shown in equation (1.190):


 [U1 ] = eiφ

[U ] = eiφn.P
2
k = [UN ]k → (1.190)

[U ] = eiφn.G
 3

[U4 ] = eiφn.D

Here we have used the notation P for the set of three Pauli matrices, G for the set
of eight Gell–Mann matrices, and D the sixteen Dirac matrices (see Appendix
2 for definitions of these matrix sets). These compact representations will be
useful when we turn to consider depolarisation effects in Chapter 2.
Depolarisation and
scattering entropy 2

This chapter is concerned with the process of wave depolarisation and its formal
description using the algebraic and geometrical tools developed in Chapter 1.
Depolarisation is inherently a stochastic process, and somewhat destroys the
transfer of vector information from source to far field. Methods for quantifying
this loss of information will be developed based on the concept of scattering
entropy, which we shall see provides a unifying concept of depolarisation across
many types of polarisation problems. InAppendix 3 we outline the basic features
of multivariate statistical signal analysis required for a description of coherence
and its relation to depolarisation in polarimetric studies. Readers not familiar
with these basic stochastic concepts should consult the Appendix for more
details.
The first important point is to draw a distinction between wave depolarisation
and crosspolarisation. These are often confused in the literature, and yet have
important and subtle distinguishing characteristics, as we now demonstrate.
The term ‘crosspolarisation’ refers to any process in wave propagation and
scattering that causes coupling between orthogonal states of polarisation. This
process can be deterministic or stochastic in nature. For example, backscatter
from a dipole oriented at θ degrees has a scattering matrix as shown in equa-
tion (2.1). Here we can see a level of crosspolarisation given by cos θ sin θ. As
a second example, consider wave propagation though a half-wave plate (that is,
a lossless birefringent plate with a thickness that generates a 180-degree phase
shift between its eigenpolarisations at the design wavelength λ). When such a
plate is oriented at θ degrees to the wave coordinates, the Jones matrix has the

Cross-polarising systems

Depolarising
systems

Fig. 2.1 The set of depolarising systems as a

subset of cross-polarisation
72 Depolarisation and scattering entropy

form shown in equation (2.2).

   
cos θ − sin θ 1 0 cos θ sin θ
[S] =
sin θ cos θ 0 0 − sin θ cos θ
 
cos2 θ cos θ sin θ
= (2.1)
cos θ sin θ sin2 θ
   
cos θ − sin θ 1 0 cos θ sin θ
[M2 ] =
sin θ cos θ 0 −1 − sin θ cos θ
 2 
cos θ − sin2 θ 2 cos θ sin θ
= (2.2)
2 cos θ sin θ sin2 θ − cos2 θ

Here again we see a coupling of energy between modes, this time given by
sin 2θ . Such an element is widely used, for example, to rotate the polarisation
of a laser beam.
The key feature relating these two examples is that there exists some deter-
ministic matrix transformation that can be used to remove the crosspolarisation.
Hence neither of these systems depolarise the incident wave.
Depolarisation is therefore a coupling of energy from deterministic into
stochastic modes of the field. It is connected with reversibility of crosspo-
larisation processes. It can be considered a noise-generating process, although
as we shall show, information can still be extracted from the stochastic modes.
Hence as shown schematically in Figure 2.1, all depolarising systems cause
crosspolarisation but not vice versa. This chapter is concerned with systems
that lie in the shaded region of Figure 2.1.
Note that here we define a depolariser as a system that causes noise genera-
tion when illuminated by a pure polarised wave (waves of the form shown in
equation (1.133)). This avoids any of the semantic issues raised, for example,
in Hovenier (2004) concerning the definition of depolarisation when consider-
ing illumination by partially polarised waves. We have already seen in Section
1.5.3 that the Lorentz transformation makes it possible even for simple point
scatterers, characterized by a single amplitude matrix [S], to both increase and
decrease the coherence of partially polarised waves according to the geometry
of the Lorentz boost. However, the conservation of zero wave entropy means
that such systems are here characterized as polarising. We now turn to consider
systems where such a conservation law no longer applies.

2.1 The wave coherency matrix

In Figure 1.11 and accompanying discussion of polarisation geometry we estab-
lished that even though plane waves are dynamic in nature, the geometrical
parameters of the polarisation ellipse are invariant with time. However, in scat-
tering and propagation through random media the geometry of the ellipse can
become a function of space/time due, for example, to motion of particles in the
scattering medium or to coherent fluctuations associated with speckle. Under
such circumstances the ellipse itself becomes a dynamic quantity. It is then
of interest to see if there exists some set of average parameters that can be
used, together with some measure of spread or variance, to describe the wave
 2.1 The wave coherency matrix 73

state. Note that here we limit attention to quasimonochromatic waves—those

for which the spectrum of fluctuations ω is small compared to ω. For an
introduction to the topic of wideband polarimetry when this is no longer true,
see Nelander (1995).
To quantify the stochastic nature of wave depolarisation, we first select an
orthogonal polarisation base x,y, and then form averages over all possible prod-
ucts of complex field components. This is achieved by defining a 2 × 2 wave
coherency matrix [J ] (Born and Wolf, 1989) from the outer product of polar-
isation vectors, as shown in equation (2.3), where from equation (A3.9) it
follows that |j12 |2 ≤ |j11 | · |j22 | and so det([J]) ≥ 0.
+ , + ,  
+ ∗T
, Ex Ex∗ Ex Ey∗ j11 j12
[J ] = E.E = + , + , = ∗ = [J ]∗T (2.3)
Ey Ex∗ Ey Ey∗ j12 j22

This provides an important link with Lorentz geometry via the Minkowski
metric, as discussed in Chapter 1. Hence [J ] is a positive semidefinite (PSD)
Hermitian matrix; that is, its eigenvalues are both real and non-negative. It has
an associated real quadratic form defined as shown in equation (2.4):
w∗T [J ] w ≥ 0 (2.4)
where w is a two-element unitary complex wave vector. We can now use this
matrix to obtain a general expression for the coherence between signals in
arbitrary p and q polarisation channels. We start by defining the two desired
receiver channels on the Poincaré sphere from their unitary w vectors, as shown
in equation (2.5):
   
cos αP cos αQ
wP = w = (2.5)
sin αP eiδP Q sin αQ eiδQ
The received signals in these components are then obtained by projection onto
the incident E vector. The ‘PQ’ coherence can then be obtained as shown in
equation (2.6):
. + ,
SP = w∗TP E SP SQ∗
⇒ γPQ = + ,+ ,
SQ = w∗TQ E SP SP∗ SQ SQ∗
+ ,
∗T w
(2.6)
w∗T
P E.E Q w∗T
P [J ] w Q
 + , + , =
w∗T ∗T w .w ∗T E.E ∗T w w∗T ∗T
P E.E P Q Q P [J ]w P .w Q [J ]w Q

Hence [J ] contains all the information required to synthesize the coherence for
arbitrary polarisation channels p and q. For this reason it is termed the wave
coherency matrix, and forms the central focus for a quantitative treatment of
wave depolarisation. Note that if p = q then the coherence is always unity as
expected (we then compare a signal channel with itself). Often p and q will be
orthogonal, when for example we wish to measure the full polarisation state of
a wave. In this case we note that if we choose an orthogonal w pair such that
the following condition applies:
[J ] wQ = λwQ (2.7)
—that is, as an eigenvector of [J ]—then from equation (2.6) the coherence will
always be zero. Hence the eigenvalue problem of [J ] is linked to the coherence
74 Depolarisation and scattering entropy

properties of the wave. To pursue this idea further, we note that for a change of
polarisation base we have E 2 = [U2 ] E, and the new coherency matrix can then
be easily derived as shown in equation (2.8):
+ , + ,
[J ]2 = E 2 .E ∗2 = [U2 ] E.E ∗T [U2 ]∗T
+ ,
= [U2 ]. E.E ∗T [U2 ]−1 = [U2 ].[J ].[U2 ]−1 (2.8)

It follows from the properties of Hermitian matrices that we can then always
find a base such that [J ] is diagonal, as shown in equation (2.9):
 
∗T
  λ 0  ∗T
[J ] = [U2 ] . [D] . [U2 ] = w w⊥ . 1 . w w⊥ λ1 ≥ λ2 ≥ 0
0 λ2
(2.9)

Here w is the eigenvector corresponding to the maximum eigenvalue λ1 , and

[U2 ] is a 2×2 unitary matrix formed with columns generated by the (orthogonal)
eigenvectors. The eigenvector is then associated with a state P on the Poincaré
sphere. Note that this state is not uniquely defined, as any component of [U ]
obtained by exponentiation of the Cartan sub-algebra of SU(2) (Cornwell, 1984;
Georgi, 1999) will be removed by the eigenvector decomposition of equation
(2.9) (see Appendix 2). Since SU(2) has a one-dimensional Cartan sub-algebra,
these ‘hidden’ variables are obtained for phase transformations of the form
shown in equation (2.10):
 
eiφ 0
[U ]cartan = (2.10)
0 e−iφ

which, as we have shown in Section 1.3.2, corresponds to the absolute phase

of the polarised wave. Hence the eigenvectors of [J ] have an arbitrary phase,
and equation (2.8) remains invariant to such changes. This is just a formal
way of saying that the coherency matrix contains information only about rel-
ative phase angles or phase differences between channels. Hence [J ] has four
independent parameters: two real eigenvalues and two angles on the Poincaré
sphere locating the point P. In our discussion of polarised plane waves we saw
that three parameters were sufficient to represent a state of elliptical polari-
sation. Hence [J ] contains one new parameter that is associated with loss of
coherence or depolarisation. To establish a general notation for this type of
parameter matching we anticipate generalization of [J ] and write a general
N × N coherency matrix in the following form (Cloude, 2001a):

[C] = (E + L) + [E + L] (2.11)

where (..) are the polarised components, [..] the depolarising parameters, and
for each, E are those parameters associated with the eigenvectors and L the
eigenvalues. From the general structure of N × N coherency matrices we then
have the following constraints (see Appendix 2):
• L + (L) = N
• E + (E) = dim(SU(N)) − −rank(SU(N))
 2.1 The wave coherency matrix 75

With this notation in place we can now write the wave coherency matrix in the
following compact form:

[J ] = (2 + 1) + [0 + 1] (2.12)

It follows from this that in order to consider the effects of wave depolarisation (N
= 2) we need only consider diagonal matrices, as the depolarisation parameter
lies entirely in the eigenvalue spectrum and not in the eigenvectors. We shall
see this is not true for higher dimensional problems. This special case of N = 2
then leads to a further important result known as the wave dichotomy.

2.1.1 The wave dichotomy

In a sense, the eigenvector decomposition states that for any wave we can find
an orthogonal base for which the numerator in equation (2.6), and hence the
coherence, is zero. Two special cases will illustrate extreme forms of this result.
If the signal is noise-like then its coherency matrix will have the diagonal form
shown in equation (2.13):
 
λ 0
[J ]noise = (2.13)
0 λ

In this case the coherence in equation (2.6) will be zero for arbitrary choice of
orthogonal p and q; that is, the signal is indistinguishable from noise, there is no
variation of coherence with polarisation, and depolarisation is complete. The
whole signal is characterized by only a single parameter: the noise variance, λ.
This represents the most extreme case of polarisation ‘memory loss’.
At the other extreme, if we consider a pure polarised wave then det([J ]) = 0
(see Section 1.3.4), and the matrix must then have one zero eigenvalue; that is,
[J ] is of the form shown in equation (2.14), where λ is in this case the squared
amplitude of the polarisation ellipse:
 
λ 0
[J ]polarised = (2.14)
0 0

In this case the coherence is unity for all choices of p and q, except when
p corresponds to the polarisation state itself. In this singular case the signal
component in the orthogonal channel is by definition zero and hence the coher-
ence function is no longer defined, and the numerator and denominator will
now both be zero. In this case the signal has zero depolarisation and is purely
deterministic in nature.
In general, of course, real signals lie between these two extremes, with
a coherence that depends on the choice of polarisation base. These examples
illustrate why the coherence is itself a bad way to characterize the depolarisation
properties of the signal. Not only does it vary with polarisation state but it
can also be singular for the case of pure polarised waves. We require instead
a basis invariant regular parameter to account for the level of depolarisation
in the signal. This can be defined in several ways—for example, from the
ratio of eigenvalues—by defining the degree of polarisation Dp as shown in
76 Depolarisation and scattering entropy

equation (2.15):
λ1 − λ2
Dp = 0 ≤ Dp ≤ 1 (2.15)
λ1 + λ 2
This function is by definition invariant to changes of polarisation base, and so
represents a fundamental property of the signal. Like coherence magnitude it
lies between 0 and 1; when Dp = 0 we have a noise signal, while Dp = 1 for
pure polarised waves.
This parameterization also allows us to derive a decomposition of the
coherency matrix into ‘polarised’ and ‘depolarised’ components by writing an
arbitrary [J ] as the sum of a polarised wave plus noise, as shown in the top row
of equation (2.16):
   
λ − λ2 0 1 0
Wave Decomposition 1 [J ] = [U ] . 1 . [U ]∗T + λ2
0 0 0 1
Wave Decomposition 2 [J ] = λ1 w.w∗T + λ2 w⊥ .w∗T
⊥
(2.16)

However, there is a second eigenvalue factorization that is equally valid. Instead

of considering a noise + signal type decomposition, we can alternatively decom-
pose the problem into an independent mixture of ‘coherent states’, as shown in
the lower row of equation (2.16). This existence of a pair of statistical models
for characterising depolarisation is known as the wave dichotomy. While the
noise + signal approach led to Dp as a measure of depolarisation, the mixture
of states approach leads to a second measure, namely the wave entropy Hw .
Entropy in general is defined as the distribution of probabilities across a set of
physical states (Wiener, 1930; Wolf, 1954; O’Neill, 1991; Brosseau, 1998).
To calculate the entropy in this case we need to generate the probabilities of
the two states in the model. As they are independent this is very easy, as the
probability of each state is just given by its normalized eigenvalue, so that in
general we can write the two probabilities of the orthogonal states as shown in
equation (2.17).
λi
Pi = 5 0 ≤ Pi ≤ 1 (2.17)
λ
Pi can be considered the probability that the state represented by the ith eigen-
vector will occur. With this definition in place, the entropy can then be defined
as shown in equation (2.18):

!
2
Hw = − pi log2 pi 0 ≤ Hw ≤ 1 (2.18)
i=1

When the entropy is zero we have zero uncertainty as to the state of polarisation;
that is, a coherence of one and a purely polarised wave. At the other extreme,
when the entropy is one then we have maximum uncertainty as to the state of
polarisation and a noise signal with a coherence of zero.
The solution to the wave dichotomy remains a matter largely of choice in
wave problems, but as we shall see for higher order coherency matrices the
 2.1 The wave coherency matrix 77

entropy approach generalizes in a straightforward way while the signal + noise

methods are plagued by ambiguities. This can be attributed to the special case
of N = 2, which has all its depolarisation effects contained in the eigenvalue
spectra (equation (2.12)).
Returning for a moment to the signal + noise approach and the degree of
polarisation, one key feature of this method is its relationship to the Stokes
vector representation of polarised waves (see Section 1.3.4). Even in the pres-
ence of depolarisation we can still expand the coherency matrix in terms of four
Stokes parameters to obtain a Stokes vector as shown in equation (2.19):

+ , + 
, 
Ex Ex∗ Ex Ey∗ 1 g0 + g1 g2 − ig3
[J ] = + , + , =
Ey Ex∗ Ey Ey∗ 2 g2 + ig3 g0 − g 1
 + , + , 
Ex Ex∗ + Ey Ey∗
 + , + , 
 Ex E ∗ − E y E ∗ 
 x y 
⇒g= + , +
 Ex E ∗ + E y E ∗
, 
 (2.19)
 y x 
 + , + , 
i Ex Ey∗ − Ey Ex∗

where now

1 2
det([J ]) = λ1 λ2 = (g − (g12 + g22 + g32 )) ≥ 0
4 0 (2.20)
Trace([J ]) = λ1 + λ2 = g0 > 0

It follows from equations (2.15) and (2.20) that

5
3
!
3
2
2
i=1 gi
gi2 = (λ1 − λ2 ) ⇒ Dp = (2.21)
g0
i=1

and so the degree of polarisation can be written as a ratio of two scalars derived
directly from the Stokes vector. The first—the g0 element—is just the total
wave intensity, while the second is the Euclidean norm of a three-vector in
the space of the Poincaré sphere. This suggests a parallel decomposition of the
Stokes vector into polarised and depolarised components, as shown in the top
row of equation (2.22). Here the first component represents a polarised wave
and a point on the Poincaré sphere. The second represents a noise signal, which
has zero everywhere except in the first element. This corresponds directly to
wave decomposition 1 in equation (2.16). The wave dichotomy means we can
also write an arbitrary Stokes vector as the sum of two orthogonal states. The
components of the two-state decomposition can be easily derived as shown in
78 Depolarisation and scattering entropy

the lower portion of equation (2.22).

   5 
53 3
 i=1 g 2
i   g 0 − i=1 gi
2

 g1   0 
Wave Decomposition 1 g =  + 
 g2   0 
g3 0
 5

 g0 + 3 2

 i=1 gi
g01 =
Wave Decomposition 2 25

 3

 g0 − i=1 gi
2
 g02 =
2
   
1 1
 √5g1   √5 −g1 
 2  3 2
 i=1 gi   i=1 gi 
3

⇒ g = g01    −g2 
 √5g32 2  + g02  √5  (2.22)
 i=1 gi
  3 2
i=1 gi
 g   
√533 2 −g3
√5 3
i=1 gi 2
i=1 gi

In summary, we have seen that there are several ways to characterize depolarised
waves. Firstly we note that while depolarisation is associated with a loss of
coherence, it is better to characterize it on the basis of the eigenvalue spectrum
of the wave coherency matrix rather than in terms of the coherence function
itself. Arising from this is the wave dichotomy, whereby the eigenvalues can
be used either as part of a signal + noise model which leads to the degree
of polarisation, or as a mixture of states model which leads to the definition
of wave entropy. Both have a simple representation in terms of the Stokes
vector.
Entropy can be viewed as an information destroying process, and with this in
mind it is interesting to characterize the various ways in which entropy can be
increased or decreased by scattering. The simplest way to increase entropy is to
consider the addition of thermal noise to a signal. Entropy can also be increased
at source, as is the case for passive remote sensing when the source is thermal
radiation. However, for active sensors the source usually has zero entropy by
design. Consequently, any change in entropy must be caused by environmental
scattering and propagation. We now turn to consider a detailed analysis of such
effects.

2.2 The Mueller matrix

The starting point for an analysis of the effects of wave scattering on depolari-
sation is to consider the wave coherency matrix obtained after wave scattering
[Js ] compared to that before, [Ji ]. Using equation (2.3) and the definition of the
scattering amplitude matrix [S] we obtain the following relationship between
the two coherency matrices, where [S] is an arbitrary 2 × 2 complex amplitude
matrix:

[Js ] = [S] . [Ji ] . [S]∗T (2.23)

 2.2 The Mueller matrix 79

Clearly [S] will have the effect of changing the eigenvalues and eigenvectors of
[J ], and hence will change the degree of wave depolarisation. By expanding the
matrix product in equation (2.23) and collecting terms of [J ] to form elements
of the corresponding Stokes vectors g i and g s we can rewrite equation (2.23)
as a linear transformation between vectors, as shown in equation (2.24):

g s = [M ] .g i (2.24)

The matrix [M ] is called the Mueller matrix, after Hans Mueller, an MIT
physicist who first derived the properties of such matrices (but seems to have
published very little in the open literature). Parallel development of a similar
approach can be found in the contemporary work of Perrin (1942).
We have already developed a geometrical interpretation of equations of the
form shown in equations (2.23) and (2.24). In equation (1.177) we showed that
the elements of [M ] are all real, despite the fact that [S] is complex. This can
be confirmed by explicit derivation of the elements of [M ] from [S], as we now
demonstrate. We can reformulate equation (2.23) in terms of Stokes vectors by
using the Kronecker matrix product (see Appendix 1), since for any set of three
matrices we can write an expansion as shown in equation (2.25):

[A][X ][B] ≡ [A] ⊗ [B]T x (2.25)

where x is a lexicographic row vectorization of the elements of [X ]:

   
x11 x12 .... x1n x11
x21 x22 .... x2n  x12 
 
[X ] =  .. ⇒x=
 : 
 (2.26)
 : : . : 
xn1 xn2 .... xnn xnn

Using this result, we can now write the vectorized wave coherency matrix of
the scattered wave as shown in equation (2.27):

   
  aa∗ ab∗ ba∗ bb∗ Ex Ex∗
 ac∗ ad ∗ bc∗ bd ∗   ∗
[S] =
a b
⇒ xs = [Y ] xi =   . Ex Ey∗ 
c d  ca∗ cb∗ da∗ db ∗  Ey Ex 
cc∗ cd ∗ dc∗ dd ∗ Ey Ey∗
(2.27)

We note that the properties of the scattered wave are determined by a 4 × 4

matrix of quadratic products of the elements of [S], arranged into a complex
matrix [Y ] as shown. In order to obtain the elements of the Mueller matrix [M ],
relating incident to scattered wave Stokes vectors, we now need to transform
the vector x in equation (2.27) into the Pauli basis using the transformation
80 Depolarisation and scattering entropy

matrix shown in equation (2.28):

     
Ex Ex∗ + Ey Ey∗ 1 0 0 1 Ex Ex∗
 Ex Ex − Ey Ey  1
∗ ∗
−1  ∗ 
g=   0 0  . Ex Ey 
 Ex Ey∗ + Ey Ex∗  = 0 1 1 0  Ey Ex∗ 
i(Ex Ey∗ − Ey Ex∗ ) 0 i −i 0 Ey Ey∗
(2.28)
     
Ex Ex∗ 1 1 0 0 Ex Ex∗ + Ey Ey∗
Ex Ey∗  1 0 i   ∗ ∗ 
⇒   0 1  .  Ex Ex∗ − Ey Ey∗ 
Ey Ex∗  = 2 0 0 1 −i  Ex Ey + Ey Ex 
Ey Ey∗ 1 −1 0 0 i(Ex Ey∗ − Ey Ex∗ )

Combining equations (2.27) and (2.28) we finally obtain an expression for the
matrix [M ] in terms of the elements of [S] and [Y ], as shown in equation (2.29):
 
m11 m12 m13 m14
m21 m22 m23 m24 
[M ] = m31 m32 m33 m34  mij ∈ 


m41 m42 m43 m44

   ∗  
1 0 0 1 aa ab∗ ba∗ bb∗ 1 1 0 0
1 1 0 0 −1  ∗ ∗ ∗ ∗ 
 .  ac∗ ad∗ bc ∗ bd ∗  0 0 1 i 

= .     
2 0 1 1 0 ca cb da db 0 0 1 −i
0 i −i 0 cc∗ cd ∗ dc∗ dd ∗ 1 −1 0 0
(2.29)

which can be expanded to yield an explicit mapping between any given [S]
matrix and the sixteen real elements of [M ], as shown in equation (2.30):
 
aa∗ + bb∗ + cc∗ + dd ∗ aa∗ − bb∗ + cc∗ − dd ∗ 2Re(ab∗ + cd ∗ ) 2Im(ab∗ + cd ∗ )
 ∗ 
1 aa + bb∗ − cc∗ − dd ∗ aa∗ − bb∗ − cc∗ + dd ∗ 2Re(ab∗ − cd ∗ ) 2Im(ab∗ − cd ∗ )
[M ] =  
2
 2Re(ac ∗ + bd ∗ ) 2Re(ac∗ − bd ∗ ) 2Re(ad ∗ + bc∗ )

2Im(ad ∗ − bc∗ )
−2Im(ac∗ + bd ∗ ) −2Im(ac∗ − bd ∗ ) −2Im(ad ∗ + bc∗ ) 2Re(ad ∗ − bc∗ )
(2.30)

2.2.1 Properties of the Mueller matrix

Matrices of the type shown in equation (2.30)—those that can be written in the
general form [M ] = f ([S])—have several interesting properties that distinguish
them from general 4 × 4 real matrices (Abhyankar, 1969; Barakat, 1981, 1987;
Kim, 1987; Cloude, 1989; Anderson, 1994; Hovenier, 1994, 2004). These are
often referred to as ‘pure’Mueller matrices in the literature. Here we summarize
their key properties and highlight their significance in terms of the scattering
of polarised waves (Mishchenko, 2000).

Mueller Property M1: if x = |det([S])| then x2 = m211 − m221 − m231 − m241

Mueller Property M2: |Tr([S])|2 = Tr([M ]) so the trace of [M] is always
non-negative
Mueller Property M3: x4 = det([M]) so the determinant of [M] can never be
negative
Mueller Property M4: if [S] → [M ] then [S]−1 → [M ]−1 if x = 0
 2.2 The Mueller matrix 81

Mueller Property M5: [S1 ] . [S2 ] → [M1 ] . [M2 ]

Mueller Property M6: α [S] → |α|2 [M ] α ∈ C. This has important implica-
tions for the reversibility of the mapping from [S] to [M ],[S] = f −1 ([M ]),
itself an important element in determining the level of depolarisation caused
by scattering.
 
1 0 0 0
0 1 0 0 
Mueller Property M7: [S]T → [4 ] [M ]T [4 ] [4 ] = 0 0 1 0 


0 0 0 −1
Mueller Property M8: [S]∗T → [M ]T  
1 0 0 0
0 1 0 0 
Mueller Property M9: [S]∗ → 4 [M ] 4 4 =  0 0 1 0 


0 0 0 −1
Mueller Property M10:

   
a b a −b
[S] = → [M ] ⇒ → [34 ] [M ] [34 ]
c d −c d
 
1 0 0 0
0 1 0 0
[34 ] = 
0 0 −1 0 


0 0 0 −1

Mueller Property M11:


  
a b d b
[S] = → [M ] ⇒ → [2 ] [M ]T [2 ]
c d c a
 
1 0 0 0
0 −1 0 0
[2 ] = 
0 0 1 0


0 0 0 1

Mueller Property M12:

 
−1 1 d −b 1
[S] = → 2 [234 ] [M ]T [234 ]
det([S]) −c a x
 
1 0 0 0
0 −1 0 0
[234 ] = 
0 0 −1 0 


0 0 0 −1

and so the inverse of [M ] can be written as follows

1
[M ]−1 = [234 ][M ]T [234 ]
x2
82 Depolarisation and scattering entropy

Mueller Property M13:

   
a b a −c
[S] = → [M ] ⇒ → [3 ] [M ]T [3 ]
c d −b d
 
1 0 0 0
0 1 0 0
[3 ] = 
0 0 −1 0 .


0 0 0 1

The relation M13 is important for application of the reciprocity theorem to scat-
tering problems (van de Hulst, 1981; Saxon, 1955). Property M12 in particular
leads to two important further results, summarized in equation (2.31):
8
1 [M ]T [234 ][M ] = x2 [234 ]
[M ]−1 = 2 [234 ] [M ]T [234 ] ⇒ (2.31)
x Tr([M ]T [234 ][M ]) = −2x2

These relations were first derived by Barakat (1981) and Simon (1987). Finally,
an important simple and widely used relationship between the elements of the
matrix [M ], called the Fry–Kattawar relation (Fry, 1981) is shown in equation
(2.32):

!
4 !
4
m2ij = 4m211 (2.32)
i=1 j=1

Although this relation is widely used in the literature, care must be taken in
its application. For example, there exist matrices that satisfy this relation but
do not have a corresponding single [S] matrix, and so for which the reverse
mapping does not exist. A simple example is diag(1,1,1,–1). Hence equation
(2.32) is a necessary but not sufficient condition for [M ] to correspond to a
single [S] (Hovenier, 1996, 2004).
All of these relations were developed with a view to establishing conditions
for the reversibility of the relation [S] = f −1 ([M ]). In general, [M ] = f ([S])
changes the degree of polarisation of the wave, but if the incident wave entropy
is zero (a purely polarised wave) then the scattered wave entropy is also zero.
This is termed the ‘conservation of zero wave entropy’, and is analogous to
the conservation of light speed in special relativity (see Section 1.5.3). This
property has to do with the question of reversibility of the mapping from [S]
to [M]; that is, [M ] = f ([S]) ⇒ [S] = f −1 ([M ])? We can use property M6
to find the conditions for the inversion to be unique, since z [S] → |z|2 [M ]
for all complex z, so [S] ⇒ [M ] ⇒ [SR ] is possible, but only where [SR ] is
the relative phase scattering matrix (that is, the scattering amplitude matrix can
be reconstructed from [M ], but only up to an arbitrary phase term). This helps
establish conditions for uniqueness, but leaves the thorny issue of existence
to be resolved. There is no guarantee that given a general 4 × 4 real matrix
it will have a corresponding single [S] matrix representation. Before dealing
with existence conditions, however, we turn to consider the special case of
backscatter and the implications of the BSA coordinate system for the Mueller
matrix.
 2.2 The Mueller matrix 83

2.2.2 The backscatter Mueller and Stokes

reflection matrix
We saw in Chapter 1 that the case of backscatter is of particular interest when
the structure of [S] and therefore [M ] simplifies, as we now show. Following
from Chapter 1, there are two versions of the backscatter [S] matrix, depending
on the coordinate system employed (wave and sensor or FSA and BSA coor-
dinates). Consequently there are two versions of [M ] to consider. Equation
(2.33) summarizes the backscatter form of [M ] for the wave coordinate system
(FSA). Equation (2.34) shows the corresponding form for the BSA coordinates.
Note that in the BSA system the matrix is real symmetric. The BSA form of
this matrix was first derived by Edward Kennaugh, at Ohio State University,
in parallel with the optical developments by Mueller and Perrin, and so it is
often termed the Kennaugh matrix, or sometimes simply the Stokes reflection
matrix, in the radar literature (Kennaugh, 1952).
In both cases we see that the sixteen elements of [M ] are reduced via the
reciprocity theorem for backscatter to only ten (actually nine; see equation
(2.38)). As [S] has only a maximum of six independent elements for backscatter,
then clearly the elements of [M ] are not all independent. Hence for a 4 × 4 real
matrix [M ] to correspond to a complex amplitude matrix [S]—to guarantee
existence of the inverse mapping [S] = f −1 ([M ])—the elements of [M ] must
satisfy some additional constraint equations:
 
a b
[S] =
−b d
 
aa∗ + 2bb∗ + dd ∗ aa∗ − dd ∗ 2Re(ab∗ − bd ∗ ) 2Im(ab∗ − bd ∗ )
 
1  aa∗ − dd ∗ aa∗ − 2bb∗ + dd ∗ 2Re(ab∗ + bd ∗ ) 2Im(ab∗ + bd ∗ )
⇒ [M ] =  
 
2  2Re(bd ∗ − ab∗ ) 2Re(−ab∗ − bd ∗ ) 2Re(ad ∗ − bb∗ ) ∗
2Im(ad ) 
2Im(ab∗ − bd ∗ ) 2Im(ab∗ + bd ∗ ) −2Im(ad ∗ ) ∗
2Re(ad + bb )∗

 
m11 m12 m13 m14
 
 m12 m22 m23 m24 
=

 (2.33)
 −m 13 −m 23 m33 m34 
m14 m24 −m34 m44
 ∗
a b
[S] =
b −d
 
1 0 0 0
 
1 0 1 0 0
⇒ [M ] =  
2 0 0 1 0
0 0 0 −1
 
aa∗ + 2bb∗ + dd ∗ aa∗ − dd ∗ 2Re(ab∗ − bd ∗ ) 2Im(ab∗ − bd ∗ )
 
 aa∗ − dd ∗ aa∗ − 2bb∗ + dd ∗ 2Re(ab∗ + bd ∗ ) 2Im(ab∗ + bd ∗ ) 
·



 2Re(ab∗ − bd ∗ ) 2Re(ab∗ + bd ∗ ) 2Re(bb∗ − ad ∗ ) ∗
2Im(−ad ) 
−2Im(ab∗ − bd ∗ ) −2Im(ab∗ + bd ∗ ) −2Im(−ad ∗ ) ∗
−2Re(ad + bc )∗

 
m11 m12 m13 m14
 
m12 m22 m23 m24 
=


 (2.34)
m13 m23 m33 m34 
m14 m24 m34 m44
84 Depolarisation and scattering entropy

Furthermore, there exists the possibility of formulating a set of Mueller matrices

that do not correspond to a single [S] matrix at all. The most extreme example
of these is the isotropic depolariser, with a matrix of the form shown in equation
(2.35):
 
1 0 0 0
0 0 0 0
[M ] =  0 0 0 0
 (2.35)
0 0 0 0

This matrix converts all Stokes vectors into a randomly polarised wave, hence its
name. However, there is no corresponding single [S] matrix. Clearly, therefore,
our analysis of depolarisation must go beyond relations like equation (2.32)
A and look more closely at the structure of Mueller matrices in general.
B The set of Mueller matrices [M ] form a subset of all possible 4×4 real matri-
C ces. Hence not all real matrices correspond to Mueller matrices and, because
of depolarisation like that shown in equation (2.35), of those that do there is
no guarantee that the inverse mapping to a single [S] will exist. This situation
is summarized graphically in Figure 2.2. It is of practical interest, therefore,
to determine two sets of conditions on a given matrix [M ]—firstly to test if it
A – the set of 4 × 4 real matrices corresponds to a Mueller matrix at all; and secondly, if it does, whether there is
B – the set of Mueller matrices
C – the set of Mueller matrices
a corresponding single amplitude matrix [S]. By doing this we may then estab-
for which [S] = f –1([M]) lish existence of the inverse mapping, and at the same time classify the different
possible types of depolarisation. A full analysis of this problem follows Section
Fig. 2.2 The set of Mueller matrices as a 2.5, but here we make some general observations about the related topic of the
subset of real 4 × 4 matrices
Stokes criterion.

2.2.3 The Stokes criterion

An important set of constraints follows directly from equation (2.24) and the
requirement that both sides must correspond to physical Stokes vectors, even in
the presence of depolarisation. The so-called ‘Stokes criterion’ has been studied
by many authors (Barakat, 1981, 1987; Fry, 1981; Gil, 1985, 1986; Cloude,
1989; Girgel, 1991; van der Mee, 1992, 1993; Givens, 1993; Hovenier, 1994,
1996), and defines valid Mueller matrices such that they satisfy the set of general
mathematical conditions shown in equation (2.36):


 g0s ≥ 0, g0i ≥ 0

2 − g2 − g2 − g2 ≥ 0
g s = [M ] g i ⇒ g0s (2.36)


1s 2s 3s
 2
g0i − g1i
2 − g2 − g2 ≥ 0
2i 3i

We can, for example, enforce these by insisting that the following quadratic
form be positive semidefinite:

g Ti [234 ] [M ]T [234 ] [M ] g ≥ 0 (2.37)

where [234 ] is defined in property M12. This set of conditions, first derived in
van der Mee (1992, 1993) and Givens (1993) in turn requires the eigenvalues of
the composite matrix [234 ] [M ]T [234 ] [M ] to be real and non-negative. This
can then be directly taken as a test of validity for candidate Mueller matrices.
 2.3 The scattering coherency matrix formulation 85

However, this is a rather complicated and, as we shall see, incomplete formu-

lation, and in the next section we derive a much simpler set of tests based on
an eigenvalue analysis of the scattering coherency matrix.
Before leaving the Mueller matrix completely, however, we return again to
the case of backscatter in equations (2.33) and (2.34). These equations give the
impression that the backscatter matrix has ten independent elements. However,
this is not true, there being only nine independent elements for the general
case. The missing constraint equation arises again from the reciprocity theorem,
which constrains the diagonal elements. In order for a general 4 × 4 real matrix
to represent a backscatter Mueller matrix it must not only have the form shown
in equations (2.33) and (2.34), but its diagonal elements must also satisfy the
following trace conditions (Mishchenko, 1995):
/
m11 − m22 − m33 − m44 = 0 (BSA coordinates)
Reciprocity ⇒ (2.38)
m11 − m22 + m33 − m44 = 0 (FSA coordinates)
The reason for this, as shown in the next section, is that the reciprocity theorem
constrains an eigenvalue of the scattering coherency matrix to be zero. In this
sense the coherency matrix formulation makes for a much simpler analysis of
depolarisation properties, and so we now turn to consider its properties in more
detail.

2.3 The scattering coherency

matrix formulation
In the previous section we saw that the study of depolarisation is fundamentally
related to a matrix of second order products of the scattering matrix, formed
into a matrix [Y ] as shown in equation (2.39):
 
    Ex Ex∗
a b a [S]∗ b [S]∗  ∗
Ex Ey∗ 
[S] = ⇒ xs = [Y ] xi = ∗ . (2.39)
c d ∗
c [S] d [S] Ey Ex 
Ey Ey∗

There are, however, two alternative ways of forming a matrix of all possible
second-order products: the covariance and coherency matrices defined as outer
products of the corresponding scattering vectors in the lexicographic and Pauli
bases (see Section 1.6), as shown in equation (2.40) (Cloude, 1985, 1986).
These matrices have a useful symmetry as, unlike [Y ], they are complex Her-
mitian, having real elements along the diagonal and complex conjugate entries
in symmetric off-diagonal positions. This reordering of the product terms has
a more direct relationship to coherence between the elements of the k vector,
hence the name coherency matrix. For example, [C] can be used to generalize
the arguments about optimum polarisations treated for [S] matrices in Section
1.4.3 to depolarising scenarios (see Tragl (1990)). Furthermore, we can exploit
the Hermitian symmetry by noting that all the 2 × 2 principal minors of [C]
and [T ] are zero, so for example, in equation (2.40) we have aa*bb* – ab*ba*
= aa*cc* – ac*ca* = 0 and so on.
For a 4 × 4 matrix there are nine such minors, and for a 3 × 3 matrix there are
four. In fact these provide just the set of required constraint equations between
86 Depolarisation and scattering entropy

the Mueller matrix elements discussed around equation (2.36). Rather than
consider these in any more detail (to investigate further see Huynen (1970,
1987) and Hovenier (1994, 1996, 2004)) we proceed to the next key step and
consider the effect of depolarisation on the form of the coherency matrix.
 
  a
a b b  ∗ 
[S] = ⇒ [C] = k L k L = 
∗T

 . a b∗ c∗ d ∗

c d c
d
 ∗ 
aa ab∗ ac∗ ad ∗
ba∗ bb∗ bc∗ bd ∗ 
=ca∗ cb∗ cc∗ cd ∗  = [C]
 ∗T

da∗ db∗ dc∗ dd ∗

 
a b
[S] = ⇒ [T ] = k P k ∗T
P
c d
 
a+d
1 a−d   
=  
 . (a + d )∗ (a − d )∗ (b + c)∗ −i(b − c)∗ = [T ]∗T

2 b + c
i(b − c)
(2.40)

We can model depolarisation as fluctuations in space or time of the elements

of the scattering amplitude matrix [S]. In general we then have the idea of a
mean [S] matrix with fluctuations characterized as shown in equation (2.41),
in which case we derive a corresponding statistical distribution of k vectors:
   
[S r, t ] = [S] + S r, t ⇒ k r, t = k + k (2.41)

The covariance and coherency matrices (and indeed the [Y ] matrix) are then
obtained as averages over the distribution, so, for example, over a population
of L independent samples we can obtain an estimate of the mean coherency
matrix as shown in equation (2.42):

!
L
[T ] = k i k ∗T
i (2.42)
i=1

In this case the principal minor relations, by the Schwarz inequality (see
Appendix 3), now have the general form of non-negative inequalities, auch
as aa∗  bb∗  − ab∗  ba∗  ≥ 0. The Mueller matrix [M ] corresponding to
[T ] is often called a sum-of-pure Mueller matrices (SPM) in the literature
(Hovenier, 2004).

2.3.1 Eigenvalue decomposition of the coherency and

covariance matrices
As [C] and [T ] are formed from the sum of component Hermitian matrices
(unlike [Y ]), they remain positive semidefinite (PSD) Hermitian, which means
 2.3 The scattering coherency matrix formulation 87

they maintain real non-negative eigenvalues and orthogonal eigenvectors, as

shown in equation (2.43):
  
λ1 0 0 0 
 λ1 ≥ λ2 ≥ λ3 ≥ λ4 ≥ 0
 0 λ2 0 0  
[T ] = [U4 ] .  
 0 0 λ3 0  . [U4 ]
∗T
⇒ [U4 ] . [U4 ]∗T = [I4 ]



0 0 0 λ4 [U4 ] = [e1 e2 e3 e4 ]
(2.43)

where λi are the real eigenvalues, and ei the corresponding complex eigenvec-
tors. Note that [C] has the same eigenvalues as [T ] but its eigenvectors, eL ,
are related to those of [T ], eP by a 4 × 4 unitary matrix, as shown in equation
(2.44):
 
1 1 0 0
  1 0 0 1 −i    ∗T
eL = ULP4 eP = √  
 e ⇒ [C] = U
 P [T ] ULP4
2 0 0 1 i LP4

1 −1 0 0
(2.44)

We have seen that for BSA reciprocal backscatter problems, the [S] matrix
is symmetric, in which case the transformation from [T ] to [C] is often
reformulated as a 3 × 3 matrix, as shown in equation (2.45):
     
√a a+d 1 1 √0
1 1
−→  2b = [ULP3 ] √ a − d  = √ 0 0 2 e P
b=c

d 2 2b 2 1 −1 0 (2.45)

⇒ [C3 ] = [ULP3 ] [T3 ] [ULP3 ]∗T

√
Note the factor of 2, which is required to maintain invariance of total scattered
power. In what follows we refer to the coherency matrix [T ], but understand
that the equations could be easily recast in the form of the covariance matrix
[C] using equation (2.44) or (2.45).
A key idea stemming from this eigenvalue decomposition is a generaliza-
tion of the wave dichotomy to higher dimensional coherency matrices. In this
case we can write the general depolarising coherency matrix as the sum of
four independent and orthogonal scattering mechanisms, as shown in equation
(2.46):

[T ] = λ1 e1 e∗T ∗T ∗T ∗T
1 + λ2 e 2 e 2 + λ3 e 3 e 3 + λ4 e 4 e 4 (2.46)

where each vector corresponds to a relative phase amplitude matrix [SR ] as

shown in equation (2.47), and λi has a direct physical interpretation in terms
of the power scattered into the mechanism represented by e. For notational
convenience we drop the brackets .., and simply refer to a general coherency
matrix as [T ] with a corresponding Mueller matrix [M ]:
 
ea  
 eb  ∗T
e=  e e = 1 ⇒ [SR ] = √1 ea + eb ec − ied (2.47)
 ec  2 ec + ied ea − eb
ed
88 Depolarisation and scattering entropy

‹[C]› - 4 × 4 Covariance

‹..›

[S]
‹..› ‹..›

Fig. 2.3 The scattering triangle, relating the

three depolarising matrices around the outside
to [S] at the centre ‹[T]› - 4 × 4 Coherency ‹[M]› - 4 × 4 Muller Matrix

While the formulation in terms of a 4 × 4 coherency matrix has several

advantages, we note from equation (2.40) that [T ] and [C] contain the same
information on averages of second-order products as [Y ]. Since [Y ] is linked
directly to the Mueller matrix [M ] it follows that there exist 1–1 mappings
between all three of these scattering matrices. The situation is summarized
graphically in Figure 2.3. All descriptors of depolarisation stem from the ampli-
tude matrix [S], and we have seen how to proceed from this starting point to
each of the three matrices [M ], [C] and [T ] by averaging. What is missing,
however, is how to go around the outside of this diagram; that is, to go from
[M ] to [T ] directly without the need to go via [S]. The relationship between
[T ] and [C] has already been derived in equation (2.44). More significant is the
relationship between [M ] and [T ], to which we now turn.
This can be derived by direct comparison between elements, and is given in
detail in equation (2.48), with the inverse mapping provided for reference in
equation (2.49):

 
t11 t12 t13 t14
t ∗ t22 t23 t24 
[T ] =  12
t ∗
 ⇒ [M ] =
13
∗
t23 t33 t34 
∗
t14 ∗
t24 ∗
t34 t44
 ∗ − i(t − t ∗ ) ∗ + i(t − t ∗ ) ∗ − i(t − t ∗ )
t11 + t22 + t33 + t44 t12 + t12 34 34 t13 + t13 24 24 t14 + t14 23 23
1 ∗ ∗
t12 + t12 + i(t34 − t34 ) t11 + t22 − t33 − t44 ∗ + i(t − t ∗ )
t23 + t23 14 14
∗ − i(t − t ∗ )
t24 + t24 13 13 
2 t13 + t13
∗ − i(t − t ∗ )
24 24
∗ ∗ )
t23 + t23 − i(t14 − t14 t11 − t22 + t33 − t44 t34 + t + i(t12 − t ∗ )
∗
34 12
∗ + i(t − t ∗ )
t14 + t14 ∗ + i(t − t ∗ )
t24 + t24 ∗ − i(t − t ∗ )
t34 + t34 t11 − t22 − t33 + t44
23 23 13 13 12 12
(2.48)
 
m11 m12 m13 m14
m m24 
 m22 m23 
[M ] =  21  ⇒ [T ] =
m31 m32 m33 m34 
m41 m42 m43 m44
 
m11 + m22 + m33 + m44 m12 + m21 − i(m34 − m43 ) m13 + m31 + i(m24 − m42 ) m14 + m41 − i(m23 − m32 )
1m12 + m21 + i(m34 − m43 ) m11 + m22 − m33 − m44 m23 + m32 + i(m14 − m41 ) m24 + m42 − i(m13 − m31 )

 
2 m13 + m31 − i(m24 − m42 ) m23 + m32 − i(m14 − m41 ) m11 − m22 + m33 − m44 m34 + m43 + i(m12 − m21 )
m14 + m41 + i(m23 − m32 ) m24 + m42 + i(m13 − m31 ) m34 + m43 − i(m12 − m21 ) m11 − m22 − m33 + m44
(2.49)

These represent the most general scattering case in the wave or FSA coordinate
system. Note that [T ] in the sensor or BSA coordinates has the same eigenvalues
as in the wave system, but the eigenvectors are related by a unitary matrix
 2.3 The scattering coherency matrix formulation 89

derived in equation (1.188) and shown again in equation (2.50):

 
0 1 0 0
0 1 1 0 0 0
eFSA = U4C eBSA =
0
e (2.50)
0 0 −i BSA
0 0 i 0

As before, interest often centres on the special case of backscatter in either

the wave or sensor coordinates. In the sensor BSA case the fourth row and
columns of [T ] become zero due to reciprocity (SHV = SVH ). In both cases [T ]
is reduced to a 3 × 3 coherency matrix, although [M ] is maintained as 4 × 4.
The backscatter mappings in the sensor (BSA) coordinates are then given as
follows:

 
t11 t12 t13 0
∗ 
t12 t22 t23 0
[T ]BSA = 
t ∗

 13
∗
t23 t33 0

0 0 0 0
 
t11 + t22 + t33 ∗
t12 + t12 ∗
t13 + t13 ∗ )
−i(t23 − t23
 ∗ ∗ ∗ ) 
1 t12 + t12 t11 + t22 − t33 t23 + t23 −i(t13 − t13 
⇒ [M ]BSA =   (2.51)
2

∗
t13 + t13 ∗
t23 + t23 t11 − t22 + t33 ∗ ) 
i(t12 − t12 
∗ )
−i(t23 − t23 ∗ )
−i(t13 − t13 ∗ )
i(t12 − t12 −t11 + t22 + t33
 
m11 m12 m13 m14
 
m12 m22 m23 m24 
[M ]BSA =
m

 13 m23 m33 m34 

m14 m24 m34 m44
 
m11 + m22 + m33 − m44 m12 − im34 m13 + im24
1 
⇒ [T ]BSA =  m12 + im34 m11 + m22 − m33 + m44 m23 + im14 

2
m13 − im24 m23 − im14 m11 − m22 + m33 − m44
(2.52)

In the wave (FSA) coordinates we have a different backscatter result, because

in this case SHV = −SVH and so the third row and column of [T ] are zero. In
this coordinate system the mappings have the following detailed form:

 
t11 t12 0 t14
∗ 
t12 t22 0 t24 
[T ]FSA =
0

 0 0 0 
∗
t14 ∗
t24 0 t44
 
t11 + t22 + t44 ∗
t12 + t12 ∗ )
i(t24 − t24 ∗
t14 + t14
 ∗ ∗ ) ∗ 
1 t12 + t12 t11 + t22 − t44 i(t14 − t14 t24 + t24 
⇒ [M ]FSA =  ∗ ) ∗ ) ∗ ) 
 (2.53)
2  −i(t24 − t24 −i(t14 − t14 t11 − t22 − t44 i(t12 − t12 
∗
t14 + t14 ∗
t24 + t24 ∗ )
−i(t12 − t12 t11 − t22 + t44
90 Depolarisation and scattering entropy

 
m11 m12 m13 m14
 
 m12 m22 m23 m24 
[M ]FSA =



 −m13 −m23 m33 m34 
m14 m24 −m34 m44
 
m11 + m22 + m33 + m44 m12 − im34 m14 − im23
1 
⇒ [T ]FSA =  m12 + im34 m11 + m22 − m33 − m44 m24 − im13 
2 
m14 + im23 m24 + im13 m11 − m22 − m33 + m44
(2.54)

There are then four important results derived directly from the eigenvector
expansion of [T ]:
1. The conservation of zero wave entropy, introduced after equation (2.32),
implies that [T ] has only one non-zero eigenvalue, and so we can state
the following important result (Cloude, 1989):

if [S] = f −1 ([M ]) exists then [M ] → [T ] ⇒ λ1 > λ2 = λ3 = λ4 = 0

(2.55)

This constitutes an efficient and complete test of a candidate Mueller

matrix: to convert it to its corresponding coherency matrix, calculate the
eigenvalue spectrum, and see how many non-zero eigenvalues it has (or
equivalently calculate the rank of [T ]). If the rank is greater than one
then the inverse mapping is not unique.
2. As an extension of this idea we can use the non-negativity of the eigen-
values of [T ] as a general test for physical Mueller matrices. It follows
that a 4 × 4 real matrix [M ] is a true Mueller matrix if, and only if,
its corresponding coherency matrix [T ] has a non-negative eigenvalue
spectrum so that

[M ] → [T ] ⇒ λi ≥ 0, i = 1, 2, 3, 4 (2.56)

3. Experimental measurements of the Mueller matrix [M ] are often dis-

torted by errors leading to non-physical results (Cloude, 1986; Brosseau,
1990) (negative scattered powers, degree of polarisation greater than
unity, and so on). We can use the eigenvalue spectrum of [T ] to devise a
filtering scheme for small errors that guarantees a matrix estimate with
stable physical properties across the whole Poincaré sphere. The idea
is to calculate the eigenvalue spectrum and zero any contributions from
negative eigenvalue contributions. The filtering algorithm then has the
following form, where sgn is the signum function, +1 for positive values,
0 for 0, and −1 for negative:
!
[Mexp ] → [Texp ] = λi ei e∗T
i ⇒ [T̂ ]
i
!1
= (1 + sgn(λi ))λi ei e∗T
i → [M̂ ] (2.57)
2
i

4. Considering the special case of backscatter, we saw that reciprocity

forced a rank reduction of [T ]. In particular, the reciprocity theorem
forces the third (wave) or fourth (sensor) row/column to be zero (see
 2.4 General theory of scattering entropy 91

equations (2.51) and (2.53)). This leads us directly to the trace condi-
tion on backscatter Mueller matrices discussed in equation (2.38). As
a special case we consider the possibility of designing wave depolaris-
ers for use in backscatter. We may, for example, consider using a very
rough surface or multiple scattering cavity with backscatter geometry.
We define the following general form for an isotropic depolariser with
0 ≤ δ ≤ 1:
 
1 0 0 0
0 δ 0 0
[M ] = 
0 0
 (2.58)
δ 0
0 0 0 δ

This represents a generalization of the example considered in equation

(2.35) (which is the special case, δ = 0) where the incident Stokes vector
is preserved in form but its degree of polarisation is reduced to δ. In the
backscatter direction, reciprocity places the following constraints on δ
from equation (2.38)):

Reciprocity
/
m11 − m22 − m33 − m44 = 1 − 3δ = 0 (sensor coordinates)
⇒
m11 − m22 + m33 − m44 = 1 − δ = 0 (wave coordinates)
(2.59)

We see that the ideal depolariser (δ = 0) is in fact impossible to realize in

backscatter (it would violate reciprocity). Note that in the BSA system
we can, however, obtain a 33% isotropic depolariser. (We return to this
idea in Section 2.5.) Consequently any depolarisation in backscattering
must be anisotropic; that is, it must transform as well as depolarise. For
this reason it is interesting to consider a more detailed parameterization
of depolarisation, as follows.

2.4 General theory of scattering entropy

In this section we formulate a general model of depolarisation that scales to
arbitrary dimension of coherency matrix N × N. The basic idea is to always
identify the ‘polarising’ contribution with the dominant eigenvector of the
matrix—the eigenvector corresponding to the largest eigenvalue. The other
eigenvectors then contribute to depolarisation with a strength given by the
remaining minor eigenvalues. By employing multidimensional unitary trans-
formations we will then be able to parameterize all possible types of system
depolarisation behaviour. We start with the general formulation and then spe-
cialize it to the three important cases for N = 2, 3 and 4. We then consider
the effects of scattering symmetries on constraining the degrees of freedom
involved in both polarised and depolarised components.
The starting point for our analysis is the idea of a unitary reduction operator:
a matrix [U−1 ] which acts to reduce the dimensionality of an N × N unitary
92 Depolarisation and scattering entropy

matrix to N − 1 × N − 1, as shown in equation (2.60) (Cloude, 1986, 1995b,

2001a):

   
1 0T 1 0T
[U−1 ] [UN ] = = (2.60)
0 UN −1 0 exp(iHN −1 )
!
M
HN −1 = hk k ⇒ h = depolarisation state vector
k=1

This operator acts to convert the first column of the coherency matrix to the
identity. The next stage is to apply such an operator to the matrix of ranked
eigenvectors ei of the coherency matrix

 
[UN ] = e1 e2 . . . eN (2.61)

and to identify the submatrix U N −1 with the depolarising aspects of the scatter-
ing process. In this way U N −1 involves continuous smooth transformation away
from the polarised reference state (the dominant eigenvector). The submatrix
U N −1 may be further parameterized in terms of an N − 1 × N − 1 Hermitian
matrix, related to the unitary transformation by a matrix exponential, and itself
conveniently expanded in terms of a set of scalar parameters being the basis
elements of the underlying algebra. This vector of coefficients we then term the
depolarisation state vector (see equation (2.60)), as it determines the nature of
the depolarisation, with different state vectors representing different types and
degrees of depolarised scattering.
This then leads us to the following notation to characterize the number
of parameters involved in polarising and depolarising components of the
decomposition of a general N × N coherency matrix T N :

T N = (E + L) + [E + L] (2.62)

where [..] are the depolarising parameters and (..) the polarising terms, and
for each, E are those parameters associated with the eigenvectors and L the
eigenvalues. From the general structure of N × N coherency matrices we then
have the following constraints:

• [L] + (L) = N
• [E] + (E) = dim(SU(N)) − rank(SU(N)) = N(N − 1)

Note that the total number of eigenvector parameters is given by the sum [E]
+ (E) = dim(SU(N)) − r(SU(N)), where dim() = N2 − 1 is the dimension of
the group and r = N − 1 is the rank of the Cartan sub-algebra or the number of
mutually commuting generators (see Appendix 2). For example, for N = 1 there
are no useful eigenvector parameters, and for N = 2 we only have two, while
for N = 4 (the most general bistatic scattering case) we have twelve parameters
available. We now look at special cases to illustrate how this classification
scheme works, and in particular how general scattering symmetries restrict the
number of parameters.
 2.4 General theory of scattering entropy 93

2.4.1 Depolarisation for N = 2 scattering systems

In the simplest non-trivial case of N = 2, each eigenvector can be parameterized
in the form shown in equation (2.63):
 
cos α
e1 = eiχ (2.63)
sin αeiφ1

The unitary reduction operator can be constructed from these parameters as

shown in equation (2.64):
    
cos α sin α 1 0   1 0
e−iχ e1 e2 = (2.64)
− sin α cos α 0 e−iφ1 0 e−iφ1

Here we see that the depolarisation subspace is now one-dimensional and

consists of a scalar phase shift. This, however, remains a ‘hidden’ parame-
ter (derived from the Cartan sub-algebra), since phase shifts of the original
polarised state are not observed when forming the coherency matrix. This leads
to the following decomposition of general two-dimensional coherency matrices
(see also equation (2.12)):

T2 = (2 + 1) + [0 + 1] (2.65)

Note that there are no depolarisation parameters in the eigenvectors, and that
all depolarisation effects lie in the eigenvalue spectrum. This leads to the
wave dichotomy and degree of polarisation ideas discussed in Section 2.1.1.
More interesting possibilities exist for higher-dimensional matrices, as we now
consider.

2.4.2 Depolarisation for N = 3 scattering systems

For the case N = 3 we can parameterize the general eigenvector as a natural
extension of the N = 2 case, as shown in equation (2.66):
 
cos α
e1 = eiχ sin α cos ψeiφ1  (2.66)
sin α sin ψeiφ2

This can be used to parameterize backscatter problems, when the reciprocity

theorem forces the coherency matrix to be 3 × 3 (see equations (2.51) and
(2.52)). From this we can define the unitary reduction operator as a cascade of
three matrices shown in equation (2.67):
   
cos α sin α 0 1 0 0 1 0 0
e−iχ − sin α cos α 0 0 cos ψ sin ψ  0 e−iφ1 0 
0 0 1 0 − sin ψ cos ψ 0 0 e−iφ2
 
  1 0T
× e1 e2 e3 = (2.67)
0 U2
94 Depolarisation and scattering entropy

In this case [T ]/[M ] have a maximum of nine parameters, and SU(3) is the gov-
erning unitary group for the eigenvectors of [T ]. However, the unitary reduction
operator means that all depolarisation effects are controlled by SU(2), which has
dimension 3 and rank 1. Hence in N = 3 backscatter the polarised/depolarised
decomposition can be written in compact form as follows:

T3 = (4 + 1) + [2 + 2] (2.68)

This shows that the eigenvectors now contain two depolarising parameters,
complemented by two real eigenvalues. The polarised component has four
eigenvector parameters and one eigenvalue (its amplitude). Note that this
decomposition makes no assumptions about scattering symmetry (other than
reciprocity). We now turn to consider the restrictions placed by scattering
symmetries on the form of the coherency and Mueller matrices in backscatter.

A 2.4.2.1 Backscatter symmetries and depolarisation

One of the most important symmetries in scattering theory is reciprocity. We
Q P have seen that the effect of this symmetry in backscatter problems is to generate
Reflect a rank reduction of the coherency matrix. In this section we consider the form
of [T ] for various other scattering symmetry assumptions in the medium, with
a view to investigating any further simplifications that may occur in the form of
[T ]. There are three primary types of symmetry to be considered (Perrin, 1942;
Rotate Van de Hulst, 1981; Nghiem, 1992; Cloude, 1996), as illustrated in Figure 2.4.
Here we see a representation of the plane of polarisation of the incident wave
and consider the scattering to be due to the sum of (independent) contributions
A
from a large number of scatterers in the scene. The first symmetry we can
Fig. 2.4 Reflection and rotation symmetry consider is reflection symmetry. Here we assume that for every scatterer at
point P there is a matching scatterer in the reflected position about some axis
AA’. Note that AA’ need not be aligned with the sensor H and V coordinates in
which case it is parameterized by a rotation angle θ.
Due to the independence assumption, the coherency matrix for this scene is
given by the sum of coherency matrices from P and Q. Assuming for the moment
that AA’ is aligned (θ = 0), then, as shown in equation (1.145), the k vectors
at P and Q are related by a minus sign in their crosspolarised components, so
that the coherency matrix has zero elements in some off-diagonal positions, as
shown in equation (2.69).

   
k0 k0
k P = k1  ⇒ k Q =  k1 
k2 −k2
   
  t11 t12 t13 t11 t12 −t13
⇒ [T ] = [TP ] + TQ = t12∗ t22 t23  +  t12
∗ t22 −t23 
∗
t13 ∗
t23 t33 ∗
−t13 ∗
−t23 t33
 
t11 t12 0
= t12
∗ t22 0 (2.69)
0 0 t33
 2.4 General theory of scattering entropy 95

In this case we see that one consequence of reflection symmetry + is that ,the
∗
coherence
+ , between co- and crosspolarised channels will be zero: S HH S HV =
∗
SVV SVH = 0, and so on. Reintroducing the idea that θ may not be zero
produces the following general form for a reflection symmetric medium:

   
1 0 0 t11 t12 0 1 0 0
[Tθ ] = 0 cos 2θ − sin 2θ  t12
∗ t22 0  0 cos 2θ sin 2θ 
0 sin 2θ cos 2θ 0 0 t33 0 − sin 2θ cos 2θ
(2.70)

Note that this will introduce apparent correlations between co- and cross-
channels due to crosspolarisation (but not depolarisation) by the axis mis-
alignment.
The second type of symmetry to consider is rotation symmetry. Here we
consider the form of [T ] that remains invariant under arbitrary rotations in the
plane of polarisation. By definition, and as an extension of the discussion around
equation (2.69), [T ] must then be formed from linear combinations of the outer
product of eigenvectors of the plane rotation matrix (see Section 4.1.1). In this
case we can therefore predict the general form of a rotation symmetric [T ] as
shown in equation (2.71):

 
a 0 0
[T ] = av1 v ∗T ∗T ∗T
1 + bv 2 v 2 + cv 3 v 3 = 0 b+c i(b − c) (2.71)
0 −i(b − c) b+c

Finally we consider the most general symmetry, formed by insisting that every
axis AA’, for all θ, has the PQ reflection symmetry. This is termed azimuthal
symmetry, and the form of [T ] can be found by adding reflection symmetric
versions of the rotation matrix of equation (2.71) to obtain equation (2.72):

   
a 0 0 a 0 0
[T ] = 0 b+c i(b − c) + 0 b+c −i(b − c)
0 −i(b − c) b+c 0 i(b − c) b+c
 
a 0 0
= 2 0 b + c 0  (2.72)
0 0 b+c

Here we have obtained a very simple diagonal coherency matrix, with two
equal diagonal values. We now turn to consider the polarised/depolarised
decomposition for each of these symmetry cases.

2.4.2.2 Depolarisation under azimuthal backscattering symmetry

This severe symmetry assumption leads directly to a diagonal coherency
(and Mueller) matrix with two degenerate coherency eigenvalues. Conse-
quently, [T ]/[M ] have only two free parameters, and the coherency/covariance
and (BSA) Mueller matrices for backscatter can be written as shown in
96 Depolarisation and scattering entropy

equation (2.73):
 
m 0 0
[T3 ] =  0 mκ 0 
0 0 mκ
 
1+κ 0 1−κ
m
⇔ [C3 ] =  0 2κ 0 
2 1−κ 0 (2.73)
1+κ
 
1 + 2κ 0 0 0
m 0 1 0 0 
⇔ [M ] =  
2 0 0 1 0 
0 0 0 2κ − 1
In this symmetry assumption the polarising/depolarising decomposition there-
fore reduces to equation (2.74):

T3az = (0 + 1) + [0 + 1] (2.74)

In this case only two parameters are required to characterize the medium (such
as VV power and HH/VV coherence or HV, but not both HV and HH/VV
coherence, as these two are now related). A convenient parameterization for
the most usual situation in remote sensing applications (a > b + c in equation
(2.72)) is provided in terms of the polarised power (m) and depolarisation or
noise parameter (κ) shown in equation (2.73).
Note that there are two important special cases. When κ = 0 we have zero
depolarisation and backscatter from a sphere symmetric scatterer (or random
collection of such), and at the other extreme κ = 1, and [T ] reduces to the 3 × 3
identity matrix. This represents the most depolarising case. Note, however, that
the Mueller matrix for this case still does not equal a noise source ideal depo-
lariser (equation (2.35)); in fact, it corresponds to a 33% isotropic depolariser.
As mentioned in the discussion around equation (2.59), this is due to the fact
that HV and VH remain correlated for all backscatter problems, and hence an
identity coherency matrix corresponds not to equal noise power in all channels,
but with the cross-channel having only half the power of the copolarised chan-
nels. This indicates that a proper analysis of polarised signals in the presence
of noise requires a full 4 × 4 coherency matrix analysis. This residual structure
implies that even in the most depolarising backscatter case there will be some
residual information on the scatterer properties, even just that following from
the reciprocity theorem.

2.4.2.3 Depolarisation under reflection symmetry

In this more complicated case, both [T ] and [M ] have six free parameters,
and the coherency matrix can be written in the general form shown in equation
(2.75), where θ is the angle mismatch between radar coordinates and the axis of
symmetry (for example the local normal in surface scattering; see Section 3.1):
   
1 0 0 t11 t12 0 1 0 0
[Tθ ] = 0 cos 2θ − sin 2θ  t12 ∗ t22 0  0 cos 2θ sin 2θ 
0 sin 2θ cos 2θ 0 0 t33 0 − sin 2θ cos 2θ
(2.75)
 2.4 General theory of scattering entropy 97

The polarised/depolarised decomposition can be written for reflection symme-

try as shown in equation (2.76):

[T ]ref = (3 + 1) + [0 + 2] (2.76)

This result shows formally that for backscatter from random media with reflec-
tion symmetry [E] = 0 and [L] = 2, so the eigenvectors contain no information
about depolarisation, only about the polarising influence of the dominant eigen-
vector. This results because the 3 × 3 matrix of eigenvectors for such media can
be written as a function of a 2 × 2 unitary matrix, as shown in equation (2.77):
   
[U2 ] 0 1 0
[U3 ] = ⇒ [U−1 ] [U3 ] = (2.77)
0 1 0 1

All the depolarisation information for such media is therefore contained in the
two minor eigenvalues of [T ]. These eigenvalues are not usually used them-
selves as measures of depolarisation in such media. Instead, widespread use is
made of two functions of the eigenvalues λi : scattering entropy and anisotropy
(Pottier, 1997), defined as in equations (2.78) and (2.79):

!
3
λi
H =− Pi log3 Pi Pi = 5 0≤H ≤1 (2.78)
λ
i=1

λ2 − λ3
A= 0≤A≤1 (2.79)
λ2 + λ 3

The entropy is zero for zero depolarisation; that is, when [T ] has only one
non-zero eigenvalue. At the other extreme, H = 1 when [T ] is diagonal and
maximum depolarisation occurs. The anisotropy is a useful parameter for
assessing the type of symmetry. For azimuthal symmetry the minor eigenvalues
are equal, and so A = 0. On the other hand, if only the smallest eigenvalue falls
to zero then A = 1, and we have significant departure from azimuthal symmetry.
This situation occurs, for example, for in-plane scattering from one-dimensional
rough surfaces and in scattering by chiral particles (see Chapter 3).

2.4.2.4 The entropy/alpha decomposition for N = 3

Despite the completeness of the full unitary reduction operator approach, in all
but the azimuthal symmetry case it requires estimation of multiple parameters
from the data. Very often a simpler approach to the parameterization of depo-
larisation effects is all that is required. The basic motivation here is to seek
a method that provides just two parameters—one indicating the total level of
depolarisation, and the other the average polarised information. In addition,
this method must cope not only with azimuthal and reflection symmetry, but
also the most general case. It is desirable that the parameters be as robust as
possible, so we require them to be polarisation basis invariant, otherwise we
would need to change the interpretation every time we changed the reference
polarisation base. This again suggests an eigenvalue decomposition of [T ],
as shown in equation (2.80), but this time, in contrast to the unitary reduction
operator approach, we seek to find averages that may be used for a reduced
98 Depolarisation and scattering entropy

parameterization (Cloude, 1995c, 1996, 1997a; Lee, 2008).

 
  λ1 0 0  ∗T
[T ] = e1 e2 e3  0 λ2 0  e1 e2 e3 (2.80)
0 0 λ3

The spread of total scattered power across the eigenvalues is evidently a good
indicator of depolarisation. This can be turned into a quantitative measure by
first normalizing the eigenvalues to unit sum, so we define three parameter Pi ,
as shown in equation (2.81). These can then be interpreted as probabilities of
the statistically independent ‘polarised’states, given by the eigenvectors ei . The
spread of probabilities can then be represented by a single scalar, the entropy
being defined as shown in equation (2.81):

λi ! 3
0 ≤ Pi = 5 ≤ 1 ⇒ H = − Pi log3 Pi 0≤H ≤1 (2.81)
λ
i=1

This then provides a suitable depolarising parameter; which is zero for zero
depolarisation (while still allowing crosspolarisation), and one for maximum
depolarisation. But what of the polarised component? From each eigenvector
we can select the scattering mechanism α as a suitable basic invariant polarised
parameter, as shown in equation (2.82).
   
e1 cos α
e = e2  = eiχ sin α cos ψeiφ1  ⇒ α = cos−1 (|e1 |) 0 ≤ α ≤ 90◦
e3 sin α sin ψeiφ2
(2.82)

However, we have three such values—one for each eigenvector, as shown in

equation (2.83):
 
cos α1 cos α2 cos α3
[U3 ] = sin α1 cos ψ1 eiδ1 sin α2 cos ψ2 eiδ2 sin α3 cos ψ3 eiδ3  (2.83)
sin α1 sin ψ1 eiγ1 sin α2 sin ψ2 eiγ2 sin α3 sin ψ3 eiγ3

The statistical interpretation suggests forming an average as a sum of the three

values, weighted by their probabilities, so we can form the average alpha
parameter as shown in equation (2.84):

α = P1 α1 + P2 α2 + P3 α3 (2.84)

Note that the form of the unitary matrix of eigenvectors shown in equation
(2.83) is overparameterized. The unitary reduction operator provides a much
more rigorous approach (for example, the columns of the unitary matrix are
not independent, as implied in equation (2.83), as they are all mutually orthog-
onal). However, the approach in equation (2.83) has a strong physical appeal
in terms of the probabilistic interpretation, and provides a useful parameteriza-
tion of the mean scattering mechanism. We have therefore derived a suitable
pair of parameters: entropy H and scattering mechanism α. To aid interpre-
tation of different types of polariser/depolariser, use can then be made of the
two-dimensional H/α or entropy/alpha plane, as shown in Figure 2.5 (Cloude,
1996).
 2.4 General theory of scattering entropy 99

H/alpha diagram for N = 3

90
Curve II
80

70

60
Alpha (degrees)

50 Feasible region

40

30
line
etry
20 mm
hal sy
mut Curve I
10 Azi

0
0 0.2 0.4 0.6 0.8 1 Fig. 2.5 Entropy/alpha diagram for N = 3
Entropy H backscatter

All reciprocal backscattering problems can be represented as points in this

plane. However, not all of the plane represents physical depolarisers. Because
of the averaging inherent in equation (2.84) there are bounds on the maximum
and minimum alpha angle we can obtain for a given entropy H. Only for zero
entropy can we obtain the full 90 degrees of alpha variation. For other values of
entropy we can calculate the maximum and minimum values of alpha to define
two bounding curves, shown as I and II in Figure 2.5. The lower bounding
curve is defined by azimuthally symmetric scattering. These have α = 0 for the
dominant eigenvector, and equal ‘spill-over’ into the α= π /2 eigenvectors by
the depolarising parameter ‘m’ as shown in equation (2.85). The upper bound
can be found in a similar manner by first ‘switching off’ the signal in the α = 0
eigenvector, and then only when the α= π/2 subspace is full, allowing signal
energy into this eigenvector. The resulting bounding curve II is also shown in
equation (2.85). Both curves tend to the same alpha point for maximum entropy
H = 1, namely π /3:

 
1 0 0
CURVE I [T ]I = 0 m 0  0 ≤ m ≤ 1
0 0 m
  


0 0 0

 [T ]II = 0 1 0  ≤ m ≤ 0.5 (2.85)


 0 0 2m
CURVE II  

 2m − 1 0 0


  0 1 0 0.5 ≤ m ≤ 1
 II =
 [T ]
0 0 1

Although the main motivation behind the H/α decomposition is to derive a

small set of key parameters to simplify classification of depolarisers, the basic
scheme is often augmented by the addition of a third parameter: the scattering
anisotropy A (equation (2.79))—again a basis invariant depolarising parameter,
100 Depolarisation and scattering entropy

as shown in equation (2.86).



 a1 = (1 − H )(1 − A)
λ2 − λ3 
a2 = (1 − H )A
A= ⇒ (2.86)
λ2 + λ 3 a3
 = HA

a4 = H (1 − A)

With entropy and anisotropy now defined, several combinations of these two
have been proposed as suitable for the enhancement of important special cases of
depolarisation. These are shown as a1 to a4 in equation (2.86). These parameters
have large values for the following cases:

a1 This parameter is large for single dominant eigenvalue scattering (H small)

with added noise (A small).
a2 This parameter is large for a strong single polarising element (H small) in
the presence of a second weaker mechanism (A large).
a3 This parameter is large for two strong mechanisms in the presence of a weak
third.
a4 This parameter is large for strongly depolarising systems with three roughly
equal eigenvalues.

This entropy/alpha approach to depolarisation was first developed for radar

backscatter, where it is often termed the Cloude–Pottier decomposition, after
the original authors (Cloude, 1996, 1997a). One advantage of this entropy/alpha
approach is that it can be scaled to different coherency dimensions, as we now
discuss for the case of reduced dimension N = 2.

2.4.2.5 The entropy/alpha decomposition for N = 2

While the entropy/alpha approach was originally designed to simplify multipa-
rameter depolarisation, as occurs in N = 3 and N = 4 problems, it can also be
applied to the simpler case of N = 2 depolarisation (Ainsworth, 2007; Cloude,
2007b). Using the interpretation of normalized eigenvalues as probabilities Pi ,
together with the fact that in N = 2 problems the second eigenvector can be
derived entirely from the principal vector using orthogonality, we obtain an
entropy/alpha parameterization of the wave coherency matrix [J ], as shown in
equation (2.87):
  
 cos α − sin αe−iδ
  
 [U2 ] =
Jxx Jxy sin αeiδ cos α
[J ] = ∗ ⇒  
Jxy Jyy 
 P
[D] = (λ1 + λ2 ) 1 0
0 P2
 π π

α2 = P1 α + P2 2 − α = α(P1 − P2 ) + P2 2

⇒ 52


H2 = Pi log2 Pi
i=1
(2.87)

Again we can define an entropy/alpha plane as shown in Figure 2.6. In this case
the upper and lower bounds can be found by considering the parameterized
 2.4 General theory of scattering entropy 101

H/alpha diagram for N = 2

90

80

70

60
Alpha (degrees)

50
Feasible region
40

30

20

10

0
0 0.2 0.4 0.6 0.8 1 Fig. 2.6 Entropy/alpha diagram for N = 2
Entropy H depolarisation

diagonal 2 × 2 matrix shown in equation (2.88):

 
1 0 mπ
0 ≤ m ≤ 1 ⇒ α min =
0 m 2(1 + m)
  (2.88)
m 0 π
0 ≤ m ≤ 1 ⇒ α max =
0 1 2(1 + m)

In this case the lower and upper curves show a symmetry about α = 45 degrees.
This can be traced to the two-dimensional eigenvector space, one of which has
α = 0 and the other π/2. In the N = 3 case we saw a 2:1 bias in favour of the
π /2 eigenvector subspace, which tends to lift the alpha range with increasing
entropy (Figure 2.5).
In this N = 2 case we also have a simple geometrical interpretation of α on
the Poincaré sphere (see Figure 1.12). In the special case that x and y in equation
(2.87) represent polarisation wave states (see Section 1.33), the angles α and δ
are then related to the orientation and ellipticity angles θ, τ of the polarisation
ellipse via a spherical triangle construction on the Poincaré sphere. However,
here we have generalized this idea so that x and y can now be arbitrary channels
of data; that is, two elements of the [S] matrix, anticipating the ideas of compact
polarimetry (discussed more fully in Section 9.3.4).
The averaging implied by the entropy/alpha approach does not pick
out the state P corresponding to the maximum eigenvalue as in the
polarised/depolarised decomposition, but instead forms an average based on
a probabilistic interpretation of making measurements on the wave and obtain-
ing polarisation X and Y with probabilities P1 and P2 respectively. Hence the
average polarisation state would have a corresponding alpha value given by α.
For example, when the coherency matrix approaches the identity (noise), then
α = π4 , being a mixture of the state X (α = 0) with its antipodal orthogonal
state Y (α= π /2).
In radar there are two important special cases when the N = 2 formalism
becomes important. It is often advantageous in radar design (from a cost, data
rate and coverage point of view) to employ a single transmitted polarisation
102 Depolarisation and scattering entropy

state and a coherent dual channel receiver to measure orthogonal components

of the scattered signal (see Chapter 9) (Souyris 2005; Raney, 2006, 2007). Such
dual polarised radars are not capable of reconstructing the complete scattering
matrix, but instead can be used to reconstruct a column of the [S] matrix or more
generally some projection. From this we can then construct a 2 × 2 coherency
matrix [J ] as an example of N = 2 depolarisation. One key decision in the
design of such radars is the best single polarisation to employ as the reference
point X on the Poincaré sphere. For example, it is shown in Chapter 4 that
circular polarisation would be a good choice, since the coherence between co-
and cross-circular polarisation can be used to estimate orientation of the scat-
terer. However, circularly polarised transmitters are not widely employed in
radar imaging systems, where there is a preference for linear polarisations. For
example, several radar systems employ horizontal (H) or vertical (V) polari-
sation transmit and receive H and V components. These radars can be used to
estimate the following forms of the coherency matrix:
+ , + , + , + ,
∗
SHH SHH ∗
SHH SHV ∗
SVV SVV ∗
SVV SVH
[JH ] = + ∗
, + ∗
, [JV ] = + ∗
, + ∗
, (2.89)
SHV SHH SHV SHV SVH SVV SVH SVH
Note that these represent important examples of the generalization of the wave
coherency matrix. We can summarize the polarimetric information content of
these new 2 × 2 matrices by relating them to the full coherency matrix [T ], as
shown in equation (2.90):
 
    1 0
1 1 1 0 + , 1 1 1 0
kH = √ k ⇒ [JH ] = k H k ∗T
H = [T ] 1 0
2 0 0 1 2 0 0 1 0 1
 
1 t11 + t22 + 2Re(t12 ) t13 + t23
=
2 (t13 + t23 )∗ t33
 
    1 0
1 1 −1 0 + , 1 1 −1 0
kV = √ k ⇒ [JV ] = k V k ∗T
V = [T ] −1 0
2 0 0 1 2 0 0 1 0 1
 
1 t11 + t22 − 2Re(t12 ) t13 − t23
=
2 (t13 − t23 )∗ t33
(2.90)
To emphasize the importance of the off-diagonal elements of these 2 × 2 matri-
ces, consider the effects of rotation about the line of sight, as occurs, for
example, in scattering from a sloped surface (see Chapter 3). To see this we
consider the form of the dual polarisation scattering vectors for coherent point
scatterers in terms of a rotation about the line of sight. These are shown in
equation (2.91) as projections of the coherent Pauli k vectors:
   
  1 0 0 k0  
1 1 1 0     1 k0 + k1 cos 2θ
kH = √ . 0 cos 2θ − sin 2θ . k1 = √
2 0 0 1 0 sin 2θ cos 2θ 0 2 k1 sin 2θ
   
  1 0 0 k0  
1 1 −1 0  1 k0 − k1 cos 2θ
kV = √ . 0 cos 2θ − sin 2θ  . k1  = √
2 0 0 1 0 sin 2θ cos 2θ 0 2 k1 sin 2θ

(2.91)
 2.4 General theory of scattering entropy 103

It follows that the coherency matrices for reflection symmetric random media
take the form shown in equation (2.92), from which we note the presence of
complex off-diagonal terms.
9 :
[JH ] = k H k ∗T
H
; <
1 |k0 |2 + |k1 |2 cos2 2θ + 2Re(k0 k1∗ ) cos 2θ |k1 |2 cos 2θ sin 2θ + k0 k1∗ sin 2θ
=
2 |k1 |2 cos 2θ sin 2θ + k1 k0∗ sin 2θ |k1 |2 sin2 2θ
9 :
[JV ] = k V k ∗T
V
; <
1 |k0 |2 + |k1 |2 cos2 2θ − 2Re(k0 k1∗ ) cos 2θ |k1 |2 cos 2θ sin 2θ − k0 k1∗ sin 2θ
=
2 |k1 |2 cos 2θ sin 2θ − k1 k0∗ sin 2θ |k1 |2 sin2 2θ
(2.92)

Only for zero tilt (θ = 0) or uniformly random θ do these diagonal terms dis-
appear. This suggests that the phase terms will be important for identifying
coherent point scatterers. Using a coherent (zero entropy) assumption we can
also relate the dual polarisation alpha parameter to the scattering matrix ele-
ments for a symmetric point (SXY = 0; see equation (4.12)) as shown in
equation (2.93):
 
 (SXX − SYY ) sin 2θ 
tan α2 =  
(SXX + SYY ) + (SXX − SYY ) cos 2θ 
  (2.93)
(SXX − SYY ) sin 2θ
δ2 = arg
(SXX + SYY ) + (SXX − SYY ) cos 2θ

In the more general case of a non-symmetric point scatterer these relations take
the form shown in equation (2.94):
 
 2SXY cos 2θ + (SXX − SYY ) sin 2θ 
tan α2 =  
(SXX + SYY ) + (SXX − SYY ) cos 2θ − 2SXY sin 2θ 
  (2.94)
2SXY cos 2θ + (SXX − SYY ) sin 2θ
δ2 = arg
(SXX + SYY ) + (SXX − SYY ) cos 2θ − 2SXY sin 2θ

One advantage of the entropy/alpha approach is how it scales naturally with

changing dimension. We have shown above how the original N = 3 idea can
be modified for the restricted N = 2 case, and so now finally turn to consider
application of the concept to the most general bistatic N = 4 scattering case.
First, however, we consider the general properties of bistatic depolarisation.

2.4.3 Depolarisation in N = 4 scattering systems

In the general bistatic scattering case, [T ]/[M ] have up to sixteen parameters
and SU(4) is the governing unitary group. SU(4) has dimension 16 and rank 4,
so [E] + (E) = 16 − 4 = 12, and [L] + (L) = 4. By application of the unitary
reduction operator, depolarisation in bistatic scattering systems is then con-
trolled by [L] = 3 eigenvalues and the SU(3) group for eigenvectors. SU(3)
has dimension 8 and rank 2, so that we can write the polarising/depolarising
decomposition in compact form as shown in equation (2.95), which shows that
104 Depolarisation and scattering entropy

there are now up to six eigenvector parameters associated with depolarisation.

[T ]bistatic = (6 + 1) + [6 + 3] (2.95)

These can be generated from the Gell–Mann matrices (seeAppendix 2) (Cloude,

1986, 1995b; Ferro–Famil, 2000). As in the N = 2 case, however, there are
several important symmetries that further reduce the number of parameters in
bistatic scattering.

2.4.3.1 Bistatic scattering symmetries

The first symmetry to consider is the reciprocity theorem. In backscatter prob-
lems this led to a symmetric BSAscattering matrix and rank-3 coherency matrix.
In bistatic problems its effect is more subtle. We begin by considering the effect
of the theorem on the [S] matrix itself (Saxon, 1955; Mishchenko, 2000). In
equation (2.96) we show that the effect of interchange of transmitter and receiver
involves a change of coordinates plus a transpose operation:

  .
E AS = S β i β s .P A
  ⇒ P B .E As = P A .E Bs
E BS = S −β i −β s .P B
Reciprocity Theorem
   T
⇒ S β i β s = S −β i −β s (2.96)

In equation (2.97) we show explicitly how the elements of the [S] matrices are
related:
   
0 1 T 1 0
[S] = S −β i −β s
−1 0 0 −1
   
a b a −c
⇒S= ⇒ Sα = (2.97)
c d −b d

Theseare just two elements of the scattering matrix group identified in equation
(1.145). As shown there, this can be interpreted in terms of the scatterer as a
180-degree rotation about the bisectrix b, as shown in Figure 2.7. This we define
as the scatterer being in its reciprocal position. Finally, we show the effect on

Scatterer
[S]

x

x
bi bs
A B
y

y
Incident Bisectrix Scattered
Fig. 2.7 Reciprocity interpreted as a rotation wave wave
about the bisectrix b
 2.4 General theory of scattering entropy 105

the associated scattering vectors in equation (2.98):

   
a+d a+d
 a−d   
k= →k = a−d  (2.98)
 b+c  α −(b + c)
i(b − c) i(b − c)

Therefore, assuming a random collection of scatterers formed from a mixed

population with equal numbers in their original and reciprocal positions, then
[T ] and [M ] have a form predicted by adding (incoherently) the [T ] and [M ]
matrices for S and S α in equation (2.97) to yield the form shown in equation
(2.99). This shows that in the bistatic case, (hv+vh) is uncorrelated with the
copolar channels, but that (hv–vh) can maintain correlation. The impact of these
correlations on the form of the Mueller matrix is shown on the right in equation
(2.99):

 
t11 t12 0 t14
9 : t ∗ t 0 t24  9 :
=  ⇔ M reciprocity
reciprocity 12 22
TFSA 0 0 t33 0 FSA
∗
t14 ∗
t24 0 t44
 
m11 m12 m13 m14
 m12 m22 m23 m24 
=
−m13 −m23
 (2.99)
m33 m34 
m14 m24 −m34 m44

Consequently, the polarised/depolarised decomposition for the reciprocal

bistatic scattering case considerably simplifies. The 4 × 4 unitary matrix of
eigenvectors can now be expressed as a 3 × 3 sub-unitary matrix, as shown
in equation (2.100). In this case we return to the N = 3 scenario, and SU(2)
governs the depolarisation subspace:

 
  [U2 ] 0 0
[U3 ] 0
[U4 ] = ⇒ [U−1 ] [U4 ] =  0 1 0 (2.100)
0 1
0 0 1

However, there is now one extra depolarising eigenvalue to consider, and so the
polarised/depolarised can be written in terms of a reduced set of ten parameters,
as shown in equation (2.101):

recip
T4 = (4 + 1) + [2 + 3] (2.101)

Following this logic there are two other symmetry combinations to consider
from the scattering matrix group. In the first we combine a scatterer with its
mirror image in the plane of scattering, defined from the incident and scattered
wave vectors k, as shown in Figure 1.14. The scattering vectors and averaged
coherency and Mueller matrices for this case then have the form shown in
106 Depolarisation and scattering entropy

equation (2.102):
   
a+d a+d
 a−d   a−d 
k=  
 b + c  → k β =  −(b + c) 


i(b − c) −i(b − c)
   
t11 t12 0 0 m11 m12 0 0
t ∗ t22 0 0  m21 m22 0 0 
→ T  =  0
12  ⇔ M  =  
0 t33 t34   0 0 m33 m34 
0 ∗
0 t34 t44 0 0 m43 m44
(2.102)
This case is again governed by SU(4), but the 4 × 4 unitary matrix can be
factored in terms of 2 × 2 unitary matrices, as shown in equation (2.103):
   
[U2 ] [0] [I2 ] [0]
[U4 ] = ⇒ [U−1 ] [U4 ] = (2.103)
[0] [U2 ] [0] [U2 ]
where [I2 ] is the 2 × 2 identity matrix, and [0] the 2 × 2 null matrix. Again
SU(2) governs the depolarisation subspace, and the polarised/depolarised
decomposition has eight parameters organized as shown in equation (2.104):
plane
T4 = (2 + 1) + [2 + 3] (2.104)

Finally we consider a combination of a scatterer and its mirror image in the

bisectrix plane, defined orthogonal to the scattering plane and including the
bisectrix vector b. In this case the scattering vectors and averaged coherency
and Mueller matrices have the form shown in equation (2.105):
   
a+d a+d
 a−d   a−d 
k=  
 b + c  → kγ =  b + c 


i(b − c) −i(b − c)
   
t11 t12 t13 0 m11 m12 m13 m14
t ∗ t22 t23 0   m12 m22 m23 m24 
→ T  =  
t ∗ t ∗ t33 0  ⇔ M  =  m13
12  
13 23 m23 m33 m34 
0 0 0 t44 −m14 −m24 −m43 m44
(2.105)
Again the SU(4) eigenvector dependence can be represented as a 3 × 3 unitary
matrix as shown in equation (2.106), so again SU(2) controls the depolarisation
subspace.
 
  [U2 ] 0 0
[U3 ] 0
[U4 ] = ⇒ [U−1 ] [U4 ] =  0 1 0 (2.106)
0 1
0 0 1
The polarised/depolarised decomposition then has ten parameters arranged as
shown in equation (2.107):

T4bi sec trix = (4 + 1) + [2 + 3] (2.107)

 2.4 General theory of scattering entropy 107

In the most general symmetry case, when all the above symmetries apply—as,
for example, for a random cloud of spheroidal particles—we obtain the averaged
bistatic coherency matrix by adding coherency matrices for all elements of the
scattering matrix group, as shown in equation (2.108):
       
a+d a+d a+d a+d
       
 a−d   a−d   a−d   a−d 
k=  → kα =   → kβ =   → ky =  
 b+c  −(b + c)  −(b + c)   b+c 
i(b − c) i(b − c) −i(b − c) −i(b − c)

   
t11 t12 0 0 m11 m12 0 0
+ symmetric ,  ∗
t12 t22 0 0 
m
 12 m22 0 0 
⇒ Twave =  ⇔ M  =  
0 0 t33 0  0 0 m33 m34 
0 0 0 t44 0 0 −m34 m44
(2.108)

This is the bistatic equivalent of azimuthal symmetry in backscatter problems.

We see that in this case the SU(4) matrix governing the eigenvectors reduces
to an SU(2) dependency as shown in equation (2.109), and so again there are
no eigenvector parameters in the depolarisation subspace.
 
[U2 ] [0]
[U4 ] = ⇒ [U−1 ] [U4 ] = [I3 ] (2.109)
[0] [I2 ]

In this case the polariser/depolariser decomposition takes on the reduced form

shown in equation (2.110):

bistatic+symmetry
[T ] −−−−−−−−−→(2 + 1) + [0 + 3] (2.110)

2.4.3.2 The scattering sphere

This is a very useful result, as this type of symmetry is quite common in envi-
ronmental remote sensing applications (see Chapter 3). Here we see that in the
most symmetric case, the polarising/depolarising decomposition can be charac-
terized by six parameters in a 2:1:0:3 cascade. Hence the principal eigenvector
can be entirely associated with polarised behaviour of the system, with a scat-
tering strength given by the largest eigenvalue of [T ]. The principal eigenvector
has only two complex elements, and hence corresponds to diagonal amplitude
matrices. Now using the SU(2)–O3 mapping (see Appendix 2) we can project all
such diagonal matrices onto the surface of a sphere in a real three-dimensional
space. This sphere shares many properties with the Poincaré sphere, but each
point now represents a scattering amplitude matrix rather than a wave state. To
distinguish the two we term this a scattering sphere.
A general polarised eigenvector can then be parameterized in terms of two
angles on this sphere, as shown in equation (2.111) and geometrically in
Figure 2.8.
108 Depolarisation and scattering entropy

2α

Fig. 2.8 Scattering sphere showing the geo- 1 0

metrical interpretation of α and δ from [S] =
0 1
equation 2.111 Scattering sphere

 
cos α
(2 + 1) → λmax (2.111)
sin αeiδ

Note that α = δ = 0 corresponds to the identify amplitude matrix, and so

any departure of these parameters from this point represents a departure from
trivial scattering behaviour. There also remain three sub-eigenvalues that fully
characterize the types of depolarisation that can occur around this polarised
system.
While it is possible to classify directly in the space of δ1 , δ2 , δ3 , it is often again
more convenient to use secondary parameters derived from the eigenvalues and
more closely related to the degree of polarisation widely used in Stokes algebra.
These parameters are defined in equation (2.112) as the scattering entropy H
and scattering anisotropies Aij :

!
4
λi
H =− Pi log4 Pi 0 ≤ H ≤ 1, Pi = 5
λ
i=1 (2.112)
λi − λj
Aij = 0 ≤ Aij ≤ 1, i>j
λi + λj

All vary between 0 and 1, with entropy being a general measure of total depolar-
isation. It is zero for non-depolarising systems, and 1 for the ideal depolariser.
The anisotropies then provide information about the variation of depolarisation
with changes in polarisation: if they are zero then there is no variation, while
as they approach 1 there exists subspaces in the polarisation domain where
depolarisation can be small. This then leads us naturally back to the idea of the
entropy/alpha decomposition.

2.4.3.3 The bistatic entropy/alpha decomposition

For backscatter problems we found that the simplified two-parameter character-
ization of the entropy/alpha plane provided a convenient way to map different
types of depolarisation. One attractive feature of this approach is that it scales to
different dimensions, including the N = 4 bistatic case. In this case the change
of dimensionality is accommodated by an extra eigenvalue and eigenvector, as
 2.4 General theory of scattering entropy 109

shown in equation (2.113) (Cloude, 2005b, 2006a):

 
λ1 0 0 0
 0 λ2 0 0  ∗T  
T  = U4  0 0 λ3 0  U4
 U = e14 e24 e34 e44
0 0 0 λ4
  (2.113)
cos α eiφ1
 sin α cos ψ eiφ2 
e4 = sin α sin ψ cos γ eiφ3 


sin α sin ψ sin γ eiφ4

The entropy/alpha parameters can now be defined as shown in equation (2.114):

!
4 !
4
H =− Pi log4 Pi 0 ≤ H ≤ 1, α= Pi αi 0 ≤ α ≤ 90◦ (2.114)
i=1 i=1

Again there are bounding curves in this plane. The upper and lower values for
α as a function of increasing entropy can be calculated as shown in equation
(2.115). Figure 2.9 shows how the valid region in the H/α plane for N = 4
bistatic scattering differs from that used in the N = 3 backscatter and N = 2
dual polarisation cases.
At low entropies there is little difference between the N = 2 and N = 3
cases, but at higher entropy values the alpha range for N = 4 is shifted to
higher values. This is a consequence of the addition of a new eigenvector, itself
having an alpha value of π/2. This acts to lift the mean alpha value for any
given entropy. In the limit of H = 1, the alpha range is reduced to a single point
at 3π/8. Again the utility of this diagram is not so much in its high fidelity
parameterization of depolarisation in N = 4 cases, but in the convenience of
representing the most general bistatic scattering problem by one polarising and
one depolarising parameter, in such a way as to maintain invariance to changes

H/alpha diagrams: N=2 (solid), N=3 (dash), N=4 (dot)

90

80

70

60

50
Alpha

40

30

20

10

0
0 0.2 0.4 0.6 0.8 1 Fig. 2.9 Summary of entropy/alpha dia-
Entropy grams for N = 2, 3, and 4 scattering
110 Depolarisation and scattering entropy

of polarisation base.
  
1 0 0 0  3mπ

 α=
0 m 0 0 2(1 + 3m)
TI = 
0
 ⇒
0 m 0 
 1 m3m
H = − log4
0 0 0 m 1 + 3m (1 + 3m)3m−1
   π
0 0 0 0 
 α=
0 1 0 0 2
 0 ≤ m ≤ 0.5 ⇒
0 0 2m 0   1 2m4m
H = − log4
0 0 0 2m (1 + 4m) (1 + 4m)4m+1
  
2m − 1 0 0 0 

3π
 0  α=
 1 0 0  0.5 ≤ m ≤ 1 ⇒ 4m +4
 0 0 1 0  1 (2m − 1)2m−1

H = − log4
0 0 0 1 (2 + 2m) (2 + 2m)2+2m
(2.115)

2.5 Characterization of depolarising systems

Finally we amalgamate the results of the previous sections to re-examine the
idea of depolarisation and its various parameterizations. For example, we saw
that the vector wave reciprocity theorem in backscatter causes a symmetry in
[S] which limits the form of the Mueller matrix (for arbitrary random scattering
problems) to that shown in equation (2.116), where we note that there is an
important additional constraint equation on the diagonal elements, leaving [M ]
with only nine rather than sixteen degrees of freedom. Reciprocity symmetry
therefore acts to limit the types of depolarisation we can observe in backscatter.
 
  m11 m12 m13 m14
a b  m12 m22 m23 m24 
[S]FSA = ⇒ [M ]FSA =  −m13 −m23 m33 m34  (2.116)

−b d
m14 m24 −m34 m44
Reciprocity ⇒ m11 − m22 + m33 − m44 = 0

The most extreme example is the isotropic depolariser, with a Mueller matrix
[MI ] of the form shown on the left-hand side of equation (2.117):
   
1 0 0 0 1 0 0 0
0 0 0 0 0 δ 0 0
[MI ] =  
0 0 0 0 ⇒ [MII ] = 0 0 δ 0
 

0 0 0 0 0 0 0 δ
 
1 0T
[MIII ] =
0 [O]3 [][O]T3
 
⇒ δ1 0 0 (2.117)
[] =  0 δ2 0 
0 0 δ3

This matrix converts all Stokes vectors into a randomly polarised wave, but
there is no corresponding single [S] matrix. This form leads to a standard
 2.5 Characterization of depolarising systems 111

generalization of the depolariser, as shown in two stages from left to right in

equation (2.117). The middle [MII ] is a partial depolariser, while the right-
hand form [MIII ] generalizes to an anisotropic partial depolariser with arbitrary
direction in Stokes space (the matrix O3 is a 3 × 3 real rotation matrix of the
Poincaré sphere). As an example of a partial depolariser, consider the special
case of forward scattering by random particles that takes the form shown in
equation (2.118):

 
1 0 0 0 
0 |ε| ≤ 1
δ 0 0
[M ] = 
0
 ⇒ |δ| ≤ 1 (2.118)
0 δ 0 
2δ − ε ≤ 1
0 0 0 ε

Again we note important physical restrictions on the diagonal elements from the
forward scattering symmetry. Given the need to always add additional constraint
equations, we can ask if all triplets δ1 , δ2 , δ3 are physically consistent, and
whether the forms in equation (2.117) exhaust all possibilities.
In this regard we have shown how we can use the coherency matrix concept
to classify depolarisers. For example, by mapping the general depolariser 
of equation (2.117) into [T ] we see that the real diagonal elements δ1 , δ2 and
δ3 are constrained by the four inequalities shown in equation (2.119). We can
interpret this geometrically by first considering a real three-dimensional space
formed by triplets δ1 , δ2 , δ3 . If we normalize the Mueller matrix to have m11 =
1, the unit cube in this space then represents the set of candidate depolarisers,
as shown on the left-hand side of Figure 2.10. Equation (2.119) then represents
the equation of four planes in this space, and all physical depolarisers must be
inside the volume bounded by these planes. The shape of this region is shown
on the right in Figure 2.10. To illustrate the sometimes subtle implications of
these constraints, we show, in equation (2.120), three candidate depolarisers,
all of which seem to be reasonable suggestions:

   
1 1 1 1 1
1 1 −1 −1 δ1 
   
1 −1 1 −1 . δ2  ≥ 0 (2.119)
1 −1 −1 1 δ3

Fig. 2.10 Section of the depolarising unit

cube (left) and restricted physical depolariser
region (right)
112 Depolarisation and scattering entropy

   
1 0 0 0 1 0 0 0
0 0.5 0 0 0 0.5 0 0 
[M1 ] = 
0
 [M2 ] =  
0 0.4 0 0 0 0.4 0 
0 0 0 0.3 0 0 0 −0.3
 
1 0 0 0
0 0.5 0 0 
[M3 ] = 
0
 (2.120)
0 −0.4 0 
0 0 0 −0.3

However, using the constraints of equation (2.119) we see that [M2 ] is not a
physical depolariser. Its corresponding coherency matrix has a negative eigen-
value, and hence no physical system can generate such a depolarisation process.
Note, however, that [M1 ] and [M3 ] are both located inside the feasible volume
and therefore correspond to realizable depolarisers.
The constraints on forward scattering and backscattering matrices in equa-
tions (2.116) and (2.118) can now also be seen as special cases of this geometry.
In forward scattering, rotation symmetry constrains δ1 = δ2 = δ, and the
following special case is thus obtained:
   
1 1 1 1 1 1 + 2δ + ε ≥ 0
1 1 −1 −1 δ 
 .  ≥ 0 ⇒ 1 − ε ≥ 0 (2.121)
1 −1 1 −1 δ  1−ε ≥0
1 −1 −1 1 ε 1 − 2δ + ε ≥ 0

We see that these inequalities give rise to the constraints shown in equation
(2.118). Equally in backscatter, the reciprocity theorem reduces [S] by relating
the crosspolarised elements, and this has the effect of reducing the rank of [T ]
from 4 to 3. Hence one eigenvalue (λ3 ) of [T ] must always be zero, even in
the presence of depolarisation. This then corresponds directly to the constraint
equation shown in equation (2.116). We see that the geometry of Figure 2.10 is
important in classifying depolarisers, and underlines the more fundamental role
played by [T ] rather than [M ] in understanding processes of wave scattering
and depolarisation.
We can further extend this analysis to include the case of a ‘general’ depo-
lariser proposed on the far right in equation (2.117), and obtained by a rotation
of the Poincaré sphere (Lu, 1996). The transformation of [T ] under such a
rotation can be derived using the mapping from rotations in a real three-
dimensional space to unitary transformations in a two-dimensional complex
space: the SU(2)–O3 homomorphism, as shown in equation (2.122).

  SU (2) − O+
1 0
g −−−−−→[S] = [U2 ] [S] [U2 ]∗T
3
g =
0 [O3 ] (2.122)
∗T ∗
⇒ [T ] = [U4B ] [T ] [U4B ] [U4B ] = [U2 ] ⊗ [U2 ]

We see now, however, that this represents only a subset of possible depolarisers,
and a more general classification should be based on the full 4×4 unitary matrix
transformations of [T ] = [U4 ] [T ] [U4 ]∗T . This underlines the importance of
general unitary transformations in the analysis of polarisation phenomena.
 2.6 Relating the Stokes/Mueller and coherency matrix formulations 113

2.6 Relating the Stokes/Mueller and coherency

matrix formulations
In this chapter we have introduced an important new formulation of polarised
scattering—the coherency matrix [T ]—and it is now of interest to formally
connect the coherency and Mueller matrix formulations via their treatment of
scattered intensity. We start by again considering a Pauli matrix expansion of
the amplitude matrix as shown in equation (2.123):
 
1 k0 + k1 k2 + ik3
E s = [S] E i = √ E i k0 , k1 , k2 , k3 ∈ C (2.123)
2 k2 − ik3 k0 − k1
We then consider a hypothetical (but always physically realizable) measure-
ment system, represented now by a complex four-element vector w, formed as
complex weights of the Pauli coefficients, as shown in equation (2.124):
p = w1 k0 + w2 k1 + w3 k2 + w4 k3 = w∗T k ∈ C
(2.124)
|p|2 ≥ 0 for w ∈ C4 w∗T w = 1
In particular we note that for arbitrary choice of w, the squared amplitude of
the projection p is always non-negative, as shown. This is important when we
consider generalization to the case of scattering from random media, when
equation (2.224) takes on the form shown in equation (2.125), where .. is now
an ensemble average.
+ , + , + ,
m = pp∗T = w∗T kk ∗T w = w∗T kk ∗T w = w∗T [T ] w ≥ 0 (2.125)
Here [T ] is the 4 × 4 scattering coherency matrix. We see that the non-negative
constraint from equation (2.124) requires [T ] to be positive semidefinite; that
is, to have a non-negative real eigenvalue spectrum. Now we have seen that
[T ] is related in a 1–1 mapping to the 4 × 4 real Mueller matrix [M ]. The
explicit mapping from [M ] into [T ] is shown in equations (2.48) and (2.49).
More formally, we can relate the sixteen-element vectors t and m, formed by
expanding [T ] and [M ] row-wise: m = [m11 , m12 , m13 … m44 ]T , by a 16 × 16
complex matrix [Q], as shown in equation (2.126). This gives the same result as
in equation (2.48), but with a more formal notation we can make use of further
analysis:
 
Q1 Q2 Q3 Q4
1 Q2 Q1 iQ4 −iQ3 
t = [Q] m ⇒ [Q] =  
2 Q3 −iQ4 Q1 iQ2 
Q4 iQ3 −iQ2 Q1
   
1 0 0 0 0 1 0 0
0 1 0 0 1 0 0 0
Q1 =   
0 0 1 0 Q2 = 0 0 0 i 


0 0 0 1 0 0 −i 0
    (2.126)
0 0 1 0 0 0 0 1
0 0 0 −i 0 0 i 0
Q3 =   
1 0 0 0  Q4 = 0 −i 0 0


0 i 0 0 1 0 0 0
114 Depolarisation and scattering entropy

For example, we can now relate the scattered intensity in the Stokes/Mueller
formulation to that in the coherency matrix form by equating expressions for
scattered intensity and using [Q], as shown in equation (2.127):
 
w∗T [T ]w = g Tr [M ] g t ⇒ w∗T ⊗ wT t = g Tr ⊗ g Tt g
 
⇒ w∗T ⊗ wT [Q] = g Tr ⊗ g Tt (2.127)
 
⇒ w∗T ⊗ wT = g Tr ⊗ g Tt [Q]−1

Hence any choice of transmit and receive Stokes vectors g t and g r can be
transformed into an equivalent w formulation and vice versa. Given a complex w
we can convert it into a weight vector for the elements of [M ], a subset of which
can then be further decomposed into a direct product of Stokes vectors. Note,
however, that from a physical point of view the most important relationship
is equation (2.124). This relationship is not so evident in the Stokes/Mueller
formulation, and so we can in some sense consider the [T ] formulation to be
more fundamental.
In the previous chapters we considered in detail an algebraic formulation
of vector wave propagation and scattering, and have used this to establish a
detailed parameterization of depolarisation phenomena. It is now time to con-
sider examples of the application of this theory to physical models for surface
and volume scattering.
Depolarisation in
surface and volume
scattering
3
In Chapter 1 we introduced the scattering amplitude matrix [S] and its trans-
formation properties, and in Chapter 2 we saw how to extend these concepts
to include the presence of depolarisation. Here we derive examples of the
explicit form of the amplitude matrix and its coherency structure for two impor-
tant canonical problems: surface scattering and volume scattering. Details of
the underlying derivations are widely available in the existing literature, with
several excellent books covering the topics required (Born and Wolf, 1989;
Ishimaru, 1991; Jackson, 1999; Kong, 1985; Lakhtakia, 1989; Tsang, 1985;
Ulaby, 1986; van de Hulst, 1981; Mishchenko, 2000). Here our purpose is
mainly to concentrate on the polarimetric properties of these solutions, and so
we proceed by first summarising the key formulae and then move on to high-
light their important polarimetric structure in terms of polarised and depolarised
component decompositions. We also concentrate on those models that a) pro-
vide a fully polarimetric solution, including phase and amplitude effects, and b)
lead to simple analytical results without the need for numerical methods such
as finite difference and finite element. This will enable us to establish a few
suitable benchmarks for later comparison with numerical experiments or more
advanced modelling techniques. Inevitably this means that we concentrate on
low-frequency solutions, where analysis is tractable. Hence we start by con-
sidering scattering by electrically smooth surfaces and by volumes composed
of electrically small particles. However, these not only have direct relevance
to low-frequency radar applications (particularly at L and P bands), but also
provide a framework by which to judge more complicated high-frequency
solutions.
In order to calculate a general [S] matrix we must solve Maxwell’s differential
equations together with appropriate boundary conditions applied on the surface
of the object. From the former we postulate polarised plane wave solutions
and, by enforcing the latter, obtain a set of equations for the unknown co- and
crosspolarised scattering coefficients in a chosen coordinate system. These can
then be solved for two orthogonal incident polarisation states to obtain the four
complex elements of the scattering amplitude matrix [S]. Key to this process
is an understanding of the behaviour of vector EM fields across a boundary
E 1 H1 D1 B1
between two materials. n

There are four sets of boundary conditions arising from the set of four n ×E1 = n × E2 n. D1 = n. D2
Maxwell equations (see equation (1.1)). At an interface between two media, E 2 H2 D2 B2
n ×H1 = n × H2 n B1 = n B2
the two curl equations require continuity of the tangential components of E and
H , while the divergence equations require continuity of the normal components
of D and B. For a general boundary we then define the local normal vector n, as Fig. 3.1 Vector boundary conditions in elec-
shown in Figure 3.1. Note that of the four sets of boundary conditions only two tromagnetic scattering
116 Depolarisation in surface and volume scattering

are independent, the others being related by Maxwell’s equations. In practice,

therefore, any two of the four can be chosen, depending on analytical conve-
nience. We now turn to consider the application of these ideas to the first of our
two themes: surface scattering.

3.1 Introduction to surface scattering

As a simple example of the above procedure, consider the classical problem of
reflection and transmission at a plane dielectric interface (Born and Wolf, 1989;
Hecht and Zajac, 1997). The upper medium has a relative dielectric constant ε1
and the lower ε2 , as shown in Figure 3.2. On the left-hand side is shown the case
for an incident electric field polarised parallel to the plane of incidence. This
corresponds to vertical polarisation, V, in radar applications, otherwise termed
‘p’ polarisation or TM for transverse magnetic wave. The magnetic field vector
H is shown at right angles to both the electric field and the wave propagation
vector β i , as follows from Maxwell’s equations for wave propagation. From
the laws of reflection and refraction, the reflected and transmitted fields are zero
everywhere except in specific directions. The reflected wave is such that the
surface normal is the bisectrix vector (angle of incidence equals the angle of
reflection), and the refracted wave obeys Snell’s law relating the transmission
angle to the incident angle and refractive index n, given by the Maxwell relation
as the square root of dielectric constant, as shown in equation (3.1) (Born and
Wolf, 1989; Hecht and Zajac, 1997):
√
n1 sin θi = n2 sin θt n = εr (3.1)

From equation (1.145) we can then generate the S β daughter problem (without
yet calculating the detailed form of the matrix), which from the symmetry of
the plane must be equal to the original mother matrix S. The only solution
to this constraint is that b = c = 0, or the matrix has zero crosspolarisation.
Hence a smooth infinite plane interface causes zero crosspolarisation. This
leaves only the two diagonal copolar elements of the amplitude matrix. Note
that this is true for any scattering angle θ (not just specular) and for arbitrary
surface roughness, as long as the roughness profile is one-dimensional (like a
corrugated rough surface).
For each polarisation combination, such as VV, we have two unknowns:
the reflected and transmitted electric field components. However, by matching

Ei n Er n
Ei Er

Hi Hr

θi θr θi θr
ε1 θr ε1 θt
x
ε2 ε2
Et
Et
Ht
Fig. 3.2 Surface reflection and transmission z
for incident V (parallel, ‘p’ or TM) and H
(perpendicular, ‘s’ or TE) polarisations V or parallel (p) incident polarisation H or perpendicular (s) incident polarisation
 3.1 Introduction to surface scattering 117

the tangential components of E and H across the boundary, two independent

equations can be obtained. This is sufficient to provide a full solution, as we now
demonstrate. First, however, we consider the perpendicular polarisation case.
On the right-hand side of Figure 3.2 we show the corresponding HH problem
(often called ‘s’ polarisation from the German senkrecht, meaning perpendicu-
lar, or alternatively TE for Transverse Electric waves) when the electric field is
now polarised perpendicular to the plane of incidence. The same procedure can
again be used to find the reflection and transmission coefficients, as follows.
We start by defining the values of the field components at the interface (z = 0)
by subscripts ‘r’ for reflected, ‘i’ for incident, and ‘t’ for transmitted. We then
choose to enforce continuity of tangential E and H fields (see Figure 3.1) to
obtain two equations for the unknown coefficients, as shown in equation (3.2):

Hrs cos θr − His cos θi = −Hts cos θt ⇒ (Hrs − His ) cos θi = −Hts cos θt
8 s
cµo
Ei + Ers = Ets (3.2)
E= n H⇒
n1 (Ei − Ers ) cos θi = n2 Ets cos θt
s

The first equation follows from the continuity of tangential electric field for ‘s’
polarisation. The second requires some manipulations, projecting the H field
onto the x axis and using the relationship between E and H components of a
plane wave from Maxwell’s equations. Note that c is the velocity of light in
free space.
Similarly, for the V or ‘p’ polarised problem, when the E field is polarised in
the plane of incidence, we can obtain two equations by enforcing continuity of
tangential E and H , as shown in equation (3.3):
8 p p p
n1 (Ei + Er ) = n2 Et
p p p (3.3)
(Ei − Er ) cos θi = Et cos θt

Both solutions for reflection and transmission coefficients can then be combined
into the reflection and transmission matrices, as shown in equation (3.4). Note
that for simplicity of notation we assume non-magnetic materials where the
relative permeability µr = 1.

3.1.1 The Fresnel equations


 n1 cos θi − n2 cos θt
  
RHH =
RHH 0 n1 cos θi + n2 cos θt
[R] = ⇒
0 RVV 
 n cos θi − n1 cos θt
 RVV = 2
n2 cos θi + n1 cos θt
 (3.4)
 2n1 cos θi
  
THH = n cos θ + n cos θ
THH 0 1 i 2 t
[T ] = ⇒
0 TVV 
 2n1 cos θi
TVV =
n2 cos θi + n1 cos θti
These four expressions are called the Fresnel equations, after the French sci-
entist Augustin-Jean Fresnel (1788–1827). Despite their relative simplicity
and longevity, they remain of fundamental importance in understanding the
interaction of polarised waves with media. The reflection coefficients can also
118 Depolarisation in surface and volume scattering

be expressed in terms of ‘exterior’ parameters θi and ε2 by using Snell’s law

(equation (3.1)) to remove θ2 , as shown in equation (3.5):
ε2 ε2 ε2
cos θi − − sin2 θi cos θi − − sin2 θi
ε1 ε1 ε1
RHH = RVV = (3.5)
ε2 ε2 ε2
cos θi + − sin2 θi cos θi + − sin2 θi
ε1 ε1 ε1

This latter form is more commonly used in radar scattering problems. Further-
more, the incident wave is often in free space, and so we can set ε1 = 1 to
further simplify the equations. Note that these equations apply even for com-
plex permittivity ε2 , such as occurs for microwave reflection from wet soils and
sea ice, for example. In this case the wave is changed in both amplitude and
phase on reflection, but equation (3.5) is still valid. We now turn to consider
the properties of such lossy material in more detail.

3.1.1.1 Fresnel equations for lossy dielectric media

In the presence of losses, the dielectric constant of the lower medium must be
written (for an exp(iωt) time variation) as a complex number εr = ε − iε ,
where the ratio tan δ = εε is called the loss tangent of the material. For exam-
ple, in a conducting medium with conductivity σ it follows from Maxwell’s
equations that J = σ E and the dielectric constant becomes complex with
wavenumber β, as shown in equation (3.6):
σ √
εr = ε − iε ≈ ε − i ⇒ β = ω0 µε = β0 n = β0 n − iβ0 n (3.6)
ωεo
Note, however, that since the material remains homogeneous, the wave polar-
isation state does not change with propagation into the material. The only
‘filtering’ of relative polarisation amplitude and phase is by the surface trans-
mission given by the Fresnel equations (3.4). However, it is of interest to study
a little further the properties of waves propagating in such lossy material. This
will lead us to some important concepts concerning penetration depth and its
relationship to volume scattering.
To study the propagation of signals into lossy material we first write the
wave vector in the material in terms of its ‘x’ (parallel to the interface) and ‘z’
(perpendicular to the interface) components, as shown in equation (3.7):

E2 = T (θ ) exp(−i(β2x x + β2z z)) (3.7)

where T (θ ) is the appropriate Fresnel transmission coefficient. We are interested

to see how such a wave propagates in the medium below the interface. The
components of the wave vector can be explicitly derived as shown in equation
(3.8), where we have assumed that the upper medium is free space and has
propagation constant β1 = β0 = 2π λ .

β2x = β1x = β0 sin θi


β2z = β2 cos θt = β22 − β12 sin2 θi (3.8)
 √
= β0 ε − iε − sin2 θi = β0 a + ib = βz − iκz
 3.1 Introduction to surface scattering 119

Evaluating the square root leads to the following result:

Et = T exp (−iβ0 sin θi x − iβz z) exp(−κz z) (3.9)
where the components κz and βz are real and defined, as shown in equation
(3.10):
&
 ' 0.5
β0 ' ε 2
2 (
βz = √ ε − sin θi 1+ +1
2 (ε − sin2 θi )2
& (3.10)
'
β0
 ' ε 2
0.5
κz = √ ε − sin2 θi ( 1 + −1
2 (ε − sin2 θi )2

This shows that the field inside the material decays exponentially from the
surface (and to first order the decay rate is proportional to frequency, so increas-
ing at higher frequencies) with coefficient κz , and that the equiamplitude and
equiphase contours are no longer the same. This is called an inhomogeneous
plane wave, and is characteristic of propagation into lossy material (those with
a loss tangent greater than zero). This idea of projecting the wave vector β o
onto surface components is also an important one, and will be used extensively
in our analysis of interferometry in Chapter 5.
Note that wave attenuation is more often expressed in terms of a power
attenuation coefficient σe with units of decibels/meter (dB/m). Equation (3.11)
can then be used for the conversion from κz .
20
σedB = κz = 8.686κz dB/m
ln(10)
(3.11)
ln(10) dB
κz = σ = 0.115σedB m−1
20 e
A key concept in such material is the penetration or skin depth, defined as the
depth at which the signal is attenuated to exp(−1) (−8.686 dB). This can be
considered a typical penetration depth into the material. In general, therefore,
the penetration depth is given in terms of the complex refractive index, as shown
in equation (3.12):
√
1 2
δ= = " 0.5 (3.12)
κz ε 2
β0 ε − sin2 θi 1+ −1
(ε − sin2 θi )2
This is the most general expression, but in the special case of very good con-
ductors with conductivity σ this expression takes the more familiar form of the
skin depth shown in equation (3.13):
"
2
δ≈ (3.13)
ωµ0 σ
As an important example of a lossy dielectric in microwave remote sensing, and
one that illustrates the underlying complexity of modelling dielectric constant,
we now summarize the main dielectric properties of soil.
120 Depolarisation in surface and volume scattering

3.1.1.2 Example: dielectric properties of soil

The interpretation of wave reflections from land surfaces involves an under-
standing of the dielectric constant of soil. This is such a basic requirement that
it is perhaps surprising to realize the complexity of this problem. At microwave
frequencies this complexity arises predominantly from the different forms that
water can take in soil, and its impact on the resulting loss-tangent.
In general there are two main contributions to the loss-tangent of a mate-
rial. The conductivity σ tends to be more important at low frequencies, but at
higher frequencies dipolar losses due to molecular absorption tend to dominate,
especially in the microwave spectrum. The predominant effect here, therefore,
is water absorption, which has a broad resonant spectrum in the microwave
region. Hence the presence of increasing water content tends to increase the
loss tangent and thus reduce the penetration depth into the soil. However, com-
plexity starts to arise when we consider that water in soil can be found in two
main forms, bound or free, and these have very different dielectric properties.
Since the proportion of bound and free water depends on soil structure, it fol-
lows that dielectric constant then depends on soil texture, and also shows a
threshold effect; that is, above a certain transition water content the dielectric
constant suddenly increases. In general, therefore, for lossy soils the loss tan-
gent tends to be large at low frequencies and reduces to a minimum around 100
MHz before increasing again to reach a plateau around 3 GHz.
In more quantitative terms, soil can be considered a complex four-phase
composite of air, bulk soil, and bound and free water, with a complex dielectric
constant given by composite models of the general form shown in equation
(3.14) (Dobson, 1985), which is valid in the frequency range 1.4–18 GHz.

ρb x  y x
εsoil
x
=1+ εss − 1 + mv εfw − mv (3.14)
ρss

where the following terms can be defined:

ρb = soil bulk density (soil/air mixture)

ρss = density of solid soil (around 2.66 g/cm3 )
εss = 4.7 − i0 = dielectric constant of dry soil
x = 0.65 (an average over 500 samples of five soil types)
y = y0 − y1 S + y2 C, S = Sand fraction, C = Clay fraction, where the
three real coefficients y0 , y1 and y2 are chosen to accommodate
changes in soil texture (see Dobson, 1985)
mv = volumetric moisture content

The imaginary component arises from the final term, defined from a Debye
model of water absorption as follows:

74.1 iσδ
εfw = 4.9 + − (3.15)
1 + i ff0 2π ε0 f

where σδ is the soil conductivity, and f0 the relaxation frequency (which varies
from 9 GHz at 0◦ C to 19 GHz at 25◦ C). Note that one of the main contributions to
soil conductivity is its salinity ‘sa ’ in ppthou with a typical derived relationship
 3.1 Introduction to surface scattering 121

f (z) = e −κ z
z

Fig. 3.3 Vertical structure function for wave

penetration into lossy material

of the form shown in equation (3.16) (Ulaby, 1986).

σδ = 0.18252sa − 1.4619e − 3s2a + 2.093e − 5s3a − 1.282e − 7s4a (3.16)

Typically, in remote sensing one is interested in estimating soil salinity or

moisture indirectly from estimates of the dielectric constant [Oh, 1992; Dubois,
1995; Hajnsek, 2003; Allain, 2003). For example, a commonly used inversion
formula for soil moisture is shown in equation (3.17) (Topp, 1980):

mv = 10−2 (−5.3 + 2.92εr − 0.055εr2 + 0.0004εr3 ) (3.17)

Hence we see that key to many remote sensing applications is estimation

of dielectric constant, and later we shall see that polarisation diversity often
provides a useful way to isolate dielectric constant in scattering problems.
In any case, we see that such lossy materials are characterized by exponential
decay normal to the surface, as shown schematically in Figure 3.3. In this sense
all homogeneous lossy media can be characterized by an exponential amplitude
function with depth, with a characteristic penetration depth given by δ. We shall
see later (in Chapter 8) that inhomogeneous materials can be characterized by
different structure functions, and how in remote sensing we can devise methods
for reconstruction of this function from scattered field data.
Finally, we note an interesting limit when κz tends to infinity and the pen-
etration depth tends to zero. This occurs, for example, for metal surfaces at
microwave frequencies. In this case the surface acts as a ‘short circuit’ to the
incoming wave with zero transmitted wave, and hence the boundary condition
on the tangential electric field requires that the total tangential electric field at
the surface is zero. For this to occur the reflected wave must be 180◦ out of
phase with the incident wave, and hence the reflection coefficient is −1. This
phase shift on reflection should be embedded in the Fresnel equations, and it
is interesting to ask how such phase shifts on reflection vary with dielectric
constant, polarisation and angle of incidence.
To investigate this, we show in Figure 3.4 an example of the Fresnel equations
evaluated for ε1 = 1, ε2 = 4. We make the following general observations from
this example:
• The magnitude of the HH channel is always greater than or equal to VV.
In fact, |VV| decreases with increasing angle of incidence and reaches
zero at the Brewster angle (for complex dielectric constants the Brewster
angle reaches a minimum of reflection but not exactly zero). This angle
is a function of the ratio of dielectric constants of the two materials, as we
now show. By considering the reflection process as scattering by electric
dipoles (see equation (1.10)) in the lower medium, we can obtain a zero
reflected signal under condition that the sum of refracted and reflected
angles is π/2. In this case the reflected component lies along the null
122 Depolarisation in surface and volume scattering

Fresnel equations evaluated for n=2

1
R HH
0.8 R VV
T HH

Reflection/transmission coefficients
0.6 T VV
0.4

0.2
0

–0.2

–0.4

–0.6
–0.8
–1
Fig. 3.4 Fresnel reflection and transmission 0 10 20 30 40 50 60 70 80 90
coefficients for εr = 4 Angle of incidence (degrees)

in the dipole radiation pattern. Using Snell’s law we can then obtain an
expression for this Brewster angle θB , as shown in equation (3.18):

π εr2
θr + θ t = ⇒ n1 sin θi = n2 cos θi ⇒ tan θB = (3.18)
2 εr1
Finally, we note that at normal incidence (θ = 0◦ ) both HH and VV have
the same magnitude but a 180-degree phase difference (see below for a
discussion of phase and coordinates), while at grazing incidence (θ =
90◦ ) both |HH| and |VV| have unit magnitude and zero phase difference.
Hence there must be a switch in phase for one of the polarisation channels
at some angle of incidence.
• The phase of the HH reflected signal is always 180 degrees, for ε2 > ε1 .
From Figure 3.2 we see that this means that the perpendicular component
(H) undergoes a 180-degree phase shift on reflection, as discussed in the
limit of perfect conductors. This is a direct consequence of the boundary
conditions and coordinate system employed. The VV channel phase has
a more complicated structure. We first note that the VV channel has zero
phase for angles of incidence less than the Brewster angle and a 180-
degree phase shift thereafter. This can be explained by inspection of the
coordinate systems employed in Figure 3.2, as follows. From Figure 3.2
we see that the ‘zero’ VV phase is actually a consequence of the same
180◦ phase change on surface reflection that is seen in the HH component
V
(consider the case as θi tends to zero). The reason it appears as a zero
H phase is that the surface coordinate definition of in-phase for polarisation
V components parallel to the plane of incidence is that their z components
are parallel.
H To further illustrate this coordinate-dependent phase notation, we show in
Incident wave Reflected wave
Figure 3.5 the coordinates used for a sensor-oriented (BSA) description of
Fig. 3.5 Sensor (BSA) coordinate system normal incidence reflection. Note that by using the same logic as above, the
for describing normal incidence Fresnel HH reflection coefficient will now be positive (with the 180-degree surface
reflection phase shift now matching the coordinate shift of the BSA system), while VV
 3.1 Introduction to surface scattering 123

will be negative. Hence we obtain the modified normal incidence reflection

matrix based on the Fresnel equations as shown in equation (3.19) (see also
equation (1.152)), which was used to obtain the specular backscatter matrix in
equation (1.153).
   
−RHH 0 n2 − 1 1 0
[R]normal = = (3.19)
0 −RVV n2 + 1 0 −1

Now, however, we are in a position to generalize this result to arbitrary angle of

incidence. The sensor coordinate scattering matrix for Fresnel reflection then
has the general form shown in equation (3.20):
 
 ε2
ε1 − sin θi − cos θi
2



 SHH = 


  
 cos θi + εε21 − sin2 θi
SHH 0
[S]sensor = A ⇒  (3.20)
0 SVV 
 ε2 ε2

 cos θ − ε1 − sin θi
2

 =
ε1 i


S VV
 ε2 ε2
ε1 cos θi + ε1 − sin θi
2

with the important special case of normal incidence (θ = 0◦ ) yielding a scalar

multiple of the identity matrix, as shown in equation (3.21):
 
1 0
[S]BSA = A (3.21)
0 1

3.1.2 Polarisation properties of surface backscatter

We now consider the important special case of surface backscatter. We begin
with the Fresnel equations and note that, aside from ‘specular reflection’ at nor-
mal incidence, the energy is all reflected away from the backscatter direction.
Hence, perfectly smooth surfaces have zero backscatter for oblique incidence.
However, most natural surfaces are not smooth, and so some backscatter sig-
nal is observed. In this section we develop the polarisation properties of the
backscattering matrix for such rough surfaces.
Backscatter n
The first modification we must make to the Fresnel equations is to allow for
the finite size of any surface element or ‘facet’. Figure 3.6 shows a schematic
representation of such an elementary scatterer. As with all scattering problems
the procedure is to estimate the currents induced on the surface, and then use θ
them as sources in the wave equation to estimate the scattering matrix. One of the
simplest methods is to employ the Physical Optics (PO) current approximation
(Ishimaru, 1991; Jones, 1989). This states that the current induced on the facet
is the same as that induced on an infinite local tangent plane. The only change L
from the Fresnel equations is then to account for the finite extent L of the Fig. 3.6 Backscatter from a square surface
scatterer. This involves evaluation of an integral over the surface similar to that ‘facet’ of dimension L
employed in the Huygens source of Figure 1.5. This result again yields a Fourier
transform relationship between the current distribution and the scattered field.
In this example we have a simple uniform current of limited physical extent
L, which by Fourier transform yields a SINC function for the scattered field,
so that in this case the BSA scattering matrix is a simple modification of the
124 Depolarisation in surface and volume scattering

Fresnel equations for an infinite plane, as shown in equation (3.22):

 
  S⊥ 0
S θ λ =
0 S|| BSA
√  
2 πL sin (βL sin θ) SHH (θ εr ) 0
= cos θ
λ βL sin θ 0 SVV (θ εr )

 ε2

 − sin2 θ − cos θ

 ε

 SHH =
1

 ε2



 cos θ + − sin2 θ
ε1
⇒ (3.22)

 ε2 ε2

 cos θ − − sin 2
θ

 ε1 ε1

 SVV =

 ε ε2

 2
 cos θ + − sin2 θ
ε1 ε1
This equation shows that, for a finite sized facet, we will always receive some
backscattered signal, and that the polarisation of the backscattered wave is given
by the Fresnel equations at the local angle of incidence. Note that if the surface
facet is large compared to a wavelength (βL >> 1) then the SINC function
falls quickly to zero with increasing θ, and so the backscattered amplitude is
very small at all but normal incidence. This is called a specular surface return,
and has a straightforward polarisation behaviour described by a multiple of
the identity matrix, as shown in equation (3.21). Such approximations are at
the heart of high-frequency rough surface scattering theory based on stationary
phase evaluation of integrals (Jones, 1989; Ishimaru, 1991). In these cases it
is no surprise, therefore, that in the absence of multiple scattering it is found
that HH = VV and HV = 0. This is trivial polarisation behaviour (although
the actual evaluation of the integrals can be involved), and is not typical of the
microwave properties of natural surfaces (see Chapter 9).
However, one way in which specular scattering can lead to a strong
backscattered signal with interesting polarisation structure is through dihedral
B
retroreflection, as shown schematically in Figure 3.7. This arises when we have
θ two surfaces with normal vectors at right angles as shown. In this case, when
a wave is incident at angle θ then it is first reflected at surface A according to
A the Fresnel equations, and then is incident on the second surface B at angle
Fig. 3.7 Retro-reflection or dihedral scatter-
π/2 − θ. This wave is again reflected according to the Fresnel equations, and
ing geometry the total angle of reflection is π , in the backscatter direction. The reflection
matrix for this dihedral problem is composed of the product of the reflection
matrices at A and B, so that we can write the matrix in terms of compounded
Fresnel coefficients, as shown in equation (3.23):
 
R R 0
[R] = [RB ] . [RA ] = HHB HHA
0 RVVB RVVA

cos θ − εA − sin2 θ sin θ − εB − cos2 θ

RHHA = RHHB =
cos θ + εA − sin2 θ sin θ + εB − cos2 θ
εA cos θ − εA − sin2 θ εA sin θ − εA − cos2 θ
RVVA = RVVB =
εA cos θ + εA − sin2 θ εA sin θ + εA − cos2 θ
(3.23)
 3.1 Introduction to surface scattering 125

Dihedral backscatter
0.8

0.6

0.4
Reflection coefficients

0.2

–0.2

–0.4 HH for n = 5
VV for n = 5
HH for n = 2
–0.6 VV for n = 2

–0.8 Fig. 3.8 Normalized BSA scattering matrix

0 10 20 30 40 50 60 70 80 90 elements for dihedral scattering for εr = 25
Angle of incidence (degrees) and for εr = 4

Making the usual coordinate corrections, the scattering matrix in the sensor
(BSA) system now has the form shown in equation (3.24):
   
S 0 R R 0
[S]BSA = ⊥ = A HHB HHB (3.24)
0 S|| 0 −RVVB RVVA

The most important consequence of this is a 180-degree phase shift between HH

and VV for a range of incidence angles centred on 45 degrees when compared to
the single reflection matrices. Figure 3.8 shows an example for the HH and VV
components of [S] for dihedral retroreflection at varying angles of incidence
for εA = εB = 2.25 and εA = εB = 9. Notice how, as the dielectric constant
increases so the width of the π phase shift zone increases. In the limit of a perfect
conductor when εr → ∞, such as metal surfaces at microwave frequencies, the
scattering matrix for a dihedral retroreflector has a constant form for all angles
of incidence, as shown in equation (3.25), which we used in equation (1.157).
 
1 0
[S]BSA = A (3.25)
0 −1

We see now that the polarimetric phase behaviour becomes more complicated
when we start to consider scattering by dielectric media, and this will motivate
the development of the ‘alpha parameter’ description of scattering in Chapter
4. First, however, we turn to another important surface scattering model, where
polarisation behaviour can be fully quantified.

3.1.3 The Bragg surface scattering model

A second important example of surface scattering is when the surface facet is
large compared to a wavelength but the surface is not smooth over the length
scale L. In this case the solution for the scattered field can be obtained as a per-
turbation of that from the underlying smooth surface, and an analytical solution
obtained for the scattering matrix in terms of an infinite series. This is termed
the small perturbation method (SPM) (Tsang, 1985). Keeping only the first term
in this series gives a good approximation to the solution when the perturbation
126 Depolarisation in surface and volume scattering

ks = 0 ks < 0.3
smooth slightly rough

Fig. 3.9 Small perturbation or Bragg rough N

∑(z i − z)
2 zi
surface scattering and definition of the rms
i =1
roughness parameter for an N -point discreti- s=
sation of the surface N−1

is limited, so that the rms roughness of the surface s is small compared to the
wavelength β = 2π/λ; that is, βs is a small quantity. In practice, βs < 0.3 is a
suitable condition, and so this constitutes a good low-frequency approximation,
typically valid at L (1.2 GHz) and P (450 MHz) bands for remote sensing of nat-
ural surfaces. Figure 3.9 shows a schematic representation of this small-scale
roughness. This is also called the vector small perturbation or Bragg scattering
model (after William Lawrence Bragg (1890–1971), who first formulated the
boundary condition required for such scattering). From our point of view the
most important observation is that the perturbed boundary conditions lead to
different polarisation coefficients from those obtained for the Fresnel case. The
scattered field from an arbitrary rough surface characterized by a height func-
tion z(x,y), illuminated by a wavelength long enough for this small perturbation
assumption to apply, is given as shown in equation (3.26) (Ulaby, 1986):

s
Epq = i 2β cos θ Bpq Ẑ (βX + β sin θ, βY )

 (1 − εr ) cos φs

 BHH =  

 cos θs + εr − sin2 θs cos θ + εr − sin2 θ





 



 (1 − εr ) εr sin θ sin θs − εr − sin2 θs εr − sin2 θ cos φs

 BVV =  



 εr cos θs + εr − sin2 θs εr cos θ + εr − sin2 θ
⇒



 − (1 − εr ) εr − sin2 θ sin φs

 BHV =  

 θ + ε − 2
θ εr cos θ + εr − sin2 θ

 cos s r sin s





 (1 − εr ) εr − sin2 θs sin φs



 BVH =  
 ε cos θ + ε − sin2 θ cos θ + ε − sin2 θ
r s r s r
(3.26)

where

1
Ẑ(βx , βy ) = z(x, y)e−i(βx x+βy y) dxdy
2π
∞

βx = −β sin θs cos φs (3.27)

βy = −β sin θs sin φs
2π
β=
λ
 3.1 Introduction to surface scattering 127

For the special case of backscatter (θs = θ and φs = π ) this yields the following
important simplification:
s
Epq = i 2β cos θ Bpq Ẑ (2β sin θ)


 cos θ − εr − sin2 θ

 B = B⊥ = = RHH


HH
θ + ε − 2
θ

 cos r sin

 0 1
⇒ (εr − 1) sin2 θ − εr (1 + sin2 θ ) (3.28)

 BVV = B|| =  = RVV

 2

 ε cos θ + ε − sin 2
θ

 r r


BHV = BVH = 0

where Ẑ is the Fourier transform of the surface profile. For backscattered power
this becomes the Fourier transform of the surface correlation coefficient or the
normalized roughness spectrum. Hence, to first order the backscatter depends
only a particular frequency component of the surface roughness spectrum
(Bragg scattering). But what of the polarisation properties of such scattering?
To proceed we make the following observations.
• The HH component equals the HH Fresnel equation, but the VV com-
ponent is very different from that predicted for Fresnel reflection. In
particular there is no longer a Brewster angle effect.
• The BSA scattering matrix for Bragg surface backscatter then has the
following form:
   
S⊥ 0 BHH 0
[S]BSA = = −i2β cos θ Ẑ (3.29)
0 S|| 0 BVV

Figure 3.10 shows the variation of the polarisation coefficients BHH and BVV
with angle of incidence for backscatter from a surface with εr = 4. Compare the
results with those for Fresnel reflection in Figure 3.4. Note that in this case the
B coefficients are not reflection coefficients. The actual level of backscattered

Bragg backscatter coefficients for n=2

7
HH
VV
6

5
Bragg coefficients

0
0 10 20 30 40 50 60 70 80 90 Fig. 3.10 Bragg coefficients for backscatter
Angle of incidence (degrees) from a rough surface with for εr = 4
128 Depolarisation in surface and volume scattering

signal depends on the component of the Fourier transform of the surface (satis-
fying the Bragg condition). Note especially the absence of a Brewster angle in
the VV channel, and also note that the magnitude of VV is always greater than
or equal to HH—the opposite trend to that observed for Fresnel reflection.
As a final observation we note that the effect of surface roughness is contained
in the Fourier transform component Ẑ, which is common to all polarisation
channels. Therefore, if we take a ratio of scattered signals, such as HH/VV, then
the effects of roughness will cancel, and the ratio depends only on the angle
of incidence and dielectric constant of the surface. This provides an impor-
tant stimulus to using polarisation diversity for surface parameter retrieval in
remote sensing. For example, we note that as the angle of incidence increases
so the difference between HH and VV increases, and that this ratio is a func-
tion of dielectric constant and hence an indicator of soil moisture or salinity.
One way to represent these variations is by using the scattering alpha parame-
ter, discussed in Chapter 4. According to this representation we can represent
the scattering ratio in a rotation invariant manner by forming the following
function:

 
 BHH − BVV  π
αb = tan −1   0≤α≤ (3.30)
B + B  2
HH VV

Figure 3.11 shows how alpha changes with angle of incidence for three cases
of dielectric constant, εr = 2, εr = 9, and εr = 81. We note that the dynamic
range in alpha available to distinguish between wet and dry surfaces increases
with angle of incidence. By taking the difference between the curves we obtain
an estimate of this dynamic range, as shown in Figure 3.12. This can be used,
for example, to assess the polarimetric calibration requirements of a sensor in
order to be able to distinguish wet and dry surfaces. We note that for angles
of incidence less than 25 degrees (typical of spaceborne radar systems) the
available dynamic range is limited to less than 5 degrees of alpha variation.

Bragg backscatter alpha parameter

45

40 n = 1.41
n=3
n=9
35

30

25
Alpha

20

15

10

Fig. 3.11 Variation of Bragg alpha angle with 0

angle of incidence for three dielectric con- 0 10 20 30 40 50 60 70 80 90
stants Angle of incidence
 3.1 Introduction to surface scattering 129

Alpha dynamic range

25
Maximum alpha difference (degrees)

20

15

10

0 Fig. 3.12 Dynamic range of alpha between

0 10 20 30 40 50 60 70 80 90 low and high dielectrics as a function of angle
Angle of incidence of incidence

Fig. 3.13 Diffuse reflection at a surface due

A to surface roughness

3.1.4 Coherent surface scattering component

Large-scale surface roughness also has an effect on the dihedral scattering mech-
anism, introduced in Figure 3.7, by reducing the level of specular or coherent
reflection from a surface at the expense of a diffuse or noncoherent component
(Ulaby, 1982). Hence the Fresnel coefficients are modified by an attenuating
multiplicative factor F that depends on surface roughness and angle of inci-
dence (of the form shown in equation (3.31)). Importantly for our purposes,
such a factor is independent of polarisation, and therefore does not change the
scattering alpha parameter. Both surfaces involved in the dihedral return will in
general be effected by such factors, but it is sufficient for us to consider a single
such reflection as shown schematically in Figure 3.13. The modified form of
reflection coefficient in the presence of rough surface scattering is then shown
in equation (3.31):

 
cos θ − εA − sin2 θ
 0 
 
F = e−2βs cos θ → [R] = F  cos θ + εA − sin2 θ 
 εA cos θ − εA − sin2 θ 
0
εA cos θ + εA − sin2 θ
(3.31)

Here s is the surface rms roughness and β the free space wavenumber 2π/λ.
Figure 3.14 illustrates how the power attenuation factor varies for typical rough-
ness variations (2, 4 and 6 cm rms) at L-band frequency (1.3 GHz). There are
130 Depolarisation in surface and volume scattering

Rough surface specular attenuation L-band (1.3 GHz)

0
2cm ms
4cm ms
–5 6cm ms

Power attenuation factor dB

–10

–15

–20

–25

Fig. 3.14 Example attenuation of surface –30

specular reflection due to surface roughness 0 10 20 30 40 50 60 70 80 90
factor F Angle of incidence (degrees)

Dihedral alpha angle (degs) for 45 degrees angle of incidence

80
60
70

80
70
75

60
80
Dielectric constant A

65 60

50
70

40
75

30

20 75
65

70 75
60

10
70 70
55 65 65 65
Fig. 3.15 Dihedral alpha parameter for a 60 60 60
22.5-degree angle of incidence and dielectric 10 20 30 40 50 60 70 80
variations in the two surfaces Dielectric constant B

three important consequences of this result for polarimetric studies. The first is
that in the presence of roughness the dihedral return, even though it is based
on a strong specular scattering effect, can be attenuated to a level where the
direct surface return can be dominant. Thus in practice we must consider the
possibility that both phenomena may be present at the same time. This will lead
us to develop so-called decomposition theorems for the interpretation of radar
scattering (see Chapter 4).
The second key idea is that any ratio of elements of the scattering matrix
is unchanged by the presence of surface roughness. As in the case of Bragg
scattering, we can then calculate the scattering alpha parameter (from equation
(3.30)) as a function of angle of incidence and dielectric constants of the two
surfaces involved. Figures 3.15 and 3.16 show examples of the dynamic range
of alpha for a shallow angle θ = 22.5◦ and a steeper angle θ = 45◦ . We note
the following key points:
 3.1 Introduction to surface scattering 131

Dihedral alpha angle (degs) for 22.5 degrees angle or incidence

80

75
35
40 30

70
50
55
45

60
70

65
60
Dielectric constant A

50
30
40 35

70
50
55

40
45

60

65

30

20
30
40 35

70
10
50
55
45

60

65
70 Fig. 3.16 Dihedral alpha parameter for a
10 20 30 40 50 60 70 80 45-degree angle of incidence and dielectric
Dielectric constant B variations in the two surfaces

1. The shallow angle alpha shows little dependence on the first dielectric
component (from the horizontal surface) but is directly related to the
second (from the vertical), due to its effective higher angle of incidence.
2. The steeper angle alpha shows more dependence on both dielectric con-
stants (horizontal and vertical surface reflections) and hence, for any
given alpha value, we can only ever obtain an estimate of the product of
the two dielectric constants.
3. The dynamic range in alpha is much larger, even for the shallow angle,
than that obtained for direct Bragg backscattering (Figure 3.12).
4. The alpha angles are nearly always greater than π/4, as expected for
dihedral scattering, but can go lower than this for very dry materials and
angles of incidence where the Brewster angle effect causes a switch in
the scattered phase.
Finally, we note that one other advantage of the alpha parameter formulation
is its invariance to rotations in the plane of polarisation. Such rotations can
occur, for example, when we consider scattering from sloped surfaces. In such
cases, although the alpha parameter can remain unchanged, crosspolarisation
is introduced to the scattering vector. We now turn to consider slope effects on
the polarisation properties of surface scattered waves.

3.1.5 The effect of surface slope on the scattering matrix

In practice the surface facet normal may have an arbitrary orientation in space
due to surface slope or large-scale roughness. If we establish a right-handed
coordinate system centred on the facet then we can write the general facet
normal vector as shown in Figure 3.17. We can further define two principal
slope components from this normal (Lee, 2000; Schuler, 2002). Rotation about
the x axis we then call range slope, defined by an angle γ . Similarly, rotation
about the y axis we call azimuth slope ψ. These slopes are direct parameters
of interest for estimation in radar remote sensing, and so we prefer to use them
rather than n in the following equations. A wave is now incident in the zy plane
132 Depolarisation in surface and volume scattering

V z
H

β n = n1 x + n2 y + n3 z

Incident wave f
coordinates

Fig. 3.17 Arbitrary surface normal orienta-

tion and range and azimuth slopes x

at an angle φ̄ to the z axis. Hence if n = (0, 0, 1) then we recover the situation

already treated above. However, in general this will not always be the case, and
we must therefore account for the effects of arbitrary local surface normal on
the scattering amplitude matrix for the facet. The key extension we now require
is to define a local tangent vector t defined in terms of the incident wave vector
β and the normal n, as shown in equation (3.32):

n×β
t=  (3.32)
 
n × β 

With this local coordinate system in place we must now modify the scattering
matrix in two ways:
• The angle of incidence θ used for evaluation of the Bragg or Fresnel
coefficients is no longer simply φ̄ in Figure 3.17, but is now defined from
the inner product between the surface normal and incident wave vector,
as shown in equation (3.33):

tan γ sin φ̄ + cos φ̄

cos θ = n.β = −n2 sin φ̄ + n3 cos φ̄ = (3.33)
1 + tan2 γ + tan2 ψ

• The combined effect of range and azimuth slopes causes an effective

rotation of the surface in the plane of polarisation through an angle χ , as
shown in equation (3.34):

cos χ = h.t  −n1
⇒ tan χ =
sin χ = − h × t  n2 cos φ + n3 sin φ̄
tan ψ
= (3.34)
sin φ̄ − cos φ̄ tan γ

The combined effect of these two angles is to modify the scattering matrix of
the facet to that shown in equation (3.35), which we see leads to the generation
of crosspolarisation from the facet:
     
  cos χ sin χ S (θ) 0 cos χ − sin χ
S n = . ⊥ . (3.35)
− sin χ cos χ 0 S|| (θ) sin χ cos χ
 3.2 Surface depolarisation 133

Indeed, by measuring the level of crosspolarisation we can in principle estimate

χ by employing an SVD of the scattering matrix and obtain remote information
about the slope of the facet. Slope also modifies the apparent alpha parameter
for the surface backscatter. The range slope component will cause an increase
(for slopes away from the radar) or decrease (for slopes towards the radar) in
the apparent alpha.
Finally, we note the effects of slope on the specular dihedral response (see
Figure 3.7). The key consequence of slope in this context is that the backscatter
ray path now no longer includes the two specular reflection mechanisms for
surfaces A and B (since the angle between the two normals is no longer π/2).
Hence the strongest return is now scattered into a small bistatic angle (given by
the slope) and is not returned in the exact backscatter direction. For this reason
the presence of slopes can attenuate the backscatter dihedral return (especially
when combined with roughness, as mentioned in equation (3.31)) and leave
only the direct surface component. The slope tolerance of the specular dihedral
component depends on many factors, such as the height of the vertical scatterer
and the radar wavelength. Instead of considering this phenomenon on a case-
by-case basis we instead prefer, in polarimetry, to model ‘surfaces’ as some a
priori unknown mixture of direct and specular dihedral scattering mechanisms,
and attempt to retrieve the properties and relative amplitudes of these from the
data itself. These are treated in Chapter 4 under the general topic of decom-
position theorems. First, however, we consider the topic of depolarisation by
surfaces.

3.2 Surface depolarisation

We have seen that surface scattering provides a diversity of polarisation
responses, varying with angle of incidence, dielectric constant and surface
slope. In particular, we have seen that slope leads to crosspolarisation in sur-
face scattering; but this is a deterministic phenomenon and can be removed,
for example, by a singular value decomposition of the scattering matrix. How-
ever, so far we have ignored any reference to depolarisation. This can and does
arise in natural surface scattering, and is difficult to quantify in the context of
rigorous scattering models (Borgeaud, 1994; Hajnsek, 2003). In this section
we consider one simple low-frequency analytic model that has been developed
to include full wave depolarisation: the extended or X-Bragg model (Hajnsek,
2003; Allain, 2003).

3.2.1 The extended or X-Bragg model

Rough surface scattering provides an important example of reflection symmetric
depolarisation, where the mean surface normal provides an obvious axis of
symmetry. In the limit of a smooth surface βs < 0.3, where β is the wavenumber
and s is the rms roughness, the small perturbation model (SPM) or Bragg
scattering applies (see equation (3.29)). In this case the backscattered electric
field is related to the component of the Fourier transform of the surface profile
‘resonant’ with the incident wave. The scattering coefficients are determined by
the boundary conditions and vary with polarisation, as shown in equation (3.36).
A key consequence of this model is that the effects of surface roughness act as
134 Depolarisation in surface and volume scattering

a scalar multiplier in all polarisation channels, and so all ratios of polarisation

remain independent of roughness and depend only on dielectric constant and
angle of incidence through the complex BHH and BVV coefficients. Such a
surface therefore has zero depolarisation (although possibly crosspolarisation
if the normal is misaligned with the wave coordinates). This is seen in equation
(3.36), where [T ] has only one non-zero eigenvalue and the scattering entropy
H is zero. The surface dielectric properties can then be represented by the
scattering mechanism α, as shown in equation (3.30). However, smooth natural
surfaces are observed to depolarise incident waves, and so this model needs
some modification. To achieve this we first go to the other extreme of a very
rough surface when βs >> 1. In this case we propose that the surface acts as
an azimuthally symmetric depolariser with a coherency matrix of the general
form shown in equation (3.37). This has scattering entropy greater than zero,
and hence causes depolarisation of the incident wave.

s
Epq = i 2β cos θ Bpq Ẑ (2β sin θ)


 cos θ − εr − sin2 θ

 BHH = B⊥ =

 cos θ − + εr − sin2 θ



 0 1
⇒ (εr − 1) sin2 θ − εr (1 + sin2 θ )

 BVV = B|| = 

 2

 εr cos θ + εr − sin2 θ




BHV = BVH = 0
(3.36)
 
a b 0
βs1
−→ [TS ] = ms b∗ c 0
0 0 0
 
  λ1 0 0  ∗T
= u1 u2 u3 .  0 0 0 . u 1 u2 u3
0 0 0
⇒ Hs = 0

We now develop a model of a depolariser based on a smooth transition between

these two limits, reducing to equation (3.36) for smooth surfaces and to equation
(3.37) for rough surfaces.
 
1 0 0
βs>>1
−→ [Ts ] = t11 0 m 0  ⇒ Hs > 0 (3.37)
0 0 m

One way to do this is to assume that the major perturbation to equation (3.36)
arises from micro-variations in surface slope as roughness increases. There is a
single parameter in the scattering vector that is influenced strongly by changes
in slope: χ , as shown in equation (3.38) (see also equation (3.35)). In this case
we can therefore propose that depolarisation is primarily caused by integration
over a distribution of χ . Even if the exact details of this distribution (p(χ )) are
 3.2 Surface depolarisation 135

unknown, the consequences for the form of the resulting coherency matrix are
important, as shown in equation (3.38):

    
1 0 0 a b 0 1 0 0
[TS ] = ms 0 cos 2χ sin 2χ  b∗ c 0 . 0 cos 2χ − sin 2χ 
0 − sin 2χcos 2χ 0 0 0 0 sin 2χ cos 2χ

 
τ= cos 2χ.p (χ) d χ 

 a bτ 0
→ [TS ] = ms b∗ τ cδ 0  ⇒ Hs > 0


δ= cos 2χ.p (χ) d χ 
2 0 0 c(1 − δ)
(3.38)

In the general case we see we must introduce two depolarising parameters δ

and τ , both of which are related to the (as yet unknown) distribution p(χ ).
However, we see that this model has the correct boundary conditions: when
δ = 1, τ = 1 (smooth surfaces), so this model tends to the zero entropy
Bragg model (equation (3.36)), while when τ tends to zero and δ tends to 0.5
(rough surfaces), so the surface tends to an azimuthally symmetric depolariser
(equation (3.37)). Significantly, between these two extremes we note that the
contribution of surface roughness and dielectric constant properties to depolar-
isation can be separated (regardless of the distribution p(χ )) by using the two
secondary parameters R and M , shown in equations (3.39) and (3.40) (Cloude,
2002a):

+ , + ,
t22 − t33 |SHH − SVV |2 − 4 |SHV |2
R (δ) = =+ , + , = 2δ − 1 0≤R≤1
t22 + t33 |SHH − SVV |2 + 4 |SHV |2
(3.39)
+ , + ,
t22 + t33 |SHH − SVV |2 + 4 |SHV |2
M (θ, εr ) = = + , = tan2 αb (3.40)
t11 |SHH + SVV |2

In this way R can be considered a roughness measure of the surface based on

its level of depolarisation—R = 0 corresponding to a very rough surface, and
R = 1 to very smooth. Such a parameter is also insensitive to variations in
surface dielectric constant, and hence is robust to changes in surface moisture
content and salinity, for example. On the other hand, M is a material indicator,
increasing with increasing dielectric constant of the surface while remaining
insensitive to roughness variations. Note that M also depends on angle of
incidence θ. As θ increases from zero (normal incidence), so the variation of
M with dielectric constant increases. R, on the other hand, is independent of θ,
and therefore does not vary with either dielectric constant or changes in angle
of incidence.
The roughness factor R behaves like a coherence amplitude, lying between
0 and 1. In fact, R equals the coherence between the LL and RR circular polar-
isations, as shown in equation (3.41). This parameter was first proposed and
136 Depolarisation in surface and volume scattering

developed as a surface discriminator in Mattia (1997).

+ ∗
,
sLL sRR
γLLRR = + ,+ ,
∗
sLL sLL ∗
sRR sRR
+ ,
(p1 + ip2 ) (−p1 + ip2 )∗
= + ,+ ,
(p1 + ip2 ) (p1 + ip2 )∗ (−p1 + ip2 ) (−p1 + ip2 )∗
8 (3.41)
p1 = SHH − SVV
p2 = 2SHV
+ 2, + 2,
|p2 | − |p1 |
−−−−−−−−−−−→ γLLRR = + 2 , + 2 , = −R
reflction symmetry
|p1 | + |p2 |

Note that in the presence of a surface with mean azimuth slope χ , the X-Bragg
coherency matrix has the modified form shown in equation (3.42):

    
1 0 0 a bτ 0 1 0 0
[TS ] = ms 0 cos 2χ sin 2χ  b τ cδ
∗ 0  . 0 cos 2χ − sin 2χ 
0 − sin 2χ cos 2χ 0 0 c(1 − δ) 0 sin 2χ cos 2χ
(3.42)

from which it follows that the revised diagonal elements of [T ] have the form
shown in equation (3.43):

t11 = a
t22 = c(δ cos 4χ + sin2 2χ ) (3.43)
t33 = c(cos 2χ − δ cos 4χ )
2

which implies that while the material index M remains invariant to azimuth
slope, the roughness indicator R varies as cos 4χ ; therefore, as mean slope
increases so the apparent roughness increases. We note, however, that by
estimating the LLRR coherence in both amplitude and phase we can com-
pensate for such slope effects (χ from the mean phase of the coherence, see
Chapter 4) and obtain a basis invariant estimate of the roughness (R from
the coherence amplitude, corrected for χ ). In this way we see that circu-
lar polarisations and their coherence are of great utility in surface scattering
studies.
As this approach is based on the low-frequency Bragg model but provides
a link to high-frequency depolarisation effects it is termed the extended or X-
Bragg surface scattering model. We can formulate this model for a geometrical
representation of its depolarisation properties in the backscatter entropy/alpha
plane as follows. If we assume for simplicity a uniform distribution of slopes
for p(χ ) with width , then the X-Bragg model takes on the explicit form
shown in equation (3.44), where the coefficients a, b and c are given in terms
 3.2 Surface depolarisation 137

of the Bragg scattering coefficients:


- sin 2 
τ = cos 2χp (χ) d χ = 

2 
 
- 1 sin 4  

δ = cos2 2χ p (χ) d χ = 1+ 
2 4
 
sin 2
 a b 0 
 2 
 sin 2   
 ∗ 1 sin 4 
→ [TS ] = ms  b c 1 + 0 
 2 2 4 
  
 1 sin 4 
0 0 c 1−
2 4
(3.44)



 cos θ − εr − sin2 θ

BHH =


a = |BHH + BVV |2 
 cos θ +
εr − sin2 θ
∗
b = (BHH + BVV ) (BHH − BVV ) ⇒ 0 1

 (ε − 1) sin 2
θ − ε (1 + sin 2
θ )
c = |BHH − BVV |2 

r r

BVV = 


2
εr cos θ + εr − sin2 θ

We can now consider the loci of H/α points generated when  varies from 0
to π for a fixed angle of incidence and varying dielectric constant εr . Such loci
are shown in Figure 3.18, for θ varying from 20 to 50 degrees. We have chosen
a high dielectric constant (εr = 75), so these loci represent the upper bounds on
α for each angle of incidence. Again we note that for each angle of incidence
the alpha angle changes only little with roughness changes, as implied by the M
parameter in equation (3.40). We note that as the angle of incidence increases,

X-Bragg model
90

80

70

60

50
Alpha

40

30

20
u = 50º
10
u = 20º u = 40º
u = 30º
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fig. 3.18 Geometrical representation of the
Entropy X-Bragg model in the entropy/alpha plane
138 Depolarisation in surface and volume scattering

so the potential depolarisation increases, and at small angles close to normal

the loci lie closely packed around the origin. At normal incidence itself, the
locus reduces to a point at the origin for all dielectric constants and roughness
variations.
A second, related approach to modelling surface depolarisation was devel-
oped in Alain (2003). In this approach the covariance matrix for the surface
backscatter is modelled in two components: a single scattering ‘coherent’
element (similar to Bragg), and a multiple scattering component, itself calcu-
lated from two surface parameters—the rms roughness and surface correlation
length—using the integral equation model (IEM) first developed by Fung
(1992). In this model, surface scattering depends not only on rms height but also
on surface correlation length Lc . Two of the most popular correlation models for
2
−r
surface studies are the Gaussian correlation function (e L2c ) and the exponential
|r|
(e− Lc ), where r is the separation of two points on the surface. This composite
model is then summarized in equation (3.45):
 s s
  m m

chhhh 0 chhvv chhhh 0 chhvv
[C] = [C]S + [C]M =  0 0 0 + 0 m
chvhv 0 
s s
cvvhh 0 cvvvv m
cvvhh 0 m
cvvvv
(3.45)

As multiple scattering becomes stronger, so the second term dominates (which

is diagonally dominant), leading to an increase in entropy and depolarisation.
Detailed comparisons between this model and X-Bragg show similar features
in the entropy/alpha plane.
So far we have dealt only with surface backscatter (N = 3 depolarisation).
We now turn to consider the treatment of the more general N = 4 bistatic
surface scattering case.

3.2.2 Polarisation effects in bistatic surface scattering

The geometry of the general bistatic surface scattering problem is shown in
Figure 3.19. An incident wave is scattered by a rough surface, and we are
interested in the polarisation properties of radiation scattered into an arbitrary
direction. The geometry of this problem is more complicated than for backscat-
ter, there being three primary angles to consider: θi , the incident polar angle; θr ,
the reflected polar angle; and φ = φr −φr , the difference in azimuthal angles for

z Reflected beam
θr dωr
Incident beam

dωi θi

φr y

φi

Fig. 3.19 Geometry of bistatic surface scat-

tering x
 3.2 Surface depolarisation 139

incident and reflected waves. In particular we are often interested in so-called

out-of-plane behaviour, when the surface normal, incident and scattered wave
vectors do not all lie in a plane. A fully polarimetric characterization of this
problem therefore involves knowledge of sixteen functions of three parameters
(the elements of [M ] or [T ]). This set defines the polarimetric bidirectional
reflectance distribution function (BRDF) f (θi , θr , φ), and its full characteri-
zation, either by measurement or theory, is a complicated process (Mendez,
1987; Priest, 2000). For this reason, various simplified physics-based models
have been proposed in the literature. One of the most common of these is the
microfacet model that we now consider (Priest, 2000).
First however, there are several important physical properties of the BRDF
to be considered. One of these relates to the integration of the function over a
hemisphere, to obtain the directional hemispherical reflectivity or DHR which
provides a normalization condition, as shown in equation (3.46):

DHR(θi ) = f (θi , θr , φ) cos θr d 2 ω = f (θi , θr , φ) cos θr sin θr d θr d φ

(3.46)

Note that the cos θr factor appears because the BRDF is defined per-unit pro-
jected area. As a consequence we note that f (θi , θr , φ) is related to σs , the
scattering cross-section per-unit illuminated area as shown in equation (3.47):
σs
f = (3.47)
cos θi cos θr
As an extreme case, a Lambertian surface is defined so that the outgoing power
per unit solid angle, per-unit projected area is independent of angle. Hence
f = 1/π for a Lambertian surface; and it also acts as an ideal depolariser
with scattering entropy H = 1, so that it has a coherency matrix of the form
shown in equation (3.48). Natural surfaces, however, do not show this type of
depolarising behaviour, and generally have entropy less than 1 (see Chapter 9).
 
1 0 0 0
0 1 0 0
[Tlambertian ] = 
0 0 1 0
 (3.48)
0 0 0 1

A second important property of the BRDF relates to the reciprocity theorem.

By exchanging θr and θi in equation (3.46) and requiring invariance of the DHR
leads to the reciprocity or exchange symmetry as shown in equation (3.49):

f (θi , θr , φ) = f (θr , θi , φ) (3.49)

Having established these basic properties of the BRDF, we now turn to consider
the microfacet model, the physics of which is rooted in geometrical optics. In
its basic form it postulates scattering from a collection of randomly oriented
microfacets (see Figure 3.6) comprising the rough surface. The geometrical
distribution of these facets, together with a local reflection law—usually the
Fresnel equations and Snell’s law—then defines the structure of the model.
One key feature of the microfacet model for our purposes is that it reduces the
complexity of the angular dependencies to a familiar similarity transformation
of the single facet scattering matrix, as we now demonstrate.
140 Depolarisation in surface and volume scattering

The starting point is to convert the three primary scattering angles into four
secondary geometrical parameters. The first is an angle , which is the local
angle of incidence (and angle of reflection) onto the facet, and is important in
determining the physical properties of scattering by a facet. This angle is defined
from the inner product of local surface normal unit vector n and direction vector
of the incident wave r i , as shown in equation (3.50):

cos  = n.r i ⇒ cos 2β = cos θi cos θr + sin θr sin θr cos (φi − φr ) (3.50)

From the local reflection law we can also relate  to the reflected wave vector
r r , as shown in equation (3.51):

r i + r r = 2 cos n (3.51)

This leads us to the second important angle, θ, being the polar angle of the local
normal, as shown in equation (3.52):
cos θi + cos θr
cos θ = (3.52)
2 cos 
Each microfacet is then characterized by its normal n, which is further assumed
to be symmetrically distributed about the z axis according to some given prob-
ability function. A commonly used example is the Gaussian surface height
distribution, which in normalized form can be written in terms of the local
surface slope (tan θ ) and the slope variance s2 , as shown in equation (3.53):
 
1 − tan2 θ
p(θ )d ω = 1 ⇒p(θ) =
2
exp (3.53)
2π s2 cos3 θ 2s2
When used in the scalar microfacet model this leads to the following form for
the function f :
 
1 − tan2 θ
f = exp R() (3.54)
8π s2 cos4 θ cos θi cos θr 2s2
where R() is a Fresnel reflectivity function for the selected scalar channel.
Note that this form of the BRDF satisfies both the normalization constraint in
equation (3.46) and the reciprocity relation in equation (3.49). Note, impor-
tantly, that this is a deterministic model. The parameters of the right-hand side
are all fixed for a given geometry and surface roughness. The geometry and
roughness effects then act as a multiplier to the scattering coefficient—itself a
deterministic quantity. Importantly this enables us to extend this model to a fully
polarimetric version by replacing R(β) by a vector formulation, as follows.
The main complication in formulating a vector BRDF is in matching the local
coordinates. There are four such systems to consider. The first is the system
defined by the incident wave direction r i and the mean surface normal in the z
direction. The second is then defined by the local surface normal n and incident
wave direction. This second is rotated about the incident direction by an angle
ηI , as shown in equation (3.55). The third coordinate system is defined by the
local surface normal n and the reflected wave direction. This has finally to be
related to the fourth: namely, that formed by the reflected wave direction and
the mean normal z. This is related to the third by a rotation about the reflected
direction by an angle ηr , as shown in equation (3.55).
 3.2 Surface depolarisation 141

The scattering matrix can then be written as a cascade of three processes:

rotation into the local system by ηI , followed by Fresnel scattering, followed
by negative rotation out of the local into the global system by ηr . Mathemati-
cally this cascade of matrices is shown in equation (3.55), which is a form we
have encountered before, being a singular value transformation of the diagonal
Fresnel scattering matrix by different left and right singular vectors.

 
 cos  − εr − sin2 
 cos θ − cos  cos θi 
 R (, ε ) =

 cos ηi = 
 hh r
sin θi sin β cos  + εr − sin2 

 cos θ − cos  cos θr 

cos ηr = 
 εr cos  − εr − sin2 
sin θr sin  R
 vv (, εr ) =
εr cos  − εr − sin2 
     
√ cos ηr sin ηr R () 0 cos ηi − sin ηi
⇒ [S] = ap . ss .
− sin ηr cos ηr 0 Rpp () sin ηi cos ηi
 
1 − tan2 θ
ap = exp
8π s2 cos4 θ cos θi cos θr 2s2
(3.55)

Having determined the form of the [S] matrix, we can now vectorize to obtain
the corresponding scattering vector k, which can be expressed compactly as a
unitary transformation, as shown in equation (3.56):
  
cos ηi − ηr 0  0  sin ηi − ηr
√   0 cos ηi + ηr  sin ηi + ηr  0 

kM = ap  
0  − sin ηi + ηr cos ηi + ηr 0 
− sin ηi − ηr 0 0 cos η i − ηr
 
Rss () + Rpp ()
 Rss () − Rpp () 
×


 (3.56)
0
0

The most important consequence of this model is that all the surface rough-
ness effects are contained in the scalar multiplier ap . Hence, like the Bragg
model for backscatter, there is zero depolarisation under this model—one of
the most commonly used for bistatic surface scattering. Consequently the scat-
tering entropy is zero and the scattering coherency matrix is rank 1, as shown
in equation (3.57):

[T ] = k M .k ∗T
M (3.57)

There is again a clear discrepancy between the model predictions and observed
depolarisation behaviour of real surfaces (see Chapter 9). In particular, as we
found with the backscatter Bragg model, there is no way to parameterize smooth
change from a zero entropy polariser to Lambertian depolariser. To overcome
this a heuristic compromise is often sought by adding a Lambertian component
to the microfacet model with parameter ad , often determined by fitting the
model to experimental data. In this case, depolarisation under bistatic surface
142 Depolarisation in surface and volume scattering

scattering can be modelled as shown in equation (3.58):

 
1 0 0 0
0 1 0 0
[T ] = k M .k ∗T  
M + ad 0 0 1 0 (3.58)
0 0 0 1

We investigate the validity of this approach in Chapter 9, when we consider the

application of these models to real surface scattering data. Now, however, we
turn to consider the second important class of depolarisation problems: namely,
those caused by volume scattering.

3.3 Introduction to volume scattering

In addition to wave reflection from boundaries between materials (surface scat-
tering), there is an important second class of polarimetric scattering phenomena
to be considered: namely, volume scattering. This occurs within inhomoge-
neous bulk material that contains local variations in dielectric properties. These
act as ‘sites’ for wave scattering that are distributed in space and act both to
increase the scattering cross section and influence propagation of the wave
through the medium via the process of wave extinction.
As well as the presence of material inhomogeneities, volume scattering also
requires significant penetration of the wave into the medium to excite these
secondary scattering centres. This latter observation leads us to a simple test to
assess the relative importance of volume versus surface scattering, based on the
effective complex permittivity of a medium εr . As shown in equation (3.59),
and as a generalization of the concept of skin depth from equation (3.13), we
can adopt an exponential structure function assumption in the material, and use
this to define a significant power penetration depth δp as shown:

εr = ε − iε ωn ωn
z − ωnc z
⇒ eiω(t− c ) = eiωt e−i = eiωt e−i
nz
√ c
z c e
n = εr = n − in
(3.59)
P(z = δp ) 1 λo λ
= ⇒ δp = =  o√ 
P(z = 0) e 2π n 2π Im( εr )


We see that the penetration can be expressed as a scale factor times the free
space wavelength, and for volume scattering to be possible we require this
factor to be much greater than unity. We see that the factor depends only on the
imaginary part of the square root of the effective dielectric constant, but often
a further simplification is made, since for most materials in microwave remote
sensing the imaginary part is much smaller than the real. We therefore have the
following common approximation:
√
ε λo ε
< 0.1 → δp ≈ (3.60)
ε 2π ε
To illustrate this relation, we now consider five important application areas in
microwave remote sensing: namely, scattering from water surfaces, ice, snow,
soil and vegetation (Ulaby, 1982, 1986; Dobson, 1985).
 3.3 Introduction to volume scattering 143

Water
Pure water has a high dielectric constant (in both real and imaginary parts;
see equation (3.15)) across the microwave spectrum, and this severely limits
the penetration depth of microwaves. Radar scattering by water can therefore
be considered a surface-dominated problem, albeit often a dynamic one with
time varying and spatially inhomogeneous roughness due to interacting wind
and current systems. Note, however, that breaking waves pose an important
example where volume scattering can occur in the water context, as these
generate a turbulent air/water mixture with a relatively low effective dielectric
constant and where wave penetration can be significant and volume scattering
effects be observed.

Land and sea ice

Pure ice has a real dielectric constant around εr = 3.15, and a low loss-tangent,
so penetration depths can extend 300 wavelengths or more, depending on fre-
quency. In this case, volume scattering can be significant in thick ice deposits
such as occur in land ice. Sea ice generally has a much larger imaginary part
due to the presence of brine, with first-year ice (εr ≈ 3.3 − i0.25) having a
higher loss than brine-deficient multi-year ice (εr ≈ 3.3 − i0.03). Penetration
depths are therefore higher into multi-year ice, and vary from 10 down to 1
wavelength.

Snow
Dry snow is a mixture of ice and air—both very low-loss materials—and so
penetration can occur to hundreds of wavelengths, with a small overall dielec-
tric constant lying in the range 1.5–2. The imaginary part of the dielectric
constant is, however, very sensitive to changes in water content, and so wet
snow (which has dielectric constant always less than 4) shows a greater variety
of penetration depths. In general, therefore, snow can behave as a surface or
volume scatterer (or both), depending on environmental conditions, viewing
geometry, and operating frequency.

Soils
As discussed in equation (3.14), soil is a composite material, the dielectric
constant of which depends on a large number of parameters. It is therefore
difficult to derive general conclusions about soil penetration. However, across
a large part of the microwave spectrum, penetration depth depends primarily
on soil moisture content, and for most mid-latitude soil types penetration is
therefore seldom more than a wavelength or so. An important exception must
be made for hyper arid sand/soil environments, where penetrations of many
wavelengths can occur. Soil is therefore similar to snow in that both surface
and volume scattering can occur, although in most mid-latitude agricultural
areas the assumption is often made of pure surface scattering.

Vegetation
Vegetation scattering is perhaps the most complex of all, and simple models for
penetration depth are often of limited use. Indeed, vegetation scattering pro-
vides an important focus for the development of polarimetric interferometry in
Chapter 6, which allows us, amongst other things, to estimate the penetration
144 Depolarisation in surface and volume scattering

depth into vegetation. It is sufficient here to point out that vegetation is often a
strong volume scattering environment, combining multiple wavelengths of pen-
etration with the presence of strong dielectric discontinuities (moist branches,
twigs, leaves, and so on), which are often larger than a wavelength, and act to
scatter the waves in a complex manner.
Surface and volume scattering effects are often discriminated on the basis
of their dependence on angle of incidence. Surface scattering generally shows
a fall-off with increasing angle, while volume effects vary only slightly with
change of angle. However, in this text we are concerned primarily with the
polarisation properties of the two phenomena. In preparation for a discussion
of the polarisation properties of volume effects, we now consider scattering by
an individual particle. In the next section we use these results with averaging
to analyse the depolarisation behaviour of random volume scattering.

3.3.1 Small particle scattering

In the previous sections we dealt with reflection and scattering at a rough inter-
face between two media. While this provides a good model for surfaces, it
does not deal with the case of volume scattering. This occurs when we have a
three-dimensional distribution of scattering centres, and arises in practice for a
description of vegetation cover and in soil and ice penetration. One approach
to modelling such effects is to consider each localized scattering centre as
a particle of known shape and material composition embedded in an homoge-
neous background. The problem then reduces to establishing the scattered fields
by the particle, and combining a large number of such elements to provide a
macroscopic description of the scattering medium.
In general, even the first stage of this process can be complicated, as
Maxwell’s equations must be solved across the boundary of the scatterer. For
spherical particles this can be achieved using an expansion of incident and scat-
tered fields in spherical wave functions, and matching of the fields in a similar
manner to that used in the Fresnel equations. This leads to a convergent solution
known as the Mie series, which can be used for calculating the full scattering
matrix of spheres of arbitrary size and material composition and for arbitrary
scattering angle  (van de Hulst, 1981; Hovenier, 2004). However, spherical
particles have a very special symmetry and hence limited polarisation response
(although they can still depolarise via multiple scattering, as demonstrated in
Chapter 9). Furthermore, they are not found in many natural media such as veg-
etation, and consequently such a symmetric solution is of limited applicability
in radar remote sensing.
A more useful approximation is to assume spheroidal particles, with one long
axis and two equal minor axes, as shown in Figure 3.20 (Jin, 1994a). Here we

z
x1 Particle
frame

x2 = x3

Fig. 3.20 Spheroidal particle geometry and

particle coordinate frame
 3.3 Introduction to volume scattering 145

can now model variations in shape from prolate (needle-like) through spherical
and into oblate (disc-like). This provides a more realistic variation of shape
for modelling in many natural media. Again the solution can be established
from Maxwell’s equations by expanding fields in terms of vector spheroidal
wave functions and matching coefficients across the boundary. However, this
solution is not nearly as convenient as the Mie series, for the following reasons
(Mishchenko, 2000, 2006, 2007):
• The computation of vector spheroidal functions is a complicated math-
ematical and numerical problem in itself, especially for absorbing
particles.
• The spheroidal functions are not orthogonal on the surface of the scatterer,
and the unknown expansion coefficients must therefore be solved using
an infinite set of algebraic equations, which in practice must be truncated
and solved numerically. Unfortunately, for large particles this system
becomes ill-conditioned, and care is needed in the numerical computation
of results.
As a result of these difficulties, several alternative numerical methods have
been developed based, for example, on finite element, finite difference, and T-
matrix approximations. These provide more robust general-purpose numerical
solutions, but are too complicated in structure for simple analytical studies.
To avoid these problems we illustrate the important polarisation physics by
considering the case of Rayleigh scattering by spheroids. Lord Rayleigh (1842–
1919) first derived an approximation for scattering in the small particle limit
when βa << 1 by assuming that the fields in and near the particle are the
same as those derived from electrostatics, and that the internal field inside the
particle is homogeneous. This approximation leads to a fully analytic solution,
even for chiral particles, which as we have seen in Chapter 1, case III, is a
case of special importance in polarimetric studies. We now consider the main
polarimetric features of this solution method.
With reference to Figure 3.20, we now assume that the particle dimensions
are small compared to the wavelength of the incident radiation so that the field
inside may be considered constant, and we can ignore any time delays for fields
to propagate across the particle. In quantitative terms this requires the following
condition:

|n| βa  1 (3.61)

where n is the complex refractive index of the particle, β is the wavenumber

2π/λ, and a is the volume equivalent sphere radius of the particle. The first
important stage in the solution is to recognize that the particle has a definite
orientation, specified by its major axis, shown as the dotted line in Figure 3.20.
This can be used to define a particle frame or coordinate system that has its z
axis aligned with the major axis.
The incident field induces a dipole moment in the particle, which then acts to
reradiate or scatter energy. We wish to allow the possibility of chiral particles
(such as a small helix) when the incident electric field can generate circulating
as well as linear currents. This means that we must allow both electric (p)
and magnetic (m) dipole moments in the particle (see equations (1.10) and
(1.11)). These moments will evidently depend on the orientation of the driving
146 Depolarisation in surface and volume scattering

field, and thus are to be represented by polarizability tensors. The moment

vectors are then related to the driving field vectors by the following matrix
equation:
     
p α ee α em E
= . (3.62)
m −α em α mm H

where α ij are 3 × 3 polarizability tensors, which are all diagonal in the par-
ticle frame. The field radiated by elementary electric (p) and magnetic (m)
dipole moments can be obtained directly from Maxwell’s equations, as shown
in equations (1.10) and (1.11), and are given explicitly in equation (3.63):
0 1 exp (−iβ r)
o
E = ω2 µo (I − rr).p − ωβo r × m (3.63)
4π r

Here r is a unit vector in the direction of the scattered wave. Before we can use
this relationship with equation (3.62) we first must transform the coordinate
system from the particle into the laboratory frame. The latter can be defined as
shown in Figure 3.21. Without loss of generality, we consider the wave incident
along the +z direction and scattering in the y-z plane in the direction r, which
makes an angle  with the z axis. We then wish to find the 2 × 2 scattering
amplitude matrix [S] for radiation polarised parallel (P) or perpendicular (S) to
the scattering plane, as shown in Figure 3.21. Using the r vector as defined in
Figure 3.21 together with equation (3.63), we can express the radiated field by
the particle as shown in equation (3.64):
  2
1 1 β exp(−iβo r)
ES = p1 − √ (m3 sin ψ − m2 cos ψ) o
εo εo µo 4π r
  2 (3.64)
1 1 β exp(−iβo r)
EP = (p2 cos ψ − p3 sin ψ) − √ m1 o
εo εo µo 4π r

r
Ψ
ES r inc = [0 0 1]
z r = [0 sin Ψ cosψ ]

EP
y

 E = [1 0 0] E S = [1 0 0]
rinc ⇒ S
 EP = [0 1 0] E P = [0 cosΨ -sinψ ]

EP

Fig. 3.21 Laboratory coordinate frame for

particle scattering ES
 3.3 Introduction to volume scattering 147

To solve this we need the components of plab = (p1 , p2 , p3 ) and mlab =

(m1 , m2 , m3 ). To convert the polarizability tensors from the particle to the lab-
oratory frame we make use of Euler angles relating one coordinate system to
the other (Goldstein, 1980). As we are dealing with spheroids, the transforma-
tion from one to the other requires only two angles, and we choose a particle
orientation angle θ and a tilt angle τ as shown in equation (3.65):

   
cos τ 0 − sin τ cos θ sin θ 0
[R] =  0 1 0  . − sin θ cos θ 0
sin τ 0 cos τ 0 0 1
 
cos τ cos θ cos τ sin θ − sin τ
=  − sin θ cos θ 0  (3.65)
sin τ cos θ sin τ sin θ cos τ

Using this rotation matrix we can then express equation (3.62) in the laboratory
frame, as shown in equation (3.66):

     inc 
plab R−1 α ee R R−1 α em R E
= . (3.66)
mlab −R−1 α em R R−1 α mm R H inc

where now we can make use of the incident electric and magnetic fields for a
TEM wave in the laboratory frame directly as

 T εo  
E inc = ES EP 0 ⇒ H inc = −EP ES 0 (3.67)
µo

Evaluation of the elements of one block of equation (3.66) will illustrate the
general form of the solution, as shown in equation (3.68):

     
cos τ cos θ − sin θ sin τ cos θ ρ1 0 0 cos τ cos θ cos τ sin θ − sin τ
     
R−1 αR = 
 cos τ sin θ cos θ sin τ sin θ  
. 0 ρ2 0 
 .  − sin θ cos θ 0 

− sin τ 0 cos τ 0 0 ρ2 sin τ cos θ sin τ sin θ cos τ
 
ρ2 + (ρ1 − ρ2 ) cos2 τ cos2 θ (ρ1 − ρ2 ) cos2 τ sin θ cos θ − (ρ1 − ρ2 ) cos τ sin τ cos θ
 
=
 (ρ1 − ρ2 ) cos τ sin θ cos θ ρ2 + (ρ1 − ρ2 ) cos τ sin θ − (ρ1 − ρ2 ) cos τ sin τ sin θ 
2 2 2 
− (ρ1 − ρ2 ) cos τ sin τ cos θ − (ρ1 − ρ2 ) cos τ sin τ sin θ ρ1 sin τ + ρ2 cos τ
2 2

(3.68)

We now have all the elements we require for a solution. We start by setting the
incident field as S or P polarised in equation (3.67), and then use equation (3.66)
with (3.68) for each block element to obtain the p and m vectors. Finally we
use equation (3.64) to derive the S and P components of the scattered field. In
this way we can calculate the scattering matrix as a function of the three angles
, θ, τ and the four pairs of polarizability constants for the particle.
Rather than calculate the full set of equations, we consider instead two
important special cases that permit further simplification of the above equations.
148 Depolarisation in surface and volume scattering

3.3.1.1 Example 1: scattering by small non-chiral particles

In this case we can set α em = α me = α mm = 0, and the equations simplify to
the following form:
     
p1 ρ2 + (ρ1 − ρ2 ) cos2 τ cos2 θ (ρ1 − ρ2 ) cos2 τ sin θ cos θ − (ρ1 − ρ2 ) cos τ sin τ cos θ E inc
     S 
p2  =  (ρ1 − ρ2 ) cos2 τ sin θ cos θ ρ2 + (ρ1 − ρ2 ) cos2 τ sin2 θ − (ρ1 − ρ2 ) cos τ sin τ sin θ  . E inc 
     P 
p3 − (ρ1 − ρ2 ) cos τ sin τ cos θ − (ρ1 − ρ2 ) cos τ sin τ sin θ ρ1 sin2 τ + ρ2 cos2 τ 0
  2
1 βo exp(−iβo r)
⇒ ES = p1
εo 4π r
  2
1 β exp(−iβo r)
EP =
εo
(p2 cos ψ − p3 sin ψ) o
4π r
(3.69)

From these equations the elements of the scattering matrix can then be directly
obtained as shown in equation (3.70):
 
SPP SSP
[S] =
SPS SSS


 SPP = (AP − 1) sin θ cos τ (sin θ cos τ cos  − sin τ sin ) + cos 

SSP = (AP − 1) cos θ cos τ (sin θ cos τ cos  − sin τ sin )
⇒

 SPS = (AP − 1) sin θ cos θ cos2 τ

SSS = (AP − 1) cos2 θ cos2 τ +1
(3.70)
where we have defined the particle anisotropy Ap as a ratio of principal
polarizabilities, so that
ρ1
Ap = (3.71)
ρ2
For simplicity we have cancelled all factors common to the elements of [S]. Of
particular interest is the case of backscatter ( = 180◦ ), when the equations
further simplify to those shown in equation (3.72):

  
 SPP = − (AP − 1) sin2 θ cos2 τ − 1

S SSP SSP = − (AP − 1) sin θ cos θ cos2 τ
[S] = PP ⇒ (3.72)
SPS SSS 
 S = (AP − 1) sin θ cos θ cos2 τ
 PS
SSS = (AP − 1) cos2 θ cos2 τ +1
This scattering matrix is expressed in the wave coordinates, and if we convert
to the sensor (BSA) system, commonly used in radar studies, we obtain the
following solution:

  
 SPP = (AP − 1) sin2 θ cos2 τ + 1

S SSP SSP = (AP − 1) sin θ cos θ cos2 τ
[S]BSA = PP ⇒ (3.73)
SPS SSS 
 S = (AP − 1) sin θ cos θ cos2 τ
 PS
SSS = (AP − 1) cos2 θ cos2 τ +1
which we see is complex symmetric, as expected. Note that this matrix has a
straightforward eigenvalue expansion, as shown in equation (3.74):
     
cos θ sin θ 1 0 cos θ − sin θ
[S]BSA = . .
− sin θ cos θ 0 sin2 τ + AP cos2 τ sin θ cos θ
(3.74)
 3.3 Introduction to volume scattering 149

which demonstrates that from measurement of the backscatter amplitude matrix

of a single particle we can determine the angle θ from the singular vectors of
[S], and from the ratio of singular values we obtain κ = AP cos2 τ + sin2 τ .
This ratio is 1 when τ = 90◦ ; that is, when viewed along its major axis, in
which case it has a circular symmetric cross-section. In this case the particle
is indistinguishable from a sphere. However, when τ = 0—when the major
axis lies in the plane of polarisation—we obtain a ratio equal to AP , the particle
anisotropy.
Interpretaton of AP in terms of the particle geometry requires evaluation of
the ratio of principal values of polarizability. This can be evaluated analytically
for small spheroids of volume V (Ishimaru, 1991; Tsang, 1985), and yields the
following result:

εo (εr − 1) V L2 + 1
εr −1
ρi = ⇒ AP = (3.75)
1 + (εr − 1) Li L1 + 1
εr −1

where L1 and L2 are shape functions given explicitly by equation (3.76):

  

 1 − e2 1 1+e x22

 e2 −1 + ln x1 > x2 = x3 e2 = 1 − (prolate)
2e 1 − e x12
L1 =  

 1+f2 1 x22

 2 1 − arctan f x1 < x2 = x3 f2 = − 1 (oblate)
f f x12
1
L2 = L3 = (1 − L1 ) (3.76)
2

Figure 3.22 shows a plot of the relationship between particle shape r = x1 /x2
and the polarizability ratio for low and high values of dielectric constant. Hence
we see that Ap is a general indicator of particle shape, although it is bounded in
value by the dielectric constant to lie between limits given by equation (3.77)

Particle shape vs polarisability ratio

20

15 er = 5
er = 50
10Log(Ap), Ap = Polarisability ratio

Ap = r
10

–5

–10

–15

–20 Fig. 3.22 Relationship between polarizabil-

–20 –15 –10 –5 0 5 10 15 20 ity ratio and shape for varying dielectric
10Log(r), r = x1/x2 constant
150 Depolarisation in surface and volume scattering

(Ablitt, 2000):

1 εr + 1
< AP < (3.77)
εr 2

Note also that to a good approximation LL12 ≈ xx21 (the error is less than 0.04
in all cases), and hence for large εr equation (3.75) yields AP ≈ xx12 , which
then becomes a direct indicator of particle shape. This again demonstrates the
importance of SVD in the analysis of the scattering matrix. In this case the
singular vectors lead to estimation of the particle orientation angle and the ratio
of singular values to the projected shape and dielectric constant of the particle.

3.3.1.2 Example 2: scattering by small chiral particles

As a second example of the small particle scattering matrix, we now consider
spheroids made from chiral material. In this case the magnetic dipole moment
contribution must be included in the radiated field. However, we can still make
some useful simplifying assumptions.
We introduced chiral materials in equation (1.59), and then defined the scalar
chiral admittance in equation (1.60). We noted that chirality is very weak for
many natural materials. Under this assumption we can assume that the param-
eter β0  is small, as suggested in equation (1.69). This then has the following
implications for the polarizability tensors in equation (3.52):

• The magnetic tensor can be set to zero α mm = 0, as the magnetic field

effects will be small compared to those induced by the electric field.
• We can ignore any chirality effects in the electric field tensor α ee , and
hence this tensor has the same form as in example 1 for non-chiral
particles.
• We must consider the effect of the mixed tensor α em , which in the simplest
case will be linearly dependent on the small parameter β0 .
With these comments in mind we postulate the following form of equation
(3.62):
  

 ρ1 0 0

α = 0 ρ2 0


  −1   ee
  
 ρ2
plab R α ee R R−1 α em R E inc 0 0
= . ⇒  
mlab −R−1 α em R 0 H inc 

 δ1 0 0


α em =  0
 δ2 0

0 0 δ2
(3.78)

From the assumption of a linear dependence on β0  we also have

√
ρ1 δ1 δ1 εr
AP = = iκ = =i β0  (3.79)
ρ2 δ2 α1 εr − 1

where κ is a dimensionless chirality parameter. Note that the α em tensor is

purely imaginary, as pointed out in equation (1.60). This allows us to rewrite
 3.3 Introduction to volume scattering 151

equation (3.78) in the simplified form shown in equation (3.80):

  

 AP 0 0

α =  0

 0 1
 lab   −1   ee
−1  
 0 0 1
p R α ee R R α em R E inc
= . ⇒  
−R−1 α me R 
inc
mlab 0 H  AP 0 0



α =  0 1 0

 em
iκ
0 0 1
(3.80)

We can now evaluate this explicitly. The result is shown in equation (3.81):
 
SPP SSP
[S] =
SPS SSS


SPP = (AP − 1) sin θ cos τ (sin θ cos τ cos  + sin τ sin ) + cos 





 +iκ (AP − 1) cos θ cos τ (sin θ cos τ − cos τ sin θ cos  − sin τ sin )





 SSP = (AP − 1) cos θ cos τ (sin θ cos τ cos  + sin τ sin )

 


 +iκ[(AP − 1) cos τ cos2 θ cos τ + cos τ sin2 θ cos  + sin τ sin θ sin  +(1+ cos )]
⇒

 SPS = (AP − 1) sin θ cos θ cos2 τ

 



 −iκ[(AP − 1) cos τ cos2 θ cos τ + cos τ sin2 θ cos  − sin τ sin θ sin  +(1+ cos )]






SSS = (AP − 1) cos2 θ cos2 τ +1



 +iκ (AP − 1) cos θ cos τ (sin θ cos τ − cos τ sin θ cos  − sin τ sin )
(3.81)

Again, as we are primarily interested in backscatter, we set  = 180◦ . At the

same time we apply the sensor coordinate corrections to obtain the backscatter
matrix in the BSA convention, as shown in equation (3.82) (Cloude, 2002b):


 SPP = (AP − 1) sin2 θ cos2 τ + 1



 −iκ (AP − 1) sin 2θ cos2 τ



 SSP = (AP − 1) sin θ cos θ cos2 τ
  

 −iκ (AP − 1) cos2 τ cos 2θ
S SSP
[S]BSA = PP ⇒ (3.82)
SPS SSS 
 S = (A − 1) sin θ cos θ cos2τ


PS P

 −iκ (AP − 1) cos2 τ cos 2θ





 S = (AP − 1) cos2 θ cos2 τ + 1
 SS
+iκ (AP − 1) sin 2θ cos2 τ

This can be written as a decomposition of the scattering matrix in the following

form:

[S] = [S]A + [S]B = R (−θ ) (S1 + S2 ) R (θ)

   
1 0 0 i
= R (−θ ) . − κ (AP − 1) cos τ
2
.R (θ)
0 AP cos τ + sin τ
2 2 i 0
(3.83)

This result shows that the presence of a small amount of chirality in the particle
leads to a decomposition of the backscattering matrix into the sum of two
152 Depolarisation in surface and volume scattering

terms: SA , the matrix for non-chiral particles; and SB , a chiral perturbation

which is manifest as a scalar multiple of one of the Pauli spin matrices.

3.3.2 Scattering by large particles

As particle size increases compared to the wavelength, then the scatter-
ing behaviour changes, with forward scattering dominating over backscatter.
In this section we consider some of the implications for the polarisation
properties of large particle scattering. Detailed analytical calculations are
generally difficult for large particles, and resort must usually be made
to numerical and approximation techniques such as the T-matrix and dis-
crete dipole approximation (DDA) (Mishchenko, 2000, 2006, 2007). Public
domain codes are available for both these advanced techniques; for DDA
see http://www.science.uva.nl/research/scs/Software/adda, and for the T-matrix
see http://www.giss.nasa.gov/∼crmim/t_matrix.html. These codes can cope not
only with single particles, but also with small aggregates such as multi-sphere
clusters, accounting for all multiple interactions between particles.
One problem that does allow some analytic insight is that for scattering
by spheres of arbitrary size and material. Small spheres of radius r << λ
act as Rayleigh scatterers, characterized by a polarizability ρ, as considered
in the previous section. They have a bistatic amplitude [S] matrix with the
characteristic dipole radiation pattern, a function of the scattering angle 
(where  = 0◦ is forward scatter, and  = 180◦ is backscatter), as shown in
equation (3.84):
 
cos  0
[S] = iβ ρ
3
(3.84)
0 1
Here VV (where V is perpendicular to the scattering plane) has a uniform
scattering pattern, showing equal back and forward scatter. HH, on the other
hand, shows a null for  = 90◦ —for lateral scattering—but maintains equality
of forward and backscattering amplitudes. As particle radius increases, two
things happen:
1. Because of the spherical symmetry, the crosspolarisation terms of [S]
remain zero; that is, the matrix is always diagonal in the H and V basis
(parallel and perpendicular to the scattering plane). Therefore, scattering
by a sphere has SU(2) transformation properties (any complex combi-
nation of HH and VV can be formed by two-element unitary weight
vectors w). Therefore, scattering by a sphere of arbitrary size and mate-
rial, scattering into an arbitrary direction, can be mapped as a point on
the scattering sphere of Figure 2.8.
2. The HH and VV scattering amplitudes become functions of , and are
complex (they have a phase difference and amplitude ratio different from
unity). For forward and backscatter cases, however, symmetry forces
the ratio to be ±1, with sign depending on coordinates used. Also, in
general terms the ratio of forward scattering to backscattering increases
with particle size, and there appears a more complicated pattern of nulls
in the scattering diagram.
In 1908, Gustav Mie (1869–1957), in a milestone paper, used Maxwell’s equa-
tions, together with the appropriate boundary conditions, to solve for these
 3.3 Introduction to volume scattering 153

two complex functions for arbitrary sphere size and material. The solution is
expressed as an infinite series, with so-called Mie coefficients an and bn , as
shown in equation (3.85) (Hovenier, 2004; Ishimaru, 1991):
 
SHH () 0
[S] =
0 SVV ()
 5
∞ 2n + 1

 SVV () = {an πn (cos ) + bn τn (cos )}

 n=1 n(n + 1)
⇒ (3.85)

 5∞ 2n + 1
SHH () =
 {bn πn (cos ) + an τn (cos )}
n=1 n(n + 1)

where the component functions πn and τn are defined as follows:

dPn (cos )
πn (cos ) =
d cos 
(3.86)
d πn (cos )
τn (cos ) = cos πn (cos ) − sin2 
d cos 
where Pn (..) are the Legendre polynomials (see equation (5.49)). The Mie
coefficients an and bn depend on the (generally complex) refractive index and
size parameter x = βr, where r is the radius of the sphere. The computation of
these coefficients can be quite laborious, especially for large particles, and there
are now available several public domain codes for their numerical calculation.
An example dataset (courtesy of Michael Mishchenko of NASA Goddard) for
a sphere of size βr = 4 and refractive index n = 1.32 (equivalent to water/ice
at optical wavelengths) is shown in Figures 3.23–3.25. Figure 3.23 shows the
phase function—a rather ambiguous term, as it contains no relation to phase
between complex numbers. Rather, it is a plot of the 1,1 element of the Mueller
matrix [M ]. From equation (2.48) we see that this is just the total scattered power
(trace of coherency matrix). Figure 3.23 therefore shows how the total scattered
power varies with scattering angle. We note the enhanced scattering in the

Phase function: sphere scattering (n = 1.32, br = 4)

102

101
M11

100

10–1

10–2
0 20 40 60 80 100 120 140 160 180 Fig. 3.23 Phase function for scattering by a
Scattering angle (degrees) large dielectric sphere (βr = 4, n = 1.32)
154 Depolarisation in surface and volume scattering

[S] matrix elements: sphere scattering (n = 1.32, br = 4)

3
HH
2 VV

Amplitude (dB)
–1

–2

–3

–4

Fig. 3.24 Amplitude of scattering matrix ele- –5

ments SHH and SVV for scattering by a large 0 20 40 60 80 100 120 140 160 180
dielectric sphere (βr = 4, n = 1.32) Scattering angle (degrees)

Scattering sphere: spherical Scattering sphere: spherical

particle (n = 1.32, br = 4) particle (n = 1.32, br = 4)

90 1
120 60
0.8
0.6
150 30
0.4
0.2
180 0

210 330

Fig. 3.25 Scattering sphere representation

240 300
for a large dielectric sphere (βr = 4, 270
n = 1.32)

forward direction ( = 0◦ ) compared to backscatter ( = 180◦ ). This is due to

the large electrical size of the particle. From a polarisation point of view we are
more interested in the two complex numbers from the diagonal of the amplitude
matrix. For example, Figure 3.24 shows how the dB amplitudes of the two [S]
matrix elements vary with scattering angle. Here we note important differences,
showing variation of polarised structure with angle. (Note, however, that for
forward scatter and backscatter the amplitude ratio is 0 dB, as expected from
symmetry.) There is still an additional parameter to consider, however: the
phase between HH and VV. The combined amplitude and phase variations can
be jointly visualized as points on the scattering sphere, as shown in Figure 3.25.
On the left is a three-dimensional representation of the variation of the [S] matrix
with scattering angle for the forward scattering hemisphere. On the right is a
polar projection of this sphere, which allows a more quantitative interpretation.
For example, we can see the scattering matrix start on the equator on the left
side for backscattering, and then progress with scattering angle before spiralling
towards the antipodal point on the right-hand side for forward scatter.
 3.4 Depolarisation in volume scattering 155

We have shown by example that even spherical scatterers can generate inter-
esting polarisation behaviour in bistatic scattering. When we extend this to
consider non-spherical particles such as spheroids and even more complicated
geometries, then the polarisation response changes even more (in particular for
backscatter, which is useful for radar studies). However, there is one issue we
have yet to consider: that such particles are usually not viewed alone, but in
clouds with some distribution over size, shape, orientation, and so on. It is then
of interest to see how much of this structure survives the averaging inherent in
such depolarisation processes. We now turn to consider such effects in volume
scattering.

3.4 Depolarisation in volume scattering

As an important example of residual information retrieval in the presence of
depolarisation with azimuthal symmetry, we now consider backscattering by
a random cloud of small spheroidal particles. The [S] matrix for an individual
such particle was derived in equation (3.74), and here we start by reformulating
the [S] matrix as a scattering vector transformation as shown in equation (3.87):
  
1 0 0 2 + X cos2 τ
k = 0 cos 2θ − sin 2θ   X cos2 τ 
0 sin 2θ cos 2θ 0
 
2 + X cos τ
2

= X cos2 τ cos 2θ  (3.87)

X cos2 τ sin 2θ

where X = (AP − 1), and AP is ratio of particle polarizabilities—itself a func-

tion of particle shape and dielectric constant. The angles θ and τ dictate the
orientation of the spheroid. Now consider a cloud of such particles with ran-
dom orientation. In this case we must first consider the coherency matrix of an
individual particle, as shown in equation (3.88):
   
2 + X cos2 τ 2 2 (2 + X cos2 τ )X ∗ cos2 τ cos 2θ (2 + X cos2 τ )X ∗ cos2 τ sin 2θ
   
[T ] =  ∗
(2 + X cos τ )X cos τ cos 2θ
2 2 X cos2 τ cos 2θ 2
X 2 cos4 τ cos 2θ sin 2θ 

 
(2 + X ∗ cos2 τ )X cos2 τ sin 2θ X ∗2 cos4 τ cos 2θ sin 2θ X cos2 τ sin 2θ 2

(3.88)

We can then average this matrix over all possible angles θ and τ . The probability
distributions p(..) for a random distribution are defined as shown in equation
(3.89):

dθ cos τ d τ π π
p (θ) = −π ≤θ <π p (τ ) = − ≤τ < (3.89)
2π 2 2 2

Clearly, in the face of such a distribution the off-diagonal elements of [T ] will

average to zero, as expected. However the three diagonal terms (the eigenval-
ues) maintain some scatterer information, as shown by explicit evaluation of
the integrals in equation (3.91). The coherency matrix for such a volume may
156 Depolarisation in surface and volume scattering

be written in the form of an eigenvector decomposition, as shown in equation

(3.90):
     
1 0 0 λ1 0 0 1 0 0
[T ] = 0 1 0 .  0 λ2 0  . 0 1 0 (3.90)
0 0 1 0 0 λ3 0 0 1

where the eigenvalues take the explicit form shown in equation (3.91) (Boerner,
1992, Chapters I–4; Cloude, 1999; Ablitt, 2000).
π
1 2 π
λ1 = t11 = π
(4 + 4X cos2 τ + X 2 cos4 τ ) cos τ d θd τ
4π −2 −π

4 4
= 2 + X + X2
3 15
π (3.91)
1 2 π 2 2
λ2 = t22 = π
X cos τ cos 2θ cos τ d θd τ =
2 4 2
X
4π −2 −π 15
π
1 2 π 2 2
λ3 = t33 = π
X 2 cos4 τ sin2 2θ cos τ d θd τ = X
4π −2 −π 15

In terms of Ap —the ratio of particle polarizabilities—the eigenvalues are given

as shown in equation (3.92):
2 2 2
λ1 = (2A2P + 6AP + 7) λ2 = (AP − 1)2 λ3 = (AP − 1)2
15 15 15
(3.92)

and hence the ratio of eigenvalues yields information about the shape and mate-
rial composition of the particles in the volume through the parameter Ap , where
Ap ≥ 0 (see equation (3.77)), as shown in equation (3.93):
√
λ2 A2 − 2AP + 1 1 (1 + 3Rλ ) ± 5Rλ (3 − Rλ )
0 ≤ Rλ = = P2 ≤ ⇒ AP =
λ1 2AP + 6AP + 7 2 1 − 2Rλ
(3.93)

Figure 3.26 shows how the eigenvalue ratio Rλ changes with change in shape
of parameter Ap . We notice three important points in the function. As Ap tends
to zero, so the ratio tends to 1/7 = 0.1429, being the limit for a cloud of oblate
spheroids. For Ap = 1 the ratio tends to zero and there is zero depolarisation
(a cloud of spheres). As Ap tends to infinity we see that the ratio tends to a
maximum value of 0.5 (dipole cloud), this being the most depolarising case.
If we seek to invert this relationship by measuring the eigenvalue ratio in
order to estimate mean particle shape, then we face an ambiguity issue as
shown in Figure 3.27. Here we see that for ratios less than 0.1429 we have both
a prolate and oblate solution for the same ratio. For ratios greater than 0.1429
and less than 0.5, we obtain a unique prolate solution. For ratios between 0.5
and 1 we violate the assumptions of this model, and must consider other types
of depolarisation or noise in the data.
A similar treatment can be used for scattering by a random cloud of chiral
(or handed) particles. In this case the scattering vector for an individual particle
 3.4 Depolarisation in volume scattering 157

Eigenvalue ratio versus particle size Ap

0.5
0.45

0.4
0.35
Eigenvalue ratio

0.3
0.25

0.2

0.15

0.1
0.05
0
–20 –15 –10 –5 0 5 10 15 20
Fig. 3.26 Eigenvalue ratio of coherency
10*log(Ap) matrix versus particle shape parameter Ap

Retrieval of particle shape for azimuthal symmetry

105

104

103

102

101
Ap

100

10–1

–2
10

10–3 Fig. 3.27 Prolate (solid)/oblate (dash) ambi-

0 0.1 0.2 0.3 0.4 0.5 guity in particle shape estimation from
Eigenvalue ratio coherency eigenvalue ratio

can be obtained from the scattering matrix (derived in equation (3.82)), and
written for rotation θ, as shown in equation (3.94):

  
1 0 0 2 + X cos2 τ
k = 0 cos 2θ − sin 2θ   X cos2 τ 
0 sin 2θ cos 2θ −2iκX cos2 τ
 
2 + X cos2 τ
= X cos2 τ (cos 2θ + 2iκ sin 2θ )  (3.94)
X cos2 τ (sin 2θ − 2iκ cos 2θ )

Working through a similar procedure as used for the non-chiral case, we find
the following result for the eigen-decomposition of the coherency matrix for a
158 Depolarisation in surface and volume scattering

chiral volume (Cloude, 2002b):

√    √ 
λ1 0 0
1 2 0 0 2 0 0
[T ] = 0 1 i .  0 λ2 0 . 0 1 −i (3.95)
2 λ3
0 i 1 0 0 0 −i 1

where the eigenvalues are now given by equation (3.96):

2
λ1 = (2A2P + 6AP + 7)
15
2
λ2 = (1 + 4κ) (AP − 1)2
15
2
λ3 = (1 − 4κ) (AP − 1)2 (3.96)
15

There are two important points raised by this result. Firstly, the eigenvectors
are now complex; that is, they are no longer just the default Pauli scattering
mechanisms. This highlights the importance of using the eigenvectors for a
full characterization of the scattering process. For example, expanding equa-
tion (3.95) and using (3.96) we obtain the following non-diagonal form of the
coherency matrix for volume scattering from a random cloud of chiral particles:
 2 
2A + 6AP + 7 0 0
2  P
[T ] = 0 (AP − 1)2 −i4κ(AP − 1)2  (3.97)
15
0 i4κ(AP − 1)2 (AP − 1)2

Secondly, the presence of chirality has caused a split in the smaller eigenvalues.
Indeed, we note that the chirality parameter is directly related to the scattering
anisotropy (see equation (2.79)), defined as the normalized difference of minor
eigenvalues, as shown in equation (3.98):

λ2 − λ3
A= = 4κ (3.98)
λ2 + λ 3

3.4.1 Volume scattering with reflection symmetry

In the azimuthal symmetry case considered above, we integrated over a random
distribution of particle orientations. Now we consider generalization to consider
an oriented volume; that is, volumes with a dominant orientation (taken as
θ = 0◦ ). To model this case, we first consider a uniform distribution of particle
orientation angles centred around θ = 0◦ , as shown in equation (3.99):

dθ cos τ d τ π π
p (θ) = −≤θ < p (τ ) = − ≤τ < (3.99)
2 2 2 2

where  now determines the width of the distribution ( = π reverting to

the random case). In this case the coherency matrix for an individual particle,
which has the form shown in equation (3.88), can be averaged to obtain a new
 3.4 Depolarisation in volume scattering 159

reflection symmetric form as shown in equation (3.100):

 
t11 t12 0
[T ] = t12
∗ t22 0
0 0 t33
  
 4 4 2 sin(2) 2 4 2

 t = 2 + X + X , t12 = X+ X
 11 3 15 2 3 15
→    

 2 2 sin(4) 2 2 sin(4)

t = X 1+ , t33 = X 1−
22
15 4 15 4
(3.100)

The information contained in these equations becomes more apparent by plot-

ting the data in the H/α plane (see Section 2.4.2.4). We begin by considering
the specific example of a dipole cloud (prolate spheroids with X >> 1). A
single such particle has H = 0, α = 45◦ , and as shown in Figure 3.28, when we
form a cloud of such particles in random orientations, so the entropy increases.
However, the average scattering mechanism remains nearly constant around
ᾱ = 45◦ , permitting us to identify the mean particle shape independently of the
angular spread of the particles in the cloud.
It is interesting to see if this useful property of the H/α diagram extends to
arbitrary particle shape. Figure 3.29 shows the results for both prolate and oblate
particles (in steps of 2 dB in the ratio AP and for distributions from 0 to π ). Note
that spheres always lie at the origin of the H/α diagram. We note that the match is
very good for prolate spheroids; the scattering mechanism changes only a small
amount with entropy increase. The oblate zone is more restricted in the H/α
plane, and shows more clearly the depolarising nature of such clouds, even for
a fixed particle orientation. This arises because in equation (3.100) we are still
assuming a random distribution of particle tilt angles τ . We see, however, that

Entropy/alpha diagram for oriented dipole cloud

90

80

70

60

50
Alpha

40

30

20

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 3.28 Variation of entropy/alpha for a
Entropy cloud of dipoles
160 Depolarisation in surface and volume scattering

Entropy/alpha diagram for oriented dipole cloud

90

80

70

60

50 Ap = 20dB

Alpha
P
40

30 Ap = 6dB

20
Ap = 2dB Q
10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 3.29 Entropy alpha variation for clouds
of prolate and oblate particles Entropy

the prolate spheroids are much more sensitive to orientation effects, and display
a much wider range of depolarisers. We identify two points of interest. The point
P represents the most extreme depolariser: a random cloud of dipoles with
H = 0.9464, α = 45◦ . This point is often taken as a model for branch/needle-
dominated forest scattering in radar remote sensing applications. For example,
it is embedded in the Freeman–Durden decomposition approach (see Chapter
4). The point Q corresponds to a random cloud of oblate spheroids (H = 0.62,
α = 20◦ ). This is more typical of leaf-dominated vegetation scattering, and
shows less depolarisation than the prolate case.

3.4.2 Bistatic volume scattering

We can also use the solution for scattering by small spheroidal particles as an
important example of bistatic volume scattering. Considering a single particle,
the scattering matrix was derived in equation (3.81), and following integration
over a random distribution we obtain the following average coherency matrix,
where  is the scattering angle ( = π corresponds to backscatter) and Ap is
the particle shape function or ratio of polarizabilities.


  t11 = z(1 − cos ψ) + (1 + cos ψ)2
t11 t12 0 0 


 t12 = (cos2 ψ − 1)
t ∗ t22 0 0

[T ] = f  12  ⇒ t22 = z(1 + cos ψ) + (1 + cos ψ)2
0 0 t33 0 


 t = z(1 + cos ψ)
0 0 0 t44  33
t44 = z(1 − cos ψ)
2A2P + 6AP + 7 2(AP − 1)2
f = z= (3.101)
30 2A2P + 6AP + 7
 3.4 Depolarisation in volume scattering 161

The eigenvalues of the coherency matrix can then be calculated as shown in

equation (3.102):

2
λ1,2 = fz + f (1 + cos ψ) ± f
2
1 + cos2 ψ + z(z − 4) cos2 ψ

λ3 = fz(1 − cos ψ) (3.102)

λ4 = fz(1 + cos ψ)

(Note that the rank ordering of the eigenvalues depends on scattering angle.)
For backscatter we note that λ4 = 0 and the eigenvalues reduce to those shown
in equation (3.96). We can use these equations to highlight two important exam-
ples. In the first we consider bistatic scattering from a cloud of spheres (Ap = 1).
In this case, f = 0.5, z = 0, and the eigenvalues have the form shown in
equation (3.103):

λ1 = (1 + cos2 ψ), λ2 = λ3 = λ4 = 0 (3.103)

which shows that there is no depolarisation at any angle in single scattering

from a cloud of spheres. (Note that for the moment we are ignoring multiple
scattering contributions.) The scattering diagram with angle ψ is shown in
Figure 3.30. The pattern arises from the combination of the uniform pattern for
polarisation perpendicular to the page and the characteristic dumbbell shape
for polarisation in the plane (dashed lines in Figure 3.30).
At the other extreme we can consider bistatic scattering from a cloud of small
dipoles (Ap >> 1). In backscatter this yielded the strongest depolarisation of
all single scattering configurations. In general bistatic scattering we obtain the

90 2
120 60
1.5

150 1 30

Incident wave 0.5

180 0

210 330

240 300
Fig. 3.30 Bistatic scattering diagram for a
270 cloud of small spheres
162 Depolarisation in surface and volume scattering

Bistatic depolarisation by a dipole cloud

0.5

Relative eigen value level

lambda 1
0.4 lambda 2
lambda 3
lambda 4
0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180
Fig. 3.31 Bistatic variation of scattering
eigenvalues for a dipole cloud Scattering angle (degrees)

following form for the eigenvalues:

  
A2p
λ1,2 = 2 + cos2 ψ ± 1 − cos2 ψ + cos4 ψ
15
A2p
λ3 = (1 − cos ψ) (3.104)
15
A2p
λ4 = (1 + cos ψ)
15

Figure 3.31 shows how the eigenvalues (normalized so that λ = 1) vary with
scattering angle. Note how for backscatter and forward scattering the coherency
matrix has rank 3, as expected from reciprocity. The other interesting point
relates to 90-degree or lateral scattering. Here we see that the minor eigenvalues
are all equal, corresponding to a noise subspace of the coherency matrix. Hence
only at this angle can we give the scattering interpretation as a single [S] matrix
(the dominant eigenvector) with a noise background. At all other angles we see
anisotropy in the eigenvalue spectrum and hence structure to the depolarisation
process.
It is useful to plot the scattering entropy variation corresponding to
Figure 3.31. Figure 3.32 shows how bistatic entropy H varies with scatter-
ing angle. We note that nowhere is depolarisation complete (H = 1), and that
for forward scatter and backscatter the entropy is smallest and hence depolar-
isation at a minimum. The reason for this is the loss of one eigenvalue to the
reciprocity theorem, which, as we noted earlier, limits the level of depolarisation
achievable in backscatter. The maximum depolarisation occurs for the lateral
(90-degree) scattering case. This then represents the strongest depolariser so
far encountered. However, we have considered only single scattering depo-
larisation. When waves encounter multiple scattering—when we account for
 3.4 Depolarisation in volume scattering 163

Bistatic entropy of a dipole cloud

1

0.95

0.9

0.85
Scattering entropy

0.8

0.75

0.7

0.65

0.6

0.55

0.5
0 20 40 60 80 100 120 140 160 180
Fig. 3.32 Bistatic variation of scattering
Scattering angle (degrees) entropy for a dipole cloud

2 2
1 3 1 3
n n

Einc E(1,n) E(1,n) Einc

Fig. 3.33 Schematic representation of multi-
ple scattering for backscatter problems

interactions between particles—then there exist possibilities for higher levels

of depolarisation, as we now consider.

3.4.3 Depolarisation in multiple scattering

In the previous sections we have seen that there are limits imposed on the
level of depolarisation generated by single scattering. In this section we briefly
consider the more complicated problem of multiple scattering, and show how
this can lead to much higher levels of scattering depolarisation (Macintosh,
1989; Bicout, 1992; Brosseau, 1994; Hovenier, 2004; Mishchenko, 2006).
However, we highlight two important features of such phenomena. Firstly,
how in backscatter the reciprocity theorem still acts, even in the presence of the
most complicated multiple scattering processes, to limit the coherency matrix
to rank 3 and hence maintain a cap on the level of depolarisation that can occur.
The second point we consider is how depolarisation increases with increasing
order of multiple scattering, to highlight some interesting anisotropies that
occur in the variation of depolarisation (van Albada, 1988; Mishchenko, 2006;
Macintosh 1989).
The process of multiple scattering is shown schematically in Figure 3.33.
Here we show an incident field interacting first with particle 1, the scattered
radiation from which then interacts with particle 2, and so on, up to order
n, before being scattered in the direction of the receiver (here backscatter).
164 Depolarisation in surface and volume scattering

Hence the coherency matrix for such a scenario is not simply given by the
sum of coherency matrices for the n individual particles, but involves mutual
interactions. These interactions give rise to additional sources of depolarisation,
as we now demonstrate.

3.4.3.1 The Mishchenko decomposition

While a detailed analysis of all these multiple interactions is complicated, there
are three groupings that tend to dominate the response. The first are just the
summed direct single scattered returns from each particle. The second are
the ladder terms in the expansion of the multiple scattering integral equations
(Mishchenko, 1992). These arise from all multiple paths from 1–2–3 and so
on, and provide a new source of depolarisation. It happens that these ladder
terms are relatively easy to calculate using the techniques of vector radiative
transfer theory (VRT) (Jin, 1994b; Hovenier, 2004). This approach is based
on a set of integral equations set up as a balance of energy loss due to prop-
agation, extinction and scattering, and yields a set of linear equations that
can be solved numerically or even analytically for some simple cases (Tsang,
1985; Hovenier, 2004). There is, however, a third important contribution not
included even in these ladder terms. These are the cyclical components of the
integral equation, and basically these combine those multiple paths through
the medium which have a high coherence between a path and its time-reversed
form, as shown on the left and right of Figure 3.33. In general, therefore, we
can write the Mueller and coherency matrices for multiple scattering problems
as a sum of three terms, called the Mishchenko decomposition—first derived
in Mishchenko (1992)—as shown in equation (3.105):

[M ] = [MS ] + [ML ] + [MC ] ⇔ [T ] = [TS ] + [TL ] + [TC ]

(3.105)

where the subscripted S is for single scattering, L for the ladder terms, and C
for the cyclical terms. Concentrating on the new polarisation effects caused
by multiple scattering and their effect on the eigenvalue spectrum of [T ], we
further rewrite the coherency matrix in the form shown in equation (3.106),
where ‘MS’ now contains all the multiple scattering contributions:

[T ] = [TS ] + [TL ] + [TC ] = [TS ] + [TMS ] (3.106)

We can shed some light on the structure of these new depolarising terms by
considering the multiple paths in Figure 3.33 and writing the multiple scattering
terms as a function of their scattering vectors, as shown in equation (3.107):
9  ∗T :
[TMS ] = k (1,n) + k (n,1) k (1,n) + k (n,1)
9 :
= k (1,n) k ∗T
(1,n) + k (n,1) k ∗T
(n,1) + k (1,n) k ∗T
(n,1) + k (n,1) k ∗T
(1,n)
9 : 9 :
= k (1,n) k ∗T ∗T ∗T ∗T
(1,n) + k (n,1) k (n,1) + k (1,n) k (n,1) + k (n,1) k (1,n)

= [TL ] + [TC ] (3.107)

We see that the ladder terms are formed from averages over conventional
coherency matrix factors. However, the cyclical terms involve a mixture of
 3.4 Depolarisation in volume scattering 165

path and time-reversed path contributions. The calculation of these latter terms
is particularly difficult, but there is an interesting relationship between ladder
and cyclical terms in the exact backscatter direction. This link arises because
of the reciprocity theorem, which relates time-reversed paths as an apparent
exchange of transmitter and receiver positions. We have already seen in equa-
tion (1.147) that the two coherent [S] matrices and corresponding scattering
vectors for the two paths are related as shown in equation (3.108):
 
    1 0 0 0
  1 0  T 1 0 0 1 0 0
S(1,n) = S(n,1) ⇒ k (1,n) =  
0 0 −1 0 k (n,1)
0 −1 0 −1
0 0 0 1
(3.108)

When we substitute this connection into equation (3.107) we see that if we

know the contribution of the ladder terms (from VRT, for example) then the
cyclical terms can be immediately estimated as shown in equation (3.109):
   
t11 t12 0 t14 t11 t12 0 t14
t ∗ t22 0 t24  ∗ t24 
[TL ] =  12  ⇒ [TC ] = t12 t22 0 
0 0 t33 0  0 0 −t33 0 
∗
t14 ∗
t24 0 t44 ∗
t14 ∗
t24 0 t44
(3.109)

Note that this means that the ladder terms by themselves violate the reciprocity
theorem, and care must be exercised when using predictions from VRT for
exact backscatter calculations from random media. It is only when we add the
‘L’ and ‘C’ terms that we obtain a rank-3 coherency matrix in equation (3.106).
Turning now to an example of depolarisation by multiple scattering in a
direction other than backscatter, we consider forward incoherent scattering
by a random cloud of particles. From symmetry arguments, the Mueller and
coherency matrices for such a scenario must have the normalized form shown
in equation (3.110):
   
1 0 0 0 1 + ε + 2δ 0 0 0
0 δ 0 0  0 1−ε 0 0 
[M ] =  
0 0 δ 0 ⇔ [T ] = 
 

0 0 1−ε 0
0 0 0 ε 0 0 0 1 + ε − 2δ

 |ε| ≤ 1
⇒ |δ| ≤ 1 (3.110)

2δ − ε ≤ 1

We see that the diagonal elements must obey some restrictions in order that the
eigenvalue spectrum of [T ] be non-negative, and that the rotational subspace
eigenvalues be equal. We saw above that single scattering from such a cloud
would yield a rank-1 coherency matrix and zero depolarisation (ε = δ = 1).
However, conditions change when we allow multiple scattering. This problem,
when formulated for a cloud of spheres, can be determined analytically using
VRT (at least for the ladder terms contributions; and here we are assuming
that the coherent forward scattered wave is completely attenuated). It can be
expressed as a function of the order of multiple scattering n, with details in
166 Depolarisation in surface and volume scattering

Brosseau, 1994), and here we state the main result, shown in equation (3.111).
We show both the Mueller and corresponding coherency matrix forms of the
result as a function of n. (Note that in this case there is no dependence on
bistatic scattering angle, as we are restricting attention to forward scattering.)
For example, for n = 0 (single scattering) we obtain a Mueller matrix with the
form of the 4 × 4 identity matrix, and a rank-1 coherency matrix with three
zero eigenvalues. These correspond to forward scattering by a sphere matrix.

  n 
1 7
1+ 0 0 0
 2 10  n 
 
 3 7 
 0 0 0 
 2 10  n 
[M ] =   ⇒ [T ]
 3 7 
 0 0 0 
 2 10  n 
 
3 1
0 0 0
2 2
2  n   n  
7 7 5
+ + 0 0 0
3 10 3 7  n   n  
 
 2 7 1 5 
3 0 + − 0 0 
 3 10 3 7  n   n  
=  
2 2 7 1 5 
 0 0 + − 0 
 3 10 3 7  n  n 
 
2 7 5 5
0 0 0 + −
3 10 7 3
(3.111)

Figure 3.34 shows how the eigenvalues of this coherency matrix vary with
scattering order n. Note that for n = 0 (single scattering) the matrix has only
one non-zero eigenvalue, as expected, while as n increases so the normalized
eigenvalues all tend to 0.25; that is, to equality and maximum depolarisation
(H = 1). Figure 3.35 shows the corresponding multiple scattering entropy as
a function of increasing order. Note that the entropy approaches 1—a perfect
depolariser for orders n > 5 (corresponding to ε = δ = 0 in equation (3.110)).
A second interesting observation concerns the anisotropy of the eigenvalues.
Two of the minor eigenvalues are always equal, as expected from symmetry,

Multiple scattering coherency eigenvalues

1

0.9
lambda 1
lambda 2
0.8
lambda 3
lambda 4
Relative eigen value level

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 2 4 6 8 10
Fig. 3.34 Variation of forward scattering
eigenvalues with scattering order n Scattering order
 3.5 Simple physical models for volume scattering and propagation 167

Multiple scattering entropy (solid) and anisotropy (dash)

1

0.9

0.8
Relative eigen value level

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 2 4 6 8 10
Fig. 3.35 Forward scattering entropy and
Scattering order anisotropy versus scattering order n

and so there is only one anisotropy, A34 , to consider. The variation of this
anisotropy with order is also shown in Figure 3.35.
Note how the anisotropy is high for low orders of multiple scattering (n > 0)
and decreases steadily with increasing order. This implies that although the
spheres depolarise, they do not do so equally for all incident polarisations. We
see from the form of the Mueller matrix that this anisotropy is due to ε < δ
for all scattering orders. This means that the multiple scattering depolarises
incident circular polarisations, with a Stokes vector of the form (1, 0, 0, ±1)
more than linear, with Stokes vector (1, ±1, 0, 0), with the difference between
the two decreasing with increasing scattering order.

3.5 Simple physical models for volume scattering

and propagation
We now turn to a different aspect of volume backscattering: namely, how to
express the total scattering by a cloud of particles as a simplified volume inte-
gral. The motivation for doing this is to enable inversion of these models from
scattered field data. In the most general case we can formulate vector scattering
in random media using vector radiative transfer (VRT) theory (Chandrasekhar,
1960; Jin, 1994; Iniesta, 2003; Hovenier, 2004). In simple geometries such as
planar stratified media, there now exist efficient methods for solving these equa-
tions to include both single and multiple scattering effects. However, resort must
often be made to advanced numerical techniques. For this reason, low-order
simplified solutions are still useful in obtaining a basic physical understanding
and to enable model inversion via parameter estimation. We illustrate this by
employing two such examples: the water cloud model (WCM) (Attema, 1978;
Ulaby, 1986), which forms the basis for many practical algorithms in radar
remote sensing; and a two-phase anisotropic vegetation model from Ulaby
168 Depolarisation in surface and volume scattering

(1986) that can be used to demonstrate the importance of oriented volume

effects in wave propagation through clouds of anisotropic particles.

q 3.5.1 The water cloud model (WCM)

Nv To illustrate how volume scattering calculations typically proceed from the
Volume
scatterers
h solution for an individual particle into a cloud or scattering ensemble, we con-
sider a simple but important example called the water cloud model (WCM), first
A
derived in Attema (1978), which represents a volume as a collection of identical
particles in random positions, with density Nv particles per unit volume. For
Fig. 3.36 Geometry of the water cloud model simplicity, each particle is assumed spherical with radius ri , and furthermore we
(WCM) ignore multiple scattering between particles. The scattering geometry is shown
in Figure 3.36. A plane wave of power density P Wm−2 is incident at angle
θ, and we seek an expression for the backscatter coefficient σo (m2 m−2 ) as a
function of the volume properties, defined as follows:
Nv (m−3 ): number of particles per unit volume
h (m): height of layer
θ (rad): angle of incidence
Under the assumption of independent scattering from each particle, and ignoring
any multiple scattering effects, we can then postulate the following models for
the volume backscattering coefficient σv (m2 m−3 ) and extinction coefficient σe
(m−1 ) of the medium, as shown in equation (3.112):

!
Nv !
Nv
σv = σPi σe = QPi (3.112)
i=1 i=1

Here, σp and Qp are the single particle backscattering and extinction cross-
sections respectively, defined as the ratios of scattered and absorbed powers to
the incident power density, and both having units of m2 . For Rayleigh scattering,
when the particles are small compared to a wavelength these parameters can
be written explicitly in terms of the particle size and material composition, as
shown in equation (3.113):
   
λ2 1 − εr 2λ2  εr − 1 
QP = QA + QS = Im (βb r)3 + (βb r)6
π εr + 2 3π  εr + 2 
 
8π 2 1 − εr 3
≈ QA = Im r (3.113)
λo εr + 2
 
64π 5  εr − 1  6
σP = r
λ4o  εr + 2 

Note the following two points:

1 The extinction is composed of two parts: an absorption (A) and a scat-
tering (S) loss. For Rayleigh scattering (with βb r small), as shown, the
absorption loss is much higher than the scattering loss, and so the latter
is often ignored.
2 The important scaling parameter is the product βb r, where βb is the
wavenumber in the background medium βb = 2π
λo εrb , and λo is the
 3.5 Simple physical models for volume scattering and propagation 169

wavelength in free space. If we assume that the background is just air

(εrb = 1) then we can express QP in terms of particle radius and complex
dielectric constant εr , as shown in equation (3.113).
With this in place we can express the backscattering coefficient as an integral
over all the particles, as shown in equation (3.114):

Pi = PA cos θ 


h/ cos θ

PA
Pr = cos θ σv exp(−2σe z)dz 


4π R2 
0
(3.114)
σv cos θ
2 Pr
⇒ σoWCM (θ) = 4πR = (1 − exp(−2σe h sec θ ))
Pi 2σe
 
σv cos θ 1
⇒ σoWCM (θ) = 1− 2 L(θ) = eσe h sec θ
2σe L (θ)

Here we see that the scattering cross-section can be expressed in terms of a

combination of propagation (σe ) and scattering (σv ), with the former appearing
σv
in two ways: in the ratio 2σ e
, and in the one-way loss factor L(θ ).
The key idea of the water cloud model is to realize that the dominant material
influence in many microwave volume scattering problems comes from the water
content mv of the volume, measured in gm−3 . Hence the idea is to consider the
volume as a cloud of spherical water particles and recalculate the backscatter in
terms of water content. To do this we need two key results. The first concerns
the relation between total extinction and water content. The starting point is to
express the water content as the product of water density (106 gm−3 ) and the
fraction of space occupied by the spherical particles, which yields the following
linear relation involving constant C1 between extinction and water content:

5
Nv 4π


mv = 10 6 3
ri 
  
i=1 3 6π × 10−6 1 − εr
 N ⇒ σe = Im mv = C1 mv
1 − εr 5 v 
 λo εr + 2
8π 2
σe = λo Im r 
3 
εr + 2 i=1 i
(3.115)
σv
The second key idea in the WCM is that the ratio 2σ e
is often independent of
water content. This follows from consideration of the simplest case, when we
have a cloud of identical particles of the same radius ri = r, and when we can
write the ratio as a constant C2 , as shown in equation (3.116):

5
Nv
σv = σPi = Nv σP 



i=1 σv σP
⇒ = = C2 (3.116)
5
Nv 
 2σe 2QP
σe = QPi = Nv QP 

i=1

Note that the assumption of identical particle size can be relaxed by including a
new function—the particle size distribution function p(r)—so that the expres-
sion for total extinction (and other parameters) takes on an integral form as
170 Depolarisation in surface and volume scattering

shown in equation (3.117):

r2

σe = p(r)Qe (r)dr (3.117)

r1

While no simple forms exist for the size distribution in vegetation scattering, for
example, results are often taken from atmospheric cloud physics, where drop
size distributions are better characterized (Ulaby, 1986). The main consequence
of allowing a particle size distribution is that the backscatter σv becomes a
quadratic function of mv . Thus, in this more complicated version of the WCM,
we can combine this with the linear relation between extinction and mv to
postulate a linear relation between the ratio and mv , as shown in equation
(3.118):

σv
≈ C3 mv (3.118)
2σe

Finally, by combining all of these ideas we can then rewrite the expression for
backscatter from a volume in terms of the water content and angle of incidence,
as shown in equation (3.119), for both the uniform particle assumption (3.119a)
and the distributed particle size case (3.119b):

σoWCM (θ) = C2 cos θ[1 − exp(−C1 mv h sec θ )] (3.119a)

σoWCM (θ) = C3 mv cos θ[1 − exp(−C1 mv h sec θ )] (3.119b)

This example illustrates how we can often formulate volume scattering prob-
lems in terms of a small number of physical parameters, and then use scattering
models to simplify the expressions. However, we note that from a polarisa-
tion point of view the predicted behaviour of the WCM is trivial. With the
assumption of independent identical spherical particles and single scattering,
the scattering matrix for each is just the 2 × 2 identity, and hence there is zero
cross- or depolarisation prediction and equal scattering (in amplitude and phase)
for the copolar channels. In the next section we shall see how to extend this
model to account for non-spherical particles. We shall also make use of this
model again in a coherent form when considering polarimetric interferometry
in Chapter 7. In order to illustrate how polarisation effects can be included
in volume scattering we turn now to consider a more complicated version of
the WCM.

3.5.2 Ulaby model: two-phase volume

propagation models
Ahost medium (with background dielectric constant εh ) with one type of particle
inclusion is called a two-phase mixture. The effective dielectric constant of the
composite medium can then be expressed in terms of three parameters: εh ,
the complex dielectric constant of the included particles εI and their volume
fraction vi , defined for ellipsoidal particles as shown in equation (3.120), where
a, b and c are the particle dimensions, and Nv is the number of ellipsoids per
 3.5 Simple physical models for volume scattering and propagation 171

2c

z h z h z h
Fig. 3.37 Three examples of volume scatter-
x x x ing: prolate (left), spheres (centre), and oblate
(right)

unit volume (Ulaby, 1986, Appendix E-4).

4
vi = πabcNv (3.120)
3
In most applications of interest (microwave vegetation and forest scattering,
for example), the volume fraction is very small (< 0.01). In what follows
we restrict attention to models applicable in this small-volume fraction limit.
We also consider a low-frequency approach that requires the particle diameter
to be less than a wavelength in the particle material; that is, for spheroids,
a = b << λ. The dielectric constant of such a composite medium is then given
by a mixture model, such as the Polder–Van Santen/de Loor formula, shown in
equation (3.121):
vi (εi − εh )
εr = εh + (3.121)
1 + Lu ( εεhi − 1)

Here the parameter Lu is the particle shape factor of the ellipsoid along its u
axis, as already encountered in equation (3.76). We now consider three types of
volume of interest: one with embedded needles or prolate spheroids as shown
on the left side of Figure 3.37; one with spheres, at the centre of the figure;
and one with discs (oblate spheroids) on the right of the figure. In all cases
the positions of the particles are assumed random, but their orientations are the
same. For each of these we can write the following special forms of equation
(3.121) for the effective dielectric constant of the medium.
a) Case 1: A volume of needles aligned along the z axis (La = Lb =
0.5, Lc = 0). In this case we have different dielectric constants for dif-
ferent polarisations. For the electric field polarised in the x, y and z
directions we have the following values:
 
(εi − εh )
εrx = εh 1 + 2vi
(εi + εh )
y (3.122)
εr = εrx

εrz = εh + vi (εi − εh )

b) Case 2: A volume of spheres (L = 1/3). In this case the dielectric constant

is independent of polarisation, and is given by the following expression:
 
(εi − εh )
εr = εh 1 + 3vi (3.123)
(εi + 2εh )

c) Case 3: A volume of discs oriented in the x-y plane (La = Lb = 0, Lc = 1).

Here again we have anisotropic behaviour, with the dielectric constant
172 Depolarisation in surface and volume scattering

being polarisation dependent, as shown in equation (3.124):

εrx = εh + vi (εi − εh )
y
εr = εrx
   (3.124)
εh
εrz = εh 1 + vi 1−
εi
The above examples are all for so-called oriented volumes, where the parti-
cles are aligned to form a crystal-type structure. However, in many practical
applications the particles show no particular alignment (forming a random vol-
ume). For random orientation the dielectric constant no longer depends on
polarisation, and the scalar mixture formula takes the following form:
vi ! 1
εr = εh + (εi − εh ) (3.125)
3 1 + Au ( εεhi − 1)
u=a,b,c

It can be seen that this involves an average over the shape functions along
the three principle particle axes. For a cloud of spheres such averaging, of
course, makes no difference, and the effective medium permittivity is the same
as equation (3.123). For the two other cases, however, we obtain the following
modified results.
d) Case 4 : For a random cloud of oblate spheroids (discs):
 
vi εh
εr = εh + (εi − εh ) 2 + (3.126)
3 εi
e) Case 5 : For a random cloud of prolate spheroids (needles):
vi (εi − εh )(5εh + εi )
εr = εh + (3.127)
3(εh + εi )
V We can now make use of these results to estimate the polarisation dependent one-
way loss factor L(w, θ ) for a slab of vegetation. Figure 3.38 shows a schematic
H o
Ez of the problem. A wave is incident at θo from the normal, and traverses a slab
of thickness h. In particular, we consider the slab to be constructed of three
h phases of material, the first—a host background—being free space, so that
εh = 1. The second phase is a random collection of discs, modelling the leafy
Fig. 3.38 Geometry of wave propagation or fine-scale structure of the vegetation. Each leaf has a complex dielectric
through a slab of vegetation constant εL . Critically, for this phase the dielectric constant is independent
of polarisation. Finally, we model the stalks or trunks as a vertically oriented
array of prolate spheroids with complex particle dielectric constant εs , as shown
schematically in Figure 3.39. These will have a higher dielectric constant for
vertical polarisation than for horizontal.
The loss factor for the composite medium can then be written as the product
Leaf of polarisation independent (for the random component) and dependent (for the
oriented volume) terms as shown in equation (3.128):
 
Stalk
L w = LL .LS w (3.128)

Fig. 3.39 Composite vegetation layer com- The random component is given by an exponential propagation factor, as shown
posed of leaf and stalk contributions in equation (3.129), where n = n − in is the complex refractive index of the
 3.5 Simple physical models for volume scattering and propagation 173

two-phase material composed of leaf + background.

− cos2πθh λ nL
LL = e 0 (3.129)

For small volume fractions this refractive index can be related to the dielectric
constant as shown in equation (3.130):

 √  ε − π hεr
nL = Im( εr  ≈ r ⇒ LL = e cos θ0 λ (3.130)
2

Finally, we use Case 4 from above, setting εh = 1 and εL >> 1 (assuming

the leaves have a high water content) to obtain an expression for the one-way
loss factor as a function of the particle volume fraction and imaginary part of
dielectric constant, as shown in equation (3.131).

2vi εL − 2π hv ε
εr ≈ ⇒ LL = e 3λ cos θ0 L L (3.131)
3

An important practical measure of the leaf coverage is the leaf-area index, or

LAI, defined as the total single-sided area of all the leaves over a unit area of
ground. From this definition we can then rewrite the loss factor in terms of LAI
as shown in equation (3.132), where tL is the thickness of the leaf particles.

hvL − 2π t LAI εL
LAI = ⇒ LL = e 3λ cos θ0 L (3.132)
tL

The oriented stem loss factors cause a differential extinction; that is, a difference
between the eigenpolarisations, which in this case are H and V polarisation. The
HH polarisation is the easiest to consider first (polarisation perpendicular to the
plane of the page in Figure 3.38). This has a dielectric constant expressed
solely in terms of the stalk volume fraction (again using the approximation that
|εs |  1) as shown in equation (3.133):
) *
y (εs − 1) 4vs εs
εr = 1 + 2vs = (1 + 2vs ) − i 2 2 (3.133)
(εs + 1) εs + ε s

This shows that the H polarised wave is only weakly influenced by the presence
of the stalk components, and the extinction in this channel will be dominated by
the leafy component given in equation (3.126) (which affects all polarisations
equally).
The VV polarisation can be obtained by projecting V onto the x and z axes
and then applying the appropriate component from Case 1 (equation 3.122). For
example, Figure 3.38 shows how to calculate the z component of the electric
field from a geometrical factor sin θo . Likewise. the x component is given by
cos θo and this component ‘sees’the same dielectric constant as HH. The vertical
component sees a different extinction given from Case 1, as shown in equation
(3.134):

εrz ≈ 1 + vs εs − 1 − ivs εs (3.134)
174 Depolarisation in surface and volume scattering

In this way the loss factors for HH and VV polarisation due to the stalk com-
ponents can be written as shown in equation (3.135). Note that for normal
incidence the effect of the vertical stalks is minimal and the extinction is
isotropic (the same in HH and VV). Only as the angle of incidence increases,
so the differential effect increases, with the stalks having maximum influence
at 90◦ or grazing incidence.
   
 4π nHH h

 HH
L ≈ exp − nHH = Im
y
εr

 λ cos θ
 o
  
Ls (w) ⇒

 4π cos2 θo nHH + sin2 θo nVV h 
L ≈ exp −
  n = Im εrz

 VV λ cos θo VV

(3.135)

This model shows how extinction propagation through oriented volumes can
be related to simple physical parameters such as volume fractions and dielectric
constants. The above model, however, is restricted to low-frequency scattering
and in particular to low albedo problems (when the ratio of scattering loss to
total extinction is small).

3.5.3 Forest extinction models

The extension of these models to more general cases is complicated by the
need to consider both scattering and absorption contributions to the extinction
(Tsang, 1985). One way to deal with such complexity is to employ empirical
models generated by fitting functions (usually polynomials) to experimental
data. For example, in forestry, using a comprehensive set of airborne data across
different forest sites and using different airborne sensors (including the P-3
(UHF), Carabas II (VHF), JPL-AIRSAR (P,L,C), and SRI(UHF)), the following
model has been proposed to represent the total two-way extinction X in dB
(Bessette, 2001):

X cos θ = a.F b (3.136)

where F is the frequency in MHz and a and b are regression parameters obtained
from a global fit across all datasets. The mean values of these two parameters
for 20, 50 and 80% of data are shown in Table 3.1. Figure 3.40 shows plots of
extinction at the 20, 50 and 80% levels versus frequency for normal incidence
(θ = 0◦ ) over the range 100 MHz to 1.3 GHz. Figure 3.41 shows the accom-
panying multiplicative scale factor for increasing angle of incidence. To find
the total two-way extinction at frequency F for a given angle of incidence θ,

Table 3.1 Regression-based extinction model parameters (from Bessette (2001))

20% 20% 50% 50% 80% 80%

Attenuation factors a b a b a b
HH 0.08 0.59 0.18 0.53 0.19 0.56
VV 0.21 0.47 0.3 0.47 0.32 0.50
 3.5 Simple physical models for volume scattering and propagation 175

Forest extinction
12

10
80%
2-Way extinction (dB)

8
50%

20%
4

0 Fig. 3.40 Example of an empirical model

0 200 400 600 800 1000 1200 1400
for predicting extinction through forest
Frequency (MHz) (dash = VV; solid = HH)

Scale factor for increasing angle of incidence

3

2.8

2.6

2.4
Extinction scale factor

2.2

1.8

1.6

1.4

1.2

1 Fig. 3.41 Scale factor to account for

0 10 20 30 40 50 60 70
increased extinction at larger angles of
Angle of incidence incidence

take the value from Figure 3.40 and multiply its dB value by the scale factor
obtained from Figure 3.41. We note the following general points:

1. The extinction increases with frequency at a rate around (6/ cos θ)

dB/decade for the 80% coverage fit.
2. The extinction for VV (shown as the dashed line in Figure 3.40) is always
greater than for HH, indicating an oriented volume effect, even in com-
plex forestry, with a slight preference for vertically oriented scatterers
in the volume.
176 Depolarisation in surface and volume scattering

3.5.4 Dual polarised surface and volume depolarisation

Active remote sensing systems often operate in dual polarisation or compact
modes (see Chapter 9) (Raney, 2006; Souyris, 2005). In this case, rather than
measurement of the full [S] matrix, the system obtains a projection formed by
using only a single transmit polarisation, typically either just horizontal H or
vertical V and dual channel (H and V) coherent receive. These systems can still
be used to partially characterize depolarisation effects by random surface and
volume scatterers by employing 2 × 2 coherency matrixes (N = 2 depolarisa-
tion) as follows. For example, Bragg surface scattering from smooth surfaces
will have 2 × 2 coherency matrices of the form shown in equation (3.137):

 
1 a + c + 2Re(b) 0
[JH ]Bragg =
2 0 0
  (3.137)
1 a + c − 2Re(b) 0
[JV ]Bragg =
2 0 0

Note that these matrices imply zero scattering entropy and α2 = 0 for all
angles of incidence and dielectric constants. We can extend this to include
depolarisation caused by surface roughness by considering the X-Bragg model.
This takes the projected form shown in equation (3.138):

   
c sin 4 sin 4
1 a + 1 + + 2Re(b) 0 
[JH ]XBragg =  2 4 4  
2 0
c
1−
sin 4 
2 4
   
c sin 4 sin 4
1 a + 2 1 + 4 − 2Re(b) 0 
[JV ]XBragg = 
4
c
 
sin 4 
2 0 1−
2 4
(3.138)

Here we see non-zero scattering entropy, but note a mixture of roughness and
dielectric constant dependence in the terms. Contrast this with the full coherency
matrix formalism where we are able to separate roughness and moisture effects.
Still, the level of depolarisation in equation (3.138) is small, and to see how
high the entropy can be we turn again to the case of volume scattering from a
random cloud of anisotropic particles. Using the notation of equation (3.92) we
see that the dual polarisation coherency matrices can be expressed in terms of
the particle anisotropy Ap , as shown in equation (3.139):

 
1 3A2p + 4Ap + 8 0
[JH ]Vol = [JV ]Vol = (3.139)
15 0 (Ap − 1)2

For dipole scatterers, Ap tends to infinity and we have the strongest depolariser.
For oblate particles, on the other hand, Ap = 0 and the depolarisation is weaker.
These two cases may be distinguished in dual polarised systems using their
 3.5 Simple physical models for volume scattering and propagation 177

limiting form of 2 × 2 coherency matrix, as shown in equation (3.140):

 
3 0
[JH ]prolate = [JH ]prolate ∝ ⇒ H2 = 0.811
0 1
  (3.140)
8 0
[JH ]oblate = [JV ]oblate ∝ ⇒ H2 = 0.503
0 1

We see that the maximum dual polarised (N = 2) entropy from such a cloud
is 0.811, and note that while dual polarised systems offer some potential for
the separation of different types of volume scattering based on their levels of
depolarisation, they remain inferior to full [S] matrix systems.
 Decomposition
4 theorems

In Chapter 1 we saw how physical boundary conditions on a perfect conductor

lead to a π phase shift on plane reflection, which when applied to the backscat-
tering amplitude matrix leads to a canonical series of the form shown in
equation (4.1):
 
1 0
[S]BSA = (4.1)
0 (−1)n+1

Here n is the order of reflection, so n = 1 is single reflection leading to the

2 × 2 identity matrix, while n = 2 is characteristic of dihedral reflections,
with π phase difference between polarisations. These lead, via vectorization,
to canonical Pauli scattering vectors of the form (1,0,0)T and (0,1,0)T . These
vectors form the building blocks for a generalization of this theory to encompass
reflections from arbitrary dielectric interfaces, as we now demonstrate.
This generalization is of key practical importance for the development of
applications in remote sensing, where variations in dielectric constant and angle
of incidence must always be considered. In this chapter we see how to extend
the vector formulation to deal with dielectric interfaces, with perfect conduc-
tors then arising as a special case. The general formulation then leads us to
consider coherent decomposition theorems (Cloude, 1985; Krogager, 1993;
Cameron, 1996). We develop these ideas in several stages, first by considering
transformations—such as rotation of frame of coordinates or phase shifts to cir-
cular polarisations—and determine their impact on a description of reflection
at dielectric boundaries. Secondly, we then generalize the vector formulation
to cope with arbitrary boundary conditions in a transformation invariant way,
which requires a generalization of the vector approach via the concept of the
alpha parameter. Finally, we show how to incorporate depolarisation effects into
this formulation by considering noncoherent decomposition theorems (Huynen,
1970; van Zyl, 1990; Cloude, 1996; Freeman, 1998; Yamaguchi, 2005).

4.1 Coherent decomposition theorems

We begin with a simple but important example. Consider how the scattering
amplitude matrix [S] changes when we rotate the transmitter and receiver coor-
dinate systems through an angle θ. In matrix form the rotation can be represented
by a similarity transformation as shown in equation (4.2):
   
cos θ − sin θ SHH SHV cos θ sin θ
[S (θ)] = (4.2)
sin θ cos θ SVH SVV − sin θ cos θ
 4.1 Coherent decomposition theorems 179

By expanding the matrices on either side in terms of the Pauli spin matrices
we can write this as a unitary vector transformation as shown in equation (4.3)
(verified by expansion and use of standard trigonometric identities):
  
1 0 0 0 SHH + SVV
1  0 cos 2θ − sin 2θ 0 
 SHH − SVV 

k (θ) = √  (4.3)
2 0 sin 2θ cos 2θ 0 SHV + SVH 
0 0 0 1 SHV − SVH

Here we note that the complex sums Shh + Svv and Shv − Svh are invariant to
rotations, which already gives them a special physical significance. Of particular
importance is the form of this transformation for backscatter problems, when
the transmit and receive coordinates coincide and θ corresponds to rotation of
the object about the line of sight. In this case, Shv = Svh in the BSA system,
and so for this case the transformation reduces to three dimensions, as shown
in equation (4.4):
   
1 0 0 SHH + SVV
1 
k (θ) = √ 0 cos 2θ − sin 2θ  . SHH − SVV  = [R]k (4.4)
2 0 sin 2θ cos 2θ 2SHV

Hence we see that while 4 × 4 unitary matrices [U4 ] are required to represent
the most general bistatic scattering case, 3 × 3 unitary matrices [U3 ] are of
importance for the special case of backscatter. In equation (4.3) we noted that
Shh + Svv is an invariant to rotation, and it is interesting to ask if there are
any other complex combinations of the [S] matrix that remain unchanged by
rotations. To see this we now consider the use of eigenvalue decompositions.

4.1.1 Roll invariance and eigenvectors

We first calculate the eigenvectors of the transformation matrix [R] in equation
(4.4). By definition these states remain unchanged (apart from multiplication
by a complex scalar, the eigenvalue) on operation of the matrix. Hence any
linear combination of these states will also preserve its form. For the rota-
tion matrix, calculation of the eigenvectors is straightforward, as shown in
equation (4.5):
     
1 0 0
1 1
[R] u = λu ⇒ u = 0 , √ 1 , √  1  λ = 1, e−i2θ , ei2θ (4.5)
0 2 i 2 −i

Importantly, these three eigenvectors form a complete set; that is, any k vector
can be expanded uniquely in terms of this mutually orthogonal set. Hence we
can use the matrices corresponding to these vectors as a basis for the expansion
of an arbitrary [S] matrix to obtain equation (4.6):
           
1 0 0 1 0 0
k = x 0 + y 1 + z  1  = f 0 + ge−i2θ 1 + hei2θ  1 
0 i −i 0 i −i
(4.6)
180 Decomposition theorems

where the complex scalars f , g and h remain invariant to rotations of the object
about the line of sight. Furthermore, we may give each vector a physical inter-
pretation in terms of a canonical scattering mechanism by converting the vectors
back into their corresponding [S] matrices, as shown in equation (4.7).
     
1 0 1 i 1 −i
[S] = f + ge−i2θ + hei2θ
0 1 i −1 −i −1
 
f + ge−i2θ + hei2θ i(ge−i2θ − hei2θ )
= (4.7)
i(ge−i2θ − hei2θ ) f − ge−i2θ − hei2θ

The first of these corresponds to the sphere symmetric specular scattering pro-
cess, while the remaining two correspond to scattering from a helix, the first
(g) with right and the second (h) with left sense. Both scatter all incident
waves into circular polarisation, and hence are not found generically as paired
polarising objects in scattering by natural media. For this reason an alternative
factorization is commonly used, designed to maintain invariance to rotations
but employing more generic scattering mechanisms.

4.1.2 Krogager and Cameron decompositions

To motivate this idea we first rewrite the [S] matrix in the circular basis as shown
in equation (4.8), and then extract the ‘dominant’ helix, leaving a remainder
term as shown.
 
hei2θ if
[Scirc ] =
if −ge−i2θ
 i2θ   
 ge 0 (h − g)ei2θ 0
  
 + if |h| > |g|
0 if 0 −ge−i2θ 0 0
= +    
if 0 
 hei2θ 0 0 0

 + if |g| > |h|
0 −he−i2θ 0 (g − h)e−i2θ
(4.8)

The remainder term is the same in both cases, and has the form of a canonical
scattering process: dihedral scattering at an angle θ. This expansion was first
proposed in Krogager (1992, 1993), and has the vector form shown in equation
(4.9). Here we see that while the sphere symmetric term has been maintained,
the left and right helical coefficients have been combined into a single domi-
nant helix scattering term and the remainder used to generate a new scattering
mechanism: a dihedral scatterer oriented at angle θ/2.

    


0 0
 (h − g)  ei2θ  + 2g cos 2θ  if |h| > |g|
   

1 
 −iei2θ sin 2θ
k = f 0 +     (4.9)

 0
0 

0
 (g − h)  e−i2θ  + 2h cos 2θ  if |g| > |h|



ie−i2θ sin 2θ
 4.1 Coherent decomposition theorems 181

This technique is often called the Krogager or SDH (sphere-diplane-helix)

decomposition, in recognition of its use of these three basic building blocks
to represent arbitrary scatterers. However, we note that these are still assumed
to be three metallic scatterers, and we cannot yet account for interactions at
dielectric interfaces. The three real non-negative components of this decompo-
sition are directly obtained from the circular polarisation components, as shown
in equation (4.10):

ks = |SRL | kD = min |SRR | |SLL | kH = ±(|SRR | − |SLL |) (4.10)

As well as providing the three real parameters of the SDH decomposition, this
technique can also yield an efficient algorithm for estimation of the orientation
of the scatterer θ (Lee, 2002; Schuler, 2002). From the form of the circular S
matrix (equation (4.8)) we see that θ can be directly related to the product of
SLL and SRR as shown in equation (4.11). In equation (4.11a) we show how the
phase of the LL/RR product is related to θ and to the phase of hg ∗ .
∗
SLL SRR = −hg ∗ ei4θ
∗
(4.11a)
arg(SLL SRR ) = π + arg(hg ∗ ) + 4θ

∗ ∗ ∗ ∗

−SLL SRR = − (SHH − SVV + 2iSHV ) SVV − SHH − 2iSHV
∗
= |SHH − SVV |2 − 4 |SHV |2 + i4Re((SHH − SVV ) SHV )
  (4.11b)
4Re((SHH − SVV ) SHV∗ )
⇒ tan−1 = arg(hg ∗ ) + 4θ
|SHH − SVV | − 4 |SHV |2
2

+ ∗
,
+ ∗
, −1 4Re (SHH − SVV )SHV
arg −SLL SRR = tan + , + , = 4θ
|SHH − SVV |2 − 4 |SHV |2
+ ∗
,
1 −1 4Re (SHH − SVV )SHV 1 2Re(t23 )
⇒ θ = tan + , + , = tan−1
4 |SHH − SVV | − 4 |SHV |
2 2 4 t22 − t33
(4.11c)

In equation (4.11b) we show how the phase of −SLL SRR ∗ can then be related to
the sum of 4θ and a scattering phase between h and g. Importantly, the product
hg ∗ is often real, and so this phase is zero. We can see from equation (4.11b)
that this will occur, for example, when SHV equals zero. This always occurs for
some θ for so-called symmetric scatterers, as discussed in the Cameron decom-
position below. More generally, however, in the presence of depolarisation we
can use averaging to estimate the average phase, as shown in equation (4.11c).
Here we can see that the average scattering phase will be zero when we are
aligned to the dominant axis in reflection symmetric depolarisers (see Section
2.4.2.3), since SHV is then uncorrelated with both SHH and SVV . In this case we
see from the equation that θ may then be easily related to the elements of the
coherency matrix [T ]. Note that while this is a robust and widely used algorithm
for estimation of θs its main drawback is the limited range of θ available, due
to the factor of 14 used in the equation.
This discussion about orientation leads us to consider an important class
of objects termed symmetric scatterers (Cameron, 1996; Touzi, 2007). These
182 Decomposition theorems

J V
Symmetry axis
H
a b a –b a 0
Jref [S] = + ⇒
Fig. 4.1 Definition of symmetric scatterers c d –c d 0 d

are defined as objects having an axis of symmetry in the plane of polarisa-

tion. When this axis is aligned with the antenna coordinates then there must
be zero crosspolarisation, as shown schematically in Figure 4.1. For a hori-
zontally polarised incident wave we see that secondary current induced above
and below the symmetry line have parallel copolarised but antiparallel crosspo-
larised components. This leads to cancellation of the crosspolarised response of
the object. Formally, the scattering matrix is formed as the coherent sum of two
terms in the scattering matrix group [S] + [Sβ ] in equation (1.145). The sum
always results in a diagonal matrix in the HV system, as shown in Figure 4.1.
For such objects, therefore, there always exists a linear basis that diagonalizes
the scattering amplitude matrix. In terms of the scattering vector it follows that
it must be possible to find θ such that the crosspolar terms go to zero. This
implies that the backscattering vector from such symmetric objects must be
constrained, as shown in equation (4.12):
       
a 1 0 0 a cos α
c =0
b  = 0 cos 2θ − sin 2θ  . b −→ k = sin α cos 2θeiφ 
c 0 sin 2θ cos 2θ c sin α sin 2θ eiφ
⇒ φHH −VV = φHV (4.12)

Here we see that it is only possible to diagonalize [S] by a rotation if the phase
of the complex HV term equals that of the HH-VV term, which again gives
special significance to the Pauli matrix expansion of [S]. This result provides us
with an alternative algorithm for estimating the orientation angle of symmetric
scatterers from the scattering vector, as shown in equation (4.13):

2SHV π π
tan 2θ = − ≤θ ≤ (4.13)
SHH − SVV 2 2

While this appears to be a better approach to the LL/RR phase (it provides
a wider range of angle estimates, for example) it is formed as the ratio of
complex entities and hence is sensitive to fluctuations or departures from the
symmetric assumption (and is unstable when SHH = SVV ; that is, for specular
surface scattering). However, there are two important ideas springing from this
observation, both of which are concerned with devising strategies for dealing
with scatterers that do not obey the symmetry constraint.
The first is called the Cameron decomposition (Cameron, 1996). The idea
here is to filter the scattering vector before applying equation (4.13). The filter-
ing is performed so as to generate the maximally symmetric component of the
complex scattering vector—to keep as much of the original vector as possible,
but to enforce the symmetry assumptions that the phase of HV and HH-VV are
equal and that the amplitude ratio gives tanθ . The former requires that HV and
HH-VV share a common complex factor ε, as shown in equation (4.14). To find
 4.1 Coherent decomposition theorems 183

ε we first calculate a residual vector k between the original and symmetric

approximation, and then choose ε and θ to minimize the norm:
 
0
θ,ε
k = b − ε cos 2θ  −→ min k ∗T k
c − ε sin 2θ
∂ (4.14)
→ = 0 ⇒ ε − b cos 2θ − c sin 2θ = 0
∂ε
∂ 
→ = 0 ⇒ |b|2 − |c|2 sin 4θ − (bc∗ + cb∗ ) cos 4θ = 0
∂θ

The Cameron algorithm can then be presented in two stages, as shown in equa-
tion (4.15). The first is to generate an estimate of θ from the ratio of real
quantities, and then to use this value to combine the complex numbers b and c
into a single term ε.
   
a a
k = b −− 
−−−−−−−−−−−−−→ sym = ε cos 2θ
max symmetic component k 
c ε sin 2θ (4.15)
bc∗ + b∗ c
tan 4θ = ε = b cos 2θ + c sin 2θ
bb∗ − cc∗

Note that the orientation estimate is then formally equivalent to the LL/RR
coherence discussed in equation (4.11). One other important parameter deriving
from this filtering idea is the degree of symmetry Dsym of a scattering vector,
which expresses the ratio of power in the symmetric component to the total
power, as shown in equation (4.l6).

|a|2 + |ε|2
Dsym = 0 ≤ Dsym ≤ 1 (4.16)
|a| + |b|2 + |c|2
2

The second approach to dealing with orientation of non-symmetric scatter-

ing vectors was first developed in Huynen (1987), and further developed in
Touzi (2007). The idea is to rotate the vector until the scattering matrix can be
expressed as the product of a diagonal form (in general the singular values of
[S]) and an ellipticity or ‘tau’ transformation matrix (see equation (1.120)). This
approach contains symmetric scatterers as a special case, when the ellipticity
transformation becomes the identity matrix. The canonical (zero rotation) form
of the [S] matrix in this approach is given by equation (4.17):
     
cos τ i sin τ λ 0 cos τ i sin τ
[S] = . 1 . λ1 , λ2 ∈ C (4.17)
i sin τ cos τ 0 λ2 i sin τ cos τ

When we vectorize this expression using the Pauli basis we obtain the following
canonical form:
    
(λ1 + λ2 ) cos 2τ cos 2τ 0 i sin 2τ λ1 + λ2
k= (λ1 − λ2 ) = 0 1 0  λ1 − λ2  (4.18)
i (λ1 + λ2 ) sin 2τ i sin 2τ 0 cos 2τ 0
184 Decomposition theorems

This can be used to estimate the unknown rotation of a general k vector by

recognising that the first and third elements of the canonical form are in phase
quadrature so that the following relation gives an expression for θ:
  c 
b Re ac
sin 2θ Re + cos 2θRe = 0 ⇒ tan 2θ = −  (4.19)
a a Re b a

While this is formally correct, again it does not allow extension to the idea of
coherence and so is susceptible to complex noise fluctuations. For this reason the
most robust algorithm for orientation estimation remains the LL/RR coherence
(equation (4.11)). The only exception to this is in low noise (low entropy)
environments or when the restricted range of the LL/RR estimate is too limiting
for the application.

4.1.3 The scattering alpha parameter

We have seen above that the circular polarisation base provides a good platform
from which to derive information about the mean orientation of the scatterer
(from the LL/RR coherence phase). In Section 4.2.6 we show that circular
polarisation is also a good base to assess the effects of mean Faraday rotation
(LR/RL coherence phase). The scattering vector formulation in the Pauli base
connects these different processes as special cases of 4 × 4 unitary transfor-
mations. This idea can be extended to consider the physical interpretation of
general unitary transformations of the scattering vector. In particular we con-
sider one important example: the alpha parameter, which will allow us finally to
consider an approach to decomposition at dielectric boundaries. The problem
with dielectrics is that they change the ratio of copolar [S] matrix elements.
Hence we need to move away from a description based on the simple ratios
implicit in the Pauli matrices; that is, we need now to consider not just rota-
tions but unitary transformations that allow us to move smoothly away from the
Pauli basis. This can be achieved by using transformations of the form shown
in equation (4.20).
  
cos α − sin α 0 0 SHH + SVV
1  sin α cos α 0 0 
 SHH − SVV 

k (α) = √  (4.20)
2 0 0 1 0 SHV + SVH 
0 0 0 1 SHV − SVH

We see that α changes the copolarised terms of the scattering matrix as required,
although it remains invariant to rotations of the measurement coordinates. The
angle is conveniently defined over the range 0 ≤ α ≤ 90◦ , and represents a
smooth change of scattering mechanism. To see this we start with the sphere
 = 0º  = 45º  = 90º
symmetric specular mechanism with target vector k = [1,0,0,0]T and corre-
sponding 2 × 2 identity matrix for [S]. This we define as the α = 0◦ boundary.
Sphere Symmetry Dipole Dihedral/Helix
Changes in α then correspond to departures in scattering from this reference
mechanism. These variations can be summarized as shown in Figure 4.2. In
the range 0◦ ≤ α ≤ 45◦ the copolarised scattering terms differ in amplitude
but are equal in phase, reaching, as an extreme point, α = 45◦ , which corre-
Fig. 4.2 Physical interpretation of the sponds to an anisotropic scatterer with only one non-zero copolarised term. A
scattering alpha angle simple example of this is a linear dipole scatterer, which has strong copolarised
 4.1 Coherent decomposition theorems 185

scatter for polarisations parallel to its axis and zero for the orthogonal case.
In the range 45◦ < α ≤ 90◦ we move into a region where the copolarised
scattering coefficients are 180 degrees out of phase. This can occur for mul-
tiple scattering (dihedrals) or asymmetric scattering such as helices. In the
extreme case of α = 90◦ these scatterers have equal amplitude but maintain
the 180-degree phase difference (metallic dihedrals or helices, for example).
The alpha angle for any [S] matrix can then be directly estimated as shown in
equation (4.21):

|SHH + SVV | π
cos α = √ 0≤α≤ (4.21)
2 |SHH |2 + 2 |SHV |2 + |SVV |2 2

One key advantage of using this angle in place of simple ratios such as SHH /SVV
is that it is invariant to rotations of the object. For example, a dipole changes
its [S] matrix elements with rotation θ (see equation (1.161)), but both numer-
ator and denominator of equation (4.21) remain invariant to θ, and hence α is
unchanged. The same is true for rotation of a dihedral or any other rotationally
dependent scatterer. Hence we can identify the scattering mechanism without
needing to know the orientation. There remains an interesting question of how
best to estimate the mean α angle in the presence of noise and fluctuations. This
cannot be answered using a simple coherence estimate as it was for the orienta-
tion θ, and requires application of the entropy/alpha decompositions presented
in Chapter 2.
To illustrate this we highlight an important special case of this alpha trans-
formation. We can, for example, use it to represent the set of all diagonal [S]
matrices with complex diagonal values λ1 and λ2 as SU(2) transformations of
the identity (corresponding to a sphere symmetric scattering mechanism), as
shown in equation (4.22):

    
λ1 + λ 2 cos α − sin αe−iδ 0 1
1 λ1 − λ2  = e sin αe
√
iφ iδ cos α 0 0
2 |λ1 |2 + |λ2 |2 0 0 0 1 0
 
|λ1 + λ2 | λ1 − λ2
⇒ cos α = √ , δ = arg (4.22)
2 |λ1 |2 + |λ2 |2 λ1 + λ 2

We can now combine the above results to propose a ‘point reduction theorem’,
as follows. According to this theorem we can express an arbitrary scattering
mechanism w (and here we restrict attention to the backscatter reciprocal three-
element case) as transformations of the sphere symmetric Pauli scatterer, as
shown in equation (4.23):

 
cos α
 
w = eiφ  iφ1 
sin α cos ψe 
sin α sin ψeiφ2
     (4.23)
1 0 0 cos 2τ 0 i sin 2τ cos αd − sin αd e−iδ 0 1
      
= eiφ  
0 cos 2θ − sin 2θ   0 1 0  
 sin αd e
iδ cos αd 0  
 0
sin 2θ cos 2θ i sin 2τ 0 cos 2τ 0 0 1 0
186 Decomposition theorems

Starting on the right-hand side we begin with a 2 × 2 unitary sub-matrix that

represents the change of (complex) singular values of the [S] matrix away from
the identity matrix. We then invoke a transformation to account for ellipticity in
the singular polarisation states—a departure from symmetric scatterers. Finally
we include a rotation of the scatterer in the plane of polarisation. On the left-hand
side we show the standard trigonometric parameterization of unitary vectors in
three dimensions. We note the following relationships:
1. The parameter α remains invariant to rotations as expected, but actu-
ally satisfies the following relation to parameters of the point reduction
transform:

cos α = cos 2τ cos αd (4.24)

Only in the case of symmetric scatterers (τ = 0) does α = αd . Nonethe-

less, α remains a useful parameter as a measure of the departure of the
scattering process from sphere symmetry.
2. The parameter ψ is only equal to the rotation angle in case of symmetric
scatterers (τ = 0). In other cases the ψ parameter contains a mixture
of dependence on rotation and ellipticity. In the non-symmetric case,
orientation is better estimated using equation (4.19).

4.1.4 Orthogonal scattering mechanisms

Very often we wish to determine the component of a scattering vector k in a
predetermined ‘direction’ w. To do this we can form the projection of k onto w
as the Hermitian inner product, as shown in equation (4.25):

s = w∗T k = reiφ (4.25)

The power in this projection can then be computed as shown in equation (4.26):

p = ss∗ = w∗T kk ∗T w = r 2 (4.26)

It is also of interest to consider the match between w and the scattering mecha-
nism embedded in k. This can be found from the normalized inner product, as
shown in equation (4.27):

w∗T k
d=   0 ≤ |d | ≤ 1 (4.27)
k 

Clearly, if the scattering mechanisms are the same then d = 1. Interestingly,

however, we can also obtain ‘orthogonality’ between mechanisms, defined
when d = 0. In this case the power in the projection will be zero. Orthogonality
is already a powerful concept in the analysis of elliptically polarised waves,
which allowed us to construct bases from paired combinations of orthogonal
polarisations. Through vectorization of [S] we can now extend this analysis to
scattering mechanisms themselves.
It is interesting to investigate the orthogonality of the canonical mechanisms
identified in the various decompositions in the previous section. For example,
the Pauli matrix set form an orthogonal set, as their vectors are all mutually
orthogonal. Hence, in addition to obvious orthogonalities such as horizontal
 4.1 Coherent decomposition theorems 187

and vertical dipoles, we can also count spheres and 45◦ and 0◦ rotated dihedrals
as orthogonal mechanisms. The eigenvectors of the rotation matrix [R] used
in equation (4.5) also form an orthogonal set. However, the SDH components
themselves do not form an orthogonal set (see equation (4.9)). The helix and
dihedral do not represent orthogonal processes.
This orthogonality is a desirable property of coherent decomposition theorem
(CDT) schemes, as it provides optimum separability in classification schemes
and forms the basis for an expansion of second-order averages in the coherency
matrix, as discussed in Section 4.2. To illustrate this we consider a sphere
mapping based on diagonal [S] matrices, as developed in Chapter 3. In this case
we consider classification of scattering vectors of the reduced form, shown in
equation (4.28):
   
cos α sin α
w1 = sin αeiδ  ⇒ w⊥ = − cos αeiδ  (4.28)
0 0

This is mathematically identical to the case of orthogonal waves in C2 and

consequently we can map each scattering mechanism as a point on the scattering
sphere (analogous to the Poincaré sphere) and define for each mechanism an
orthogonal one, as shown in equation (4.28). These two mechanisms will have
antipodal points on the sphere, and α and δ represent the spherical triangle
coordinates of the points. This result has two important consequences. The first
is that we can define each point on the sphere using a real three-vector—the
equivalent of the Stokes vector for waves. The second that is we can define a
coherency matrix between the two complex singular values and extract a mean
scattering mechanism in the presence of noise. We now turn to consider the
machinery required to deal with fluctuations in the [S] matrix elements, and
consider the general problem of depolarisation. First, however, we consider the
issue of orthogonality at dielectric interfaces.

4.1.5 Orthogonality of scattering mechanisms in

natural terrain
A key question arising from the concept of scattering matrix orthogonality is
its probable occurrence in scattering from natural media. This is particularly
important in remote sensing applications when we often wish to separate depo-
larising surface and volume scattering components. In particular, we saw in
Chapter 3 that either one of the dihedral or direct surface scattering mecha-
nisms can be dominant, or indeed both can occur simultaneously in scattering
from natural terrain, depending on the topography, vegetation structure and sur-
face roughness. In this section we look at the degree to which these processes
remain orthogonal over a wide range of variations in dielectric constant.
If we define wB as the scattering vector for Bragg surface backscattering
at angle θ from a surface with dielectric constant ε r , and likewise the corre-
sponding dihedral mechanisms (at the same angle and dielectric constant of the
surface as wD ), then the degree of orthogonality Q can be defined (in dB) from
the magnitude of an inner product, as shown in equation equation (4.29):
 
Q = 20 log10 w∗TD wB
 (4.29)
188 Decomposition theorems

The smaller Q, the better the isolation between the mechanisms. Small Q better
wB wD enables separation of the mechanisms when both are present in the same scat-
B tering problem, as they then emerge as separate eigenvectors of the coherency

matrix [T ]. The scenario we are considering is an arbitrary combination of sur-
face and dihedral scattering, as shown schematically in Figure 4.3. Note that
A
the relative amplitude of the mechanisms is not so important in this context
Fig. 4.3 Schematic of combined direct and as is their relative polarimetric properties. Figures 4.4 and 4.5 show the factor
specular dihedral surface returns Q for the same two angles of incidence used in Figures 3.15 and 3.16. Note
that Q is smaller for the larger angle of incidence. For the shallower angle, Q
can still be small (−10 dB or so), but for dry materials (with a small εr < 20)

Orthogonality between Bragg/dihedral mechanisms (dB) at 22.5 degrees

80

–14
–12
.4
.6

.8
.2
70

10
60
Dielectric constant A

50

–14
–12
.6
.4

.8
.2

40

30
10

20 –14
–12
.4

.6

.8
.2

10
10 –12
10
10 20 30 40 50 60 70 80
Fig. 4.4 Magnitude of Q (equation (4.23))
for 22.5-degree angle of incidence Dielectric constant B

Orthogonality between Bragg/dihedral mechanisms (dB) at 45 degrees

80
20
25

15

70

60
Dielectric constant A

50
20

15
25

40
2520

30 15
15

15
20 20
30 26
20 20
26 25
10 20 25
25
–15 20
–15 2025
–10
10 20 30 40 50 60 70 80
Fig. 4.5 Magnitude of Q (equation (4.23))
for 45-degree angle of incidence Dielectric constant B
 4.2 Incoherent decomposition theorems 189

at shallow angles we note that the separation is the poorest at around −6 dB.
Again we note, for the shallow angle, that it is the dielectric constant of the
second surface B that causes the largest variation in Q.
The larger angle of incidence has a wide range of combinations of dielectric
constants where Q is better than −20 dB. Hence we note that separability of the
two mechanisms wD and wB , based on orthogonality, is largely independent of
dielectric constant variations and improves with increasing angle of incidence.
This discussion of orthogonality, of course, has the largest impact in depo-
larisation problems, when the orthogonal mechanisms can be easily separated
as eigenvectors of the Hermitian coherency or covariance matrix. This brings
us to consider the set of issues raised by incoherent decompositions.

4.2 Incoherent decomposition theorems

We have seen that it is often useful to represent the scattering coherency matrix
[T ] (or equivalently the Mueller matrix [M ]) as a sum of composite elements,
so in general we can write an expansion of the form shown in equation (4.30):

!
R !
R
[T ] = [Ti ] ⇔ [M ] = [Mi ] (4.30)
i=1 i=1

As a simple example, consider a signal + noise decomposition as shown in

equation (4.31), where n is the noise variance.
 
1 0 0 0
0 1 0 0
[T ] = [T1 ] + n [TN ] = [T1 ] + n 
0 0 1

0
0 0 0 1
  (4.31)
1 0 0 0
0 0 0 0
⇔ [M ] = [M1 ] + n [MN ] = [M1 ] + n 
0

0 0 0
0 0 0 0

As a point of caution we note again that n may be caused by depolarisation (by

a Lambertian surface, for example) as well as by system noise effects. Only in
rank-3 scattering (reciprocal backscattering, for example) can we estimate the
system noise n from the difference between crosspolarisation channels as the
smallest eigenvalue of the HV/VH 2 × 2 coherency matrix shown in equation
(4.32) (Hajnsek, 2001):

1 + ∗
, + ∗
,
n̂ = SHV SHV + SVH SVH
2
 
+ ∗
, + ∗
,2 + ∗
,+ ∗
,
− SHV SHV − SVH SVH + 4 SHV SVH SVH SHV (4.32)

This can then be used as a method of noise filtering, for example, by inverting
equation (4.31):

[T1 ] = [T ] − n̂ [TN ] ⇔ [M1 ] = [M ] − n̂ [MN ] (4.33)

190 Decomposition theorems

In this section we look at generalizations and extensions of this idea. Such rep-
resentations are called incoherent decomposition (ICDs), because the addition
of coherency matrices implies addition of power, which in turn implies there is
no phase coherence between the elements. When this is not the case we must
employ coherent decomposition theorems (CDTs)—a description of which was
given in the previous section. There are two key elements to any ICD:
1. The choice of order R in equation (4.30). In principle we are free to
choose any order, but in practice there are several constraining factors.
The first is recognition of the fact that each composite matrix must have
at least one free parameter to describe its form and structure (n for the
noise element in equation (4.31), for example). On the other hand we
have seen that a general coherency or Mueller matrix has at most sixteen
free parameters, and so this means that R ≤ 16. Further restrictions apply
from various symmetry constraints. The most important of these is the
reciprocity theorem in backscatter, for which the coherency matrix has
rank 3 and only nine free parameters. In most backscatter applications,
therefore, R ≤ 9 (a notable exception being the treatment of backscat-
ter Faraday rotation, which is nonreciprocal, and has R ≤ 10). Other
symmetries may apply on a case-by-case basis: for example, backscat-
ter reflection symmetry (a common case in microwave remote sensing),
for which R ≤ 6, and the most severe case, being backscatter azimuth
symmetry with R ≤ 2.
2. The nature of the component elements of the expansion Ti . This leads to
a further general partitioning of the decomposition as follows. The idea is
often to force one or more components of the ICD to correspond to a rank-
1 coherency matrix (or equivalently a Mueller matrix that corresponds to
a single amplitude [S] matrix). Consequently, we can write the general
ICD in the form shown in equation (4.34), where P ≤ R is the number
of rank-1 components:

!
P !
R
  !
P !
R
 
[T ] = [Ti ] + Tj ⇔ [M ] = [Mi ] + Mj (4.34)
i=1 j=Q i=1 j=Q

Each rank-1 component requires up to a maximum of seven (4 × 4

coherency) or five (3 × 3 coherency) parameters. If each 5 component
actually requires Qi parameters, then it follows that Q = Pi=1 Qi + 1.
A natural extension of this concept is to consider each element of the expansion
a reflection symmetric depolariser. This then represents the most general case
found in the literature. There are two important special cases of this approach,
as follows.

4.2.1 The Huynen decomposition: rank 1

scattering + noise decompositions
It is interesting to speculate on the existence of a generalization of the signal +
noise decomposition in equation (4.31). In particular we now seek decompo-
sition into a single rank 1 component (the ‘signal’) [T1 ] plus a remainder [TN ]
 4.2 Incoherent decomposition theorems 191

(considered as ‘noise’), as shown in equation (4.35):

[T ] = [T1 ] + [TN ] ⇔ [M ] = [M1 ] + [MN ] (4.35)
There certainly exist interesting special cases of this type of decomposition, as
first pointed out by Chandrasekhar (1960) in his work on radiative transfer the-
ory. Following this example we consider lateral scattering (scattering through
90◦ bistatic angle) by a random cloud of small spheroids (equation (3.101)) to
obtain a coherency matrix of the form shown in equation (4.36), where Ap is
the ratio of polarizabilities or particle shape parameter. As shown, this can then
be reformulated as a rank-1 scattering mechanism (dipole scattering k D ) plus
a ‘noise’ term, although in this case the ‘noise’ is not thermal but is caused by
depolarisation due to the particle anisotropy (for example, the noise term goes
to zero for spherical particles, Ap = 1).
     
√1 √1 0 0 λ1 0 0 0 √1 − √1 0 0 1
 21 2
   12 2
  
− √ √1 0 0  0 λ2 0 0 √ √1 0 0 1 −1
[T ] = 

2 2 

 2 2  kD = √  
 0 0 1 0  0 0 λ2 0  
 0 0 1 0
 2
0


0 0 0 1 0 0 0 λ2 0 0 0 1 0
 
1 0 0 0
 
0 1 0 0
= (λ1 − λ2 ) k D k D + λ2 
∗T
0 0
 (4.36)
 1 0

0 0 0 1
 
1 0 0 0
 
1 1 0 1 0 0
= (2A2P + 6AP + 7)k D k ∗T
D + (AP − 1)2 



15 15 0 0 1 0
0 0 0 1

Problems arise, however, when we try to generalize this idea to an arbitrary

coherency matrix. The basic issue is that there is an infinite number of ways of
extracting a rank 1 scattering mechanism from [T ]. To see this, consider starting
with an arbitrary mechanism w. We can then always write the decomposition
into a single mechanism plus remainder, as shown in equation (4.37):
[T ] = tw ww∗T + [TN ] tw = w∗T [T ] w (4.37)
Some authors (most notably in the Huynen decomposition (Huynen, 1970,
1987)) have proposed solving this problem by insisting that the remainder term
satisfy some invariance properties under a change of polarisation base. This then
constrains the choice of w for the single mechanism. To see how this works we
begin by noting the importance of the null-space of the remainder matrix. By
definition, [TN ] lies in a subspace of the full polarisation space. Hence there
always exist vectors v, spanning the null-space of the matrix, defined as shown
in equation (4.38):
[TN ] v = 0 (4.38)
Now, insistence that the remainder matrix [TN ] maintains its form under a
unitary change of base (that its null space is unchanged) requires that the null
space v is constrained to be an eigenvector of the change of basis matrix, as
shown in equation (4.39):
[UB ] [TN ] [UB ]∗T v = 0 ⇒ [UB ] v = λv (4.39)
192 Decomposition theorems

Huynen, for example, considered reciprocal backscatter problems (a rank-3

coherency matrix) and insisted that the remainder was invariant to rotations
about the line of sight through an angle θ; that is, he assumed an explicit form
for the change of base matrix, as shown in equation (4.40):
       
1 0 0 1 0 0
1 1
[UB ] = 0 cos θ − sin θ  ⇒ v = 0 , √ 1 , √  1  (4.40)
0 sin θ cos θ 0 2 i 2 −i

It follows that there are still three choices for the null space, given by the three
eigenvectors of the change of basis matrix. Huynen selected the first on physical
grounds and suggested the following general form, which now bears his name:
the Huynen decomposition:
   
t11 t12 t13 0 0 0
[T ] = [T1 ] + [TN ] = t12
∗ t22 − n22 t23 − n23  + 0 n22 n23 
∗
t13 ∗ − n∗
t23 t33 − n33 0 n∗23 n33
23
(4.41)

We clearly see the null space in the remainder term corresponding to the first
eigenvector in equation (4.40), and the elements nij are selected so as to force
the first matrix to have rank 1. However, by choosing one of the other eigenvec-
tors in equation (4.40) for the null space we would produce a different rank 1
matrix in the expansion, and hence we see that although this approach does con-
strain selection of the rank-1 component it does not provide a unique solution.
This approach does, however, raise the issue of orthogonal spaces and their
importance in ICDs. We now turn to consider an alternative formulation, which
solves the uniqueness problem by considering a general eigenvalue expansion.

4.2.2 The Cloude–Pottier decomposition

The discussion in the previous section pointed out several uniqueness problems
in attempts to extract a single ‘dominant’ rank-1 coherency matrix from a gen-
eral depolarising system. Paradoxically it is easier to expand any depolarising
system as a sum of four rank-1 components, as shown in equation (4.42):

[T ] = t1 [T1 ] + t2 [T2 ] + t3 [T3 ] + t4 [T4 ]

(4.42)
⇔ [M ] = m1 [M1 ] + m2 [M2 ] + m3 [M3 ] + m4 [M4 ]

Cloude and Pottier first proposed this idea in Cloude (1996), and the decom-
position therefore bears their name. Proof of the uniqueness of this expansion
follows directly from the Hermitian nature of the coherency matrix, which can
be expanded in terms of its orthogonal eigenvectors and real eigenvalues, as
shown in equation (4.43):
8
!
4 λ1 ≥ λ2 ≥ λ3 ≥ λ4 ≥ 0 ∈ 
[T ] = λi ei e∗T
i (4.43)
i=1 e∗T
i ej = 0

There are four important ideas stemming from this approach.

 4.2 Incoherent decomposition theorems 193

4.2.2.1 Eigenvector decomposition: dominant scattering mechanisms

We see that the decomposition in equation (4.42) is unique if we associate
ti = λi [Ti ] = ei e∗T
i . It follows that if we want to identify a ‘dominant’
and unique rank-1 scattering mechanism we should choose the eigenvector
corresponding to the maximum eigenvalue of the coherency matrix (Cloude,
1989). In this way we can finally write a basis invariant and unique expansion
into a rank-1 plus noise decomposition, as shown in equation (4.44):

!
4
[T ] = λ1 [T1 ] + [TN ] [TN ] = λi ei e∗T
i (4.44)
i=2

In this case the null space of [TN ] is defined by the dominant eigenvector e1 ,
and by definition this remains invariant to all unitary transformations of the
problem, and not just the rotation invariance required in the Huynen approach.
We note that this approach also has the advantage of automatically scaling with
varying rank of the coherency matrix. In backscatter reciprocity, for example,
the smallest eigenvalue is always zero, and the most general expansion is there-
fore provided in terms of three components, as shown in equation (4.45). The
extension to rank 2 and rank 1 depolarising systems follows immediately.

[T ] = t1 [T1 ] + t2 [T2 ] + t3 [T3 ] ⇔ [M ] = m1 [M1 ] + m2 [M2 ] + m3 [M3 ]

(4.45)

4.2.2.2 Eigenvector secomposition and contrast optimization

This eigenvector expansion has several useful applications other than identify-
ing the dominant rank-1 scattering process. For example, a common problem
in radar imaging is the enhancement of contrast between two scatterers (Novak,
1989). In this case we consider two depolarising systems with corresponding
coherency matrices [TA ] and [TB ], and wish to find the single scattering mech-
anism w that maximizes the contrast between the two systems. The contrast
can be defined as a ratio of intensities, given by the ratio of Hermitian forms
Q, as shown in equation (4.46). Optimization of contrast can then be formu-
lated using the Lagrangian shown, and the single scattering mechanism that
maximizes this ratio is obtained as an eigenvector of the product of matrices,
as shown in equation (4.46):

w∗T [TA ] w max Q

Q= −→ L = w∗T [TA ] w − λ(w∗T [TB ] w − 1)
w∗T [TB ] w
(4.46)
∂L
⇒ = [TA ] w − λ [TB ] w = 0 ⇒ [TB ]−1 [TA ] w = λw
∂w

In this case, therefore, the best single mechanism to isolate in each matrix
is w, and not the individual eigenvectors ei . In this case we make use of the
following ICD:

[TA ] = tA ww∗T + [TNA ] tA = w∗T [TA ] w

(4.47)
[TB ] = tB ww∗T + [TNB ] tB = w∗T [TB ] w
194 Decomposition theorems

4.2.2.3 Eigenvector decomposition and CFAR detection

As a second important example of eigenvector decompositions we consider the
problem of radar backscatter detection in the presence of depolarising ‘clutter’;
that is, unwanted background signal (Ioannidis, 1979; Novak, 1989; Wanielik,
1992). In this case we make a measurement of the scattering matrix [S], and
wish to set a threshold in order that the false alarm rate be constant (that is
the rate of error associated with making a detection when actually no target
is present) in the presence of depolarising Gaussian clutter (see Appendix 3),
forming a so-called constant-false-alarm-rate, or CFAR, detector. In this case
we first vectorize a sample into the form k as shown in equation (4.48), and
then assume for the clutter background alone that the vector has a multivariate
normal distribution as shown. In this case the CFAR implementation involves
calculating a metric Q as shown, and deciding on detection with constant errors
based on the setting of a threshold ‘x’. Note that the metric Q involves the inverse
coherency matrix for the background depolarising system, and so assumes we
have some knowledge of this matrix, either by measurement or modelling.
 
a
vectorise k = b ⇒ p(k) = 1 −1
e−k [T ] k
T
[S] −− −−−→
c π det([T ])
3

⇒ Q = k ∗T [T ]−1 k > x (4.48)

It is instructive to formulate this detection process in terms of the eigenvector
decomposition of the coherency matrix. Equation (4.49) shows how Q can be
expressed as a function of the eigenvalues and eigenvectors of [T ].
 
 2 1 1 1
Q = k  w∗T e1 e∗T + e e ∗T
+ e e ∗T
w
λ1 1
λ2 2 2 λ3 3 3
 2    
k   2 k 2  ∗T 2 k 2  ∗T 2
= w .e  +
∗T w .e  + w .e  (4.49)
1 2 3
λ1 λ2 λ3
We now see an interesting case where the optimum solution lies in the smallest
eigenvalue. The best detection arises when k (with corresponding mechanism
w) is orthogonal to the first and second eigenvalues (w∗T ei = 0) and lies parallel
to the smallest eigenvector w∗T e3 ≈ 1. In this case (λ3 small) the contrast
between the ‘target’ and ‘clutter’ is maximized, and the Q factor is large even
when the absolute value of k is small, leading to so-called sub-clutter visibility
of the target. In this case it is best to characterize the depolarising ‘clutter’ using
the following ICD:
[T ] = λmin e3 e∗T
3 + [TN ] (4.50)

4.2.2.4 The entropy/alpha decomposition

The three previous applications have all sought to identify a single eigenvalue/
eigenvector to solve an optimization problem. However, we have seen in
Section 2.4 that we can also identify an average scattering mechanism from the
eigenvectors and associate it with a degree of disorder, the entropy. This gives
us an invariant two-parameter characterization of the depolariser represented
by [T ]. Furthermore, we have seen that we can then plot such paired values in
a plane: the entropy/alpha diagram (see Section 2.4.2.4).
 4.2 Incoherent decomposition theorems 195

This can then be used to classify different types of scattering behaviour.

This approach has been applied, for example, as an unsupervised classification
method in imaging radar polarimetry (Pottier, 1997; Cloude, 1997a), and as
a physics based pre-processor for more general statistical classification tech-
niques based, for example, on the Wishart distribution (Ferro-Famil, 2001; Lee,
1999, 2008).
These examples all illustrate the importance of considering the full eigen-
value spectrum of the coherency matrix, and how in some applications there is
useful information even in the smaller eigenvalues and eigenvectors. Another
way to systematically exploit this information is to consider an approach to
ICDs based on physical scattering models, as we now consider.

4.2.3 Model-based incoherent decompositions

One important alternative approach to ICDs is to use physical models of scat-
tering depolarisation to determine the number and parameterization of each
component (van Zyl, 1989; Freeman, 1998; Dong, 1998; Yamaguchi, 2005).  D
S V
The general starting point for this approach is to identify the main components
to any backscattering scenario, as summarized schematically in Figure 4.6.
This model in its simplest form involves a generic two-layer approach, with a Fig. 4.6 Schematic representation of the var-
volume layer of particles (vegetation, snow, and so on) above a non-penetrable ious scattering contributions used in model-
based decompositions
boundary or surface. The total average backscatter is then determined from a
3 × 3 coherency matrix composed of three main elements, as shown in equa-
tion (4.51). The first component—S in Figure 4.6—is direct backscatter from
the underlying surface, recognising that its scattering behaviour is modified by
propagation through the top layer, which may cause attenuation of the surface
response but can also act to distort the polarisation behaviour of the surface.
For relatively smooth surfaces the small perturbation or Bragg scattering model
is often used to model this component. (Although alternatives such as the X-
Bragg or IEM can be easily included (see equation 3.42), they always incur the
expense of adding more parameters and hence make inversion more difficult.)
[T ] = [TS ] + [TD ] + [TV ]
 
cos2 αs sin αs cos αs eiφs 0
= mS sin αs cos αs e−iφs sin2 αs 0
0 0 0
 
cos2 αd sin αd cos αd eiφd 0 (4.51)
+ mD sin αd cos αd e−iφd sin2 αd 0
0 0 0
 2 
2AP + 6AP + 7 0 0
+ mV  0 (AP − 1)2 −i4κ(AP − 1)2 
0 i4κ(AP − 1) 2 (AP − 1)2
As shown in equation (3.36), Bragg scattering acts as a strong polariser with
zero depolarisation and (for a flat surface) zero crosspolarisation. Thus Bragg
scattering contributes a rank-1 coherency matrix with three unknown param-
eters, as shown in equation (4.51), where α s < π/4 depends on dielectric
constant and angle of incidence. The second component—D in Figure 4.6—
involves a multiple scattering interaction between the surface reflection and
196 Decomposition theorems

volume-scattering elements that return the reflected signal back to the observer.
While this combination can often be complex and a source of wave depolari-
sation, it is often assumed dominated by a double specular reflection process
or simple dihedral reflection from dielectrics, rather than a reflection/scattering
combination. For this to occur the volume needs to be populated not only by
Dihedral/double bounce
scattering
small particles (leaves, branches, and so on) but also by some electrically large
scatterers (vertical tree-trunks, for example), in which case the reflection from
the surface (modelled by the Fresnel equations, modified by the Rayleigh coef-
ficient to account for surface roughness; see equation (3.31)) is followed by
a second Fresnel reflection from the surface of the large scatterer, as shown
Fig. 4.7 Geometry of dihedral backscattering schematically in Figure 4.7. This reflection can be written in terms of the angle
mechanism of incidence using the constraint that the final angle of reflection must be 180◦
for backscatter, shown in equation (4.52):


 cos θ − εr1 − sin2 θ

 RH 1 =

 cos θ + εr1 − sin2 θ







 εr1 cos θ − εr1 − sin2 θ


  RV 1 =


RH 1 RH 2 0 εr1 cos θ + εr1 − sin2 θ
[SD ] ∝ → (4.52)
0 −RV 1 RV 2 


 sin θ − εr2 − cos2 θ

RH 2 =



 sin θ + εr2 − cos2 θ





 εr2 sin θ − εr2 − cos2 θ


RV 2 =
εr2 sin θ + εr2 − cos2 θ

In this case the dihedral component can be modelled as a rank-1 polarising

element with three parameters, as shown in equation (4.51), with αd > π/4
depending now on the angle of incidence and the two dielectric constants of
surface and reflector.
The third component—V in Figure 4.6—is direct volume scattering from
the top layer itself. By assuming this layer has azimuthal symmetry it can
be modelled as a random cloud of spheroidal particles with shape parameter
Ap , and chirality κ, the coherency matrix for which was derived in equation
(3.97). This contribution acts as a three-parameter depolariser with coherency
matrix, as shown in equation 4.51. An obvious extension is to include multiple
scattering as discussed in Section 3.4.3.
In total, this model generates (at least) nine parameters for only six observa-
tions or measurements (the three diagonal elements of the coherency matrix plus
one off-diagonal complex and one imaginary term). This is important because
we often try to use such models for inversion or parameter estimation; that is,
to compare the predictions of the model against the measurement and adjust the
parameters until the difference between the two is minimized. We then take the
parameters used for the minimization as estimates of the true values. However,
when the number of unknowns exceeds the number of observations we face
problems of uniqueness; that is, many different combinations of the parameters
could achieve a match against the limited measurements, and so we are unable
to select the true solution. It is therefore of interest, in such model-based ICDs,
to try to more closely match the number of model parameters to the number of
 4.2 Incoherent decomposition theorems 197

measurements. There are, of course, many ways to do this, but several common
assumptions emerge in practice.

4.2.4 The Freeman–Durden decomposition

One common set of approximations appears in the Freeman–Durdan decompo-
sition, as follows (Freeman, 1998). The first idea is to ignore volume chirality
(κ = 0), and also to select a fixed value for Ap —the particle shape in the volume
term. This is often assumed to correspond to a cloud of prolate spheroids or
dipoles (Ap  1). This then reduces the unknowns to seven. The second feature
is then to assume that in practice one or other of the surface or dihedral responses
dominates, and so we can set the minor mechanism to have a known α value
(αs = 0 if the dihedral is dominant, or αs = π/2 if the surface is dominant)
without too much loss of accuracy. This again reduces the number of unknown
model parameters by two. In this way we obtain a balanced system with five
model parameters and five unknowns, which can now be inverted. In summary,
the Freeman–Durden decomposition—a special case of equation (4.51), and
an example of an ICD into two rank-1 components plus one rank-3—can be
written in terms of the coherency matrix, as shown in equation (4.53):
 
t11 t12 0
[T ] = t12∗ t
22 0 
0 0 t33
   
2 0 0 cos2 αs cos αs sin αs eiφs 0
= mv 0 1 0 + ms cos αs sin αs e−iφs sin2 αs 0
0 0 1 0 0 0
 
cos2 αd cos αd sin αd eiφd 0 /
 −iφ  αs = 0, φs = 0
+ md cos αd sin αd e d sin αd
2 either
0 −−−−→
αd = 0, φd = 0
0 0 0
(4.53)

As mentioned, the main reason for adopting this simplified form is to enable
inversion of the model directly from data, as follows. The first step is to use
the HV channel to directly estimate the volume scattering component mv , as
shown in equation 4.54:

mv = t33 (4.54)

We can then estimate the α parameter, under the hypothesis of a dominant

scattering mechanism, by first calculating its apparent value from the coherency
matrix (here called the Freeman alpha value, α F ) and then assigning the model
parameters to α F according to a π/4 threshold, as shown in equation (4.55):


 αd = αF , αs = 0







 t22 − t33 π

⇒ md = sin2 α , ms = t11 − 2t33 − md cos αd if αF >
2

t22 − t33 d 4
tan αF = ⇒ (4.55)
∗
t12 .t12 
 π

αs = αF , αd =

 2





 t − 2t π
⇒ ms = 11 33
, md = t22 − t33 − md sin2 αs if αF <
cos2 αs 4
198 Decomposition theorems

Also shown in equation (4.55) are equations used to estimate the scattering
power terms ms and md . Finally, the last parameter, the scattering phase, can
be estimated as shown in equation (4.56):

 π

φd = arg(t12 ), φs = 0 if αF >
4
(4.56)

φs = arg(t12 ), φd = 0 π
if αF <
4

One attractive feature of this model is that it represents a direct decompo-

sition of the total backscattered power into three components, as shown in
equation (4.57):

Ptotal = PS + PD + PV = ms + md + 4mv (4.57)

Note, however, that the model inversion does not guarantee non-negative
estimates of power. In the presence of noise or violations of the underlying
model assumptions, some of the power components could be estimated as
negative—an undesirable feature of such models.

4.2.5 Generalized Freeman–Durden decompositions

The Freeman–Durden approach is clearly not unique, and we can choose to
simplify the general model of equation (4.51) in many different ways. Sev-
eral examples exist in the literature: for example, Freeman modified his own
approach (Freeman 2007) by relaxing the random assumption for the volume
scattering and instead assigning the particles an orientation distribution with
a single parameter to be estimated from the data. A second approach is that
due to Yamaguchi (2005). The main additive features of his technique are to
include non-symmetric features such as scattering from helical type scatter-
ers, as occurs in chiral materials, or to consider more general volume scattering
terms, allowing for a relaxation of the azimuthal symmetry assumption to reflec-
tion symmetry. Rather than present an exhaustive survey of these methods, we
return to the key idea of target orthogonality to show how the eigenvector
approach can be used to enhance such models.
In the Freeman approach nothing was mentioned about orthogonality of the
surface and dihedral component, but we have seen that for any surface mecha-
nism with parameter α s the orthogonal scattering mechanism has α = π/2−αs ,
which lies in the dihedral regime. Therefore, another approach to reducing the
unknowns in equation (4.51) is to postulate that the surface and dihedral mecha-
nisms are orthogonal. This allows us to express αd in terms of αs and thus reduce
the parameter count by one. When added to the relaxed Freeman assumption
of azimuthal symmetric scattering for the volume term and the loss of one
phase angle, we again obtain a balanced system with five parameters and five
measurements. This model—which we can call a hybrid Freeman/eigenvalue
technique—has the compact form shown in equation (4.58), where α is now
determined as predominantly surface or dihedral scattering, depending on the
dominant eigenvalue of a 2 × 2 sub-matrix as shown, where FP = 2 corresponds
 4.2 Incoherent decomposition theorems 199

to the Freeman–Durden model.

     
cos α 0 1 sin α 0 1 Fp 0 0
     
[T ] = ms − sin α  . cos α − sin α 0 + md cos α  . sin α cos α 0 + mv  0 1 0
0 0 0 0 1
   
ms cos2 α + md sin α cos α sin α(md − ms ) 0
2
Fp mv 0 0
   
=  cos α sin α(md − ms ) md cos2 α + ms sin2 α 0 +  0 mv 0 
0 0 0 0 0 mv
 
Fp 0 0
 
⇒ [T ]SD = [T ] − mv  0 1 0
0 0 1
     
cos α sin α 0 ms 0 0 cos α − sin α 0
     
= − sin α cos α 0 .  0 md 0 .  sin α cos α 0 (4.58)
0 0 1 0 0 0 0 0 1

This model can then be inverted by first calculating ms and md as eigenvalues

of the rank 2 matrix TSD , and then using these to obtain estimates of the α
parameter by first parameterising the eigenvector as (1,e2 )T , then solving for
the complex number e2 , and finally normalizing the eigenvector by 1 + |e2 |2
to obtain α, as shown in equation (4.59):
mv = t33

(t11 + t22 − (Fp + 1)t33 ) ± (t11 − t22 − (Fp − 1)t33 )2 + 4 |t12 |2
md ,s =
2
) 
 2 *− 12
 t12 
αd ,s = cos−1  1 +  

 (4.59)
t22 − t33 − md ,s

Note also that this approach provides a means for avoiding negative powers by
keeping only the non-negative eigenvalue spectrum of TSD . Also, if we assume
one of ms , md is always zero then we can estimate the volume parameter Fp .
We can also use this model to estimate the ratio of surface-to-volume scattering
components, termed µ, which, as we shall see in Chapter 7, is an important
parameter in the development of polarimetric interferometry. According to the
model, when we select the crosspolarisation channel HV, the surface-to-volume
scattering ratio µ is by definition zero, as the surface components are all con-
strained into the rank-2 upper portion of the coherency matrix. On the other
hand, to find the surface-to-volume ratio for the largest surface component we
can proceed as follows. First we select the maximum ‘surface’ component from
the maximum of the pair ms and md as defined in equation (4.59). (Note that we
are including both direct surface and dihedral returns in the ‘surface’ compo-
nents to distinguish them from the pure volume scattering component.) From
this we can then calculate the alpha parameter of the maximum component, as
shown in equation (4.60):
"  1
 2 − 2
  
mmax = max(md , ms ) ⇒ αmax = cos−1  1 + 
t12    
t22 − t33 − mmax 

(4.60)
200 Decomposition theorems

If we then select a scattering mechanism w based on this alpha parameter, we

can obtain an expression for the desired ratio, as shown in equation (4.61):
 
cos αmax  
[T ] = mmax − sin αmax  . cos αmax − sin αmax 0
0
 
sin αmax  
+ mmin cos αmax  . sin αmax cos αmax 0
0
 
Fp 0 0
+ mv  0 1 0
0 0 1
 
cos αmax
w =  sin αmax  ⇒ w∗T [T ] w = mmax + mv (1 + (Fp − 1) cos2 αmax )
0
surface mmax 1
⇒µ= = (4.61)
volume mv (1 + (Fp − 1) cos2 αmax )

Notice that this ratio is not the maximum surface-to-volume ratio. This can
always be obtained by using a w calculated as a generalized eigenvector of
the matrix [TV ]−1 [TSD ] (see equation (4.46)). The general solution to this is
complicated, but if we make the assumption that there is one dominant surface
component (either ms or md ), then we can obtain a direct estimate of the maxi-
mum surface-to-volume ratio by finding the largest eigenvalue of [TV ]−1 [Tmax ],
as shown in equation (4.62):
 
 cos2 αmax cos αmax sin αmax 
mmax  − µ
 Fp Fp =0
 
mv cos α sin αmax − µ 
max sin αmax
2
(4.62)
 
mmax 1
⇒ µmax = sin2 αmax + cos2 αmax
mv Fp

It is also interesting to note that we can also define another small value of
surface-to-volume scattering (in addition to the HV channel) by selecting the w
vector orthogonal to the maximum surface component. In this case we obtain
a µ value, given as shown in equation (4.63):
 
− sin αmax
w =  cos αmax  ⇒ w∗T [T ] w = mmin + mv (1 + (Fp − 1) sin2 αmax )
0
surface mmin
⇒ µmin = = (4.63)
volume mv (1 + (Fp − 1) sin2 αmax )

In the introduction to the general model of equation (4.51) we mentioned that

the surface response is actually modified by propagation through the upper
medium. So far, however, we have ignored such propagation effects, on the
 4.2 Incoherent decomposition theorems 201

assumption that the volume is random and hence acts as a scalar attenuation of
all polarisations equally, which has no effect on the alpha parameters of surface
and dihedral components. However, in some cases this assumption is not true,
and leads us to yet another class of decompositions—but this time based on a
multiplicative expansion. We now turn to consider such ideas.

4.2.6 Propagation distortions in model-based

decompositions
To consider the effects of wave propagation on depolarising systems (Azzam,
1978), we begin by considering the simplest case of an homogeneous propaga-
tion channel (see Section 1.2.4), in which the orthogonal eigenpolarisations for
propagation in the medium are aligned with the [S] matrix measurement basis
a,b. In this case there is by definition no cross-coupling between eigenstates,
and so the propagation matrix is diagonal. The effect of wave propagation into
the medium followed by scattering, and then propagation out of the medium,
can then be represented as a matrix product, as shown in equation (4.64):
 −iβ z     −iβ z 
e a 0 Saa Sab e a 0
[S]observed = . .
0 e−iβb z Sba Sbb 0 e−iβb z
/ (4.64)
βa = βo na = βo na − iκa
βb = βo nb = βo nb − iκb

where βo = 2π λ is the free space wavenumber. Note that βa and βb are in

general complex, relating to both phase shifts and attenuation by the medium.
It is convenient, in what follows, to factor the propagation matrix by extracting
the mean propagation constant, as shown in equation (4.65):
 
exp(−iβa z) 0
0 exp(−iβb z)
  
  (βa −βb )
−i(βa + βb ) exp −i 2 z 
0
 (4.65)
= exp z
2 0 exp i (βa −β
2
b)
z

The first step in understanding the effects of propagation on scattering entropy

is then to vectorize this matrix expression. The best way to do this is in the
lexicographic basis, when we obtain the form shown in equation (4.66):
  
exp (νz) 0 0 0 Saa
 0 1 0 0   
k L (z) = e−i(βa +βb )z   Sab  = [PL ].k
 0 0 1 0  Sba  L

0 0 0 exp (−νz) Sbb

υ = −i(βa − βb ) (4.66)

Note that the crosspolarised channels are influenced by the mean propagation
constant, while the copolar channels are also influenced by the differential
propagation. We can combine these effects into a single 4 × 4 matrix [PL ] as
202 Decomposition theorems

shown. The final step is then to express the matrix [PL ] in the Pauli base, so
we can finally relate propagation to the coherency matrix formulation. This is
shown in equation (4.67), where we see we can express the matrix in terms of
elementary hyperbolic functions.
   
1 0 0 1 1 1 0 0
1 1 0 0 −1 0 0 1 −i
[PP ] = 

 [PL ]  
2 0 1 1 0 0 0 1 i 
0 i −i 0 1 −1 0 0
 
cosh νz sinh νz 0 0
 sinh νz cosh νz 0 0
= e−i(βa +βb )z 
 0
 (4.67)
0 1 0
0 0 0 1

As an important special case, if the medium is random (if it has azimuthal

symmetry), then the differential propagation constant is zero, and the propaga-
tion matrix [PP ] reduces to a multiple of the identity matrix. This is the form
implicit in the Freeman decomposition. In this case the surface component mS
is attenuated by the mean extinction in the medium, but its polarimetry remains
unchanged. A second important example is when the medium is lossless (zero
extinction) but acts as a retarder (with differential phase shifts). In this case ν
is purely imaginary, and the propagation matrix has the special form shown in
equation (4.68):
 
cosh iνz sinh iνz 0 0
 sinh iνz cosh iνz 0 0
[PP ] = e−i(βa +βb )z 
 0

0 1 0
0 0 0 1
 
cos(νz) i sin(νz) 0 0

−i(βa +βb )z i sin(νz) cos(νz) 0 0
=e  (4.68)
 0 0 1 0
0 0 0 1

In general, however, the coherency matrix of the scatterers beneath the

propagation channel will have a distorted form, as shown in equation (4.69):
 
[T (z)] = [PP ] [T ] [PP ]∗T = e−2(κb +κa )z [P(τ )] [T ] P τ ∗
 z (4.69)
τ = υz = (κb − κa ) − iβ0 (na − nb )
cos θ0

where z is now the slant range coordinate, related to z , the vertical coordinate,
by θ o , the angle of incidence (see Figure 3.38). We note again the attenuation
of amplitude by mean wave extinction, but also note that since P is not gen-
erally diagonal then there results some distortion to the apparent polarimetric
parameters of [T ].
To illustrate this we consider the effects on the apparent scattering param-
eter α for a scatterer viewed through a medium with zero retardence but total
 4.2 Incoherent decomposition theorems 203

Alpha distortion due to differential extinction

90

80

70
Observed alpha angle

60

50
9 bB
40

30 6 dB
3 dB
0 dB
20

10

0 Fig. 4.8 Distortion of the alpha parameter

0 10 20 30 40 50 60 70 80 90
due to differential extinction in the volume
True alpha angle layer

differential extinction . By changing variable we can then express (from equa-

tion (4.67)) the propagation transformation directly in terms of the differential
extinction in dB and in terms of the tangent of the new alpha parameter, as
shown in equation (4.70) (see equation 3.11):

     
cos α −s cosh A sinh A cos α
=e .
sin α sinh A cosh A sin α
(4.70)
sin α + cos α. tanh(0.1151 ∗ )
⇒ tan α =
cos α + sin α. tanh(0.1151 ∗ )

Figure 4.8 shows how this distortion behaves for varying differential extinction.
We see that for  = 0 dB we have zero distortion as expected (this is the
basic assumption of the Freeman decomposition). However, as  increases
we see that the range of apparent alpha becomes compressed until, when the
differential extinction becomes very large, we obtain an apparent alpha of 45◦
for all scatterers. This corresponds to all scattering being filtered by a polariser
with an orientation given by the dominant eigenpolarisation.
Finally, we consider the more general case when the eigenpolarisations are no
longer matched to the backscattering matrix basis. In this case we must modify
equation (4.63) to include a unitary congruent transformation by a 2 × 2 unitary
matrix, as shown in equation (4.71):

     −iβ z 
e−iβa z 0 Saa Sab e a 0
[S]observed = [U2 ] . [U2 ]T .
0 e−iβb z Sba Sbb 0 e−iβb z
/
βa = βo na − iκa
(4.71)
βb = βo nb − iκb
204 Decomposition theorems

This leads to a modified vectorization of the problem, as shown in equation

(4.72):
 
Saa
  Sab 
k L (z) = e−i(βa +βb )z [PL ] U4B  
Sba  = [PL ].k L
B
(4.72)
Sbb
 
where the 4 × 4 unitary matrix U4B is a function of two parameters: α w and
δ, defining the coordinates of the eigenpolarisation on the Poincaré sphere. It
can be obtained by direct expansion, as shown in equation (4.73):
 
cos αw − sin αw e−iδ 0 1
[U2 ] = ⇒ [U2 ] ⊗ [U2 ] = U4L
sin αw e iδ cos αw
 
cos2 αw − cos αw sin αw e−iδ − cos αw sin αw e−iδ sin2 αw e−i2δ
0 1 cos α sin α eiδ cos2 αw − sin2 αw − cos αw sin αw e−iδ 
 w w 
U4L =  
cos αw sin αw e iδ − sin αw
2
cos αw
2 − cos αw sin αw e−iδ 
sin2 αw ei2δ cos αw sin αw eiδ cos αw sin αw eiδ cos2 αw
(4.73)

Now converting to the Pauli basis we obtain the following form for the unitary
matrix, as shown in equation (4.74):
   
1 0 0 1 1 1 0 0
0 1 1 1 0 0 −1 0 1 0 0 1 −i
  L  
U4P =   U  
2 0 1 1 0  4 0 0 1 i 
0 i −i 0 1 −1 0 0
 2 
cos αw + sin αw cos 2δ
2
i sin2 αw sin 2δ 2i sin αw cos αw sin δ 0
 −i 2
α 2 α − 2
α −2 sin αw cos αw cos δ 0
 sin w sin 2δ cos w sin w cos 2δ 
= 
 2i sin αw cos αw sin δ 2 sin αw cos αw cos δ cos2 αw − sin2 αw 0
0 0 0 1
(4.74)

Note, for example, when δ = 0—when we consider only linear eigen-

polarisations—then this matrix reduces to the standard rotation matrix, as shown
in equation (4.75):
 
cos2 αw + sin2 αw 0 0 0
 P  0 cos 2 α − sin 2 α −2 sin αw cos αw 0
U4 = 

w w 
0 2 sin αw cos αw cos2 αw − sin2 αw 0
0 0 0 1
 
1 0 0 0
0 cos 2αw − sin 2αw 0
=0 sin 2αw
 (4.75)
cos 2αw 0
0 0 0 1

We can represent the distorting effects of wave propagation through an arbi-

trary (homogeneous) layer on the backscattering coherency matrix, as shown
 4.2 Incoherent decomposition theorems 205

in equation (4.76):

   ∗T  
[T (z)] = e−(κb +κa )z [P(τ )] U4P (αw , δ) [T ] U4P (αw , δ) P(τ ∗ )
 z (4.76)
τ = νz = (κb − κa ) − iβo na − nb
cos θo

Finally we turn to consider an important practical example of the utility of the

vectorization scheme: to analyse the effects of Faraday propagation rotation
distortion on target decomposition. Such effects must be considered, for exam-
ple, in the analysis of low-frequency radar Earth observation data collected
from space when propagation through the ionosphere cannot be ignored (see
Section 1.2.2.3).

4.2.6.1 Scattering vector formulation of Faraday rotation

The scattering vector formulation can also be used to further analyse the effects
of Faraday rotation (see equation (1.58)) (Bickel, 1965; Wright, 2003; Freeman,
2004). In this case the scattering matrix is transformed by the same rotation
matrix on the left and right sides, as shown in equation (4.77):

   
  cos ψ sin ψ SHH SHV cos ψ sin ψ
Sψ = (4.77)
− sin ψ cos ψ SVH SVV − sin ψ cos ψ

By expansion and vectorization in terms of the Pauli matrices this can be written
as a unitary transformation of the scattering vector, as shown in equation (4.78):

  
cos 2ψ 0 0 − sin 2ψ SHH + SVV
1  0 1 0 0  SHH − SVV 
k (ψ) = √  
 
 SHV + SVH  (4.78)
2 0 0 1 0
sin 2ψ 0 0 cos 2ψ SHV − SVH

This relationship can be used to estimate the Faraday rotation by recognising

that for calibrated reciprocal backscattering, SHV –SVH is zero on the right-hand
side of this equation. Hence any observed difference between SHV and SVH
must be due to Faraday effects, and from equation (4.78) it follows that the
Faraday rotation can be found from the ratio shown in equation (4.79):

SHV − SVH
tan 2ψ = (4.79)
SHH + SVV

One problem with practical use of equation (4.79), however, is that it involves
the ratio of two complex quantities on the right and a real quantity on the left,
and hence it is susceptible to data fluctuations and residual calibration errors.
It would be more robust to devise an algorithm that involves the ratio of real
quantities so that averaging may be used. We can use the scattering vector
formulation combined with a 4 × 4 unitary transformation into the circular
basis to solve this as follows.
206 Decomposition theorems

We begin by relating the lexicographic circular basis scattering vector to the

Pauli expansion by a 4 × 4 unitary matrix, as shown in equation (4.80):
     
SLL 0 1 i 0 SHH + SVV
SLR   1  
  = √1  i 0 0  √1 SHH − SVV  = [U4circ ] k (4.80)
SRL  2 i 0 0 −1  2 SHV + SVH 

SRR 0 −1 i 0 SHV − SVH

This can be derived from the circular change of base transformation in equation
(1.162). This can then be used to transform equation (4.80) into the circular basis
by a unitary similarity transformation, as shown in equation (4.81):
     
SLL (ψ) cos 2ψ 0 0 − sin 2ψ SLL
SLR (ψ)  0 1 0 0  SLR 
    [U4circ ]∗T  
SRL (ψ) = [U4circ ]  0 0 1 0  SRL 
SRR (ψ) sin 2ψ 0 0 cos 2ψ SRR
   
1 0 0 0 SLL
0 exp (−i2ψ) 0 0  

=  . SLR  (4.81)
0 0 exp (i2ψ) 0 SRL 
0 0 0 1 SRR

Note, significantly, that the copolarised terms (LL and RR) are not distorted by
Faraday rotation. If again we assume that the underlying scatterer is reciprocal
(SLR = SRL on the right-hand side of equation (4.81)), as in backscatter BSA
problems, then the Faraday rotation angle can simply be found from the phase
of the product of crosspolarised terms, as shown in equation (4.82):

1 ∗
ψ= arg(SRL SLR ) (4.82)
4

which lends itself to averaging in the presence of data fluctuations by estimating

the complex coherence between the cross-channels LR and RL (see Chapter 2
for a general discussion of coherence in polarimetry). Therefore, if the Faraday
distorted linear scattering matrix [S F ] is measured by a sensor, then the average
Faraday rotation component may be obtained directly from equation (4.82), as
shown explicitly in equation (4.83):
; <
F − S F )(S F + S F )∗ )
−2Re((SHV
tan 4ψ =  F 
VH

HH VV
 (4.83)
S + S F 2 − S F − S F 2
HH VV HV VH

Note the mathematical similarity to the estimation of orientation using copo-

larised circular channels (equation (4.11)), although the two formulae describe
very different physical phenomena.
In general we have seen that circular polarisation provides the best basis for
a description of orientation effects (both reciprocal and non-reciprocal). Polari-
metric phase and coherence play an important role in such studies. However, we
have seen that in scattering from random media, entropy effects act to destroy
 4.2 Incoherent decomposition theorems 207

such phase information and hence reduce our ability to exploit coherence. There
is, however, a second way to generate phase information from active systems:
by the use of interferometry. Here we shall see that new possibilities arise for
the control of phase, even in random media problems, and this, when combined
with polarimetry, will lead us finally to a general formulation of how best to
exploit the ‘orientation memory’ effect in scattering from random media.
 Introduction to radar
5 interferometry

We saw in Chapters 3 and 4 that the polarisation properties of scattering from

natural media, especially when volume scattering is involved, can lead to sig-
nificant levels of depolarisation (high scattering entropy). This depolarisation
has two characteristic features. Firstly, it can be very high (entropies above
0.9, even for single scattering), which limits the accuracy of polarimetric phase
estimation due to the high speckle associated with high entropy (see Appendix
3 and López-Martínez (2005)). The second and more important issue is that the
level of depolarisation is determined entirely by the structure of the medium
and is independent of the sensor. There is no sensor control of entropy. On
the other hand, radar interferometry, as we shall show, provides a means of
controlling entropy by choice of sensor configuration (baseline). This, when
combined with polarimetry, then gives us the ability to tune a sensor for a par-
ticular environment and to obtain good phase estimates, even in the presence
of strong depolarisation. Before considering the combination of polarimetry
with interferometry, we first review the basic features of single-channel radar
interferometry (Bamler, 1998; Franceschetti, 1999; Kampes, 2006).

5.1 Radar interferometry

We have seen that a scattered single-channel radar signal may be represented
by a complex scalar, representing its amplitude a and phase φ, as shown in
equation (5.1):

s1 = a1 eiφ1 = a1 e−i(β2R1 +φS1 ) = a1 e−i( λ 2R1 +φS1 )

2π
(5.1)

Here we have further decomposed the phase of the signal into two parts—the
first a propagation phase that depends on the distance between the radar and
the scattering point (R1 ), and the second a scattering phase that depends on
the detailed nature of the scattering process. In polarimetry we concentrated
on looking at the changes of scattering phase by combining two signals with
different polarisations but collected at the same point in space, so cancelling
the range phase term, as shown in equation (5.2):
.
s1 = a1 e−i( λ R1 +φS1 )
4π

⇒ s1 s2∗ = a1 a2 ei(φS2 −φS1 ) (5.2)

s2 = a2 e−i( λ R1 +φS2 )
4π

We have seen in the previous chapters how to characterize fluctuations in the

polarimetric phase due to depolarisation in the scattering process. In particular
we have seen that in the presence of random volume scattering from anisotropic
 5.1 Radar interferometry 209

particles this depolarisation can be very high (especially in the case of a random
volume of extremely prolate spheroids), leading to a noisy polarimetric phase.
It would be useful to be able to counter this depolarisation in some way. Radar
interferometry provides such an option, as we now show.
S1
In radar interferometry we begin again with the reference signal s1 of equation
B
(5.1). However, rather than diversify polarisation for a fixed position, we now db
S2
keep polarisation constant but vary spatial position to collect a second signal
s2 from a displaced point, separated from the first by a spatial baseline B, as 
R2
shown in Figure 5.1. Note that there are inherently two ways to achieve this.
R1
The first—‘single-pass interferometry’—occurs when we use two radars at the
H
same instant in time to collect signals s1 and s2 . Alternatively, it is often more
convenient (and more cost effective) to use a single radar system to collect a P z
signal at position 1 and then move the radar to position 2 at some time later
to collect s2 . In the airborne/space radar context this is called ‘repeat-pass h0
y
interferometry’, and the time difference between the data collections is called
the temporal baseline TB (in seconds), to be contrasted with the spatial baseline Fig. 5.1 Baseline geometry of radar interfer-
B (with units of metres). Note that we have another degree of freedom for ‘single ometry
pass’ sensors in choosing either to transmit and receive from both points (the
dual transmitter mode described in Figure 5.1 and often termed ‘ping-pong’), or
just to transmit from 1 and receive on both 1 and 2 (called the single transmitter
or standard mode). In this case we can avoid the expense and complexity of a
second transmitter. However, in this case the factor of 4π in equation (5.3) is
halved to 2π , since the phase difference now only depends on R1 − R2 rather
than 2R1 − 2R2 . In what follows we consider the dual transmitter ping-pong
mode to illustrate the main form of governing equations.
In either case, when we form the complex product s1 s2∗ for interferometry we
cancel the scattering phase terms (assuming they are the same; we shall see later
how good this approximation is in practice) and keep the geometrical phase. In
effect we now obtain a signal phase which depends only on the difference in
range between the two positions, R = R1 − R2 , as shown in equation (5.3):
.
s1 = a1 e−i( λ 2R1 +φS1 )
2π

⇒ s1 s2∗ = Ae−i( λ R)

4π
(5.3)
s2 = a2 e−i( λ 2R2 +φS2 )
2π

The key radar observable is now the interferometric phase φ, as defined in

equation (5.4):

4π
φ = arg{s1 s2∗ } = − R + 2πN N = 0, ±1, ±2, . . . (5.4)
λ
Note that this interferometric phase is 2π ambiguous; that is, for every shift of
half a wavelength λ/2 in range difference R the phase difference repeats itself.
We therefore face problems if we wish to invert this equation and determine R
from a measurement of phase, as there are multiple solutions. This is called the
‘phase unwrapping problem’ (Bamler, 1998), and its solution requires the use
of additional information, such as use of phase from nearby scattering points
to generate spatial phase gradients and force continuity, or the use of multiple
baselines (data from three or more spatial positions).
In what follows we consider three different types of radar interferometry
used in remote sensing: across-track, differential, and along-track modes.
210 Introduction to radar interferometry

5.1.1 Across-track interferometry

The next key step is to recognize that from a measurement of phase and knowl-
edge of the components of baseline B we can estimate the height of a scattering
point h0 (see Figure 5.1). This phase-to-height conversion was the original
motivation for the development of radar interferometry (Graham, 1974). This
conversion arises directly from the geometry of the triangle PS1 S2 shown in
Figure 5.1. From the cosine law applied to this triangle we can then derive the
following relationship between the height at point P, h0 , the system geometry,
and the interferometric phase, as shown in equation (5.6):
R22 − R21 − B2
sin (θ − δb ) = ⇒ h0 = H − R1 cos θ (5.6)
2BR1
From this, knowing B (often to an accuracy of mm), the baseline angle δb , and
the two ranges R1 and R2 , we can obtain an estimate of θ for the point P. This
then enables calculation of height at the point P above some datum H, as shown
in equation (5.6).
When direct measurements are made of range from two positions, this pro-
cess is called radar stereo–grammetry. In this case, the accuracy of the final
height estimate depends primarily on the accuracy of the range difference mea-
surements. It is this range difference parameter that can be measured much more
accurately using interferometry. This consequently leads to a much higher accu-
racy estimate of the angular position, and hence of the height of the scatterer.
To see this, consider setting R2 = R1 + R in equation (5.6), and assuming
that R1  B and R1  R, to obtain the following relationship between phase
and height z = ho .
R2 B2 φλ
B sin (θ − δb ) = R + − ≈ R =
2R1 2R1 4π
(5.7)
λφ
⇒ sin (θ − δb ) = ⇒ z = H − R cos θ
4π B
This gives the phase and position of a single scatterer. In the presence of a
second scatterer we obtain a change of phase obtained from equation (5.7), as
shown in equation (5.8):

φ = 4π λ B cos (θ − δb ) θ
z = R sin θ θ − R cos θ
4π B cos (θ − δb ) 4π B cos (θ − δb )
⇒ φ = R + z (5.8)
λR tan θ λR sin θ
 
4π Bn z 1
= 1+
λR tan θ R cos θ
where Bn is the normal component of the baseline (see Figure 5.7). For scatterers
lying in a plane, z = 0, and the phase gradient can then be related to the normal
component of the baseline Bn as
∂φ 4π Bn
= (5.9)
∂R λR tan θ
This is called the flat-earth component of the phase, and can be used to either
estimate the baseline from the phase gradient over flat terrain or, since generally
 5.1 Radar interferometry 211

z z

R R
ysin
 
zcos
y y Fig. 5.2 Approximate ray-path geometry for
P phase estimation

a high-frequency component of the phase signal, it can be removed by the

process of flat-earth removal (discussed in Section5.1.1). When this is removed
this leaves us with a simple linear relationship between phase and elevation z.
Another informative way to derive this relation—employing a local coordi-
nate system around the point P, and which also leads to a linear relationship
between phase and height of the form φ = βz h0 , where the scale factor β z is
called the interferometric wavenumber—is as follows. We first consider con-
struction of a local coordinate system around the point P, as shown in Figure 5.2.
The y axis represents the local surface tangent, and z the local vertical. Con-
sidering first the y axis, the extra contribution to phase for a point P, separated
locally from the origin by distance y, is approximately ysinθ , as shown in
Figure 5.2.
Using this we can then write the phase for signals scattered from point P at
ends 1 and 2 of a baseline, as shown in equation (5.10a):
.
s1 = a1 e−i λ (R+y sin θ1 )
4π

⇒ s1 s2∗ = Aei (sin θ2 −sin θ1 )

4π
λ (5.10a)
s2 = a2 e−i λ (R+y sin θ2 )
4π

θ1 +θ2 θ2 −θ1 4π θ
s1 s2∗ = Aei y(sin θ2 −sin θ1 ) ) sin( ) y cos(θ )
4π 4π
λ = Aei λ
2y cos( 2 2 ≈ Aei λ

(5.10b)
θ +θ θ −θ
i 4πλθ
s1 s2∗ = Ae −i 4π
λ
z(cos θ2 −cos θ1 )
= Ae i 4π
λ
2z sin( 1 2 2 ) sin( 2 2 1 )
≈ Ae z sin(θ )

(5.10c)

In case the angular difference to the point θ = (θ2 − θ1 ) is small (that is,
B  R), we can further simplify this expression, as shown in equation (5.10b).
Similarly, for points P shifted in the z direction (the right-hand side of Figure
5.2) we have an extra phase term, −z cos θ, which leads to a corresponding
interferometric phase, as shown in equation (5.10c). Combining these two, we
obtain the following expression for the total interferometric phase as a function
of y and z around the point P.

4π θ
φif (y, z) = (y cos θ + z sin θ) (5.11)
λ

Currently we see that the phase varies not only with height of the point P (the
coordinate z) but also for shifts in surface position (y). This somewhat disturbs
our desire to have a clean phase-to-height conversion. We can remove this
problem and obtain our desired result by shifting the frequency of the signal
collected at position 2 before forming the interferogram, using a process called
range spectral filtering, as follows.
212 Introduction to radar interferometry

5.1.1.1 Range spectral filtering

One way to remove the surface (y coordinate) phase dependence in equation
(5.11) is to generate an interferogram, for each frequency f in the spectrum of
signal 1, using a shifted frequency component of signal 2, as shown in equation
(5.12) (Gatelli, 1994):

φ = arg(s1 (f )s2∗ (f + f )) (5.12)

In order to completely cancel the y-dependence of phase in equation (5.11)

for arbitrary position P, we must employ a shift f based on geometry θ,
as shown in equation (5.13). This, then, yields a scale factor β z equal to the
desired vertical sensitivity of the interferometer: namely, ∂φ
∂z .

θ
if f = f
tan θ
2θ 4π θ
⇒ β = β =
tan θ λ tan θ
⇒ φ = φif + β (z cos θ − y sin θ)
4π θ 4π θ (5.13)
⇒φ= (y cos θ + z sin θ) + (z cos θ − y sin θ)
λ λ tan θ
4π θ ∂φ
⇒φ= z= z = βz z
λ sin θ ∂z
4π θ
⇒ βz =
λ sin θ
z This transformation can be given a simple geometrical interpretation if we
s1
employ a two-dimensional wavenumber representation of interferometry as
shown in Figure 5.3. Here we show the complex signal from the first sensor
position as a vector in the β y ,β z plane with magnitude equal to 2π/λ and angle
 
s2 θ to the point P. The second position (before frequency shift) then yields a
 vector with the same length but rotated about the origin by θ, as shown. If we
y consider the y components only, then in order to give s2 the same projection on
cos the y axis as s1 requires a radial shift in wavenumber by β (which is negative
if θ is positive) before forming the interferogram.
Fig. 5.3 Wavenumber space geometrical
Furthermore, this idea can be generalized to the bistatic surface scattering
interpretation of radar interferometry
case, using a concept known as the memory line, first described in Le (1998).
This can be motivated by considering the geometrical representation of a general
bistatic scattering geometry in wavenumber space, as shown in Figure 5.4. Here
we see that a bistatic angle ψ leads to a shift of the point representation in wave-
space in both angle and radial coordinates. To see this, just consider the general

z
r
r1  r2
Real space Wave-space
y


c c
Fig. 5.4 Geometry of generalised bistatic r = 2b cos f=u+
radar in real and wavenumber domains 2 2
 5.1 Radar interferometry 213

c
phase term as shown in Figure 5.5. Now consider forming an interferogram 2|b|cos
2
ib.r i|b|(r1 + r2)
between signals for two different bistatic geometries with corresponding wave- e =e ⇒ |b|
2 
c
space coordinates r1 ,φ 1 and r2 ,φ 2 . In order that the phase difference between |b|
signals has no y-dependence and depends only on height, then equal projections
onto the y axis are required, which in turn requires that equation (5.14) holds: Fig. 5.5 Bistatic triangle construction for
radar interferometry
 
ψ ψ
|β| cos cos θ + =C (5.14)
2 2
Consequently, we can now change any of the three parameters, θ, ψ and β,
B

but as long as the product shown in equation (5.14) remains constant then the
interferometric phase will be a function only of height. Equation (5.14) is the
equation of a line—the so-called memory line of the interferometer—and all
systems lying along this line satisfy our desired phase dependency. We see that
for the special case of backscatter (ψ = 0) this constraint reduces to βcosθ being
constant, which is the origin of the frequency shift introduced in equation (5.13).
B

Returning now to the case of backscatter, we can obtain a useful approxima- u =
R
tion for θ as found in equation (5.13) in the special case R  B (as occurs
for spaceborne radar for example). In Figure 5.6 we show how the angular 
baseline width θ can be approximated from the geometry as θ ≈ B⊥ /R.
Note that this expression does not include the absolute baseline B, but only its
component perpendicular to the range line. This leads us to consider the various
spatial components of the baseline, as follows. Fig. 5.6 Baseline geometry triangle

5.1.1.2 Baseline components

It is often convenient to perform a geometric decomposition of the baseline
vector as shown in Figure 5.7, which includes various important components.
The horizontal (BH ) and vertical (BV ) can be simply defined in terms of the
baseline orientation δb , as shown, while the parallel and perpendicular com-
ponents require knowledge of both the baseline angle and the mean angle
of incidence. The main conclusion from Figures 5.6 and 5.7 is that we can
often approximate the vertical sensitivity of an interferometer in terms of the
baseline-to-wavelength ratio, as shown in equation (5.15):
4π θ 4πB⊥ 4π cos(θ − δb ) B
βz = ≈ = (5.15)
λ sin θ λR0 sin θ R0 sin θ λ
We can therefore adjust the sensitivity of the interferometer by adjusting the
B/λ ratio. This provides us with the ability to ‘tune’ the performance of the

V
S1
B
BH = Bcosdb
H B
BV = Bsindb
S2

B
= Bsinf = Bsin ( p – u + db) = Bcos(u – db)
2
p
B
= Bcosf = Bcos ( – u + db) = Bsin(u – db)
u 2

P Fig. 5.7 Summary of baseline components

214 Introduction to radar interferometry

sensor to different applications—a feature not available in radar polarimetry,

and one which we shall see is very important in the development of polarimetric
interferometry.
An important measure of the sensitivity of the interferometer is the height
corresponding to a phase shift of 180◦ —the π -height. This can be derived from
equation (5.15), as shown in equation (5.16):

π λ sin θ λ R0 sin θ
hπ = = ≈ (5.16)
βz 4θ 4 B⊥
Note that since Ro  B, this height is much greater than a wavelength. We
shall see later how in differential interferometry we can obtain much higher
sensitivities when the π -height itself reduces to λ/4. Before leaving this topic
we highlight three important issues.
1. The frequency shift used in equation (5.12) is actually a function of the
local angle of incidence at the point P. Hence, in the presence of topographic
variations it is affected by range slopes. For example, if the terrain is sloped
towards the radar then the effective angle of incidence required to determine the
frequency shift is reduced. In the extreme case, when the surface slope equals
the radar angle of incidence we ultimately determine effective normal incidence
onto the surface and lose sensitivity completely. In this case we obtain so-called
‘blind’ angles for the interferometer. In general, therefore, in the presence of
range slope η we must modify the expression for spectral shift, as shown in
equation (5.17):
cB⊥
f = − (5.17)
R0 λ tan (θ − η)
As an example we consider evaluation of the required spectral shift for typi-
cal values of a space radar (ERS-1) operating at 23◦ of incidence at C band
(λ = 5.66cm) and transmitting a signal with bandwidth 15.5 MHz with a slant
range of 840 km. Figure 5.8 shows the required frequency shift as a function
of surface slope. We note the following important features:
(a) As slope increases from zero, the required frequency shift eventually
exceeds the bandwidth of the pulse (f /W = 1). This is the start of
a band of ‘blind angles’, and the width of this band is defined from
this point until the slope increases to the extent that the frequency shift
returns to the pulse bandwidth. This high angle range, however, is
plagued by layover distortion in radar imaging, whereby points at the
top of a slope are at closer range to the radar than those at the foot of
the slope.
(b) In the negative range slope direction (away from the radar) the required
shift slowly decreases to zero for grazing incidence on the surface.
Thereafter the slope disappears into shadow, and there is no longer a
measurable radar backscatter return.
Correct application of the frequency shift concept therefore requires detailed
knowledge of the terrain slope. This can be obtained, for example, if a reference
digital elevation model (DEM) is available, from which the range slopes can
be calculated by spatial differentiation. In this way an adaptive variation of
spectral shift can be performed. In the absence of such topographic information,
 5.1 Radar interferometry 215

Baseline = 200m, R = 840km

20

15

Lay-over
10
Frequency Shift (MHz)

Shadow
5

0
Blind angles
–5

–10

–15

–20
–80 –60 –40 –20 0 20 40 60 80
Fig. 5.8 Example of frequency shift versus
Surface slope (degrees) surface slope

W-f
dashed = spectrum of s1
solid = shifted spectrum of s2

Fig. 5.9 Schematic representation of range

W spectral shift

however, a compromize shift corresponding to the mean slope can be used,

recognising that there will be some errors in the presence of severe topography.

5.1.1.3 Critical baseline

We see from equation (5.15) that we can always increase the sensitivity of the
interferometer by increasing the baseline. There is, however, a limit to this
process, as eventually the frequency shift required will exceed the available
bandwidth of the radar signal. If this bandwidth is W Hz, then by equating the
shift to W we can obtain an expression for the critical baseline, as follows:
 
cB⊥,crit  λWR0 tan (θ − η) 
W =− 
⇒ B⊥,crit =   (5.18)
R0 λ tan (θ − η) c 

from which we see that it is the product of bandwidth times wavelength λW

that is important in setting the maximum sensitivity of the interferometer. Thus
for a given geometry, a system with 10 MHz of bandwidth at 600 MHz (0.5-m
wavelength) has the same critical baseline as a 100-MHz bandwidth system at
6 GHz.
Note also that this frequency shift leads to a reduction in useful bandwidth
and hence to a reduction in range resolution of the radar system. This arises
because with a shift there is a reduced common overlap between the spectra at
ends 1 and 2 of the baseline, as shown schematically in Figure 5.9. The slant
216 Introduction to radar interferometry

range resolution of an interferometer r, normally equal to c/2W for a single

radar sensor, then has the modified form shown in equation (5.19):
c
r = (5.19)
2(W − f )

5.1.1.4 Flat earth removal

Returning to the general expression for interferometric phase (equation (5.8)),
one parameter of importance is the fringe frequency with slant range fs , defined
again in equation (5.20):

φ = 2βR ≈ 2βB|| = 2βB sin θ φ = 2βB cos θ θ
⇒
z = h − R cos θ z = R0 sin θ θ − R cos θ
4π B⊥ 4π B⊥
→ φ = R + z (5.20)
λR0 tan θ λR0 sin θ

1 ∂φ  2B⊥
→ fs =  =
2π ∂R z=0λR tan θ0

This is interpreted physically as the rate of change of phase across a surface

for which z = 0—a ‘flat earth’. For example, for a space radar operating
at a 23 degree angle of incidence at C-band (5.66=cm wavelength) at a slant
range of 840 km, and with a perpendicular baseline of 200 m, this leads to a
phase rate of 1 cycle/50 m (which is only around five slant range cells for a
15-MHz bandwidth typical of space radars). This high-frequency phase term is
therefore often removed at the start of interferometric analyses, so generating
a constant phase for flat terrain, which then helps emphasize any topographic
variations that may be present. It also facilitates coherence estimation, as we
show in the next section. It can always be returned to the phase signal at the
end of processing.
Flat-earth processing involves multiplication of the interferogram by a con-
jugate phase signal to cancel the flat-earth components, as shown in equation
(5.21). Here r is the slant range coordinate of the radar.

s1 s2∗ e−i2π fs r (5.21)

This is called ‘flat earth’ removal for interferometric processing. Note that it
is not a replacement for spectral shift processing, and its sole purpose is to
remove the high-frequency modulation of the underlying phase to facilitate
further processing and interpretation. Note that for airborne radar or other large
swath geometries, when the angle of incidence varies across the range swath,
the spatial frequency of the flat-earth signal is no longer constant but increases
with decreasing angle of incidence, leading to a ‘chirp’ or variable frequency
reference. In this case the same procedure can be used for flat-earth removal,
but fs must be evaluated more carefully by accurate evaluation of the θ term.
We now briefly consider two important variations on radar interferometer
design. The first—called ‘differential interferometry’—is designed to maximize
sensitivity to changes in repeat-pass sensors; while the second—along-track
interferometry (ATI)—is designed to sense the velocity rather than position
vector of scatterers.
 5.1 Radar interferometry 217

s1s2 d
d

R2
H
R1

z Fig. 5.10 Geometry of differential interfer-

ometry

5.1.2 Introduction to differential interferometry

An important special case of radar interferometry arises for zero spatial base-
lines, B = 0, but non-zero temporal baseline TB . In this case, from equation
(5.3) we expect the interferometric phase to be zero for all surfaces. However,
this ignores any motion of the surface that may have occurred between passes
of the sensor. Figure 5.10 shows a schematic representation of this situation.
In the solid line is shown a surface measured from position Q at time t1 to
obtain s1 . The dashed line is the new position of the surface at time t2 = t1 + TB
when measured again by the sensor at Q to obtain s2 . The interferogram formed
as s1 s2∗ will have non-zero phase because of the radial component of shift of
the surface. If different parts of the surface move by different amounts then
we obtain a phase map of these displacements in the interferogram. Equation
(5.22) shows how such an interferometer measures the projection of the surface
displacement vector d onto the line of sight.
4π 4π  
φ= (R2 − R1 ) = d sin (θ − δd ) (5.22)
λ λ
Hence we see that there is no sensitivity to movement perpendicular to the line
of sight (when θ = δd ). Note that in this case there is no baseline scaling and
the sensitivity is maximum, given by the full wavenumber β. If we define the
quantity d = d  sin (θ − δd ) as an equivalent height to the offset baseline case,
then the π -displacement is given as
λ
dπ = (5.23)
4
If the wavelength is a few centimetres and the phase accuracy is around 5◦ ,
then this can lead to sensitivities of the order of mm. There are, however, two
major problems with differential interferometry:
The first is difficulty in obtaining a true zero spatial baseline (exact repeat-
pass). Hence in practice there is always some spatial baseline B to consider,
which inevitably contributes a phase component sensitive to topography, as
considered earlier. Following flat-earth removal, the total interferometric phase
can therefore be more accurately written, as shown in equation (5.24):
4πB⊥ z 4π
φ = φtopo + φdisplacement = − + d
λR0 sin θ λ (5.24)
⇒ φdisplacement = φ − φtopo
There are two ways to combat this topographic dependency. The first is to
use a reference DEM to estimate the topographic phase component, and to
218 Introduction to radar interferometry

then remove its contribution from the phase signal, as shown in the lower
portion of equation (5.24). One potential problem with this approach is that
the reference DEM employed will have some errors, and these can propagate
into the displacement estimate. For example, if the DEM has an elevation error
z then this translates into an equivalent phase error of φ z which in turn is
interpreted as a ground motion dz , as shown in equation (5.25):

4π B⊥ λ B⊥
φz = z ⇒ dz = φz = z (5.25)
λR0 sin θ 4π R0 sin θ
This can be expressed more conveniently in terms of the critical baseline for
a sensor with bandwidth W operating at centre frequency f0 , as shown in
equation (5.26):

B⊥ B⊥ W z
dz = z = (5.26)
R sin θ Bcrit f0 cos θ

Clearly the error can be minimized by reducing the baseline, and in the limit
of zero spatial baseline the topographic error is removed. The second approach
(when a reference DEM is not available) is to employ three (or more) passes to
eliminate the topographic phase dependence. If n ≥ 3 passes are available then
we can define a set of interferometric parameters, as shown in equation (5.27):

- time baselines tn−m = tn − tm

- spatial baselines B⊥,n−m (5.27)
- interferometric phases φn−m

We can then decompose the set of phases φ n−m into a part proportional to the
spatial baselines (topography) and a part proportional to the temporal baselines
(displacement at constant velocity). To illustrate this we consider the simplest
case of three passes, yielding complex signals s1 , s2 and s3 . The first signal,
s1 , is then used as a ‘master’ to generate interferograms with both s2 and s3 .
The first baseline B12 we wish to be dominated by topographic effects, and
so should be as large as possible, with a small temporal baseline so as to
minimize displacement effects. The second baseline, B13 , however, should be
dominated by temporal effects, and so should combine a small spatial with
long temporal baselines. With this combination the displacement phase can be
directly estimated as shown in equation (5.28). (Note that if the baseline ratio is
too large (>4) then phase unwrapping may be required of the φ 12 interferometric
phase before scaling.)

B13
φdisplacement = φ13 − φ12 (5.28)
B12
The second problem faced in differential interferometry is the effect of wave
propagation between the sensor and the surface. The propagation of microwaves
through the atmosphere causes a phase shift due to variations in refractive index.
For repeat-pass sensors the changing atmosphere causes a change in this phase
and hence a phase error in the interferogram. Such propagation effects can be
large, for example, for a spaceborne radar at C-band, as atmospheric delays
can cause an error of half a fringe (π radians), or in extreme cases up to three
 5.2 Sources of interferometric decorrelation 219

fringes. For low-frequency radars, phase shifts due to propagation through the
ionosphere can cause similar problems (Freeman, 2004). In general, therefore,
the phase of an interferogram can be written in component form, as shown in
equation (5.29):

φ = φflat + φtopo + φdisplacement + φpropagation

4πB⊥ r 4π B⊥ z 4π
=− − + d + φpropagation (5.29)
λR0 tan θ λR0 sin θ λ
Note that the propagation phase does not depend on baseline and hence can-
not be removed by baseline diversity. It is embedded as an error source in the
displacement phase. Recently there have been several techniques proposed for
separating the propagation from displacement phase by employing the former’s
distinct lack of temporal correlation combined with high spatial correlation aris-
ing from the fractal nature of the underlying atmospheric phase screen (Kampes,
2006). However, this method requires the acquisition of a large number of
passes, and hence takes us beyond the bounds of a basic introduction. Instead
we turn to consider the third important type of interferometer: ATI.

5.1.3 Along track interferometry (ATI) s1

s2
The third important interferometric configuration to be considered is along Platform Velocity vr
track interferometry, or ATI. This is a single-pass configuration with two radar
systems displaced with a spatial baseline parallel to the direction of motion of
the platform, as shown schematically in Figure 5.11. The key idea is that in
this configuration the spatial baseline and platform velocity combine to obtain v
a short temporal baseline t (typically of the order of 10–100 msecs), during
which the scatterer (moving itself with velocity v) will move, and thus cause a P
change in range, which leads to a phase shift.
Fig. 5.11 Geometry of along-track interfer-
We can then quantify the relationship between interferometric phase and ometry
velocity as shown in equation (5.30):
4π 4π ∂R 4π B
φ= R = t = vLOS (5.30)
λ λ ∂t λ vr
where vLOS is the line-of-sight component of the velocity vector v of the point
P. Hence ATI remains blind to velocities parallel to the platform motion. Such a
technique can be used to measure ocean currents and glacier motion, as well as
the speed of point scatterers such as ships and land vehicles. One key limitation
of this idea is the maximum temporal baseline that can be used. Decorrelation
effects in the scatterer eventually lead to a loss of coherence (discussed in
the next section). For this reason the temporal baseline needs to be designed
with a measure of the typical scatterer decorrelation time in mind. This brings
us to consider decorrelation and its relation to an important new observable,
interferometric coherence.

5.2 Sources of interferometric decorrelation

In the previous section we showed how interferometric phase can be related
to several important surface parameters (height, velocity, displacement, and so
220 Introduction to radar interferometry

on), depending on the configuration used. However, so far we have ignored the
influence of noise and its impact on phase estimation. In polarimetry we saw
that noise arises from depolarisation and is manifest as an increase in scattering
entropy. In this section we first consider a formalism to include noise effects
in radar interferometry, and then consider a set of various potential sources of
noise (Zebker, 1992). Some of these are system related (signal-to-noise ratio,
for example), but others are related to wave scattering effects (volume and
baseline decorrelation in particular) and hence can be considered analogous
to wave depolarisation effects in polarimetry. By considering such coherent
scattering in detail, we will see how we can then turn the noise problem around
and use the interferometric coherence as a new radar observable to help estimate
surface and volume scattering parameters. This will then lead us to consider, in
the next chapter, combinations of polarimetry with interferometry.
We start with a general expression for interferometric phase φ as the sum
of ‘signal terms’ φ if and a noise term φn , (equation (5.31)) characterized by
its statistical moments. Generally the noise term will have zero mean, but is
characterized by non-zero standard deviation σ φ .

φ = φif + φn (5.31)

To see the impact of such stochastic fluctuations in the phase on surface param-
eters, consider the important special case of surface height estimation using
across-track radar interferometry. As shown in equation (5.32), the phase vari-
ance σ φ leads to a scaled height variance, derived from the relation of phase
to changes in slant range σ R and then using equation (5.15) to relate range to
height via the normal baseline.

λ λ R0 sin θ R0 sin θ λ
R= φ ⇒ σR = σφ ⇒ σh ≈ σR ≈ σφ (5.32)
4π 4π B⊥ B⊥ 4π

This relation can be used to estimate errors in surface height based on system
parameters (baseline geometry and angle of incidence) and the phase variance.
To proceed, we need to further investigate the different ways in which phase
noise can be generated in radar interferometry. To do this we first relate noise
variance to an underlying coherence.
We start by employing a coherency matrix formulation of radar interferom-
etry, as shown in equation (5.33). Here again we can define a useful secondary
parameter: the interferometric coherence, with a magnitude between 0 (pure
noise) and 1 (pure signal).
+ 2 , + ∗ ,  + ∗,
|s1 | s1 s2 s1 s2
[T2 ] = + , + 2 , ⇒ γ̃ = + ,+ , (5.33)
∗ |s2 |
s2 s1 |s1 |2 |s2 |2

From Appendix 3 it then follows that the phase variance can be related to the
coherence by the following Cramer–Rao bounds (Seymour, 1994):
"
1 − |γ |2 1 − |γ |2
σφ ≤ σ|γ | ≤ √ (5.34)
2L |γ |2 2L
 5.2 Sources of interferometric decorrelation 221

where L is the number of independent samples used in forming the average

<..>. Note that often the coherence itself is estimated from the data (Touzi,
1999), in which case it has an estimation variance defined using the Cramer–Rao
bound also shown in equation (5.34). We now take a closer look at the origin
of these fluctuations. We look at four factors: signal-to-noise ratio, temporal
decorrelation, baseline, and volume decorrelation.

5.2.1 Signal-to-noise decorrelation

The two complex signals s1 and s2 can first be decomposed into signal (a) and
noise (n) terms, as shown in equation (5.35):

s 1 = a + n1
(5.35)
s2 = a + n2

Now we must invoke some assumptions about the statistical distribution of the
noise terms. The most common assumption, based on the central limit theorem,
is that the noise terms are complex Gaussian random variables and hence have
uniform phase distributions and are uncorrelated both with the signal (a) and
with each other. Under this assumption the coherence can be evaluated as shown
in equation (5.36):

|a|2 SNR
γsnr = = (5.36)
|a| + |n|
2 2 1 + SNR

Here we see a simple relation between the signal-to-noise ratio (SNR) and
coherence. As the SNR tends to infinity (zero noise) then the coherence tends
to unity, while if the SNR tends to zero then the coherence also tends to zero.
Figure 5.12 shows how the coherence is related to SNR (expressed in dB).

Noise decorrelation
1

0.9

0.8

0.7

0.6
Coherence

0.5

0.4

0.3

0.2

0.1

0
–30 –20 –10 0 10 20 30
Fig. 5.12 Relationship between signal-to-
SNR (dB) noise ratio and coherence
222 Introduction to radar interferometry

5.2.2 Temporal decorrelation

A second key model employed for a noise source is to assume that the noise has
constant amplitude but a Gaussian distribution of phase. This model is more
appropriate when we have a signal in the presence of an unwanted ‘clutter’
background. In this case we can use the following identity for the average
phase difference between stochastic signals with Gaussian phase statistics:
 9 :
s1 = a + meiφ1 σφ2
⇒ e−i(φ1 −φ2 ) ≈ e− 2 (5.37)
s2 = a + me iφ2

This then enables us to calculate an expression for the coherence, as shown in

equation (5.38):

|a|2 + |m|2 e−σφ SCR + e−σφ

2 2

γ = = (5.38)
|a|2 + |m|2 SCR + 1

where SCR is now the signal-to-clutter ratio. The most important application
of this model is to temporal decorrelation in repeat-pass interferometry. In this
case the ‘clutter’ noise is caused by motion of scatterers (such as wind-driven
vegetation) between passes. If the rms motion along the line of sight is δ rms ,
then the phase variance and hence coherence in equation (5.38) can be simply
related to this shift, as shown in equation (5.39):
16π 2
4π SCR + e− λ2 δrms
σφ = δrms ⇒ γ = (5.39)
λ SCR + 1
Importantly, we see that this coherence depends on the ratio of rms motion to
wavelength, and hence for a given shift the effect on coherence is worse for
higher frequencies. This drives us to consider lower frequencies to minimize
the effects of temporal decorrelation in repeat-pass interferometry (Hagberg,
1995; Askne, 1997, 2003, 2007).
Note that in the general case we can have a combination of these statistical
effects, such as temporal decorrelation γ t in combination with noise decorre-
lation γ snr . In this case the coherence is formed from products of triple sums,
as shown in equation (5.40). The most important consequence of this is that
coherence always decomposes in a multiplicative series of component terms
(to be contrasted with polarimetric decomposition which led to expansion as a
sum of component terms).

|a|2 + |m|2 e−σφ
2
s1 = a + meiφ1 + n1
⇒γ = 2
s2 = a + meiφ2 + n2 |a| + |m|2 + |n|2
|a|2 + |m|2 e−σφ
2
|a|2 + |m|2
= . = γt γsnr
|a|2 + |m|2 |a|2 + |m|2 + |n|2
(5.40)

Although only shown for a combination of two components in equation (5.40),

the same argument can be used for coherence of an arbitrary mixture of inde-
pendent terms. For example, in addition to SNR and temporal effects there
are always some additional sources of coherence loss due to processing errors.
 5.2 Sources of interferometric decorrelation 223

Typically in radar applications the two signals s1 and s2 are formed from co-
registered synthetic aperture radar (SAR) images (often collected at different
times of a repeat orbit), and in practice it is impossible to exactly match the
two radar signals (see Chapter 9). There will always be some small residual
fractional offset in range and azimuth pixel size δ rg and δ az between the two
images (Krieger, 2005). This causes a coherence loss component given by
equation (5.41):

sin π δrg sin π δaz

γproc = (5.41)
π δrg π δaz

Current processing accuracies are limited to about 1/10 of a pixel in both range
and azimuth, and so we see that this error is independent of baseline and has
a value around 0.97. This error becomes particularly significant in the case of
small spatial baselines, where it can become the dominant source of decorre-
lation. To combine this error source with the other two we simply extend the
decomposition of equation (5.40), as shown in equation (5.42):

γ = γSNR γt γproc (5.42)

This approach gives us the ability to consider different independent decor-

relation sources and include them in the final expression for coherence in a
straightforward (multiplicative) way. In particular, there are two more important
scattering-based decorrelation sources to be considered: baseline and volume
decorrelation, which we now consider in turn.

5.2.3 Baseline decorrelation

Surface scattering can give rise to an important source of coherence loss termed
baseline decorrelation (Zebker, 1992; Gatelli, 1994). The origin of this process
can be found in the ideas of frequency shift, as discussed in equation (5.13).
There we showed that the expression for interferometric phase (before spectral
shift filtering) contains a dependence on the y or surface coordinate of the
scattering point. Therefore, if we have a distribution of scattering points within
a range cell they will add coherently to yield some resultant complex return.
However, when we shift position to the other end of the baseline we obtain
a slightly different coherent sum from the same set of points (simply because
the surface component of the wave vector has changed). This fluctuation in
the complex sum for surface scatterers leads to a loss of coherence, as we
now show.
We can immediately see one important additional benefit of performing the
spectral shift before interferogram formation. If the spectral shift is applied,
then by definition the contributions from both ends of the baseline have the
same surface component of wavenumber and hence the same coherent phase
addition for surface scatterers. Following spectral filtering the coherence equals
1, and baseline decorrelation is removed. From our discussion around equation
(5.18) we see that this will be possible up to a maximum baseline, called the
critical baseline, after which the spectral overlap will be zero and we obtain
zero coherence. Thus the baseline decorrelation is given by the ratio of shifted
spectral overlap to total bandwidth W . This results in the expression for baseline
224 Introduction to radar interferometry

decorrelation shown in equation (5.43):

Bcrit − B⊥ B⊥ cB⊥
γB = =1− =1− (5.43)
Bcrit Bcrit W λR0 tan (θ − η)
We reiterate that this decorrelation occurs only if spectral filtering is not applied.
By employing a spectral shift we can always ensure that γB = 1 (up to a
maximum separation of the critical baseline, although note from equation (5.19)
that the price to pay for this shift is that the range resolution reduces).
Before leaving this topic, we note one important scenario that always gener-
u = l R = c ates unit baseline coherence, independent of spatial baseline and spectral shift.
2
2W

= R
This is when the resolution cell contains only a single point scatterer. To see this,
tanu consider an alternative interpretation of critical baseline in terms of an effec-
tive scattering diagram, as shown schematically in Figure 5.13. Here we show
a surface resolution element, which for a distributed surface scatterer (shown in
Fig. 5.13 Geometric interpretation of
grey) has a spatial extent bounded by the bandwidth of the radar pulse W . This
baseline decorrelation spatial segment has an apparent projected size ⊥ perpendicular to the line of
sight, as shown in Figure 5.13. This projected surface element radiates back
to the radar (the process of scattering), and has an effective beam width given
by θ, as shown. The critical baseline then occurs when this beamwidth fails
to enclose both points 1 and 2 of the baseline. However, for a point scatterer
(shown as the black disc), the spatial extent is not governed by the bandwidth
but by the spatial size of the scatterer. In the limit of a point target ⊥ is a delta
function, and hence θ becomes very large. The wide beamwidth therefore
encloses all pairings of baseline end points 1 and 2, the critical baseline tends
to infinity, and there is zero baseline decorrelation. This observation leads to
the permanent scatterer (PS) technique in radar interferometry (Kampes, 2006),
where high-accuracy positional information can be obtained from radar inter-
ferometry by restricting attention to point targets only—such as occur in urban
areas, where there are many point-like man-made structures (see Chapter 9).

5.2.4 Volume decorrelation: the Fourier–Legendre series

In the previous section we saw how a random distribution of scatterers in a
surface plane can cause decorrelation and loss of interferometric coherence
through baseline (also called geometric) decorrelation. However, by employing
range spectral filtering over terrain with known surface slope, we are always
able to remove this decorrelation source (up to a limit given by the critical
baseline). In a similar manner we note that a vertical distribution of scatterers
will also cause a loss of coherence. This is termed volume decorrelation, as
it often originates from volume scattering by layers of vegetation or snow/ice
above the surface (Hagberg, 1995; Treuhaft, 1996).
However, one key distinguishing feature of volume decorrelation is that it
is not possible to remove its effect by range spectral filtering. We saw from
equation (5.13) that we can always choose k to remove the y but not the
z dependence of interferometric phase. Therefore, two scatterers separated
by a distance z will always have a phase difference given by the vertical
wavenumber β z , as shown in equation (5.44):
4π θ 4π B⊥
φ = βz z = z≈ z (5.44)
λ sin θ λR0 sin θ
 5.2 Sources of interferometric decorrelation 225

If we have a general variation of scattered power with z given by a vertical

structure function f (z), the lower bound of which is at z = zo , and the upper
bound of which is at z = z0 + hv , where hv is the height of the layer, then the
interferometer will see a complex signal given by the weighted sum of contri-
butions, as shown in equation (5.45), from which we can obtain an expression
for the interferometric coherence, as shown in equation (5.46):

zo +hv z = z − z0 hv
s1 s2∗ = f (z)eiβz z dz −−−−−−−−−→ eiβz zo f (z )eiβz z dz (5.45)
zo 0
- hv
f (z )eiβz z dz
γ̃ = e iβz zo 0
- hv = eiβz zo |γ̃ | ei arg(γ̃ ) (5.46)
0 f (z )dz

Note that this is a complex coherence; that is, it has phase as well as magnitude,
and part of the phase arises from the integral of the structure function f (z)
shown in the numerator. This real non-negative function allows for arbitrary
profile of scattering between the bottom and top of the layer (Cloude, 2006b).
This relation shows that there is a direct relationship between the observed
coherence and vertical structure properties of the scattering layer. For example,
the height of the layer is found in the limits of the integral, the phase of the
surface, while not equal to the phase of the coherence, is contained therein, and
finally the structure function f (z) influences the coherence in both amplitude
and phase.
Special cases of the structure function are often used in practice: for example,
constant scattering amplitude or an exponential to more accurately model wave
extinction effects in the layer (Treuhaft, 1996). Here we first develop a general
theory of volume decorrelation based on arbitrary structure functions, and then
specialize our discussion to these important special cases. The approach we use
is to expand the bounded function f (z) in a Fourier–Legendre series, as follows.
We first normalize the range of the integral in the numerator by a further change
of variable, as shown in equation (5.47):

hv zL = 2z
hv − 1
1
f (z )e iβz z
dz −−−−−−− −−−→ f (zL ) eiβz zL dzL (5.47)
0 −1

We then rescale variation of the real non-negative function f (z) so that if 0 ≤

f (z) ≤ ∞ then f (zL ) = f (z) − 1 and −1 ≤ f (zL ) ≤ ∞. Critically, we
can now develop f (zL ) in a Fourier–Legendre series on [−1,1], as shown in
equation (5.48):
!
f (zL ) = an Pn (zL )
n
(5.48)
2n + 1 1
an = f (zL )Pn (zL )dz
2 −1

where the first few Legendre polynomials of interest to us are given explicitly
as shown in equation (5.49). Figure 5.14 shows plots of these functions for
hv = 10 m. The first represents a simple uniform distribution, while the sec-
ond includes linear variations, then quadratic and so on, with the higher-order
226 Introduction to radar interferometry

Legendre polynomials
10
P0
9 P1
P2
8 P3
P4
7 P5
P6
6

Height
5

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Fig. 5.14 The Legendre polynomials from
zeroth to sixth order Relative density

functions offering ever-higher resolution of functional variation. In this way

any function can be represented over the interval from z = 0 to z = hv by the
‘spectrum’ of real parameters an .

P0 (z) = 1

P1 (z) = z
1 2
P2 (z) = 3z − 1
2
1 3
P3 (z) = 5z − 3z
2 (5.49)
1 4
P4 (z) = 35z − 30z 2 + 3
8
1 5
P5 (z) = 63z − 70z 3 + 15z
8
1 
P6 (z) = 231z 6 − 315z 4 + 105z 2 − 5
16

The numerator and denominator of the general expression for coherence can
now be written as shown in equation (5.50):

hv hv i βz hv 1 βz hv
f (z )eiβz z dz = e 2 (1 + f (zL ))ei 2
zL
dzL
0 2 −1
(5.50)
hv hv 1
f (z)dz = (1 + f (zL ))dzL
0 2 −1
 5.2 Sources of interferometric decorrelation 227

from which it follows that the coherence can be written as shown in

equation (5.51):

-1 i βz2hv zL
iβz z0 i βz2hv −1 (1 + f (zL ))e dzL
γ̃ = e e -1
−1 (1 + f (zL ))dzL
 
-1 5 (5.51)
−1 1 + an Pn (z L ) eiβv zL dzL
n
= eiβz z0 eiβv  
-1 5
−1 1 + an Pn (zL ) dzL
n

By expanding the series and collecting terms, this equation can be rewritten in
simplified form, as shown in equation (5.52):

-1 -1 -1
(1 + a0 ) −1eiβv zL dzL + a1 −1 P1 (zL )eiβv zL dzL + a2 −1 P2 (zL )eiβv zL dzL + · · ·
γ̃ = e iβz z0 iβv
e -1 -1 -1
(1 + a0 ) −1 dzL + a1 −1 P1 (zL )dzL + a2 −1 P2 (zL )dzL + · · ·

(1 + a0 )f0 + a1 f1 + a2 f2 + ..an fn
= eiβz z0 eiβv
(1 + a0 )
ai
= eiβz z0 eiβv (f0 + a10 f1 + a20 f2 + · · · ) ai0 = (5.52)
1 + a0

Note that evaluation of the denominator is simplified by using the orthogonality

of the Legendre polynomials. Evaluation of the numerator involves determina-
tion of the functions fn , which are straightforward integrals employing repeated
use of the following identity:

 
n βz eβz nz n−1 n(n − 1)z n−2 (−1)n n!
z e dz = z −
n
+ · · · (5.53)
β β β2 βn

As an example, equation (5.54) shows detailed calculation of the first two terms
in the series. The first, corresponding to the zeroth-order Legendre polynomial,
just yields a SINC function, while the first order linear polynomial gives a
slightly more complicated function.

 1
1 1 eiβv z 1  sin βv
f0 = eiβv z dz = = eiβv z − e−iβv z =
2 −1 iβv −1 i2βv βv
1   1
1 eiβv z 1
f1 = zeiβv z dz = z−
2 −1 iβv iβv −1 (5.54)
1  1 
= eiβv z + e−iβv z − eiβv z − e−iβv z
iβv (iβv )2
 
sin βv cos βv
=i −
βv2 βv
228 Introduction to radar interferometry

For reference we give the explicit form of all these functions up to sixth order
in equation (5.55).

sin βv
fo =
βv
 
sin βv cos βv
f1 = i −
βv2 βv
 
3 cos βv 6 − 3βv2 1
f2 = − + sin βv
βv2 2βv3 2βv
    
30 − 5βv2 3 30 − 15βv2 3
f3 = i + cos βv − + 2 sin βv
2vβv3 2βv 2βv4 2βv
 
35(βv − 6)
2 15
f4 = 4
− 2 cos βv
2βv 2βv
 
35(βv4 − 12βv2 + 24) 30(2 − βv2 ) 3
+ + + sin βv
8βv5 8βv3 8βv
 
−2βv4 + 210βv2 − 1890 30βv4 − 840βv2 + 1890
f5 = i cos βv + sin βv
βv5 βv6
 
42βv4 − 2520βv2 + 20790)
f6 = cos βv
βv6
 6 
2βv − 420βv4 + 9450βv2 − 20790)
+ sin βv (5.55)
βv7

We note the following important points:

1. The even index functions are real while the odd are purely imaginary.
We note also that the unknown coefficients an are all real.
2. The functions vary only with the single parameter β v , which itself
is defined from the product of two parameters: height hv , and the
interferometric wavenumber β z .

Graphs of these functions are shown in Figure 5.15. We see that the first
is a ‘SINC’ relation between coherence and increasing height–baseline prod-
uct. This is the expected functional relationship for scattering by a uniform
layer. However, we see that as the height–baseline product increases so the
other functions become more important. We can conclude, therefore, that the
interferometric coherence is sensitive to changes in the structure function f (z).
There are two special cases of structure function of particular importance due
to their widespread use in the literature. We now turn to consider these in more
detail.

5.2.4.1 Special case 1: the uniform profile

If we assume f (z) = 1—a constant structure function—then all the higher order
Legendre coefficients are zero, and the coherence becomes a function only of
 5.2 Sources of interferometric decorrelation 229

Legendre coherence function

1
Re(f0)
Im(f1)
0.8 Re(f2)
Im(f3)
Re(f4)
0.6 Im(f5)
Re(f6)

0.4

0.2

–0.2

–0.4
0 0.5 1 1.5 2 2.5 3
Fig. 5.15 Coherence basis functions for
kz*h/2 Legendre expansion

height, given by a complex SINC function, as shown in equation (5.56):


hv
sin βz2hv
γ̃ = eiβz z0 eiβz 2 βh
(5.56)
z v
2

There are two important features of this model. Firstly it shows that volume
scattering provides a phase offset, given in this case by half the volume height.
Hence in the presence of volume scattering the interferometric phase no longer
represents the true surface position but is offset by a bias. For vegetated terrain
this is called vegetation bias, and provides an error source in the use of radar
interferometry for true surface topography mapping. We see that the only way
to minimize this effect is to employ small baselines so that the product β z hv
remains small. However, this reduces the sensitivity of the interferometer and
is difficult to sustain over forested terrain, where hv can reach up to 50 m or
more. Note that we cannot simply use the phase of the interferogram to estimate
volume height hv , since the total phase involves addition of an unknown phase
shift due to the lower bound of the volume (z0 ). Only if we can provide an
estimate of this lower bound can we then use the phase to estimate height. We
shall see in Chapter 8 how to provide such an estimate.
The second key feature of the SINC model is that the coherence amplitude
falls with increasing height and hence the phase variance increases with hv . Note
that there is no effect of the lower bounding surface on coherence amplitude
(assuming range spectral filtering has been employed), and in principle we can
therefore use an estimate of measured coherence amplitude to estimate height
(for a known baseline). In particular, for short baselines we can expand the SINC
function in a series and obtain a useful direct height estimate from coherence,
as shown in equation (5.57):
"
sin x x2 24(1 − |γ̃ |)
x1⇒ ≈1− ⇒ hv ≈ (5.57)
x 6 βz2
230 Introduction to radar interferometry

However, as we shall see in the next section, this approach is sensitive to

variations in the actual structure function of the volume. The SINC model is
really valid only for very small height–baseline products, and for moderate
baselines higher-order terms in the Legendre expansion of f (z) can no longer
be ignored. In fact, as we shall see in Chapter 8, we can turn this idea around
and design offset baselines to enhance the higher-order terms and thus enable
parameter estimation for the layer.
Nonetheless, this SINC model is commonly used, especially by radar system
designers, who wish only to assess the relative importance of volume decorre-
lation in the overall coherence budget for an interferometer. Finally, we note
that this model contains no polarisation dependence at all. The volume decor-
relation and phase bias of the SINC model are functions only of the height hv .
We shall see, however, that the higher-order terms of the Legendre expansion
are sensitive to changes in wave polarisation, and this will suggest the develop-
ment of polarimetric interferometry for parameter estimation. First, however,
we turn to consider a second important special case: the exponential profile.

5.2.4.2 Special case 2: the exponential profile

A second important structure function is the exponential—widely used to model
the physical effects of wave propagation through a volume scattering layer
(Treuhaft, 1996, 2000a; Papathanassiou, 2001). This is in accordance with the
water cloud model described in Section 3.5.1. According to this idea, contribu-
tions from the top of the volume are weighted more strongly in the coherence
calculation than those deeper into the volume, as the latter experience a smaller
incident signal due to wave extinction, combining the physical effects of wave
attenuation due to absorption of energy by the volume and scattering loss due
to the presence of particles. The combined effect of these two processes can be
represented by a one-way power loss extinction coefficient σ e with natural units
of m−1 , but often expressed in engineering units of decibels per meter (dB/m).
Note that two systems of units can be related using equation (5.58) (compare
this with equation (3.11), for amplitude extinction). In addition we note that in
radar applications there is a two-way propagation channel, and so the signal is
attenuated both on the way in and out of the volume (see Figure 5.16). Hence the
total extinction is 2σe . Finally, we must also account for the increased attenua-
tion path length through the medium when illuminated at an angle of incidence
θ 0 , as shown in Figure 5.16.
10σe
σedB = ≈ 4.34σe ⇒ σe ≈ 0.23σedB (5.58)
ln(10)
Rather than expand the exponential function in a Legendre series, it is eas-
ier in this case to explicitly evaluate the coherence integrals, as shown in

z = z0 + hv
o
2sez
f (z) = e coso
z = z0
Fig. 5.16 Exponential structure function
 5.2 Sources of interferometric decorrelation 231

equation (5.59):
- hv 2σe z

iβz z0 mv 0 e cos θo eiβz z dz

γ̂ = e - hv 2σe z
mv 0 e cos θo dz
2σe eiβz z0 hv 2σe z
= eiβz z e cos θo dz
cos θo (e2σe hv / cos θo − 1) 0 (5.59)



2σe
(e p2 hv − 1) p1 =
p1 cos θo
= f (hv , σe ) = eiβz z0
p2 (ep1 hv − 1)   2σe
p2 = + iβz
cos θo

This example illustrates the important new idea that the coherence in general
depends not only on the volume depth hv but also on the shape of the vertical
structure function. The exponential model essentially allows a one-parameter
model for variation of structure (via σ e ). High extinction implies an effective
scattering layer at the top of the volume, such as a high-elevated forest canopy,
for example. Figure 5.17 shows an example of how the coherence varies for an
exponential profile with varying σ e and depth hv . We have selected a baseline
corresponding to β z = 0.1567 (which corresponds to a zero of the SINC model
at 40 m), and considered a 45-degree angle of incidence. Note that for zero
extinction we again obtain, as a special case, the SINC model.
However, as extinction increases so the coherence increases for a given
height. This arises physically as the effective scattering volume is being
squeezed into a smaller and smaller region close to the top of the volume
as extinction is increased. This can be confirmed by plotting the phase of the
coherence. We first define the fractional phase centre height Pc from the interfer-
ometric phase φ as Pc = hvφβz . Figure 5.18 shows how Pc varies with extinction

Volume decorrelation vs. Extubctuib (betaz = 0.1567)

1

0.9

0.8

0.7

0.6
Coherence

0 dB/m
0.5
0.125 dB/m
0.4 0.125 dB/m

0.3 0.75 dB/m

0.2

0.1

0
0 5 10 15 20 25 30 35 40
Fig. 5.17 Volume decorrelation versus
height (m) height for various extinctions
232 Introduction to radar interferometry

Phase height vs. extinction (betaz = 0.1567)

1

0.9

0.8

0.7

0.6

Coherence
0.5

0.4
0 dB/m
0.3 0.125 dB/m
0.2 0.125 dB/m
0.75 dB/m
0.1

0
0 5 10 15 20 25 30 35 40
Fig. 5.18 Phase centre height versus height
for various extinctions Height (m)

90 1
120 60
0.8

0.6
150 30
0.4

0.2

180 0 dB/m
0

0.75 dB/m
0.125 dB/m
210 330

Fig. 5.19 Representation of complex volume

240 300
coherence variation inside unit circle for vary-
ing extinction 270

and height. Note that for the special case of zero extinction we obtain a phase
centre halfway up the layer as expected in the SINC model. However, as the
extinction increases we see that the phase centre moves towards the top of the
layer, approaching Pc = 1 in the limit of infinite extinction.
We have seen from Figures 5.17 and 5.18 that the coherence amplitude and
phase variations are linked. Indeed, it is instructive to visualize both at the same
time by employing the coherence diagram representation (see Appendix 3).
Figure 5.19 shows how the complex coherence varies inside the unit circle
in the complex coherence plane for three extinction values and layer depths
varying from 0 to 40 m (using the same parameters used in Figures 5.17 and
5.18). Here we see that the SINC model spirals quickly to the origin—to zero
 5.2 Sources of interferometric decorrelation 233

coherence—while the high-extinction cases show gentler spirals with more

rapid phase variation around the unit circle.

5.2.5 Summary: coherence decomposition

We have seen in this chapter that the interferometric coherence may be decom-
posed into a product of terms, the most important of which are shown in equation
(5.60), where we define the following important components:

γ̃ = eiφs γSNR γt γproc γs γ̃v (5.60)

γSNR Decorrelation due to additive noise in the signals.

γt Temporal decorrelation due to motion of scatterers between passes in
repeat-pass interferometry.
γproc Loss of coherence due to processing errors associated, for example, with
image misregistration in radar imaging.
γs Baseline or surface decorrelation. This depends on the nature of the
surface scattering (point scatterers or random surface scattering), but
can always be removed (set equal to 1) by employing range spectral
filtering.
γ̃v Volume decorrelation. This is a complex coherence, in that unlike the
other terms it distorts both the mean and standard deviation of the
interferometric phase.
We have developed a general method for predicting the volume coherence for
a given structure function using a generalized Fourier–Legendre expansion, and
considered in detail the important special cases of a uniform and exponential
profile.
In particular we have seen that interferometric coherence can be controlled
through baseline selection, even for scattering from random media. This gives
us the ‘entropy control’ missed with polarimetry alone. The next step is to
incorporate polarisation effects into radar interferometry. In Chapter 6 we show
how to do this in a formal mathematical way before exploring some of the
physical models used in Chapter 7.
 Polarimetric
6 interferometry

In this chapter we formally combine the topics of polarimetry and interfer-

ometry. Our purpose is to establish a general framework for describing the
formation and analysis of interferograms for arbitrary choice of transmit and
receive wave polarisations (Papathanassiou, 1997). This will lead us to study
the variation of interferometric coherence with polarisation, and ultimately to
develop methods for coherence optimization (Cloude, 1997b), for investigat-
ing the dynamic range of interferometric coherence variation with polarisation.
This will then lead us, in Chapter 7, to apply the optimization procedures to
surface and volume scattering scenarios in the same way as for polarimetry
alone in Chapters 3 and 4.

6.1 Vector formulation of radar interferometry

To generate a vector interferogram we require two key ideas (Cloude, 1997b,
1998). The first is that an interferogram is always formed between two complex
scalars, representing the amplitude and phase of scattered fields at ends 1 and
2 of a spatial or temporal baseline. We therefore need some general way to
project the vector polarisation matrix data onto a complex scalar quantity. The
standard way to do this is through a Hermitian inner product of vectors s = x∗T
y. This has the advantage that it directly yields a scalar phase related to the
differences between x and y.
The second key idea is that we can always select an arbitrary polarimetric
scattering mechanism using the w vector formulation, introduced in Chapter
2 and shown again for N = 1, 2, 3 and 4-dimensional scattering in equation
(6.1). Conventional single channel ‘scalar’ interferometry then makes use of
w(1), dual and compact polarimetry w(2), full backscatter polarimetry w(3),
and bistatic polarimetry w(4).
 
cos α eiφ1
w(1) = eiφ w(2) =
sin α eiφ2
 
  cos α eiφ1 (6.1)
cos α eiφ1  sin α cos ψ eiφ2 
w(3) = sin α cos ψ eiφ2  w(4) =  
 sin α sin ψ cos γ eiφ3 
sin α sin ψ eiφ3
sin α sin ψ sin γ eiφ4

Combining these two ideas leads us to the following general procedure for gen-
erating a vector interferogram. We first project the complex scattering vectors
k 1 and k 2 , measured at ends 1 and 2 of the baseline, onto the conjugate of
 6.1 Vector formulation of radar interferometry 235

the desired polarimetric scattering mechanisms w1 and w2 (which importantly

may be different polarisations at either end of the baseline). These projections
provide two complex scalars s1 and s2 , representing the complex scattering
components that can then be combined into an interferogram, from which we
can estimate the corresponding phase using a standard Hermitian inner product
for complex vectors, as shown in equation (6.2):

s1 = w∗T
1 .k 1 ⇒ φ = arg(s s∗ ) = arg(w ∗T .k k ∗T w ) (6.2)
s2 = w∗T
2 .k 2
1 2 1 1 2 2

This expression is quite general, applying to N = 1, 2, 3 or 4 depolarisation

problems. The most common form used in radar is the N = 3 case (backscat-
ter with reciprocity), explicitly shown for reference in the Pauli base in
equation (6.3):

(shh
1 + s1 ) (shh
1 − s1 ) √ 1 
∗
s1 = w1,1 √ vv ∗
+ w1,2 √ vv
+ w1,3 2shv = w1 .k 1 
∗ ∗T 

2 2
(shh
2 + s2 ) 2 − s2 )
(shh √ 2 
 (6.3)
∗
s2 = w2,1 √ vv + w2,2 ∗ ∗
√ vv + w2,3 2shv = w∗T 
2 .k 2 
2 2
⇒ φ = arg(s1 s2∗ ) = arg(w∗T ∗T
1 .k 1 k 2 w 2 )

Before proceeding, one important required piece of housekeeping is that we

do not want the phase of the interferogram to depend on the arbitrary phase
difference between the complex vectors w1 and w2 , and we therefore enforce
the additional normalization constraint shown in equation (6.4):

φw = arg w∗T 1 w2 = 0 (6.4)

This is automatically satisfied if we choose w1 = w2 , but in the general case

must be explicitly enforced by modifying w2 , as shown in equation (6.5):
∗T
w2 → e−i arg(w1 w2 )
w2 (6.5)

In this way we can include polarimetry with interferometry in a consistent, com-

plete and logical manner for any dimension of depolarisation. Importantly, this
same approach can then be combined with averaging to predict the coherence
of the interferogram for polarisations w1 and w2 , as shown in equation (6.6):
.
s1 = w∗T
1 .k 1  E(s1 s2∗ )
⇒ γ̃ w1 , w2 = 0 ≤ |γ̃ | ≤ 1 (6.6)
s2 = w∗T
2 .k 2
E(s1 s1∗ ). E(s2 s2∗ )

6.1.1 Generalized coherency matrix formulation

We can reformulate this procedure for complex coherence estimation using
matrices, as shown in equation (6.7). The advantage of doing this is that we
can then easily extend the idea to multiple baselines and also provide a for-
mal link with our ideas about wave depolarisation and coherence in different
dimensions. The basic idea is to stack the coherent scattering vectors for each
end of the baseline, k 1 and k 2 , into a single column vector. The coherency
matrix is then formed from the average product of this vector with its conjugate
236 Polarimetric interferometry

transpose, called [2 ], where the subscript 2 now refers to the number of spatial
positions used.
=  >  
k 1  ∗T  T11 12
[2 ] = . k1 k ∗T =
k2 2 ∗T
12 T22
w∗T
1 12 w 2
⇒ γ̃ (w1 , w2 ) =  
w∗T
1 T11 w 1 . w∗T
2 T22 w 2
(6.7)

For N-dimensional depolarisation problems [2 ] is a 2N × 2N Hermitian

matrix. In the general multibaseline case, where M spatial positions are
available, the matrix [M ] becomes MN × MN in size.
We can make one further structural observation about the general Hermi-
tian matrix [M ]. It is always composed of N × N sub-matrices, as shown
in equation (6.8), where the M diagonal blocks T ii represent the polarimet-
ric information at each of the M spatial positions. The information in these
matrices can be interpreted using any of the depolarisation techniques (such as
entropy/alpha) discussed in Chapters 2 and 4.
 
T11 12 ... 1M
   ∗ T22 ... 2M 
T1 12  12 
[1 ] = [T ] → [2 ] = → [M ] =  . .. .. 
∗12 T2  .. .
..
. . 
∗1M ∗2M . . . TMM
(6.8)
 
Here our interest centres more on the new N × N complex matrices ij ,
which contain information related to the variation of interferometric phase with
polarisation. These matrices are neither Hermitian nor unitary, and hence have
a general 3 × 3 complex structure. We can see that these block elements play
an important and separable role in determining the coherence, as shown at right
in equation (6.7). Under a unitary change of base of the scattering vector k by
an N × N unitary matrix, [2 ] then transforms as shown in equation (6.9):
     
T11 12 UN 0 T11 12 UN∗T 0
k = [UN ] k ⇒ =
∗T
12 T22 0 UN ∗T
12 T22 0 UN∗T
(6.9)

From here we can then start by considering single baseline polarimetric inter-
ferometry (SBPI) [2 ], which involves measurements at only two separated
spatial/temporal positions. Here the T ii matrices can still have polarimetric
dimension N = 1, 2, 3 or 4, but the matrix 12 contains additional information
about the variation of interferometric coherence. This procedure can then be
easily extended to multiple baselines (and frequencies), as shown in equation
(6.8) (Ferro-Famil, 2001, 2008).
Considering the special but important case of radar backscatter N = 3, there
are several special cases of unitary change of base U 3 to be distinguished.
Equation (6.9) is expressed in the linear Pauli basis (see equation
√ (6.3)). If we
wish to convert to the standard linear lexicographic base of HH, 2HV and VV,
 6.1 Vector formulation of radar interferometry 237

to predict the interferometric coherence in these channels, then we can employ

the following unitary matrix in equation (6.9) (see equation (2.45)):
 
1 1 √0
1 
UN = ULP3 =√ 0 0 2 (6.10)
2 1 −1 0

We can then focus attention on a general change of wave base states: movement
of a reference point P over the surface of the Poincaré sphere. If the spherical
triangle coordinates of P are α w and δ w (see Figure 1.12), then the general
unitary matrix for use in equation (6.9) takes the form shown in equation (6.11):
 
cos αw
P= ⇒ [U3 ] = [U3L ][ULP3 ]
sin αw eiδw
 √ 
cos2 αw − 2 cos αw sin αw e−iδw sin2 αw e−i2δw
 √ √ 
⇒ [U3L ] =  2 cos αw sin αw eiδw cos2 αw − sin2 αw − 2 cos αw sin αw e−iδw 
√
sin αw e w
2 i2δ 2 cos αw sin αw e w
iδ cos αw
2

(6.11)

For example, if we want to convert from the linear H,V √ basis to left and right
circular L,R so as to obtain the matrix in the basis LL, 2LR,RR, then we would
set α w = π /4, δ w = π /2 in equation (6.11) and obtain the composite change of
basis matrix shown in equation (6.12). This can then be used in equation (6.9)
to express all matrices in the circular basis.
 √     
0 1 1 √1 −1
2i √ 1 1 √0 √0 1 i
1 1
U3circ =  2i √0 2i √ 0 0 2 = √  2i 0 0
2 2 2
−1 2i 1 1 −1 0 0 −1 i
(6.12)

Note, however, that equation (6.11) does not represent the most general unitary
transformation. It has only two free parameters, while the general 3 × 3 unitary
matrix has nine (see Appendix 2). These extra degrees of freedom are gener-
ated by combining triplets of orthogonal scattering mechanisms, as used in the
eigenvector decomposition of [T ], for example (see Section 4.22). Starting with
an arbitrary mechanism, we have five degrees of freedom; the second then must
be orthogonal to the first, and so has 5 − 2 = 3 parameters. The third must then
be orthogonal to the first two, and so has 5 − 4 = 1 parameter, producing nine
in total. If these are then combined into a special unitary matrix (det(U 3 ) = 1)
this reduces to eight parameters. The set of general unitary transformations is
then governed by the eight Gell–Mann matrices, as shown in equation (6.13)
(see Appendix 2) (Cloude, 1995b; Ferro-Famil, 2000).
  det(U3 )=1
[U3 ] = w1 w2 w3 −→ [U3 ] = exp(iφn.G) (6.13)

The key conclusion is that arbitrary unitary matrices in the change of base
formulation—equation (6.9)—can be given a clear physical interpretation in
terms of triplets of polarimetric scattering mechanisms. The change of wave
polarisation base then forms only a subset of these matrices through equation
238 Polarimetric interferometry

Polarisation selection/w vector w1 w2 w3

HH Ö2 Ö2 0
HV 0 0 1
VV Ö2 –Ö2 0
HH+VV 1 0 0
HH+VV 0 1 0
2HV 0 0 1
LL 0 Ö2 Ö2i
LR 1 0 0
Fig. 6.1 Example scattering mechanisms
used for POLInSAR RR 0 –Ö2 Ö2i

(6.11). This will be important when we come to consider coherence optimization

in Section 6.2, as we can then allow unconstrained search through all available
parameters of the unitary matrix.
While these formal unitary transformations are useful for analytical manip-
ulations, in practice we are very often concerned only with direct evaluation
of the coherence (equation (6.7)) for different polarisations. In this case we
can first estimate the matrices in a fixed basis (the Pauli basis of equation
(6.3), for example), and then use diversity of w to generate the different polar-
isations. The weight vectors w1 and w2 then define user-selected scattering
mechanisms at ends 1 and 2 of the across-track baseline. Figure 6.1 shows
some important examples of the weight vector w = (w1 ,w2 ,w3 )T for coherence
estimation in the commonly used linear, Pauli and circular bases. This table can
be used with equation (6.7) to generate interferograms in different polarisation
channels.
Consequently, in applications we first need to estimate the composite matrices
of equation (6.8) from the radar data itself. Given L samples of MN dimensional
scattering vectors u, the estimate [Z] of [] is then conveniently formed using
a maximum likelihood (ML) estimator, as shown in equation (6.14):

1 ! ∗T
L
[Z] = uj uj (6.14)
L
j=1

For finite L there will be errors in this estimate relating to higher-dimensional

forms of coherence bias, as discussed in Appendix 3. To illustrate this, consider
a numerical example from single baseline polarimetric radar interferometry
(SBPI), when MN = 6, as shown in equation (6.15):
 
2 0 0 1.8 0 0
π
0 0.6ei 4 
 1 0 0 0 
 π 
0 0 1 0 0 0.4ei 2 
[2 ] = 
1.8

 (6.15)
 0 0 2 0 0 
 π 
0 0.6e−i 4 0 0 1 0 
π
−i 2
0 0 0.4e 0 0 1

This matrix corresponds physically to scattering by a random dipole cloud

with polarisation dependent complex interferometric coherences of magnitude
0.9, 0.6 and 0.4 respectively, and with separated interferometric phase centres
 6.1 Vector formulation of radar interferometry 239

Numerical estimation of coherence (mean of 256 realizations)

1

0.9

0.8

0.7

0.6
Coherence

0.5

0.4

0.3

0.2

0.1

0
5 10 15 20 25 30 35 40 45 50
Fig. 6.2 Estimation of coherence triplet of
Number of looks (L) equation (6.15) versus number of looks L

of 0, π /4 and π /2. If we now consider numerical estimation of these three

coherence amplitudes as a function of number of looks L, using the Monte
Carlo data simulation technique described in Appendix 3, we obtain the typical
convergence shown in Figure 6.2. Each point in this graph (for a fixed L) is
obtained as the mean of 256 realizations of the randomized estimation process.
We see a small noise variation due to the finite sampling, but can still see
the general behaviour expected of coherence estimation: namely, coherence
bias for a small number of looks, which reduces as L increases. This bias also
increases with decreasing coherence, so we see the 0.9 channel has little bias,
and we obtain accurate estimates for a small number L > 5 looks. The low 0.4
channel, however, shows much slower convergence, and requires in excess of
25 looks for ‘good’ estimation.
This bias is due to finite sampling, which also has an impact on the esti-
mation of depolarisation parameters of the cloud. For example, the scattering
entropy of the cloud of dipoles is H = 0.946 (see Figure 3.29). We can estimate
this entropy again as a function of L by isolating the T11 component of the
estimated coherency matrix, performing an eigenvalue analysis and then calcu-
lating entropy. When we do this as a function of number of looks L, we obtain
the estimates shown in Figure 6.3. Note that in this case the entropy estimate is
underestimated for a small number of looks, and only slowly converges to the
correct value (we obtain around 5% relative error for L > 20 looks). We note
that this bias depends on the underlying entropy. For low entropy scattering
(with one dominant eigenvalue) the convergence accelerates, and the multi-
looking requirements are much reduced. From this point of view a dipole cloud
represents an extreme case of depolarisation, and hence represents an upper
bound on the bias issues for single scattering problems.
Nonetheless, such numerical biases have to be considered when dealing with
practical questions of the variation of coherence with polarisation, as clearly
the apparent dynamic range will always depend on L and we must ensure that
240 Polarimetric interferometry

Estimated scattering entropy of dipole cloud

1

0.9

0.8

Scattering entropy
0.7

0.6

0.5

0.4
0 10 20 30 40 50
Fig. 6.3 Estimated scattering entropy versus
number of looks for dipole cloud Number of looks (L)

L is sufficiently large to minimize any numerical bias. We now turn to consider

the issue of quantifying the range of coherence variation with polarisation by
employing systematic optimization techniques based on Lagrange multipliers.

6.2 Coherence optimization

A fundamental question of importance in polarimetric interferometry is to deter-
mine the maximum interferometric coherence change with polarisation. If it
changes only slightly, then polarimetry plays only a weak role. On the other
hand, if the coherence varies strongly with polarisation then this indicates
important changes in the relative positions of scattering mechanisms, which
we can then exploit for parameter estimation. A quantitative approach to this
estimation can be made based on the mathematics of optimization theory to
which we now turn (Cloude, 1997b, 1998; Tabb, 2001, 2002a, 2002b; Pascual,
2002; Colin, 2005, 2006; Neumann, 2008).
We first investigate this question by using a formal Lagrange multiplier opti-
mization process as follows. Our starting point is the general expression for
complex coherence, conveniently written in terms of sub-matrices, as shown
in equation (6.16):.

 w∗T
1 12 w 2
γ̃ w1 , w2 =  → max |γ̃ | (6.16)
w1 w2
w∗T ∗T
1 T11 w 1 .w 2 T22 w 2

This represents the coherence obtained when polarisation w1 is used at the first
and w2 at the second end of the baseline. In general, therefore, it combines both
interferometric and polarimetric contributions to coherence. Our objective is
to find the extreme values of the magnitude of this function. There is a slight
complication in that equation (6.16) is a complex function, and so we must
 6.2 Coherence optimization 241

decide whether to maximize the absolute value, the phase, or real and imaginary
parts. These various choices open up different forms of optimization as we now
consider.

6.2.1 Unconstrained optimization

We start by determining which scattering mechanisms w1 and w2 maximize the
magnitude of the interferometric coherence (Cloude, 1997b). To answer this,
we set up a (complex) Lagrangian function L, as shown in equation (6.17).
This function comprizes the numerator of the coherence constrained by two
Lagrange parameters λ1 and λ2 , which permit variation of the numerator while
keeping the denominator constant. In this way we can find the extreme values
of the magnitude LL* by setting the complex partial derivatives of L and L* to
zero, as shown in equation (6.17):
 
L = w∗T ∗T ∗T
1 12 w 2 + λ1 w 1 T11 w 1 − 1 + λ2 w 2 T22 w 2 − 1
 ∂L

 ∂w∗T = 12 w2 + λ1 T11 w1 = 0

(6.17)
⇒ 1
∗

 ∂L
 ∗T = ∗T ∗
12 w 1 + λ2 T22 w 2 = 0
∂w2

This yields a set of coupled equations for the unknown vectors w1 and w2 and
the Lagrange multipliers λ1 and λ2 . There are now two important options for
solution of these equations. In the most general case we allow w1 and w2 to
be different and so allow full polarisation diversity. In this case we can find a
solution to the coupled equations as a pair of eigenvalue problems, as shown
in equation (6.18):
 −1  2
w1 =w2
T22 ∗T −1 ∗
12 T11 12 w 2 = λ1 λ2 w 2 K1 w1 = νw1 = γopt  w1
−→ λ1 = λ2 = γ̃opt ⇒  2
 −1 −1 ∗T
T11 12 T22 12 w1 = λ1 λ∗2 w1 K2 w2 = νw2 = γopt  w2
(6.18)

This shows that the optimum scattering mechanisms can be obtained from
eigenvalue equations involving composite products of the elements of [y ].
Furthermore, the two Lagrange multipliers are complex but equal. To show this
we left multiply the top derivative equation in (6.17) by w∗T
1 and the lower by
w∗T
2 , and use the normalization condition on the Hermitian forms on the right-
hand side of L in equation (6.17) to show that λ1 =λ2 . Note that for backscatter
problems there are then three optimum values corresponding to the square mod-
uli of the three eigenvalues of K. Hence the 3 × 3 matrices K 1 and K 2 have the
same non-negative real eigenvalue spectra in the range 0 ≤ υ3 ≤ υ2 ≤ υ1 ≤ 1,
but different and non-orthogonal eigenvectors, as neither K 1 nor K 2 are gen-
erally Hermitian or unitary matrices. The maximum coherence is given by the
square root of the largest eigenvalue ν 1 with corresponding scattering mecha-
nisms given by the eigenvectors. Note that the optimum complex coherence can
be found by first calculating the eigenvectors w1 and w2 from equation (6.18),
phase normalizing using equation (6.4), and then using them directly in equa-
tion (6.16). Alternatively we may calculate the optimum directly by solving
the following generalized eigenvalue problem, obtained by a straightforward
242 Polarimetric interferometry

rewriting of the derivative equations in (6.17) and using the fact that λ1 = λ2 ,
as shown in equation (6.19):
     
0 12 w1 T11 0 w1 ∗T
= λ ⇒ γ̃opt = λmax e−i(arg(w1 w2 ))
∗T
12 0 w 2 0 T 22 w 2
(6.19)

where λmax is the eigenvalue with maximum modulus. The advantage of this
formulation is that it scales naturally to the multi-baseline case, as first shown
in Neumann (2008). When M-tracks are available we must use the generalized
coherency matrix [M ] shown in equation (6.8). In this case the unconstrained
optimization problem can be formulated by generalising the Lagrangian to a
sum of numerators with constrained denominators, leading to the generalization
of equation (6.19), as shown in equation (6.20):

!
M !
M M 
!
L= w∗T
i ij w j + λ w∗T
i Tii w i − 1
i=1 j=i+1 i=1
     
0 12 . . . 1M w1 T11 0 ... 0 w1
 ∗T 0 . . . 2M   w2   0 T22 ... 0   w2 
 12     
⇒ . .. .. ..   ..  = λ  .. .. .. ..   .. 
 .. . . .   .   . . . .  . 
∗T
1M ∗T
2M . . . 0 w M 0 0 . . . TMM wM
(6.20)

Note that here λmax now corresponds to the weighted sum of optimized
coherence moduli—a type of average across all the baselines.

6.2.1.1 SVD interpretation of unconstrained optimization

We can obtain a useful physical interpretation of this optimization process by
reformulating it as a singular value decomposition (SVD) (see Appendix 1). The
starting point for this is to realize that we can always pre-whiten the polarimetric
scattering vectors; that is, we can transform them into a base with the identity as
a coherency matrix, corresponding to ‘white’ noise. This can be achieved using
a transformation involving the square root of the actual polarimetric coherency
matrix, which can best be evaluated in terms of its matrix of eigenvalues [D]
and eigenvectors [U ], as shown in equation (6.21). This represents a change
of polarisation base given by the matrix [U ] followed by a weighting of the
channels by the reciprocal of the square root of eigenvalues.
 
1 0 ··· 0
+ , 0 1 · · · 0
 
k n = T −1 k = D−1 [U ] k ⇒ k n .k n∗T = IN =  . . .
 .. .. . . ... 

0 0 0 1
 
−1 −1
⇒  = T11 12 T22
  8 ∗T
I3   w2 = λ1 λ∗2 w2
[2 ]noise = ∗T ⇒
 I3 .∗T w1 = λ1 λ∗2 w1
(6.21)
 6.2 Coherence optimization 243

More significant is the effect of this transformation on the polarimetric inter-

ferometry sub-matrix [12 ]. The transformation does not generate noise in the
interferogram, but yields a structured matrix
as shown in the lower part of
equation (6.21). The optimum states w1 and w2 are then given as the left and right
singular vectors of the matrix
as shown. They can be obtained from standard
eigenvalue problems for the Hermitian matrices

∗T and
∗T
respectively.
Coherence optimization can therefore be considered a problem in singular value
decomposition of the matrix
—physically the result of pre-whitened noise
interferometry between the polarimetric channels.

6.2.2 Constrained optimization

A second important form of coherence optimization first imposes the additional
constraint that w1 = w2 —that the scattering mechanisms at either end of the
baseline are equal (Tabb, 2001, 2002a; Colin, 2005, 2006). This is often sup-
ported by the physical argument that for small baselines the optimum scattering
mechanisms, in the absence of temporal changes, should be equal. However,
as we shall see, there are also good numerical as well as physical reasons for
adopting this approach in many applications.
In the constrained case the general optimization equations of (6.17) simplify
as shown in equation (6.22):
8
w1 =w2 12 w + λ1 T11 w = 0 
−→ ⇒ (T11 + T22 )−1 12 + ∗T ∗
12 w = −(λ1 + λ2 )w
∗T ∗
12 w + λ2 T22 w = 0
8 
[H ] = 12 12 eiφ + ∗T
12 e
−iφ
−1
[T ] [H ]w = λ (φ) w
[T ] = 12 (T11 + T22 )

max|λ(φ)| w∗T
opt [H ]w opt
−→ wopt ⇒ γopt = (6.22)
w∗T
opt [T ]w opt

We see again an eigenvalue equation, but this time based on averages of the
sub-matrices. However, one drawback of this approach is, as shown on the
right-hand side of equation (6.22), that it maximizes only the real part of the
eigenvalue and so represents a phase-sensitive optimization. Hence this finds
only a local maximum, and to find the true global optima we need to introduce
a free phase parameter exp(iφ). By then repeating the optimization in equation
(6.22) for different values of φ we can then obtain the global maxima. The
general procedure for constrained optimization is then summarized in the lower
portion of equation (6.22).
The optimization process is then formally equivalent to a mathematical prop-
erty called the numerical radius of an N × N complex matrix [A] (Murnaghan,
1932; Li, 1994; He, 1997; Mengi, 2005). This itself is defined from the field of
values F(A), as defined in equation (6.23):
2 ? ? 3
F(A) = w∗T [A]w, w ∈ C N , ?w? = 1 (6.23)

The numerical radius is then the radius of the smallest circle that contains the
field of values, as defined in equation (6.24):

r(A) = max {|z| , z ∈ F(A)} (6.24)

244 Polarimetric interferometry

In our context we can formally relate the numerical radius to coherence opti-
mization by first generating the constrained form of the Lagrangian function
LC , as shown in equation (6.25):

LC = w∗T 12 w + λ w∗T T w − 1 (6.25)

This is almost in the form required, and needs only a pre-whitening transfor-
mation to remove the polarisation dependence of the constraint equation (the
factor T). Using the square root of T as a basis transformation, we then obtain
the following modified form:

T −1 w ⇒ Lc = w∗T −2
12 T − 2 wn + λ w∗T
1 1
wn = n T n wn − 1 (6.26)

The optimization is therefore equivalent to finding the numerical radius of

the transformed polarimetric interferometric matrix
introduced in equation
(6.20), as shown in equation (6.27):

= T −1 12 T −1 ⇒ γopt = r() (6.27)

There are many theorems and algorithms in the mathematics literature dealing
with the concept of numerical radius (Murnaghan, 1932; Li, 1994; He, 1997;
Mengi, 2005). Unfortunately there are no general analytical solutions avail-
able, but various numerical iterative algorithms have been proposed—one of
which involves exactly the phase transformation and repeated eigensolution
approach represented in equation (6.22). Furthermore, it leads to a third impor-
tant approach to optimization, based not on coherence amplitude but on phase
difference or coherence separation, as we now consider.

6.2.3 Maximum coherence separation and the

coherence region
In the previous two sections we considered methods for finding the polarimetric
scattering mechanisms w1 and w2 that maximize the interferometric coherence
magnitude. Since the local phase variance in an interferogram is inversely
proportional to coherence, this optimization will, by definition, lead to the
interferogram with minimum phase noise. This important analytical result is
somewhat marred by the practical issues of coherence bias, as discussed in
Appendix 2 and demonstrated by example in Figure 6.2.
There is, however, a completely different approach to the optimization pro-
cedure. Instead of concerning ourselves with the local phase variance, we
often seek a pair of scattering mechanisms w1 and w2 that maximize not the
coherence amplitude but the separation of complex coherences in the complex
plane. Physically these might then represent separated phase centres in a veg-
etation layer, for example (see Chapter 7) (Flynn, 2002; Tabb, 2002a). The
first approach to this problem (Tabb, 2002a) was to develop an algorithm for
maximising the phase separation, without regard to the coherence magnitude.
However, this can cause problems when dealing with low coherence regions,
as it can be sensitive to any noise in the data. A slightly modified approach is to
consider the maximum separation of complex coherence values. This approach
leads to a useful algorithm for application in the RVOG and related models (see
Chapters 7 and 8). For this reason we consider this algorithm in more detail.
 6.2 Coherence optimization 245

This ‘optimization of separation’ can be conveniently formulated using the

constrained approach where w1 = w2 . In this case we found that the following
eigenvalue equation could be used to maximize the real part of the complex
coherence:
max|λmax (φ)−λmin (φ)|
φ
[T ]−1 [H (φ)]w = λ (φ) w −→ γ̃opt (6.28)
The desired maximum difference is then given by the maximum difference
between eigenvalues of this matrix. Again we need to employ a phase trans-
formation φ to ensure that we secure the global optimum separation. In this
case we obtain a pair of w vectors, wa and wb —one from each eigenvector
corresponding to the max/min eigenvalues. The two complex coherences can
then be explicitly evaluated, as shown in equation (6.29):

w∗T
a [H ]w a 
γ̃1 = ∗T 
wa [T ]wa 
⇒ γ̃ = |γ̃1 − γ̃2 | = γ̃opt (6.29)
w∗T [H ]wb 
γ̃2 = b∗T 

wb [T ]wb
In summary, we have seen that there are three main approaches to coherence
optimization in polarimetric interferometry:
1. The unconstrained amplitude optimization provides the most general
mathematical solution, yielding the minimum phase variance interfer-
ogram across independent polarimetric variations at either end of the
baseline.
2. The constrained amplitude approach yields a slightly sub-optimum solu-
tion, but one constrained to keep the polarimetry constant at either end
of the baseline.
3. The constrained approach also yields a complex separation optimization
to find the two scattering mechanisms with maximum interferometric
separability inside the unit circle.

6.2.3.1 The coherence region

We can provide a useful geometrical interpretation of these various concepts
using the coherence diagram. This is a unit circle representation of coherence
in the complex plane (Figure 6.4). The first concept we can then consider is
that of the coherence region inside this diagram (Flynn, 2002). For any given
polarimetric interferometry matrix 2 there will be some sub-region of the
whole unit circle that encloses all possible values of coherence (for all states
w). This is called the coherence region of the matrix 2 . We will see that in
some cases this region may in fact shrink to a point, while in others it can
include large parts of the circle. In general the shape and size of the region
are determined by the nature of the scattering processes. We will see later, in
Chapter 7, how to predict the limiting shape of the region for various canonical
surface and volume scattering problems. Then, in Chapter 8, we will show how
to use knowledge of the region shape to estimate important physical parameters
(such as vegetation height) from radar data.
First we demonstrate how the boundary of the region can be computed numer-
ically for the constrained case (w1 = w2 ) using the eigenvalue equation derived
246 Polarimetric interferometry

90 1
120 60
0.8

0.6
150 30
0.4
φ
0.2

180 0
Coherence region

210 330

240 300
Fig. 6.4 Definition of the coherence region
of a polarimetric interferometric matrix 2 270

in equation (6.22). For each value of φ this eigenvalue equation yields the
extreme values (through the maximum and minimum eigenvalues) of the real
part of the coherence. For each of these eigenvalues there corresponds an eigen-
vector, which can be used to estimate a corresponding complex coherence, as
shown in equation (6.30). These two coherences then define two points on
the boundary of the coherence region, as shown schematically in Figure 6.4.
Here we show an example elliptical coherence region (see equation (6.31)), and
show, for a specified value of φ, how we obtain two samples of the boundary.
By varying φ in the range 0 ≤ φ ≤ π we can then reconstruct the whole bound-
ary. This gives us a systematic way to visualize the boundary for an arbitrary
coherency matrix 2 .
 
[H ] = 12 12 eiφ + ∗T 12 e−iφ
[T ]−1 [H ]w = λw
 [T ] = 12 (T11 + T22 )


 w∗T
max 12 w max (6.30)
γmax (φ) = w∗T T w
λmax ,wmax 
−→ max max
⇒
λmin ,wmin 
 w∗T  w
−→ 
γmin (φ) =
min 12 min
∗T
wmin T wmin
As a more specific example, consider again the 2 matrix shown in equa-
tion (6.15). This has a region dominated by three points: the three diagonal
complex values of 12 . Figure 6.5 shows the corresponding triangular region
for this matrix defined by the three diagonal values as vertices. As we vary
the polarisation over all possible mechanisms the interferometric coherences
will always be contained within this triangle. The boundaries of the region
therefore define the various optimum values. In this case the constrained and
unconstrained maximum amplitudes are equal to the white vertex. The separa-
tion optimization yields the white and black vertices, with a maximum phase
difference of π /2.
 6.2 Coherence optimization 247

90 1
120 60
0.8

0.6
150 30
0.4

0.2

180 0

210 330

240 300
Fig. 6.5 Example coherence region for
270 example matrix in equation (6.15)

This provides a simple numerical example to illustrate the various optimiza-

tion schemes. The real utility of these algorithms, however, is their application
in the more general case, when the 12 matrix is full. Before considering
such cases and their relationship to scattering theory, we first develop a useful
subspace interpretation of the information contained in a full 12 matrix.

6.2.4 Subspace coherence region analysis: the SVD and

Schur decompositions
The field of values concept (equation (6.23)) applies to arbitrary matrix dimen-
sion, but takes on a particularly simple form for 2 × 2 complex matrices. The
field of values of any 2 × 2 matrix is an ellipse (for a formal proof see Mur-
naghan (1932)). More precisely, let a general 2 × 2 matrix A be defined as show
in equation (6.31):
   
a b ∗T λ1 δ
A= ⇒ A = [U2 ] [U2 ] (6.31)
c d 0 λ2

This matrix can, by Schur’s theorem (see Appendix 1), always be written in
terms of an upper diagonal form and a unitary matrix [U ], as shown on the
right-hand side of equation (6.31). Here λ1 and λ2 are the eigenvalues of A.
It then follows that the field of values of A is an ellipse with two foci given
by λ1 and λ2 and minor axis length |δ|. The corresponding major axis length
is given by |λ1 − λ2 |2 + |δ|2 . For the special case that δ = 0 (in which case
the matrix A is termed ‘normal’—it can be diagonalized by a unitary trans-
formation) we obtain a linear field of values, varying along a line stretching
between the two eigenvalues. We shall see in Chapter 7 that such a limiting
case plays an important role in the description of mixed surface and volume
scattering.
248 Polarimetric interferometry

There are clearly an infinite number of ways of generating such 2 × 2 matrices

in polarimetric interferometry, simply by choosing a pair of polarisation vectors
wX and wY (and in the unconstrained version, a different pair wW and wZ for the
other end of the baseline). These are then used to project the scattering vector
data k at each end of the baseline 1 and 2, which are used to generate a 4 × 4
projected (p) polarimetric interferometric coherency matrix, as shown (for the
constrained case) in equation (6.32):
9 ∗
: 9 ∗
: 9 ∗
: 9 ∗
:
 sx1 sx1 sx1 sy1 sx1 sx2 sx1 sy2
sx1 = w∗T 1
x k   

 9 : 9 : 9 : 9 :
1  1 1∗ ∗ 
sy sy   T p p 
  ∗ ∗
  
sy = wy k  |wx |=wy =1
1 ∗T  sy sx sy1 sy1 sy1 sx2 1 2
  11 12
−→ [J ] = 9 : 9 : 9 : 9 : = p∗T
∗T 2  2 1∗ ∗  
p
sx = wx k 
2
  sx sx sx2 sy1
∗
sx2 sx2
∗
sx sy 
2 2 T

  
12 22

2 9 : 9 : 9 : 9 :
sy2 = w∗T
y k ∗ ∗ ∗ ∗
sy2 sx1 sy2 sy1 sy2 sx2 2
sy sy 2

(6.32)

From this we can then generate a pre-whitened 2 × 2 matrix
p as shown in

equation (6.33):
1 p p  p
T= T11 + T22 ⇒ p = T −1 12 T −1 (6.33)
2
The field of values of this matrix (for all projection vectors) is always an ellipse.
The projection vectors can be chosen in different ways. One way is to use
compact polarimetry (see Section 9.3.4), whereby a single transmit polarisation
and (generally different) dual polarised receiver are used. A second is to employ
physical modelling of the scene to isolate a subspace of polarisations where the
desired phenomena (dihedral scattering in vegetation, for example) are isolated.
However, a third important way is to start with the full Quadpol [S] matrix data,
and then identify suitable subspaces by employing the Schur decomposition
itself.
Our starting point for a general subspace analysis is the 3 × 3 pre-whitened
polarimetric interferometric matrix
as defined in equation (6.34). The sin-
gular vector (SVD) and Schur techniques may then be directly related to
our two main approaches to optimization—the SVD suitable for an uncon-
strained approach to polarimetric interferometry, and the Schur for a constrained
optimization approach. These ideas are summarized in equation (6.34):
  


s1 0 0
[U ]∗T  0 s
 0  [V3 ] SVD

 3 2
 0 0 s3
p
 = T −1 12 T −1 =   (6.34)

 λ δ12 δ13


1

 [U ]  0 λ2 δ23  [U3 ] Schur
∗T
 3
0 0 λ3

In SVD we allow different vectors at either end of the baseline and obtain
the singular values s1 and so on, which we have shown in equation (6.20)
correspond directly to optimum coherences. In the Schur approach, however,
we constrain the decomposition to a single unitary matrix (corresponding to
equal vectors at either end of the baseline), which leads to an upper triangular
3 × 3 matrix, also shown in equation (6.34). However, following the Schur
 6.2 Coherence optimization 249

approach we can now consider a set of 2 × 2 sub-matrices of this upper diagonal

form in the knowledge that each will have an elliptical coherence region as
discussed above. For example, we can consider the subspace formed by the
pairings 1,2, 1,3 or 2,3 in equation (6.34). This can be useful if, for example,
there is noise in part of the subspace we wish to remove, or if we are seeking the
subspace with the most linear coherence region based on physical modelling
such as RVOG (see Section 7.4.2).
We can make the link between this approach and the general projection ideas
of equation (6.32) by noting that the unitary matrix [U ] obtained in the Schur
decomposition can be written as a set of three column vectors, u, corresponding
to projection vectors w by a basis transformation, as shown in equation (6.35):
1
w 1 = T − 2 u1
  1
[U3 ] = u1 u2 u3 ⇒ w = T − 2 u (6.35)
2 2
1
w3 = T−2 u 3

By setting the pair x,y equal to 1,2, 1,3, and 2,3 in equation (6.32), we then pro-
vide a link between the general Schur decomposition and projection approach.
Note that for each pairing we can directly calculate the shape of the coher-
ence region analytically, since it is always elliptical, with two foci λx and λy ,
   2  2

minor axis length δxy and major axis length λx − λy  + δxy  . As a sim-

ple example, consider again the matrix shown in equation (6.15), with a region
illustrated in Figure 6.5. In this case the
matrix has the following simple
form:
 
0.9 0 0
 π 
 =  0 0.6ei 4 0  (6.36)
π
0 0 0.4ei 2

Here the three subspace regions reduce to line segments joining the pairs of
eigenvalues, as shown in Figure 6.5. We now return to the issue of numerical
bias in the context of these new coherence optimization techniques.

6.2.5 Numerical bias in coherence optimization

In the previous section we developed some useful analytical results concerning
the issue of coherence optimization and its impact on determining the dynamic
range of interferometric coherence variation with polarisation. In this section
we briefly consider a practical issue, to be considered when applying these ideas
to measured radar data: the impact of coherence bias in matrix estimation, and
how it impacts on estimation of the optimum coherences (Touzi, 1999).
In practice we often have no knowledge of the detailed form of the coherency
matrix 2 , and must instead estimate it from experimental data. Adopting a
maximum likelihood approach, estimates can be made for the three sub-matrices
involved by averaging the scattering vector data, as shown in equation (6.37):

1! 1! 1!
L L L
T̂11 = k 1i .k ∗T
1i , T̂22 = k 2i .k ∗T
2i ,
ˆ 12 =
 k 1i .k ∗T
2i (6.37)
L L L
i=1 i=1 i=1
250 Polarimetric interferometry

The question now is, what is the influence of the number of samples (‘looks’ in
radar imaging terms) L on the coherence optimization algorithms? As L → ∞
we should obtain the true matrices (since in this limit the matrices converge to
their correct values), but for small L we can expect overestimation of coher-
ence and hence distortion of the coherence region. To analytically study the
effects of L on optimum coherence estimation is a difficult task (see Touzi,
1999; Lopez-Martinez, 2005), and here we therefore employ some illustrative
numerical simulations based on use of Monte Carlo simulations (see Appendix
3 for details) to illustrate the nature of the problems involved, and to form
some general conclusions about bias effects in optimization methods. We again
make use of the random volume scattering example shown in equation (6.15),
and this time use the simulated data to estimate the matrices using equation
(6.37) before applying the various constrained and unconstrained optimization
algorithms.
Figure 6.6 shows the results of applying the unconstrained optimization algo-
rithm to the estimated matrices (again each point formed from an average of
256 coherence estimates). We note that the bias issues are more severe than for
the standard coherence estimation (shown dashed, for reference, in Figure 6.6).
This reflects the underlying higher dimensionality of the general unconstrained
optimization process. Direct coherence estimation implies that we have a priori
knowledge of the w vectors—in this case just the Pauli scattering vectors—
and so can project before we undertake the coherence estimation to obtain the
improved convergence shown in the dashed lines in Figure 6.6. However, for
the unconstrained optimization process we not only have to estimate coherence
values but also do not know the projection vectors themselves. These too must
be estimated from the data. Hence we need to estimate a larger number of param-
eters from the data itself. This increased dimensionality requires an increased
number of looks for convergence. This provides a qualitative explanation of
the increased bias seen in Figure 6.6.

Optimum coherence (mean of 256 realizations)

1.1

0.9

0.8
Coherence

0.7

0.6

0.5

0.4

Fig. 6.6 Coherence bias in matrix estima-

0 20 40 60 80 100
tion of optimum coherence triplet in equation
(6.15) Number of looks (L)
 6.2 Coherence optimization 251

90 1
120 60
0.8

0.6
150 30
0.4

0.2

180 0

210 330

240 300 Fig. 6.7 Distortion of coherence region of

equation (6.15) arising from coherence bias
270 in matrix estimation method

We note also that the triplet of optimum states (the three eigenvalues of
[K]) have different bias issues. The first and second optima are overestimated
for small L, but the smallest optimum value is actually underestimated. This
correlated bias behaviour between eigenvalues means that the apparent dynamic
range of coherence variability with polarisation is overestimated for small L.
We see that it takes in excess of L = 50 looks for the estimate bias to settle
down, but note again slow convergence even beyond this point.
This overestimation of dynamic range is also apparent in the estimation of
the coherence region. If we employ coherency matrix estimates for L = 6 looks
and then L = 50 looks we obtain typical region estimates as shown in Figure 6.7.
Here, in grey we show the estimated boundary region for L = 6, and note its
overestimation compared to the true region (shown as the black triangle). For L
= 50 we see a much better estimate (shown in black), more accurately reflecting
the bounds of the true coherence region, and hence prociding a better estimate
of the true dynamic range.
We now turn to consider how the physical structure of surface and volume
scattering controls the size and shape of the coherence region.
 The coherence of
7 surface and volume
scattering

In this chapter we investigate in more detail the shape and structure of the
limiting form of the coherence region for vector surface and volume scatter-
ing problems. In Section 6.2.3.1 the coherence region is defined as the region
in the complex plane bounding the variation of interferometric coherence with
polarisation. Here we extend this idea to define a related concept: the coherence
loci, defined as the curves traced out by variation of interferometric coherence
with physical parameters of a scattering model (Papathanassiou, 2001; Cloude,
2003). Our objective is to relate the coherence loci as a limiting form of the
coherence region (as the number of looks tends to infinity) in order to estab-
lish strategies in Chapter 8 for using polarimetric interferometry for physical
parameter estimation.
We begin by looking at simple models of surface and volume-only scattering.
We then consider extension of these ideas to multilayer media which, impor-
tantly, will allow us to consider combinations of surface and volume effects.
We then look in detail at two important models widely used in the literature for
interpreting coherence diagram: the random-volume-over-ground, or RVOG
(Treuhaft, 1996, 2000a; Papathanassiou, 2001; Cloude, 2003), which is closely
related to an interferometric version of the water cloud model IWCM (Askne,
1997, 2003, 2007) (see Section 3.5.1); and the oriented-volume-over-ground,
or OVOG (Treuhaft, 1999; Cloude, 2000a). Both these models are character-
ized by having a small number of independent physical parameters, often fewer
than observables in the scattered field, so enabling consideration of methods
for estimation of these parameters from data (see Chapter 8).
As we shall see, both of these models (RVOG and OVOG) make assumptions
about the vertical variation of scattering in the layered media (through the struc-
ture function), which naturally leads us to consider a more general approach
termed coherence tomography (Cloude, 2006b, 2007a) that permits arbitrary
structure function and allows an efficient parameterization of the dependence
of coherence of changes in structure.
In general terms we note that the coherence loci must somehow be related to
variation of the vertical structure function f (z) with polarisation. For example,
if the scatterers in a scene do not change relative amplitude as w changes,
then the structure function, whatever shape it has, will be constant, and the
coherence will be constant with polarisation, so yielding a point coherence
loci. This point can then be stretched to a radial line by adding polarisation-
dependent temporal, or SNR, decorrelation, but the underlying physics will
be determined by a point in the complex diagram. We now investigate this
relationship between coherence loci shape and structure function variations in
more detail, for surface and volume scattering scenarios.
 7.1 Coherence loci for surface scattering 253

7.1 Coherence loci for surface scattering f (z)

The first issue we face is in defining surface scattering in the context of inter-
ferometry. By definition, surface scattering occurs at a discontinuity between
two media, and hence a good model for its vertical structure function would f (z) = (z – zo)
be a Dirac delta function located at the interface between the media, as shown
schematically in Figure 7.1, where the surface is clearly located at position
z = zo . However, we have seen that in microwave remote sensing of natural z
surfaces there is always some penetration of the wave into the lower medium,
Fig. 7.1 Idealized vertical structure function
depending on the effective dielectric constant (see Section 3.1.1.1), and hence for surface scattering
it is of interest to consider the circumstances under which this delta function
assumption is supported. In the context of interferometry, what is important
is not so much the absolute penetration depth but its value scaled to βz —the
vertical sensitivity of the interferometer. Hence by combining the definition of
penetration depth δp (equation (3.12)) with the baseline dependence of verti-
cal wavenumber βz , we can obtain a relationship between effective complex
material constant εr = ε − iε and baseline geometry, as shown in equation
(7.1). Here we set a threshold of 0.1 radians for the product, as this represents
a typical interferometric phase shift due to wave penetration of only 5◦ and a
maximum volume decorrelation of only 0.998. These are within the bounds
of estimation error for typical interferometer geometries, and thus represent a
somewhat arbitrary but realistic threshold for the delta function assumption to
apply.

β z δp
< 0.1
2
√ √ √
4π θ λ ε 2θ ε 2B⊥ ε
⇒ . = ≈ < 0.1
λ sin θ 2π ε sin θ ε Ro sin θ ε
√
ε Ro sin θ
⇒ < (7.1)
ε 20B⊥

If this inequality is satisfied (and we further assume that range spectral filtering
has been employed to remove any baseline decorrelation (see Section 5.1.1.1)),
the interferometric coherence can then be estimated as shown in equation (7.2),
where we have also, for the moment, ignored any temporal or SNR decorrelation
terms.

- hv
δ (z) eiβz z dz
γ̂ = e iβz zo o
- hv = eiβz zo = eiφo (7.2)
o δ (z) dz

We see a simple result, with a coherence of unity and phase depending on

the surface elevation. Turning now to the polarisation dependence, as we vary
polarisation so the backscatter amplitude from the surface will change. We
can propose this variation of backscatter as a reflection symmetric depolariser
(see Section 2.4.2.3), which has a polarimetric coherency matrix as shown in
equation (7.3). In the absence of temporal and SNR decorrelation we can now
calculate the optimum coherence values from the corresponding [K] matrix, as
254 The coherence of surface and volume scattering

shown in equation (7.3):

−1 −1 ∗T
K = T11 12 T11 12
 −1    −1  
t11 t12 0 t11 t12 0 t11 t12 0 t11 t12 0
∗  ∗  ∗  ∗ 
= t12 t22 0  eiφo t12 t22 0  . t12 t22 0  e−iφo t12 t22 0 
0 0 t33 0 0 t33 0 0 t33 0 0 t33
 
1 0 0
 
= 0 1 0 (7.3)
0 0 1

Here we see that [K] is just the identity matrix, indicating an interferometric
coherence of unity for all polarisation states. This is just a consequence of
the fact that although the absolute level of the delta function in the structure
function f (z) varies with changes in polarisation, its position remains fixed at
the surface boundary z = zo , and hence the coherence remains the same for all
polarisations. In this case the corresponding coherence region shrinks to a point
on the circumference of the unit circle (the angular position of which depends
on βz zo ).
In practice this result must be extended to include variations in SNR with
polarisation. The backscatter power from smooth surfaces can be very low,
especially in the crosspolarised channel, and this variation will be apparent as
a polarisation-dependent coherence, as shown in equation (7.4):

w∗T [T11 ] w 1
s/n = ⇒ γsnr (w) = (7.4)
n 1 + w∗T [Tn 11 ]w

where the noise power n can be estimated directly in reciprocal backscatter

problems as the smallest eigenvalue of the HV/VH N = 2 coherency matrix,
as shown in equation (7.5) (Hajnsek, 2001):
  
1 + ∗
, + ∗
, + ∗
, + ∗
,2 + ∗
,+ ∗
,
n= SHV SHV + SVH SVH − SHV SHV − SVH SVH + 4 SHV SVH SVH SHV
2
(7.5)

Temporal decorrelation can, of course, occur with surface changes between

passes in repeat-pass interferometry. However, there is no strong reason why
such changes should occur in a polarisation-sensitive way, and so we can realis-
tically model such effects as a scalar multiplier applied equally to all polarisation
channels. In this way our final expression for the polarisation dependence of
coherence in surface scattering scenarios takes the form shown in equation (7.6):

γ̂ = γt γSNR (w)eiφo (7.6)

This corresponds to a coherence loci given by a radial line segment in the

complex plane, as shown schematically in Figure 7.2. Note that although the
coherence amplitude can vary with polarisation, the average phase of the coher-
ence is constant, geometrically implied by the radial nature of the coherence
loci in the coherence plane. In this case the maximum coherence corresponds
to the polarisation with maximum signal-to-noise ratio, representing one end
of the loci, as shown in Figure 7.2. The other boundary of the loci corresponds
 7.2 Coherence loci for random volume scattering 255

90 1 90 1
120 60 120 60
0.8 0.8
0.6 0.6
150 30 150 30
0.4 0.4
0.2 0.2

180 0 180 0

210 330 210 330

Fig. 7.2 Coherence loci for surface scatter-

240 300 240 300 ing: ideal case (left), and with SNR and
270 270 temporal decorrelation (right)

similarly to the polarisation with minimum SNR. The whole line position (and
scaled line length) is dictated by the temporal decorrelation γt . At extremes,
when γt = 1, the line maximum can approach the unit circle at the point φo as
shown. When γt = 0, the whole line reduces to a point at the origin.

7.2 Coherence loci for random volume scattering

We now turn to consider determination of the coherence region for volume
scattering. We begin with the strongest polarisation symmetry assumption:
azimuthal symmetry, which leads to a random-volume approximation, for
which the polarisation coherency matrix is diagonal and of the general form
shown in equation (7.7):
 
1 0 0
[T ] = mv 0 s 0 0 ≤ s ≤ 1.0 (7.7)
0 0 s

where the absolute scattering cross-section mv is given, for example, by the

water cloud model (see Section 3.5.1), and s depends on particle shape and
varies for single scattering in the range 0.5 (dipole cloud) to 0 (spheres). This
strong symmetry assumption also has an important impact on the shape of the
coherence loci, as we now demonstrate.
Ignoring, for the moment, temporal and SNR effects, we can calculate the
optimum coherences from the [K] matrix, as shown in equation (7.8):
−1 −1 ∗T
K = T11 12 T11 12
   
1 1
0 0   0 0  
 t11   t11 
1    t11 0 0   t11 0 0
1    1  1   
= 0 0  I2  0 t22 0   0 0  I2∗  0 t22 0 
I1  t22  I1  t22 
 1  0 0 t33  1  0 0 t33
0 0 0 0
t33 t33
 
 2 1 0 0
 I2   
=   0 1 0 (7.8)
I1
0 0 1
256 The coherence of surface and volume scattering

Z We note that [K] is again a multiple of the identity matrix, indicating that
Top of layer zo + hv
the coherence does not change with polarisation and, as we found for surface
scattering, the coherence loci reduces to a point in the coherence diagram.
However, unlike surface scattering, the point does not lie on the unit circle.
Instead it lies within the circle at a point determined by the complex volume
f (z) = fv(z) decorrelation caused by the structure function f (z) (see Section 5.2.4). The
integral factors I1 and I2 in equation (7.8) can be expressed in terms of the
Surface position zo Legendre expansion of the structure function, as shown in equation (7.9). The
key consequence of the random symmetry assumption is that the Legendre
coefficients are independent of polarisation, and so the structure function, which
Fig. 7.3 Schematic representation of an can be arbitrary, as shown in Figure 7.3, must remain invariant to changes in
arbitrary vertical structure function polarisation.

hv


I2 = eiβz zo f (z )e iβz z
dz 


0  I2
⇒ γ̃ =

 I1
hv 

I1 = f (z )dz 
0
βz hv (1 + a0 )f0 + a1 f1 + a2 f2 + · · · + an fn
= eiβz z0 ei 2 (7.9)
(1 + a0 )

The effect of signal-to-noise ratio will be similar to that found for surfaces; that
is, to provide a polarisation-sensitive radial shift of the coherence towards the
origin. However, since volume scattering is generally more depolarising than
surface scattering, the variation of scattered power with polarisation will be
less, and hence SNR effects less important, than they are for surfaces. On the
other hand, temporal decorrelation can be much more important for volume
scattering, especially in vegetation applications, due to its susceptibility to
wind-driven motion on short time-scales. To further complicate issues, the
effects of temporal changes may not be uniform across the structure function.
For example, wind-blown motion may affect the top of the vegetation layer
more than the lower regions. To accommodate this we can modify the coherence
integrals to include in the numerator (I2 ) a new temporal structure function g(z),
as shown in equation (7.10):

hv 
I2 = eiβz zo g(z )f (z )e iβz z
dz 



0 I2
⇒ γ̃ = (7.10)
hv 
 I1


I1 = f (z )dz 
0

The function g(z) will vary between 0 and 1, being zero in regions of maximum
change and 1 for zero change. In terms of the Legendre expansion, g(z) will of
course have its own expansion coefficients, which in general will be different
from those of f (z), and hence the effect of temporal decorrelation must formally
be evaluated as a product of Legendre series in the numerator I2 . In the simplest
case (and the one most often used in the literature) we can assume that g(z) =
γt —a constant function with height, in which case the overall coherence can
 7.2 Coherence loci for random volume scattering 257

90 1 90 1
120 60 120 60
0.8 0.8
0.6 0.6
150 30 150
0.4 0.4 30
f = f0 + fb
0.2 0.2

180 0 180 0

210 330 210 330

Fig. 7.4 Coherence loci for random volume

240 300 240 300 scattering: ideal case (left), and with com-
270 270 bined SNR and temporal decorrelation (right)

be expressed as shown in equation (7.11):

βz hv (1 + a0 )f0 + a1 f1 + a2 f2 + · · · + an fn
γ̃ = γSNR (w)γt eiβz z0 ei 2 (7.11)
(1 + a0 )
Figure 7.4 summarizes the coherence loci for random volume scattering. Again,
as for surface scattering, it is represented by a radial line in the complex
coherence diagram.
There are two important differences between the surface and volume coher-
ence regions. The maximum coherence in the surface case (in the absence
of temporal and SNR effects) was on the unit circle (the point on the left of
Figure 7.3). However, in volume scattering, even in the limit that γt = γsnr = 1,
the maximum coherence no longer lies on the unit circle but somewhere inside—
the exact location depending on the baseline geometry and importantly on the
structure function of the volume scattering. The second difference is the pres-
ence of phase bias in the volume scattering case. The phase of the coherence
does not correspond to the bottom of the layer, but is offset by a term φb , the
value of which depends on the structure function. These observations provide
us with our first important link between coherence and important structural
parameters. We now consider two special cases: first the exponential structure
function, and then issues related to orientation effects in the volume.

7.2.1 Special case I: the exponential profile

We have seen that under azimuthal polarisation symmetry, the structure func-
tion for random volume scattering can be arbitrary, as long as it the same in
all polarisations. However, one special case is of interest because of its rela-
tion to physical models of propagation through an homogeneous layer. This
is the exponential profile, used in deriving the water cloud model (WCM) for
backscatter in Section 3.5.1 and in the study of volume decorrelation in Section
5.2.4.2. In this case we can evaluate the complex coherence explicitly, without
the need for a Legendre expansion, as shown in equation (7.12). We note that
the volume decorrelation is now a function of just two physical parameters: the
height of the layer (hv ), and the mean extinction σe . This gives us two physical
parameters to locate the coherence point in the complex plane. As this point is
specified by two measurements (amplitude and phase) there is a good match
between observables and unknown parameters. However, the match is spoilt by
258 The coherence of surface and volume scattering

the addition of the unknown phase of the surface φ(zo ). This acts essentially as
a new physical parameter (the location of the bottom of the layer) that must also
be estimated from the data. Hence we now have three unknowns and only two
observations. Nonetheless, this concept of reducing the number of parameters
required to describe the structure function so as to better match the number of
observations is an important one. We shall see in the case of layered media
that it leads us to a convenient solution for estimation of physical parameters
from data.
 I2
γ̃ (w) = γSNR w γt
I1
- hv
e− cos θo
2σe hv 2σe z
 e cos θo eiβz z dz
= γSNR w γt 0
- hv
e− cos θo
2σe hv 2σe z

0 e cos θo dz
 2σe eiφ(zo ) hv 2σe z
= γSNR w γt / θ
eiβz z e cos θo dz
cos θo (e e v
2σ h cos o − 1)
0

 2σe

 p=

 cos θ
 p ep1 hv − 1
= γSNR w γt where 1 p = p + iβz
p1 ephv − 1 


βz = 4π θ ≈ 4π Bn

λ sin θ λH tan θ

= γSNR w γt γ̃v (7.12)

One important form of the exponential structure approximation occurs when

we let the depth of the layer tend to infinity. In this case, taking the limits of
equation (7.12), and ignoring for the moment temporal and SNR decorrelation,
we obtain the following special form of the coherence for an infinitely thick
half-space.
 2σe

 p=

 cos θ
p ep1 hv − 1  lim hv →∞ p
γ̃ = eiβz zo p1 = p + iβz −→ γ̃ = eiβz zo eiβz hv
p1 e − 1 
ph v  p1


β = 4π θ ≈ 4π Bn
z
λ sin θ λH tan θ
1
⇒ γ̃ e−iβz (zo +hv ) = (7.13)
βz cos θ
1+i
2σe
In this case it makes more sense to shift the phase origin to the top of the volume
rather than the lower, in which case we obtain the following expression for the
complex coherence:
1
γ̃ (h∞ ) = (7.14)
1 + i βz2σ
cos θ
e

This is a function of only one physical parameter: the mean extinction. The
coherence loci for this model has a simple form, forming a semicircle in the
coherence plane, starting on the unit circle at the top phase reference for infi-
nite extinction (set to zero phase for convenience in Figure 7.5), and moving
 7.2 Coherence loci for random volume scattering 259

Coherence loci for infinite volume

90 1
120 60
0.8

0.6
150 30
0.4

0.2

180 0

210 330

240 300 Fig. 7.5 Coherence loci for semi-infinite ran-

dom volume scattering medium with varying
270 extinction

towards the origin (zero coherence) for zero extinction, as shown by the line in
Figure 7.5.
This represents one of the simplest possible models for volume decorrela-
tion, and when combined with a measured coherence (shown as the point in
Figure 7.5), provides a means for estimation of the mean extinction in the vol-
ume from the coherence magnitude of a single polarisation interferogram. (We
saw in the water cloud model (WMC), in Section 3.5.1, that this extinction is
often directly related to the water content mv , and hence we can often use this
extinction as a proxy for water content.) However, the assumption of an infi-
nite depth restricts this approach to applications where layer thickness greatly
exceeds wave penetration depth. Important examples are thick land-ice (Dall,
2003; Sharma, 2007) and high-frequency penetration of vegetation and snow;
but in general terms the assumptions of this model are not robust, and layer
thicknesses are often small compared to penetration, so that scattering from the
underlying bounding surface cannot be ignored. Treatment of these scenarios
will require multilayer scattering models to be developed in the next section, but
first we consider an important variation on the exponential structure function
assumption: the case of oriented volume scattering.

7.2.2 Special case II: oriented volume scattering

In many agricultural crop and ice remote sensing problems, the scatterers in a
volume may have residual orientation correlation due to their natural structure
(stalks in a wheat field, for example). The propagation of signals through such a
volume can no longer be assumed to be isotropic. Clearly, polarisations parallel
and perpendicular to the mean orientation axis will suffer different extinctions.
We considered a coherency matrix formulation of such propagation effects in
Section 4.2.6, and noted that essentially we again need to make an exponential
structure function approximation through such volumes, but now one where the
260 The coherence of surface and volume scattering

exponential coefficient itself varies as a function of polarisation. In this section

we consider the coherence loci for such oriented volume scattering (Treuhaft,
1999; Cloude, 2000a; Ballester-Berman, 2005, 2007).
In such cases the volume has two eigenpolarisation propagation states x and
y (which for an homogeneous channel (see Section 1.2.7) will be orthogonal).
Only along these eigenpolarisations is the propagation simple, in the sense that
the polarisation state does not change with penetration into the volume. If,
however, there is some mismatch between the wave polarisation coordinates
and the medium’s eigenstates, then there arises a complicated situation in which
the polarisation of the incident field changes as a function of distance into the
volume. Here we demonstrate that the coherence optimizer always obtains a
matched solution, and is thus useful in the application of parameter estimation
schemes to oriented volume scattering problems. It also leads to determination
of the coherence loci for such cases.
Essentially we now assume that the medium has backscatter reflection sym-
metry about the (unknown) axis of its eigenpolarisations (rather than azimuthal
symmetry as in the random volume case), and so we obtain a polarimetric
coherency matrix [T ], and from this the covariance matrix [C], for backscatter
(using the unitary transformation as derived in equation (7.16)), as shown in
equation (7.15):
   
t11 t12 0 −1
c11 0 c13
[C]=[ULP3 ][T ][ULP3 ]
[T ] = t12
∗ t22 0 ←→ [C] =  0 c22 0  (7.15)
0 0 t33 ∗
c13 0 c33
 
1 1 √0
1
[ULP3 ] = √ 0 0 2 (7.16)
2 1 −1 0

We can then also relate the [K] optimization matrices from equation (6.14) in
the two representations, as shown in equation (7.17):
−1 −1 ∗T
KT = T11 12 T11 12
−1 −1 −1 −1 −1 −1
⇒ KC = (ULP3 T11 ULP3 )(ULP3 12 ULP3 )(ULP3 T11 ULP3 )(ULP3 ∗T
12 ULP3 )
−1 −1
⇒ KC = ULP3 KT ULP3 ↔ KT = ULP3 KC ULP3 (7.17)

We make one further assumption: that the eigenpolarisations x and y are ortho-
gonal linear states; but we do allow for a mismatch in the angle between these
states and the radar coordinates by an angle ψ. We can now obtain an expression
for the coherency matrix [T11 ] = [T22 ] and interferometry matrix [12 ] for an
oriented volume extending from z = z0 to z = z0 + hv as vector volume
integrals shown in equations (7.18) and (7.19) (see equation (4.69):
8 .
hv (σx +σy )z
iφ(zo ) ∗
[12 ] = e R(2ψ) e iβz z
e cos θo P (τ ) TP(τ )dz R(−2ψ) (7.18)
0
8 .
hv (σx +σy )z
∗
[T11 ] = R (2ψ) e cos θo P (τ ) TP(τ )dz R (−2ψ) (7.19)
0
 7.2 Coherence loci for random volume scattering 261

where for clarity we have dropped the brackets around matrices and define the
following terms:
 
1 0 0
R(ψ) = 0 cos ψ sin ψ  (7.20)
0 − sin ψ cos ψ
   
 cosh τ sinh τ 0 t11 t12 0
P(τ )TP τ ∗ =  sinh τ cosh τ 0 t12 ∗ t22 0 
0 0 1 0 0 t33
 ∗ ∗

cosh τ sinh τ 0
×  sinh τ ∗ cosh τ ∗ 0 (7.21)
0 0 1
  z
τ = νz = κy − κx − iβo nx − ny (7.22)
cos θo
where κx,y are the amplitude extinction coefficients of the volume for x and y
polarisations. Note that if we cannot align the radar coordinates with the vol-
ume then the matrix term R(2ψ), which multiplies the whole matrix integral
expression inside the brackets, causes a coherent mixing of terms that is dif-
ficult to interpret. We will show that the polarimetric optimizer automatically
aligns the radar to the oriented volume. This result follows from knowledge of
the explicit form of the matrix [K], which for this problem enables direct calcu-
lation of its eigenvalues and eigenvectors, and hence optimization parameters
in closed form.

7.2.3 Optimum coherence values for oriented

volume scattering
To account for the effects of wave propagation on the polarimetric response of
an oriented volume, it is simpler to employ the covariance matrix [C] in the x/y
basis rather than the coherency matrix [T ]. Initially we set ψ = 0; that is, we
assume that the radar and medium eigenpropagation coordinates are aligned. In
this case we can explicitly invert the polarimetric covariance matrix as shown
in equation (7.23):
   
c11 I1 0 c13 I2 c33 I4 0 −c13 I2
1
C11 = 0 c22 I3 0  ⇒ C11
−1
=  0 f
c22 I3 0 
∗ I∗ f ∗ I∗
c13 2 0 c33 I4 −c13 2 0 c11 I1
(7.23)
∗ I I ∗ ) = (c c − c c∗ )I I , and similarly for
where f = (c11 c33 I1 I4 − c13 c13 2 2 11 33 13 13 1 4
the polarimetric interferometry matrix we can write the following factorization:
 
c11 I5 0 c13 I6
12 = ULP3 12 ULP3 −1
= eiφ(zo )  0 c22 I7 0  (7.24)
∗
c13 I8 0 c33 I9

The volume integrals I1 − I9 are defined in terms of the complex propagation

σ
constants βx = β0 nx − iκx = β0 nx − i σ2x and βy = β0 ny − iκy = β0 ny − i 2y for
262 The coherence of surface and volume scattering

the two eigenpolarisations and βz , the interferometric wavenumber, as follows:

h h ∗
h
I1 = e2σx z dz I2 = e2i(βy −βx )z dz I3 = e2(σx +σy )z dz
0 0 0
h h h ∗
I4 = e2σy z dz I5 = eiβz z e2σx z dz I6 = eiβz z e2i(βy −βx )z dz
0 0 0
h h h
−2i(βy∗ −βx )z
I7 = eiβz z e(σx +σy )z dz I8 = eiβz z e dz I9 = eiβz z e2σy z dz
0 0 0
(7.25)

Hence the first part of the optimization matrix [KC ] has the following form,
which is diagonal if I4 I6 − I2 I9 = I8 I1 − I2∗ I5 = 0.

 
c33 I4 0 −c13 I2  
eiφ(z0 )   c11 I5 0 c13 I6
−1  f  
C11 12 =  0 0  0 c22 I7 0 
f  c22 I3 
∗ I
c13 0 c33 I9
∗ I∗
−c13 0 c11 I1 8
2
(7.26)

From equation (7.25) we can easily show that both equations are satisfied for
arbitrary medium parameters, as we have the following relationships:

h ∗
I 4 I6 = e2σy z eiβz z e2i(βy −βx )z dz = I2 I9
0
(7.27)
h
2σx z iβz −2i(βy∗ −βx )z
I8 I1 = e e e dz = I2∗ I5
0

Hence the product C −1 −1 −1 ∗T

11 12 is diagonal. It follows that K c = C 11 12 C 11 12
is also diagonal, which confirms that the optimum coherences are obtained
when the radar coordinates are aligned with the medium axes. Furthermore,
we can also find expressions for the complex diagonal values (the optimum
coherences), as shown in equation (7.28):

∗ I I )
(c11 c33 I4 I5 − c13 c13 2 8 I4 I5 2σx eiφ(zo ) h 2σx z
γ̃1 = ∗ = = f (σx ) = eiβz z e cos θo dz
(c11 c33 − c13 c13 )I1 I4 I1 I4 cos θo (e2σx hv / cos θo − 1) 0

I7 (σx + σy )eiφ(zo ) h (σx +σy )z

γ̃2 = = f (σx , σy ) = eiβz z e cos θo dz
I3 cos θo (e(σx σy )hv / cos θo − 1) 0
∗ I I ∗)
(c11 c33 I1 I9 − c13 c13 6 2 I1 I9 2σy eiφ(zo ) h 2σy z
γ̃3 = ∗ = = f (σy ) = eiβz z e cos θo dz
(c11 c33 − c13 c13 )I1 I4 I1 I4 cos θo (e2σy hv / cos θo − 1) 0
(7.28)

By using the relationship between [T ] and [C], the eigenvectors of K c and

K T = T −1 −1 ∗T
11 12 T 11 12 are orthogonal, and given as shown in equation (7.29).
Any mismatch between the radar and medium principal axis (the angle ψ) can
 7.2 Coherence loci for random volume scattering 263

now be corrected by an inverse rotation of the eigenvectors of K T , as shown in

equation (7.30):
     
1 0 1
eigenvectors 1   1  
KT ←→ w1 = √ −1 w2 = 0 w3 = √ 1
2 0 1 2 0

     
1 0 0
eigenvectors
KC ←→ w1 = 0 w2 = 0 w3 = 1 (7.29)
0 1 0

     
1 0 1
eigenvectors 1     1  
R(ψ)KT R(−ψ) ←→ w1 = √ − cos 2ψ  w2 =  sin 2ψ  w3 = √ − cos 2ψ 
2 2
sin 2ψ cos 2ψ − sin 2ψ
(7.30)

We see that the eigenvectors of [K] are orthogonal for oriented volume scat-
tering, and also contain information about the orientation of the medium’s
eigenpolarisations, while the eigenvalues are related to the coherences for the
corresponding eigenwave extinctions. As expected from physical arguments, g~yy > g~xy > g~xx –> Ο
the highest (lowest) coherence is obtained for the polarisation with the high-
est (lowest) extinction. The higher the extinction, the less penetration into the
volume and hence the lower the effective volume decorrelation. This connec-
tion between the structure function and optimum coherences is summarized in Fig. 7.6 Summary of physical interpretation
Figure 7.6. Importantly, this provides our first example of an extended, non- of optimum coherence triplet for oriented
trivial coherence loci. Figure 7.7 shows how the loci, defined by the three volume scattering
optimum coherence points, can be constructed. Note the following important
features:
1) The three optima are rank ordered in coherence amplitude (radius inside
the unit circle), with the highest coherence associated with the highest
extinction propagation channel.
2) The three optima are also ranked in phase. If we take the phase of the
bottom of the layer (z = zo ) as reference (shown as the black point

90 1 90 1
120 60 120 60
0.8 0.8
0.6 0.6
150 30 150
0.4 0.4 30

0.2 0.2

180 0 180 0

210 330 210 330

Fig. 7.7 Coherence loci for oriented volume

240 300 240 300 scattering: ideal case (left), and including
270 270 SNR or temporal decorrelation (right)
264 The coherence of surface and volume scattering

90 1
120 60
0.8

0.6
150 30
0.4

0.2

180 0

210 330

Fig. 7.8 Coherence loci for an infinite half 240 300

space with oriented volume scattering and
varying extinctions 270

on the unit circle in Figure 7.7) then the lowest coherence amplitude is
always closest in phase to this point, followed by the crosspolarised chan-
nel, and then the highest coherence point always has the highest phase
shift.
3) The loci extension is caused by physical changes in volume decorrelation
with polarisation, and so far we have ignored effects due to temporal and
SNR decorrelation. These can be included in the analysis by allowing
radial shifts, so extending the loci towards the origin, forming a new loci
bounded by the dotted lines on the right side of Figure 7.7.

As an important special case we can consider the coherence region and loci for a
semi-infinite oriented volume. Figure 7.8 shows the semicircular loci developed
for the infinite random volume (equation (7.14)), and superimposed we show the
three optima for the oriented volume. The coherence corresponding in Figure
7.6 to the polarisation with maximum extinction has the highest coherence and
position closest to the top of the layer (which now corresponds to zero phase
in this diagram). The minimum extinction lies furthest in phase, while the
crosspolarised channel has a coherence intermediate in phase and amplitude
between the maximum and minimum extinctions.
From a parameter estimation perspective the finite slab oriented volume (OV)
model is interesting. We see that we have six observables (the phase and ampli-
tude of coherence in three optimum polarisations), and yet we have only four
unknowns (the layer depth, two values of extinction and phase of the bottom
of the layer). This is a good starting point for developing robust algorithms for
estimating these parameters from experimental data (see Chapter 8). However,
there is one major problem to be addressed, in that we have so far ignored
the presence of a ‘hard’ boundary behind the layer. In radar applications this
is often a soil or rock surface beneath vegetation or snow/ice layer, and as we
shall see, this considerably distorts the coherence loci shape. To consider such
 7.3 The coherence loci for a two-layer scattering model 265

issues we need to extend our approach to consider the coherence region for
two-layer scattering problems.

7.3 The coherence loci for a two-layer u

z = zo + hv z
scattering model
Layer 1
In this section we consider coherent scattering from a two-layer medium as
shown schematically in Figure 7.9. A wave is incident at angle θ to the normal, z = zo y
Layer 2
and first impinges on the top layer, which we assume is a volume scatterer z = zo – d
of depth hv . The bottom of this layer has position zo , defining the boundary
Fig. 7.9 Schematic representation of the
between the two layers. Below this extends a second medium, with depth d .
geometry of a two-layer scattering problem
In what follows we shall assume that the mean dielectric constant of layer 2 is
much greater than that of layer 1, and that d >> δp , the penetration depth in
layer 2. This has two important consequences for our analysis. Penetration into
layer 2 is small compared to the depth of layer 1. In addition, we assume that
the penetration into layer 2 is small enough to make the baseline scaled factor
small; that is, βz δp < 0.1 (see equation (7.1)). In this case there is no significant
volume decorrelation from layer 2. Although there may be negligible volume
decorrelation from layer 2, the large contrast in mean dielectric constant across
the boundary at z = zo implies that there will be a strong surface reflection. We
conclude that the influence of the second layer is to act as a hard boundary behind
the volume. As we shall see, however, this leads to significant complexity in
the coherence region, mainly because the reflection and scattering from this
boundary is polarisation sensitive, and also because of the complexities caused
by multiple scattering, as we now consider.
Figure 7.10 shows a schematic representation of the four principal scatter-
ing mechanisms to be considered. On the left is shown the two principal direct
mechanisms: volume backscattering from layer 1, and surface backscatter from
the rough boundary at z = zo . The first point to make is rather obvious from
this diagram, but evidently the surface component is seen through layer 1, and
hence the backscatter will depend not only on surface properties but also on
two-way propagation through layer 1. Even in these simple mechanisms we see
that the responses from the two layers are coupled in the final solution. This
coupling effect becomes more apparent when we add multiple scattering mech-
anisms shown on the right-hand side of Figure 7.10. In the simplest case we can
consider second-order interactions whereby backscatter can occur through two
cascaded specular reflections—first from the surface, and then from elements
in the scattering volume. Note that we actually have two scenarios to consider
in such effects—the first running from A to B, and then the time-reversed path
from B to A. It is the combination of these two mechanisms that maintains a

Direct Direct Surface/volume Surface/volume/surface

surface volume A B

Fig. 7.10 Schematic representation of single

and multiple scattering contributions in two-
layer volume-over-surface problem
266 The coherence of surface and volume scattering

symmetric backscatter matrix for reciprocal media (see Section 3.4.3). Impor-
tantly, even though these are second order, because the two can be specular
(forward scattering with angle of incidence equal to angle of reflection), such
second-order scattering contributions can be as large as, or larger than, the
direct backscatter mechanisms themselves. Hence we cannot ignore them for
a full development of the coherence loci for this two-layer problem. As we are
considering coherent scattering, we must also concern ourselves with the phase
of such second-order mechanisms.
P
Figure 7.11 shows a schematic representation of typical second-order scatter-
 ing from a volume scattering element P at height hp . We are concerned with the
S
P phase of the second-order signal from P (not the direct return) compared to R.

This will depend on the range difference relative to a direct surface return com-
hp Q R
ing from the surface at point R as shown. Point Q, being the specular reflection
Q R point on the surface, defines a triangle PQR as shown. It follows from the geom-
etry of this triangle, shown enlarged on the right-hand side, that the distance
Fig. 7.11 Ray geometry explanation of phase QS + SP = 2SR. Combining this with recognition that PQ represents a wave
centre location for dihedral scattering front of the incident plane wave, which by definition is an equi-range contour,
leads us to conclude that the range difference between P and R is zero for all
heights hp . This important result implies that in backscatter geometry, the phase
difference between the second-order and direct surface scattering components
is always zero.
There is, however, one further complication to be considered. If we now
extend our analysis to consider across-track radar interferometry, then the
second-order scattering behaves very differently for single and dual transmitter
modes. In dual transmitter mode (which includes repeat-pass interferometry as
a special case) we transmit and receive separately from each end of the base-
line B. From each position the second-order scattering effects (being exactly
in backscatter geometry) behave as shown in Figure 7.11, and consequently
the phase difference across the baseline will be exp(iβz zo )—the same as for
the direct surface scattering component from point R. Therefore, in dual trans-
mitter mode the second-order scattering effects behave like an effective and
additional surface component, with a phase corresponding to the underlying
surface position and a coherence of unity, indicating zero volume decorrela-
tion, even though the distributed volume is involved in the scattering mechanism
(see equation (7.31)).

γ̃sv2TX = eiφ(zo ) (7.31)

Note that this all follows from the special geometry of the triangle PQR in
Figure 7.11. In terms of modifications of the structure function, we note that
the second-order effects add an additional delta function contribution at z = zo .
Note also that from a polarimetric point of view, the second-order components
have a polarisation signature (scattering α > π/4) very different from direct
surface scatter (α < π/4). Hence, for dual transmitter modes we conclude
that second-order surface–volume interactions and direct surface scattering are
separable in polarimetry but not in interferometry.
Now consider the case of single transmit/dual receive interferometry. By
definition this is a configuration that involves a small but non-zero bistatic
scattering angle δθ, as shown schematically in Figure 7.12. For the end of
the baseline operating with both transmit and receive modes, the second-order
 7.3 The coherence loci for a two-layer scattering model 267

P P

hp hp


Q R Q R

Fig. 7.12 Ray diagrams for bistatic dihedral

Surface-volume component Volume-surface component scattering phase contributions

scattering will again have an exact backscatter geometry, and the effective
phase centre lies on the surface at R (shown by the dash line in Figure 7.12).
However, for the end of the baseline with a receive-only mode we obtain a
height-dependent phase shift due to the small bistatic geometry. As the volume
scattering elements are distributed over a range of heights from 0 to hv , there
will consequently be a volume decorrelation effect in this mode. To analyse
further, consider separately the geometry of the surface–volume and volume–
surface contributions shown in Figure 7.12. On the left we show the ray path
for the surface–volume term. The scattering into a small bistatic angle δθ gives
rise to a height-dependent phase given by equation (7.32). (Note the factor of
2π in place of 4π , because we are considering the single transmitter case.)

2π sin θ δθ 2π sin θ δθ
φSV (hp ) = hp = sin2 θβz hp φVS (hp ) = − hp
λ λ
= − sin2 θβz hp (7.32)

Here βz is the vertical wavenumber of the interferometer (assuming range spec-

tral filtering has been employed), and hp is the height of the volume scattering
element (see Treuhaft, 1996, 2000a).
A similar argument applies to the volume–surface component shown on the
u
right-hand side of Figure 7.12. However, this time we must consider the small
bistatic angle δθ as arising from the surface specular point rather than the vol- P
ume scatterer at P. In order to relate this to a height-dependent phase, we note
that the surface scattering appears, from the baseline point of view, to come hp
from a virtual point Pm which lies beneath the surface as shown in Figure 7.13,
Q R
so that the distance PQ equals Pm Q. This has the effect of changing the sign of
the interferometric phase as shown on the right in equation (7.32). These phase
variations will combine and lead, via integration across the full depth of layer –hp
1, to volume decorrelation. To calculate an expression for this decorrelation, u
we consider it as arising from an effective vertical profile function f2 (z), which Pm
in general will have a shape different from that of the single scattering volume
return fv (z). However, since the path length for second-order scattering through
Fig. 7.13 Ray diagram for effective dihedral
layer 1 is invariant to the height hp of P (and equal to 2hv / cos θ ), it follows scattering contributions
that for an homogeneous layer (with an exponential profile fv (z), for example)
the total extinction suffered by the wave is independent of the height hp (that
is, scattering elements at the top of layer 1 suffer propagation loss equal to
those from the bottom of the layer). Under these circumstances f2 (z) will be a
268 The coherence of surface and volume scattering

uniform profile ranging from –hv to hv from which we can calculate the decor-
relation caused by second-order scattering interactions for single transmitter
configurations as a SINC function, as shown in equation (7.33):
- hv i sin2 θβ z -0
dz + −hv ei sin θβz z dz
2

iφ(zo ) 0
e z

γ̃sv1TX =e -h
2 0 v dz
- hv 2
θβz z
iφ(zo ) −hv
ei sin dz
=e
2hv
sin(sin2 θβz hv )
= eiφ(zo ) (7.33)
sin2 θβz hv

Note that the mean phase centre still lies on the surface at z = zo (as it does
for the dual transmitter case). However, we now have a radial shift towards the
origin of the coherence diagram, with the amplitude of the coherence decreasing
with increasing depth of layer 1.
We have seen in the above that second-order interactions cause some com-
plexity in the analysis of coherent scattering from two-layer media. This is
further compounded when we realize that in theory there is an infinite cas-
cade of such higher-order surface–volume interactions to be considered. For
example, on the far right of Figure 7.10 is a typical third-order interaction of
surface–volume–surface scattering. Fortunately, for random media, such higher
order interactions are usually very small compared to first and second order.
Physically we can see that this arises because the backscatter level of such high
order interactions is determined by a cascaded product of small quantities. The
surface–volume–surface interaction, for example, involves the product of two
surface reflections, which will be small, all multiplied by the backscatter rather
than specular forward scatter from particles in the volume, which will generally
be smaller. Add to this the increased effective propagation distance inside the
lossy material of layer 1, and we can see, at least qualitatively, how interactions
higher than second order can often in practice be ignored (although there are
some notable exceptions such as scattering from complex man-made structures
such as bridges and buildings). Nonetheless, for most remote sensing applica-
tions this result will allow us to justify the use of simpler second-order models
in deriving the coherence loci for such problems.
First, however, for completeness we consider in detail the coherence prop-
erties of third-order surface–volume–surface contributions. For simplicity we
P
consider only the dual transmitter configuration. Figure 7.14 shows a schematic
of the geometry concerned. Again we are interested, for interferometry, in how
hp the effective range difference to a point P across the baseline varies as the height
u
R
hp is changed. In this case we see that the third-order effect has the same range
Q
variation as scattering from a virtual point Pm located a distance hp beneath the
surface, as shown. Therefore, even in the dual transmitter case we now obtain
–hp a height-dependent phase that will cause (complex) volume decorrelation and
a loss of coherence amplitude combined with a (negative) phase bias.
Pm We can calculate the level of this coherence by realizing that it has a cor-
responding structure function f3 (z) that is extended below the surface into the
Fig. 7.14 Ray diagram to locate phase centre range 0 < z < −hv . From this structure function we can then calculate the
for third-order scattering contribution corresponding complex coherence from an integral of the general form shown
 7.3 The coherence loci for a two-layer scattering model 269

z z z

z = z 0 + hv z = z0 + hv z = z0 + hv

fs2TX (z) fsv2TX (z)

fv2TX (z)
f(z) f(z) f(z)

z = z0 – hv z = z0 – hv z = z0 – hv

z z

z = z0 – hv z = z0 – hv

2TX
fsvs (z) fsv1TX (z)
f(z) f(z)

z = z0 – hv z = z0 – hv Fig. 7.15 Structure functions for various

contributions to two-layer scattering

in equation (7.34). Note that the superscripted N represents the number of trans-
mitters (1 or 2), and we have extended the range of the coherence integral from
−hv to hv to accommodate calculations for the virtual scattering points.
- hv NTX
iφ(zo ) −h
f (z )eiβz z dz
γ̃iNTX = e - v i  (7.34)
 hv NTX 
 −hv fi (z )dz 

Figure 7.15 summarizes the form of the structure functions for various com-
ponents of the two-layer scattering problem. In the first diagram we show the
direct volume term, fv2TX , which has some arbitrary shape, bounded by the
surface and the top height of the layer. Next we show the corresponding direct
surface return, fs2TX , which is a simple delta function located on the surface. For
the two-transmitter scenario this delta function also matches the second-order
surface–volume scattering contribution, fsv2TX . However, the single-transmitter
case has a uniform structure function extending across the full range, as shown
in the lower diagram fsv1TX . Finally we show the structure function for third-
2TX . This has a totally negative extent, but again can lead
order scattering fsvs
to volume decorrelation with an arbitrary structure function bounded by the
surface and top layer.
Of more general interest is the way in which these components combine to
provide the overall coherence variation for a two-layer problem. To see this we
need to incorporate all the mechanisms into a single generalized polarimetric
interferometric formulation. The starting point is to define the coherent scat-
tering vector k as the sum of contributions at ends 1 and 2 of the baseline, as
shown in equation (7.35). In these expressions, [P] is the vector propagation
270 The coherence of surface and volume scattering

matrix through layer 1 (see Section 4.2.6).


k 1 = k v1 + [Ps ]k s1 + [Psv ]k sv1 + [Psvs ]k svs1 + · · ·  averaging
 −→
k 2 = k v2 + [Ps ]k s2 + [Psv ]k sv2 + [Psvs ]k svs2 + · · · 
 + ∗T , + , + , + ,
 ∗T ∗T ∗T ∗T ∗T
 T11 = k 1 k 1 = k v1 k v1 + [Ps ] k s1 k s1 [Ps ] + [Psv ] k sv1 k sv1 [Psv ] + · · ·

⇒ + ∗T , + ∗T
, + ∗T
, ∗T
+ ∗T
, ∗T

12 = k 1 k 2 = k v1 k v2 + [Ps ] k s1 k s2 [Ps ] + [Psv ] k sv1 k sv2 [Psv ] + · · ·


T11 = TV + Ps Ts P ∗T + Psv Tsv P ∗T + · · ·
s sv
⇒ (7.35)
12 = v + Ps s P ∗T + Psv sv P ∗T + · · ·
s sv

We can then combine these vectors with averaging (which removes all
cross-products between mechanisms—an expression of independent scatter-
ing mechanisms) to express the generalized polarimetry and interferometry as
the sum of component matrices, as shown in equation (7.35). From these matri-
ces we can then determine a general expression for the observed coherence as
a function of polarisation, as shown in equation (7.36):
 w∗T 12 w w∗T (v + Ps s Ps∗T + Psv sv Psv
∗T + · · · )w
γ̃ w = ∗T = ∗T
w T11 w w (Tv + Ps Ts Ps∗T + Psv Tsv Psv
∗T + · · · )w

m0v (w)γ̃v (w) + ps (w)m0s (w)γ̃s (w) + psv (w)m0sv (w)γ̃sv (w) + · · ·
=
m0v (w) + ps (w)m0s (w) + psv (w)m0sv (w) + · · ·
(7.36)
Here we have rewritten each term as a product of three components: its nor-
malized radar cross section mo , its coherence contribution γ̃ , and a propagation
factor p that attenuates each contribution according to the propagation paths
involved. Note that we can write this expression as the product of total radar
backscatter times total observed coherence, as shown in equation (7.37):
γ̃ (w)m0 (w) = m0v (w)γ̃v (w) + ps (w)m0s (w)γ̃s (w)
+ psv (w)m0sv (w)γ̃sv (w) + · · · (7.37)
This gives us a procedure for deriving the coherence loci for two-layer prob-
lems, by which we first need to calculate the three elements for each mechanism,
and then combine them as shown in equation (7.36). As mentioned earlier,
for lossy layers this series converges quickly, and we need not consider the
complexity of scattering higher than second order. To illustrate this we now
develop three particular forms of this model that are widely used in the lit-
erature: a coherent two-layer version of the water cloud model (IWCM),
the closely related random-volume-over-ground or RVOG model, and the
oriented-volume-over-ground or OVOG model.

7.4 Important special cases: RVOG, IWCM

and OVOG
An important class of models can be generated by making the following
assumptions about the two-layer scattering problem:
 7.4 Important special cases: RVOG, IWCM and OVOG 271

1. Assume dual transmitter operation only (including repeat-pass as a

special case), so removing the coherence loss due to surface–volume
interactions. In this case the direct surface and surface–volume multiple
scattering contributions both have structure functions given by a Dirac
delta function.
2. Assume an exponential structure function for the direct volume return.
This amounts to the physical assumption of a layer of uniform density
characterized by a mean wave extinction σe , which may nonetheless be
a function of polarisation.
3. Assume that the layer is lossy enough and the surface rough enough that
third- and higher-order interactions can be ignored.
By allowing polarisation dependence of extinction we are essentially assuming
that layer 1 is an oriented volume, and this leads to the most complicated form of
such two–layer scenarios: the oriented-volume-over-ground (OVOG) model.
Before considering this case, however, we first develop a pair of models based
on the simpler assumption that the propagation is scalar and does not depend
on polarisation.
By assuming a random volume for layer 1, it follows that the propagation fac-
tors simplify, as they become independent of polarisation and are a function only
of a single mean extinction coefficient σe . There are two important models that
make use of this approach: the random-volume-over-ground (RVOG) model
(Treuhaft, 1996; Papathanassiou, 2001; Cloude, 2003), and the interferometric
water cloud model (IWCM) model (Askne, 2003, 2007). They differ primarily
in their assumptions about the importance of temporal decorrelation. In RVOG
it is common to assume that γt = γsnr = 1, which indicates a dominance of
volume decorrelation over all other sources; while for IWCM it is commonly
assumed that γt is dominant. RVOG is therefore better suited to single-pass or
low-frequency large spatial/low temporal baseline repeat pass interferometry,
while IWCM has been applied mainly to high-frequency small spatial/large
temporal baseline applications. We now examine the polarisation dependence
of each of these models, with a view to deriving their coherence loci.

7.4.1 The random-volume-over-ground (RVOG) model

z = z0 + hv
In the RVOG approach the structure function for the two-layer problem reduces
to the simple form shown in Figure 7.16. Note that the second-order (dihedral)
scattering effects are included as a coherent addition to the direct surface return.
Importantly, the polarisation dependence of coherence is now restricted to a
single term, as we now demonstrate. The RVOG model leads to a coherence as
shown in equation (7.38): z = z0

γ̃ (w)m0 (w) = m0v (w)γ̃v + ps m0s (w)eiφ(zo ) + psv m0sv (w)eiφ(zo )

f(z)
−iφ(zo )
iφ(zo ) m0v (w)γ̃v e + ps m0s (w) + psv m0sv (w)
⇒ γ̃ (w) = e Fig. 7.16 Composite structure function for
m0v (w) + ps m0s (w) + psv m0sv (w)
RVOG model
γ̃vo + µ(w)
⇒ γ̃ (w) = eiφ(zo ) (7.38)
1 + µ(w)
where µ is the ratio of effective surface-to-volume scattering. We also note that
the volume decorrelation component does not depend on polarisation, and is
272 The coherence of surface and volume scattering

given explicitly as shown in equation (7.39):


 2σe
p1 =
p1 (ep2 hv − 1)  cos θo
γ̃v e−iφ(zo ) = γ̃vo = (7.39)
p2 (ep1 hv 
− 1)  2σe
p2 = + iβz
cos θo

Significantly, only the parameter µ changes with polarisation. This is the ratio
of effective (sum of direct and second-order) surface-to-volume scattering. We
can develop an explicit form for µ as shown in equation (7.40):

 ps m0s (w) + psv m0sv (w)

µ w = (7.40)
m0v (w)

This can be further simplified by realizing that the propagation factors for direct
surface and second-order interactions are the same and given by equation (7.41):

ps = psv = e−2σe hv sec θ (7.41)

Moreover, we can express the volume scattering contribution in the denominator

of equation (7.40) as a function of the scalar extinction, layer depth and angle
of incidence (all independent of polarisation), and a polarisation-dependent
scattering cross-section, as shown in equation (7.42):

 cos θ  
mov w = 1 − e−2σe hv sec θ mv w (7.42)
2σe

Here mv has a corresponding diagonal coherency matrix, as shown in equation

(7.43), where s depends on the mean particle shape in the volume (s = 0 for
spheres 0.5 for prolate spheroids).
 
1 0 0
mv (w) = w∗T [Tv ] w = mHH +VV w∗T 0 s 0 w 0 ≤ s ≤ 0.5 (7.43)
0 0 s

From this we see that the cross-section can vary in the range mHH +VV ≤
mv (w) ≤ smHH +VV . For a dipole cloud, for example, we note there is only
a 3-dB variation of RCS with polarisation. The factor µ, however, can have a
much wider dynamic range than this, as shown in equation (7.44):

 2σe (m0s (w) + m0sv (w))e−2σe hv sec θ

µ w =
mv (w) cos θ (1 − e−2σe hv sec θ )
2σe m0s (w) + m0sv (w)
= (7.44)
cos θ (e2σe hv sec θ − 1) mv (w)

Here we see that the numerator depends directly on the variation of ‘effective’
surface scattering with polarisation. We can assume that this has reflec-
tion symmetry and a corresponding variation with polarisation, as shown in
 7.4 Important special cases: RVOG, IWCM and OVOG 273

equation (7.45):

m0s (w) + m0sv (w) = w∗T ([Ts ] + [Tsv ])w

= w∗T [Tes ] w
 
1 t12 0
= mHH +VV w∗T t12
∗ t22 0 w (7.45)
0 0 t33

Here the subscripted es denotes the effective surface components. The polarisa-
tions that maximize and minimize the µ ratio will be of interest in establishing
the coherence loci for RVOG. To find these we need to solve the following eigen-
value equation arising from optimization of the µ ratio, as shown in equation
(7.46) (see Section 4.2.2.2):

w∗T [Tes ] w
max ⇒ [Tv ]−1 [Tes ] wopt = λwopt (7.46)
w∗T [Tv ] w

Explicitly, we then obtain the following optimum ratio values as a function

of the mean particle shape and normalized effective surface coherency matrix
elements.
 −1  
1 0 0 1 t12 0
mhh+vv
s 0 s 0 t ∗ t22 0 
12
mhh+vv
v 0 0 s 0 0 t33
  " 
 2

 m hh+vv t22 t22 4 |t12 | 
2


 λ1 = shh+vv max 1 + ± 1− +

 2mv s s s



  " 
 2
⇒ λ = s m hh+vv t t 4 |t |2
12  (7.47)
min 1 +
22 22

 2 ± 1− +

 2m hh+vv s s s


v



 m hh+vv t33

λ3 = shh+vv
mv s

We note two important features of this solution:

1) Firstly, the optimum eigenvectors (of which there are three from equation
(7.46)) are not orthogonal (since T −1
v T es is neither symmetric nor Her-
mitian). Contrast this with the case of polarimetric interferometry for an
oriented volume, which yields a set of three orthogonal scattering mech-
anisms (see equation (7.29)). This is often an important signature of the
presence of multilayer scattering effects in polarimetric interferometry.
2) The second important point to note is that the optimum µ values are
given by the eigenvalues λ1 , λ2 and λ3 . The ratio λ1 /λ3 then gives a
measure of the maximum dynamic range of µ with polarisation. We
shall see below that this also impacts on the size of the coherence loci
for the RVOG model. Note that these optima will not in general occur for
a fixed polarisation basis (the Pauli basis, for example), as the structure
of [Tes ] will change with surface conditions (roughness, moisture, and
so on). Therefore, if we wish to make use of these optimum values for
274 The coherence of surface and volume scattering

parameter estimation, for example, then we need to employ a more adap-

tive processing approach, based on coherence optimization, to be able
to exploit these extreme values. We are now in a position to determine
the coherence loci for the RVOG model as follows.

7.4.2 Polarisation coherence loci for RVOG

The shape of the coherence loci for the RVOG model is best developed by
first rewriting the expression for RVOG coherence (equation (7.38)) in the
following form:
 
µ(w)
γ̃ (w) = eiφ(zo ) γ̃vo + (1 − γ̃ vo )
1 + µ(w)
= eiφ(zo ) (γ̃vo + F(w) (1 − γ̃vo )) (7.48)

Here we have deliberately isolated the polarisation dependence in a single term,

F(w). This factor is real non-negative, and lies in the range 0 ≤ F w ≤ 1,
with limits occurring at one end for pure volume scattering (µ = 0), and at
the other by pure surface scattering (µ = ∞). Hence F(w) is directly the
fraction of (effective) surface scattering in the observed signal. With γ̃vo a fixed
complex number, independent of polarisation, this, then, is the equation of a
straight line in the complex plane, going through the points γ̃vo and eiφ(zo ) , as
shown in Figure 7.17. The coherence loci for the RVOG model is therefore a
straight line.
Note that in practice not all of this line will be visible from experimental data,
and it is here that the dynamic range of µ becomes important. In reality there
will only ever be visible some limited segment of this line, corresponding to the
variation of F from µmin to µmax . Note importantly, however, that this line is not
radial, as the volume coherence is always complex, and thus there is a phase as
well as an amplitude variation with polarisation. Note also that the variation of

90 1
120 60
0.8
g~vo 0.6
150 30
0.4

0.2

180 0if (z )
e o

210 330

240 300
Fig. 7.17 Coherence loci for RVOG model 270
 7.4 Important special cases: RVOG, IWCM and OVOG 275

0.9 f = 0º

0.8

0.7 f = 90º
0.6
Coherence

0.5

0.4

0.3

0.2

0.1 f = 180º

0
–30 –20 –10 0 10 20 30
Fig. 7.18 Variation of coherence magnitude
Mu (dB) in RVOG model for various phase angles

coherence amplitude with increasing µ is not monotonic. As µ increases from

zero the coherence passes through a local minimum, the position of which
can be calculated exactly from the coherence expression in equation (7.48), as
shown in equation (7.49). The value of µ that produces the minimum coherence
(point of closest approach of the line to the origin of the coherence diagram) is
given by µ = 0, except when the parameter Fmin in equation (7.49) is greater
than zero. This is a function not only of the volume coherence amplitude, but
also its phase. Figure 7.18 shows an example for volume coherence of 0.8 with
30-degree steps in phase from 0◦ to 180◦ , at which point the line passes through
the origin and hence the minimum coherence is zero.

γ̃ (w) = eiφ(zo ) (γ̃vo + F(w) (1 − γ̃vo ))

 2
⇒ L = γ̃ (w) = a + bF + cF 2
dL b |γ̃vo |2 − Re(γ̃vo )
⇒ = b + 2cF ⇒ Fmin = − = 
dF 2c (1 − γ̃vo ) 1 − γ̃vo∗
   
   |γ̃vo |2 − Re(γ̃vo )   Im(γ̃ ) 
  vo 
 
⇒ γ̃ (w)min = γ̃vo +  (1 − γ̃vo ) =  
 (1 − γ̃vo ) 1 − γ̃vo∗ ∗ 
  1 − γ̃vo
) *
|γ̃vo |2 − Re(γ̃vo )
⇒ if Fmin > 0 → µmin = 10 log10 (7.49)
1 − Re(γ̃vo )

We see that for small volume phase shifts φ the minimum is given by small
µ < −30 dB, but as the phase increases so that Re(γ̃vo ) > |γ̃vo |2 or φ >
cos−1 (|γ̃vo |), then the minimum coherence occurs for higher values of µ. Figure
7.19 shows how the µ for minimum coherence changes for the example shown
in Figure 7.18. Note that for phase angles up to cos−1 (0.8) = 37◦ the minimum
276 The coherence of surface and volume scattering

–5

–10

Mu (dB)
–15

–20

–25

–30
0 20 40 60 80 100 120 140 160 180
Fig. 7.19 µ required for minimum coherence
in RVOG model versus phase Phase angle (degrees)

coherence is given by the volume only (µ = 0) point. However, for phase shifts
above this the minimum occurs for a mixture of surface and volume scattering.
For high phase shifts the minimum occurs for an almost equal mixture of surface
and volume scattering.
However, we have seen that the RVOG assumes an exponential structure
function in the volume, and hence the coherence amplitude and phase are not
independent quantities. In fact the phase centre for the volume-only component
of RVOG must lie between halfway and the top of layer 1 (see Figure 5.18).
In other words, for RVOG it follows that the phase of the interferogram for the
volume-only channel φ ≥ βz2hv . Coupled to this is the realization that for RVOG
the coherence amplitude can never be less than the zero extinction limit; that
is, |γ̃vo | ≥ sinc( βz2hv ).
This then begs the question as to whether the minimum coherence for the
RVOG model can ever be given by µ = 0. For this to be possible the following
inequality must apply:

  
βz hv βz hv βz hv
φmin < cos−1 (|γ̃min |) ⇒ < cos−1 sinc 0≤ ≤π
2 2 2
(7.50)

However, this inequality is never satisfied for the RVOG model, and hence we
conclude that for RVOG the minimum coherence is never given by the µ = 0
volume scattering channel. There is always some mixture of surface and volume
scattering that combines to produce a minimum. This rather surprising result
follows from the assumed form of the structure function (an exponential). This
exposes a weakness of the RVOG model: that its assumption of a uniform layer
with a simple extinction profile ignores any variations due to vertical structure
in volume scattering. We now turn to consider, in more detail, how structure
variations are dealt with in RVOG.
 7.4 Important special cases: RVOG, IWCM and OVOG 277

7.4.3 Structural ambiguity in RVOG z

Top of layer
Before leaving the RVOG model we first consider one important extension:
its generalization to arbitrary volume structure functions. By maintaining the
assumption of a random volume in layer 1 and polarisation-dependent delta f (z) = fv(z) + m2(w) d(z – zo)
function contributions from the effective surface components, we can generalize Surface scattering
the RVOG approach to arbitrary structure functions, as shown schematically in
Figure 7.20. Importantly, this modification maintains the line as coherence loci,
but it changes the relationship between the phase and coherence amplitude of Fig. 7.20 Structure function for general struc-
the volume only (µ = 0) point. This we call a structured-volume-over-ground tured volume-over-ground or SVOG model
(SVOG) model, which has the general form shown in equation (7.51):
 
µ(w)
γ̃ (w) = eiφ(zo ) γ̃vo + (1 − γ̃vo )
1 + µ(w)
- hv
fv (z )eiβz z dz
γ̃vo = eiβz zo 0 - h (7.51)
0 fv (z )dz
v

(1 + a0 )f0 + a1 f1 + a2 f2 + · · · + an fn
= eiβz z0 eikv
(1 + a0 )

For this more general SVOG model the relationship between coherence
amplitude and phase is not as restrictive as it is for RVOG, and so the min-
imum coherence can indeed be the volume-only coherence, under the general
condition that the phase φ < cos−1 (|γ̃vo |).
The classical RVOG model can be made to accommodate changes in struc-
ture by varying the extinction over a sufficiently wide range. However, this
gives rise to a structural ambiguity in RVOG, in that we can fit RVOG to
non-RVOG situations simply by adjusting the model parameters. A simple
example of this ambiguity is shown in Figure 7.21. Here we show a simple
case of vertical structure: a layering of the volume, with scattering from a thin
top layer, and surface reflection from a position separated from this volume
by a gap.
The coherence model for this three-layer problem can be written as shown
in equation (7.52). This still has a linear coherence loci, but the coherence
amplitude of the volume-only channel will be small, while its phase will be

z = hv

Constant
number hv, 1
density
z=0

1 >> 2
z = hv

hv, 2
Vertical
structure z = hc

Fig. 7.21 Vertical structural ambiguity in the

z=0 RVOG model
278 The coherence of surface and volume scattering

large. It is possible to fit RVOG to this structure (as shown in the upper portion
of Figure 7.21). However, to explain the combination of high coherence and
large phase offset we need to employ an effective extinction for the medium
that is much larger than the actual value.
 
µ(w) 
γ̃ (w) = e iβz zo
e iβz hc
γ̃vo + 1 − eiβz hc γ̃vo
1 + µ(w)
- hv 2σe hv

hc e cos θ eiβz z dz
γ̃vo = - hv 2σe hv
(7.52)
hc e cos θ dz

7.4.4 The coherence loci for IWCM

z = zo + hv
In this section we consider a model closely related to RVOG, but one that
arose independently out of generalizations of the water cloud model (WCM;
Temporal see Section 3.5.1) (Askne, 1997, 2003, 2007). This model, called the interfer-
stability
function ometric water cloud model (IWCM), places more emphasis on the temporal
changes in volume and surface scattering, and was developed for representing
z = zo the coherence observed by repeat-pass, high-frequency, small spatial baseline
radar systems. The model shares the exponential structure function for volume
scattering and assumed random volume for layer 1 of RVOG (although in its
f (z) most general form it further splits the effective extinction into wave extinc-
tion in the canopy and the fraction of gaps between the vegetation; see Askne
Fig. 7.22 Composite structure function for (2003)). However, it explicitly includes a vertical temporal stability function,
the IWCM model
as shown schematically in Figure 7.22. This function is 1 where the volume is
stable, and 0 where unstable. While arbitrary stability functions (and their corre-
sponding Fourier–Legendre expansions) can be envisaged, usually the simplest
assumption of a uniform pulse function (zeroth-order Legendre function) with
amplitude γtv for the volume component and γts and γtsv for the surface ele-
ments, is taken. Equation (7.53) shows the general form of coherence for this
model:

γ̃ (w)m0 (w) = m0v (w)γtv γ̃vo eiφ(zo ) + ps m0s (w)γts eiφ(zo )

+ psv m0sv (w)γtsv eiφ(zo ) (7.53)

There are two forms of this model that deserve special attention. In the first
we consider its form for short spatial/long temporal baselines and high radar
frequency, where volume decorrelation is very small and temporal effects for
both surface and volume dominate. We shall call this the high-frequency or
HF-IWCM. In the second form—more closely linked to RVOG—we consider
longer spatial/shorter temporal baselines and low radar frequency when tem-
poral effects are mixed with significant levels of volume decorrelation. This we
call the low-frequency or LF-IWCM.
In the first case the βz value is small and the mean wave extinction is
high (because of the high-frequency approximation), and the volume decor-
relation can therefore be approximated by a unitary phase shift, as shown in
 7.4 Important special cases: RVOG, IWCM and OVOG 279

equation (7.54):

 2σe
p1 =
p1 (ep2 hv − 1)  cos θo
γ̃vo =
p2 (ep1 hv − 1) 
 2σe
p2 = + iβz
cos θo
lim σe →∞ p1 iβz hv 1
−→ γ̃vo ≈ e = eiβz hv (7.54)
p2 1 + i βz 2σ
cos θo
e


θo
−i βz 2σ
cos θo iβz hv − cos
≈e e e iβz hv
=e 2σe

With this in place the HF-IWCM takes the following form:


mv w cos θ  cos θo
γ̃ (w)m0 (w) = γ̃tv 1 − e−2σe hv sec θ eiβz (hv − 2σe )
2σe
+ e−2σe hv sec θ γ̃es m0es (w) (7.55)

or in terms of coherence it can be written thus:

γ̃ (w)
mv (w) cos θ  cos θo
γ̃tv 2σe 1 − e−2σe hv sec θ eiβz (hv − 2σe ) + e−2σe hv sec θ γ̃es m0es (w)
= 
mv (w) cos θ
2σe 1 − e−2σe hv sec θ + e−2σe hv sec θ m0es (w)
cos θo
γ̃tv eiβz (hv − 2σe ) + γ̃es µ(w)
= (7.56)
1 + µ(w)
Again we see that the parameter µ—the surface-to-volume scattering ratio—is
important in determining the coherence loci for this model. The loci in this
case is a triangle, with two vertices on the unit circle and one at the origin
(γtv = γtes = 0), as shown in Figure 7.23. An important simplified case of
HF-IWCM arises in the limit γtv = 0 and γes = 1. This occurs in vegeta-
tion problems, for example, when wind-driven temporal change destroys the
coherence completely from layer 1, while the underlying surface scattering
contributions show no change. In this case we move along the radial line OP in
Figure 7.23, for which the coherence loci depends entirely on µ, as shown in
equation (7.57):
µ(w)
γ̃ (w) = eiφ(zo ) (7.57)
1 + µ(w)
Note that this has same form of coherence variation with polarisation as SNR
decorrelation, as demonstrated in Figure 7.24 (compare this with Figure 5.12).
In the second form of this model—the LF-IWCM—we consider the limit
of larger spatial/lower temporal baselines and low radar frequency combined
with volume-only temporal decorrelation γtv due, for example, to short-term
wind-blown effects in vegetation cover. This model is closely related to RVOG,
since the volume coherence can no longer be approximated by a phase shift.
280 The coherence of surface and volume scattering

90 1
120 60
0.8

0.6
150 Q 30 if (z ) ib (1– cosuo )
0.4 e o e z 2se

0.2

180 0 0
P eif (zo)

210 330

240 300
Fig. 7.23 Coherence loci for the IWCM
model 270

0.9

0.8

0.7

0.6
Coherence

0.5

0.4

0.3

0.2

0.1

0
Fig. 7.24 Coherence variation with surface- –30 –20 –10 0 10 20 30
to-volume ratio (µ) for the IWCM model Mu (dB)

The two models—RVOG and LF-IWCM—can be connected by rewriting the

latter in the following form:

γ̃ (w) = eiφ(zo ) (γtv γ̃vo + F(w) (1 − γtv γ̃vo )) (7.58)

Here we see that the coherence loci remains a non-radial line segment, but that
the fixed point representing the volume is shifted towards the origin by the scale
 7.4 Important special cases: RVOG, IWCM and OVOG 281

90 1
120 60
0.8
g~v
0.6
150 30
0.4

gtg^v 0.2

if (zo)
180 0 e

210 330

240 300
Fig. 7.25 Coherence loci for the RVOG
270 model with temporal decorrelation

factor γtv . This amounts to a rotation and stretch of the coherence line about
the surface topography point, as shown in Figure 7.25 (Papathanassiou, 2003).
In conclusion, we have shown that the coherence loci for a two–layer random
volume over ground scattering problem is a line segment in the complex plane.
This line is radial when temporal effects dominate, and shifts to a non-radial
phase variant line as volume decorrelation becomes more important. In both
cases we note two important features. The first is the importance of the ratio µ,
being the ratio of effective surface-to-volume scattering. The second key point
is that the line passes through the unit circle at the surface phase point. We shall
see later that this provides us with a method for correcting for vegetation bias
in radar interferometry by line fitting and by finding this intersection.

7.4.5 The coherence loci for OVOG

In the previous section we considered the case when layer 1 is a random volume,
and showed that the coherence loci is then a straight line in the complex plane. A
natural extension of this approach is to consider layer 1 as an oriented volume.
z = zo + zv
In this case the polarimetry becomes more complex, as discussed in Section
7.2.2. However, we shall see that the coherence loci may still be obtained as a
simple extension of the RVOG approach.
The OVOG model maintains the assumption of a uniform layer with an expo-
f(z) = fv(w, z) + m(w) d(z – zo)
nential structure function, but is characterized by a pair of eigenpolarisations for
propagation through the medium. These orthogonal states then define a triplet of
structure functions, as shown schematically in Figure 7.26. The states with high-
est extinction XX and lowest extinction YY are separated by the crosspolarised z = z0
channel XY. The effective surface components (shown as a line at z = zo ) are
viewed through the polarisation filter of volume 1, which distorts their appar-
ent polarimetry (see Section 4.2.6). In the presence of combined surface and Fig. 7.26 Composite structure function for
volume scattering we must now consider a triplet of coherence formulae—one the OVOG model
282 The coherence of surface and volume scattering

for each eigenpolarisation combination, as shown in equation (7.59):

γvo (2σx , hv ) + µxx

γ̃xx = eiφ = eiφ (γ̃vo
xx
+ Fxx (1 − γ̃vo
xx
))
1 + µxx

iφ γvo σx + σy , hv + µxy xy xy
γ̃xy =e = eiφ (γ̃vo + Fxy (1 − γ̃vo )) (7.59)
1 + µxy

iφ γvo 2σy , hv + µyy yy yy
γ̃yy =e = eiφ (γ̃vo + Fyy (1 − γ̃vo ))
1 + µyy

Here both volume and surface scattering have polarimetric coherency matrices
with reflection rather than azimuthal symmetry, and the surface-to-volume scat-
tering ratios include the effects of propagation distortion. The dynamic range
of µ can be developed using a modification of the procedure used in equation
(7.46), as shown in equation (7.60):

w∗T [P] [Tes ] [P]∗T w

µopt → max ⇒ [Tv ]−1 [P] [Tes ] [P]∗T wopt = λwopt
w∗T [Tv ] w
(7.60)

Where [P] is a propagation distortion matrix (see Section 4.2.6). While it is

now not so easy to develop an analytic solution for the eigenvalues of this
optimization, we can obtain an estimate of the coherence loci by using a simple
geometrical argument, as follows.
Each term in the triplet of coherences in equation (7.59) has the same form
as the RVOG model, and thus corresponds to a line segment in the complex
plane. Furthermore, as these eigenstates bound the oriented volume solution
(see Section 7.2.2) it follows that the loci must be contained within the triplet
of lines defined in equation (7.59). Figure 7.27 shows the resulting triangu-
lar coherence loci for the OVOG model. The coherence for each polarisation

90 1
120 60
0.8
c = f ( x, y, hv)
0.6
150 30
0.4

0.2

180 0

210 330

240 300
Fig. 7.27 Coherence loci for the OVOG
model 270
 7.4 Important special cases: RVOG, IWCM and OVOG 283

is constrained to move up and down its own straight line inside the unit cir-
cle, depending on the µ ratio. The three lines for the eigenstates define the
boundary of this region, coming to a focus at the ground topography point
φ, and having a spread ψ, as shown in Figure 7.27. Importantly, this spread
depends on the differential extinction in the volume layer and not on the µ val-
ues or surface topography. In the special case that ψ = 0 we again obtain the
random-volume-over-ground RVOG model. We also note that the OVOG model
requires that the crosspolarised XY coherence line lies between the XX and
YY lines.

7.4.6 The oriented-volume-under-ground (OVUG) model z = z0

Finally, we consider an important variation of the OVOG model, applicable to
cases where scattering occurs from the top surface of layer 1 and at the same
time hv tends to infinity, so that we can effectively ignore the influence of f(z) = fv(w,z) + (w) (z-zo)
scattering from the layer 1–2 interface. Such a model can be used for analysis
of thick layers, as occur, for example, in high-frequency land-ice applications,
where scattering from the top air-ice interface usually dominates that from the
bottom ice/rock interface. This scenario is summarized in Figure 7.28, in which
is shown the corresponding structure function. The coherence function for this
problem can be derived in a similar manner to the OVOG model, and is shown in
equation (7.61). The key difference here is the phase of the volume term, which Fig. 7.28 Composite structure function for
the OVUG model
now lies below the surface reference rather than above it (equation (7.62)).
γ̃ (w) = eiφ(zo ) (γ̃∞
∗ ∗
(w) + F(w)(1 − γ̃∞ (w))) (7.61)
∗ 1
γ̃∞ (w) = (7.62)
1 + i β2σz cos θ
e (w)

The coherence loci for this problem can be obtained as an extension of the OV
region, as shown in Figure 7.29. Here again we see a region formed by three
bounding lines for the eigenpolarisations emanating from the surface point, with
variations along each line given by the fraction of surface-to-volume scattering,
F(w). Again the RVUG or random volume version of this model is obtained
as a limiting case when the extinctions become equal and we obtain the single
line coherence region, as shown on the right-hand side of Figure 7.29.

Coherence loci for infinite volume Coherence loci for infinite volume

90 1 90 1
120 60 120 60
0.8 0.8
0.6 0.6
150 30 150
0.4 0.4 30

0.2 0.2

180 0 180 0

210 330 210 330

240 300 240 300 Fig. 7.29 Coherence loci for the OVUG
270 270 model (left) and OVOG model (right)
 Parameter estimation
8 using polarimetric
interferometry

In the previous chapter we developed the form of the backscatter polarimetric

z = z0 + hv z
interferometric coherence loci for a two-layer scattering model. We now turn to
consider algorithms for the inverse problem for such a case; that is, we consider
Layer 1
methods for estimation of parameters of the two-layer model from observations
of the coherence variation with polarisation (Papathanassiou, 2001; Cloude,
z = z0 y
Layer 2
2000b, 2000c, 2003; Stebler, 2002; Ballester-Berman, 2005; Praks 2007). We
z = z 0- d
start by identifying the key parameters of interest by reference to the schematic
Fig. 8.1 Geometry of two-layer scattering diagram shown in Figure 8.1. Based on this we can identify the following
problem important parameters of interest in remote sensing:
1. The position of the bottom of the layer, zo (or its associated interfer-
ometric phase, φo ). This if often called the underlying or true surface
topography or ground position in vegetation and snow/ice applications.
2. The depth of layer 1, hv , which may correspond to the height of
vegetation or depth of a snow layer, depending on the application.
3. In some applications (such as land-ice penetration), interest centres on
the position (phase) of the top of layer 1, especially when its depth tends
to infinity (Dall, 2003; Sharma, 2007). In this case the main application
is to compensate the penetration depth into the layer so as to locate the
true surface position.
4. The structure function in layer 1, f (z). For exponential models such as
RVOG and OVOG this amounts to estimation of a pair of extinction
coefficients, while in more general terms it amounts to estimation of the
Fourier–Legendre spectrum of the structure function for the layer.
5. The surface-to-volume scattering ratio, µ. This function, when com-
bined with the total backscatter cross-section, can be used to separate
scattering contributions from layers 1 and 2 and hence isolate volume or
surface scattering for further study (Cloude, 2004, 2005a).
We now turn to consider estimation techniques for each of these in turn.

8.1 Surface topography estimation

There are three basic approaches to the estimation of underlying surface topog-
raphy (Papathanassiou, 2001; Sagues, 2000; Cloude, 2000c, 2003). The
simplest is an extension of conventional interferometry, employing the phase of
an interferogram for some selected surface-dominated polarisation vector ws .
The second approach is to employ two polarisation states to remove phase bias
 8.1 Surface topography estimation 285

from the top layer. Finally, we can use multiple polarisations and least squares
correction to phase bias. We now consider each of these in turn.
In the simplest case the phase of the surface component can be estimated
directly from the coherence, as shown in equation (8.1):

φ̂ = arg(γ̃wS ) 0 ≤ φ̂ < 2π (8.1)

By subsequently employing phase unwrapping, the surface topography can

then be estimated. The precision of this estimate (given by the height variance)
depends on baseline and the coherence amplitude of the interferogram, as shown
in equation (8.2):

&  
'
R0 sinθ λ ' 1 − γw 2
σh ≈ σφ σφ ≤ (  2
s
(8.2)
B⊥ 4π 
2L γws 

where the Cramer–Rao bound (minimum value) of the phase variance σφ for
a given number of looks L is given on the right-hand side of equation (8.2)
(Seymour, 1994). This, of course, resorts to conventional interferometry in the
limiting case of bare surfaces (hv = 0), but in other cases is made complicated
by the phenomenon of phase bias.
We have seen that the volume coherence for layer 1 is complex and hence
contributes a phase offset from the surface itself, and so equation (8.1) will
generally overestimate the surface position for RVOG and OVOG, and under-
estimate it for OVUG. Hence it is clear that polarisation ws should be chosen
so as to minimize this bias and optimize the accuracy of the estimate. A second
objective must be to choose ws to also maximize the SNR, so as to minimize the
decorrelation due to noise and hence optimize the precision. In RVOG, OVOG
and OVUG these requirements amount to maximization of µ, the surface-to-
volume ratio. The best polarisation to use would therefore be that given by
equation (7.47) or (7.60). However, as we have no a priori knowledge of the
separate volume and surface component coherency matrices, we cannot make
direct use of this equation. Instead we must employ an indirect solution, as we
now investigate.
The problem is that there is no single polarisation that always maximizes µ.
For a bare surface (hv = 0) at low frequencies when the Bragg surface scatter-
ing model is valid, a good choice is VV, as HH has less scattered power and
thus lower SNR, and HV is zero. A better choice still is HH+VV, as the zero
polarimetric phase difference leads to an even better SNR than VV. Clearly, the
optimum would weight the Bragg scattering matrix elements to maximize the
SNR. For higher frequencies and rougher surfaces the depolarisation increases,
the difference between HH and VV becomes less, and polarisation plays less of
a role in bare surface parameter estimation. In general, therefore, an unbiased
coherence optimizer would provide a suitably adaptive solution. For bare sur-
faces the constrained optimizer of equation (6.22) would provide a good choice
(as shown again in equation (8.3)). Note, however, that in order to implement
such an optimizer we require access to full scattering matrix data, so as to be
286 Parameter estimation using polarimetric interferometry

ˆ 12 .
able to estimate the component matrices T̂11 , T̂22 , 
 
[ ˆ H] = 1  ˆ 12 eiφ + 
ˆ ∗T e−iφ
−1 ˆ
[T̂ ] [H ]w = λ (φ) w 
2 12
[T̂ ] = 1 T̂11 + T̂22
2
(8.3)
max|λ(φ)| w∗T ˆ
opt [H ]w opt
−→ wopt ⇒ γopt =
w∗T
opt [T̂ ]w opt

The above analysis breaks down, however, in case layer 1 has non-negligible
thickness. We can set a suitable threshold on the product of wavenumber and
layer thickness to estimate this breakpoint, such as βz hv < 0.1 (see equation
7.1), for the surface approximation to hold. If the product exceeds this threshold
then we require a different strategy to optimize estimation of surface topography
as follows.
When the layer thickness can no longer be ignored, we face complica-
tions arising not only from phase bias and increased volume decorrelation,
but a change in scattering mechanism. This arises especially when the dihedral
second-order scattering is dominant. In this case HH is often preferred to VV
(the opposite of the bare surface case), as it has a higher specular reflection
coefficient at the surface, and VV will in this case have a lower µ. By the
same reasoning, the Pauli choice HH–VV is sometimes selected in preference
to HH+VV, as this has an even higher µ than HH, due to the 180-degree polari-
metric phase shift that occurs in the case of a dominant second-order scattering
scenario.
A second problem also arises in the case that layer 1 is a random volume,
when it acts to depolarise the scattered wave with a high entropy. This implies
that the volume scattering ‘contaminates’ every polarisation vector w, and con-
sequently that it is impossible to find a candidate ws which contains surface-only
scattering. This means that the phase bias due to volume scattering in layer 1
will be present across the whole of polarisation space. The best we can do is
again try to select the ws with maximum µ. However, there is no longer a guar-
antee that the coherence amplitude optimizer of equation (8.3) will correspond
to the maximum µ (see the discussion in Section 7.4.2). While the optimizer
will still guarantee the highest coherence and hence the highest precision, it no
longer guarantees the highest accuracy, because of the presence of phase bias.
To proceed further we need to consider methods for the removal of this bias.

8.1.1 Phase bias removal

We can make use of the SVOG (with RVOG as a special case) model to remove
the phase bias and improve the accuracy of surface topography estimation. We
have seen that the SVOG model predicts linear coherence loci in the complex
plane. Importantly, this line intersects the unit circle at the desired surface
topography point eiφ . Therefore, if we start by selecting an arbitrary polarisation
w1 and evaluate its interferometric coherence γ̃1 it will lie somewhere on this
line, generally displaced from the desired unit circle point by some unknown
phase bias. However, if we now choose a second polarisation state w2 , and the
only condition we set on w2 is that it have a higher surface-to-volume scattering
ratio than the first, so that µ2 > µ1 , then it follows that we can find the unit
 8.1 Surface topography estimation 287

circle point from a line fit as follows. The idea is to use γ̃1 as a fixed point on
the line and relate γ̃2 by a scale factor F2 along the line towards the unit circle.
In this way the two coherences can be related as shown in equation (8.4). We
see that the desired ground topography point is embedded in these equations,
and we can solve for it directly.

γ̃1 γ̃2 − γ̃1 (1 − F2 )
⇒ eiφo = 0 ≤ F2 ≤ 1 (8.4)
γ̃2 = γ̃1 + F2 (eiφ − γ̃1 ) F2
Here we see that the phase term is obtained as a weighted average of the two
complex coherences. If F2 tends to unity then it represents the desired surface
point, and γ̃2 is taken as the solution. However, in general there will be phase
bias present in both channels (because of depolarisation in layer 1), and hence
this mixture formula is required to compensate for this bias. There remains a
problem in that to solve for the surface topography we first require an estimate
of the factor F2 . This can be obtained directly from the estimated coherences
γ̃ˆ1 and γ̃ˆ2 by forming the product eiφ e−iφ = 1, using equation (8.4) to obtain a
quadratic. Taking the root that makes F positive, we obtain the solution shown
in equation 8.5:
φ̂ = arg(γ̃ˆ2 − γ̃ˆ1 (1 − F2 )) 0 ≤ F2 ≤ 1
√
−B − B2 − 4AC (8.5)
AF2 + BF2 + C = 0 ⇒ F2 =
2
2A
A = |γ̃1 |2 − 1 B = 2Re((γ̃2 − γ̃1 ).γ̃1∗ ) C = |γ̃2 − γ̃1 |2
Note that unlike the simple phase algorithm (equation (8.1)), this requires
coherence estimates in both amplitude and phase, and hence is susceptible
to non-compensated errors in coherence such as those due to SNR or tempo-
ral effects. Also, of course, the estimates of coherence themselves have some
variance due to their stochastic nature and the finite number of looks L used
in the estimator (the coherence region). For stability of the phase estimate we
therefore need to ensure that F2 is as large as possible. (If F2 tends to zero so
that the two points are close together, then large errors can result.) Some care
is therefore required in the selection of w1 and w2 . There are three strategies
used in making an appropriate selection:
In physics-based selection we use our understanding of surface and volume
scattering to select the two channels. For example, the low-frequency Bragg
surface model predicts zero or (for higher-order forms of the model) very low
levels of crosspolarisation HV from a flat surface. On the other hand, volume
scattering from a cloud of anisotropic particles can yield high levels of HV (for
a dipole cloud only 3 dB below the maximum RCS). Hence HV is often chosen
as a candidate for the w1 channel. The w2 channel can likewise be selected on
the assumption that specular second-order scattering is dominant, and so HH
or HH–VV are good choices as they are likely to satisfy the requirement that
µ2 > µ1 . In summary, the direct physics-based approach produces allocations
of the two channels such as those shown by the two examples in equation (8.6):
γ̃1 = γ̃HV γ̃2 = γ̃HH −VV
or (8.6)
γ̃1 = γ̃HV γ̃2 = γ̃HH
288 Parameter estimation using polarimetric interferometry

Note that the lower option is particularly well suited to dual polarised active
systems that can transmit only linear H polarisation but receive H and V
components (see Chapter 9).
If more specific information is available about the scattering problem to
hand, based, for example, on direct EM scattering model simulations, then these
assignments can of course be modified as appropriate. Although such selections
may match very well a specific application or dataset, they are generally not
sufficiently robust for widespread application. For this reason we turn to a
second approach, based instead on phase optimization, that adapts itself to
variations in the data.

8.1.2 Coherence separation optimization

We have seen that phase bias can be removed under the assumptions of the
SVOG model by fitting a line between two coherence values. Furthermore,
best results will be obtained if we employ two polarisation states with the
maximum difference in µ, as these will be less sensitive to fluctuation noise in
the coherence region estimates for a fixed number of looks L. Under the SVOG
model, µ impacts directly on the phase centre of the interferogram, so that as µ
decreases so the phase bias increases. This indicates that the optimum pairing of
polarisation vectors w1 and w2 to choose would be those corresponding to the
ends of the linear coherence loci in the complex plane. One way to estimate these
is to employ coherence separation rather than coherence amplitude optimization
(see Section 6.2). In this method we employ fully polarimetric data to calculate
the following eigenvalue problem (see Section 6.2.3).

max|λmax (φ)−λmin (φ)|

φ
[T ]−1 [H (φ)]w = λ (φ) w −→ γ̃opt

w∗T
a [H ]w a 
γ̃1 = 

w∗T  (8.7)
a [T ]w a 
⇒ γ̃ = |γ̃1 − γ̃2 | = γ̃opt
w∗T [H ]wb 



γ̃2 = b∗T 
wb [T ]wb

For each phase angle φ we find the distance between the maximum and min-
imum eigenvalues. By finding the maximum of this distance as a function of
φ we then automatically align the solution with the axis of the linear coher-
ence region and use these as an estimate of the coherence loci bounds. We then
find the polarisation scattering mechanisms wa and wb from the corresponding
eigenvectors. From these we can derive the two coherence γ̃max and γ̃min for
use in the topographic phase estimation algorithm of equation (8.5).
However, we face a potential problem with all these line-fitting ideas to ensure
that we always choose the correct rank ordering of the two coherences, remem-
bering that we must ensure that µ2 > µ1 to find the correct topography point.
In fact this is a general problem with all phase bias removal algorithms based on
the assumption of a linear coherence loci. By definition, a line intersects the unit
circle at two points (see Figure 8.2), one of which is the true topographic phase,
while the other represents a false solution obtained for the line fit technique by
exchanging the rank order of coherences. There are several ways to resolve this
 8.1 Surface topography estimation 289

90 1
120 60
eifmax 0.8

0.6
150 30
g^ max 0.4

0.2
g–min
eifmin
180 0

210 330

240 300
270 Fig. 8.2 Unit circle surface phase ambiguity

rank-ordering dichotomy, and two common techniques use physical arguments

based on scattering theory or a comparison of the interferometric bias levels of
the two solutions. We now consider both of these.
In the physical approach we again employ our expectations for the nature of
polarimetric scattering in the channel with the highest µ value. For example, the
surface-dominated channel should have a scattering vector close to the form ws
shown in equation (8.8), where α is less than π/4 for direct surface scattering,
and greater than π/4 for dihedral second-order scattering.

 T      ∗T 
ws = cos α sin αeiϕ 0  µ >µ  γ̃1 = γ̃min if w∗T  
max w v < w min w v

−→
s v

 T  γ̃ = γ̃  ∗T   ∗T 
wv = 0 0 1 1 max if wmax wv  > wmin wv 
(8.8)

Similarly, the orthogonal state wv should match the volume scattering (and have
a lower µ). In this way we can develop an algorithm for assigning the optimum
phase states in the correct rank order for the line fit algorithm, as shown on the
right-hand side of equation (8.8).
In the second interferometric approach we decide on rank ordering by
employing the phase difference between the calculated unit circle intersection
point and assumed low µ coherence. Knowing that layer 1 is above (or below)
the surface allows us to calculate these phase differences using the same clock-
wise (or anticlockwise) rotation around the coherence diagram. If we repeat
this for both rank permutations, then one of the phase shifts will be much larger
than the other and can be rejected (especially if it is known that the layer depth
is less than the π height of the interferometer). If we define φmax as the unit
circle phase estimate obtained when we propose γ̃max as the high µ channel
estimate, and likewise φmin when we propose γ̃max , then we can decide on the
290 Parameter estimation using polarimetric interferometry

most likely rank ordering as follows:

 8
max = arg(γ̃min e−iφ̂max ) γ̃1 = γ̃min if max < min
→ (8.9)
min = arg(γ̃max e−iφ̂min )  γ̃1 = γ̃max if max > min

So, for example, in Figure 8.2, if we assume that layer 1 is above the surface for
anticlockwise phase rotation, then max = 270◦ and min = 90◦ , and therefore
according to equation (8.9) we would select γ̃1 = γ̃max on the assumption that
the layer thickness should be less than the π height of the interferometer.

8.1.3 Total least squares (TLS) surface topography

estimation
So far we have considered line fit techniques that employ only two coherence
values, selected either on the basis of scattering physics or by employing coher-
ence optimization. A more robust version of this approach is to employ multiple
polarisation channels (N > 2) and use a least squares line fit to the multiple
complex data points. In this way we can avoid problems of any selected pair of
points becoming too close, and thus minimize errors in the surface topography
estimation.
We start by using the linear coherence loci assumption to generate a linear
relationship between the real and imaginary parts of coherence, as shown in
Figure 8.3. The problem then reduces to estimation of the two coefficients M
and C. In ordinary least squares (LS) estimation we would find the M and
C that minimize the sum of squares of the vertical distance between line and
data points, y, as shown in Figure 8.3. However, this assumes that noise is
only found in one coordinate, whereas for coherence estimation both real and
imaginary parts are subject to statistical fluctuations (see equation (5.34)). A
better approach, therefore, is to employ a total least squares solution (TLS) that
accounts for errors in both x and y. Geometrically, the TLS approach amounts
to using a different measure of distance: the perpendicular distance Ri , at right
in Figure 8.3, and related to y as shown in equation (8.10):

y y
Ri = y cos θ = =√ (8.10)
1 + tan2 θ 1 + M2

~
y=Im(g)
~ ~
Im(g) = MRe(g) + C
y = Mx+C

Ri
y
~
x = Re(g)

Fig. 8.3 Total least squares line fit

 8.1 Surface topography estimation 291

If we then make the simplest assumption that the unknown fluctuation errors
are the same in x and y (the modifications, if they are not, are straightfor-
ward, but complicate the notation), the function to be minimized now has the
following form:

! 1 !
R2i = (yi − C − Mxi )2 (8.11)
1 + M2
i i

This differs from the conventional LS approach only in the pre-multiplier, which
is itself a function of M . We can then find the stationary points of this function
by differentiation, to yield the following:

∂R 1 !
0= = −2(yi − C − Mxi )
∂C 1+M 2
i

∂R 1 !
0= = −2(yi − C − Mxi )xi (8.12)
∂M 1 + M2
i

2M !
− (yi − C − Mxi )2
(1 + M )
2 2
i

From the first term we obtain a direct solution for the estimate of C as
follows:
) *
1 ! !
Ĉ = yi − M̂ xi = ȳ − M̂ x̄ (8.13)
N
i i

Then, by substituting the first equation in the second and collecting terms we
obtain an estimate of M as the root of a quadratic, as shown in equation (8.14):

−c1 ± c12 − 4c2 c0
c2 M̂ 2 + c1 M̂ + c0 = 0 ⇒ M̂ = (8.14)
2c2

where the three coefficients are defined in terms of the data points as
follows:
!
c0 = − (xi − x̄)(yi − ȳ)
i
!@ A
c1 = (xi − x̄)2 − (yi − ȳ)2 (8.15)
i
!
c2 = (xi − x̄)(yi − ȳ)
i

This then provides us with a method for fitting a line to an arbitrary number
of polarisation channels. We can then use the estimates of M and C to find the
two unit circle intersection points as shown in Figure 8.4. These two points can
292 Parameter estimation using polarimetric interferometry

90 1
120 60
0.8

0.6
150 30
0.4

0.2

180 0

210 330

240 300
Fig. 8.4 Example of total least squares line
fit to complex coherence data 270

be found explicitly in terms of M and C as shown in equation (8.16):


. −M̂ Ĉ ± M̂ 2 − Ĉ 2 + 1 


x +y =1
2 2
xp =
⇒ 1 + M̂ 2 → eiφ = xp + iyp
y = M̂ x + Ĉ 


yp = M̂ xp + Ĉ
(8.16)

Clearly, this TLS approach will suffer from the same rank-ordering dichotomy
encountered in Section 8.1.2. Indeed, in the TLS case this problem is arguably
more serious, as there is no easy way to employ the physical selection process
described in equation (8.8). As there are multiple polarisation channels, and
not just two being used, it is difficult to decide which are surface- or volume-
dominated. For this reason the TLS approach is often combined with the phase
approach of equation (8.9) to decide which phase point to use as the topography
estimate.
Finally, we note that some care is required in the choice of N polarisations to
ensure that there is sufficient diversity of µ along the line segment in the com-
plex plane. For this reason, typical selections involve the three Pauli channels
(HH+VV, HH–VV and HV), augmented by the linear channels HH and VV,
as well as the three optimum states from either constrained or unconstrained
algorithms to produce a sample set in the region N = 8–12 for estimation.

8.1.4 OVOG: surface topography with differential

extinction
In the previous section we outlined various algorithms for the removal of veg-
etation bias and consequent estimation of true surface phase in polarimetric
interferometry. The main assumption behind these techniques was that of a
linear coherence loci, which we have seen implies that layer 1 scatters with
 8.1 Surface topography estimation 293

azimuthal symmetry. While this may be a valid assumption for many applica-
tions, it is violated for an important class of problems when layer 1 displays
reflection scattering symmetry and behaves as an oriented volume. In this case
we have seen that the coherence loci is formed by a fan of three lines, emanating
from the unit circle topography point (see Figure 7.27). In theory, therefore, we
cannot strictly apply the above algorithms to the OVOG model. However, there
are two important classes of OVOG applications that deserve special mention
(Ballester-Berman, 2005, 2007; Lopez-Sanchez, 2007).
The first involves applications with only weak differential extinction, in
which case the fan angle ψ (see Figure 7.27) is small and the OVOG region
approaches a straight line, or at the other extreme, high differential extinction
combined with small minimum extinction. This latter scenario is very important
in that it can lead to a wide dynamic range in µ, with the low extinction channel
dominated by surface scattering and the high extinction by volume scattering.
In this case the line fit approach of TLS or phase optimization still provides a
good, if approximate, solution to surface topography estimation, as shown in
Figure 8.5. However, in all OVOG cases it must be realized that there is an
additional source of error due to the separation of volume terms in the complex
plane, and this can lead to large errors if extra care is not taken to make full
use of the µ spectrum for the problem. In this case, therefore, it is good to
use either the TLS with a wide diversity of polarisations, or the coherence
separation optimization technique to ensure that the maximum and minimum
µ values are being fully exploited.
Another class of OV problems of interest are for layers of effective infinite
depth, when we can assume that µ = 0 in all polarisation channels. In this
limit we essentially obtain the oriented volume or OV problem as a limiting
case of OVOG. We have seen that the coherence loci for this problem is a
semicircle (Figure 7.8), and we can devise a simple algorithm for top surface
phase estimation based on a circle fit to the data, as follows. The first point to
note is that the circle must intersect the unit circle and the origin, and hence has

90 1
120 60
0.8

0.6
150 30
0.4

0.2

Large m
180 0

210 330

240 300
Fig. 8.5 Line fit for topography estimation
270 under the OVOG model
294 Parameter estimation using polarimetric interferometry

Coherence loci for infinite volume

90 1
120 60
0.8

0.6
150 30
0.4
c = po + iqo
q 0.2
⇒ f = tan−1 ( o )
po
180 c 0

210 330

240 300
Fig. 8.6 Topography estimation for the OV
model 270

a fixed radius of 0.25. The second point is that the topographic phase is simply
related to the coordinates of the centre C of the circle as shown in Figure 8.6. By
combining these two observations we can set a linear least squares formulation
for the two unknown coordinates of C as shown in equation (8.17):

(p − po )2 + (q − qo )2 = r 2
r 2 = po2 + qo2 = 0.25
   2 2 
pxx qxx   pxx + qxx
p
⇒ 2 pxy qxy  . o = pxy
2 + q2 
xy (8.17)
qo 2 + q2
pyy qyy pyy yy

⇒ [A]x = b
⇒ x̂ = ([A]T [A])−1 [A]T b

Here we make use of the real and imaginary parts of the coherence in the three
eigenpolarisation combinations XX, XY and YY to fit the best-constrained
circle to the data. This then allows us to estimate the top surface position,
even though we are assuming there is no scattering from this interface and only
volume scattering is occurring. This is useful when either the top surface is very
smooth or there is a small dielectric contrast between free space and layer 1.
Note that this technique does not work if µ > 0 in any channel, as this has
the effect of pulling the coherences off the circle and towards the topography
point. In this case we must resort to the approximation used in Figure 8.5.
We have seen that there are several possibilities for using a priori assumptions
about the coherence loci for the two-layer problem, to devise algorithms for
estimation of the surface topography and hence to effect phase bias removal. We
now turn to consider a similar approach to the estimation of a second important
physical parameter: the height of the top layer.
 8.2 Estimation of height hv 295

8.2 Estimation of height hv

In this section we consider algorithms for the estimation of the top layer height
hv using single baseline polarimetric interferometry (Cloude, 2001b, 2001c,
2003; Papathanassiou, 2001, 2005; Stebler, 2002; Yamada, 2001; Praks, 2007).
The approach will be to assume particular forms for the coherence loci for
the two–layer problem, exploit knowledge of the topographic phase from the
previous section, and use various scattering models to obtain an estimate of hv
from complex coherence.
One of the simplest approaches to this problem is to use the phase difference
between interferograms as a direct estimate of layer depth (Cloude, 1998). In
general terms we then estimate the coherence in two polarisation channels: wv ,
which is volume scattering only and has a phase centre near the top of the layer;
and ws , which is surface dominated and has a phase centre near the surface. By
forming the phase difference between these interferograms and scaling by the
interferometric wavenumber βz we obtain the following estimation algorithm:
arg(γ̃wV γ̃w∗s ) 4π θ
ĥv = , βz = (8.18)
βz λ sin θ
Here again the arg(..) function is defined in the range 0 to 2π. Although this
is a simple algorithm to implement it has some severe drawbacks in that the
layer depth estimate so obtained is generally underestimated. The problems
stem from the difficulty in finding polarisations with phase centres at the top
and bottom of the layer. We have seen that because of depolarisation in layer
1 there will be some volume scattering present in all polarisation channels,
and so the phase centre of ws , for example, will always be located above the
true surface (due to phase bias as discussed in Section 5.2.4). Likewise, we
have seen that the phase centre of the volume scattering component can lie
anywhere between halfway up and the top of layer 1, only reaching the top
in case of infinite extinction in the RVOG model, or more generally a vertical
structure function which is a delta function at z = zo + hv .
The phase bias issue for the surface channel can be compensated somewhat
by using our estimate for true surface topography, so that equation (8.18) takes
the modified form shown in equation (8.19):

arg(γ̃wV e−iφ̂ )
ĥv = (8.19)
βz

Note that φ̂ can be estimated either from the data itself or from some external
source such as a reference digital surface model (DSM). We can further improve
this algorithm by matching it to the optimization process used in equation (8.7).
In particular we can make use of the optimum coherence furthest in phase from
the surface topography point as the ideal wv channel. This then acts to maximize
the height of the phase centre of wv in layer 1. However, there still remains
the problem of compensating the volume scattering channel for variations in
structure function f (z). For example, if the structure function is uniform then
the phase optimum will still only reach halfway up layer 1, and so the phase
estimate of equation (8.19) will be only one half the true layer depth.
To try to resolve this, we note that γ̃wv is complex and so has two degrees
of freedom, of amplitude as well as phase. However, we have so far made use
296 Parameter estimation using polarimetric interferometry

only of the phase information. The idea is therefore to try to use the coherence
amplitude of γ̃wv to help compensate for variations in the structure function to
obtain a better estimate of hv . We shall see that there are various ways of doing
this, but a good starting point is to use a Fourier–Legendre expansion of the
structure function f (z) in terms of an infinite series with coefficients aio , which
are then related to the coherence as shown in equation (8.20) (see Section 5.2.4
for a derivation of the functions fi ).

ai βz hv
γ̃ = eiφ eikv (f0 + a10 f1 + a20 f2 + ...) ai0 = , kv = (8.20)
1 + a0 2

8.2.1 First-order inverse coherence model

In order to use the infinite series of equation (8.20) with experimental data we
must first truncate the series at some finite order. The simplest non-trivial case
is to truncate at first order, as shown in equation (8.21):
  
i(kv +φ0 ) sin kv sin kv cos kv
γ̃ = e eikv iφo
(f0 + a10 f1 + R1 ) ≈ e + ia10 −
kv kv2 kv
(8.21)

where R1 is the truncation error. We shall consider the typical magnitude of R1

in the next section, but for the moment we set R1 = 0. With this approximation
in place, we then have a model with two observations (the amplitude and phase
of γ̃ ) and three unknowns: φ0 , kv = βz hv /2, which is a function of the unknown
height hv and known baseline βz , and a10 , a normalized linear Legendre coeffi-
cient. Hence we have more unknowns than observations, and so in order to be
able to invert the model we require additional information. Varying the polari-
sation, while keeping the wavelength, sensor geometry and baseline constant,
provides a convenient way to add measurement diversity without adding new
parameters. Indeed, it is reasonable to assume that βv , φ0 , f0 and f1 all remain
invariant to changes in polarisation and only the Legendre coefficient a10 can
change, reflecting changes in the structure function with polarisation. In the
most general case we can then consider adding several polarisation channels to
the model of equation (8.21), but in the simplest we require just two, with scat-
tering mechanisms w1 and w2 , providing four observations and four unknowns,
as shown in the following pair of equations:
  
γ̃ w1 = ei(kv +φo ) (f0 + a10 w1 f1 ) = eiφ (f0 + a10 w1 f1 )
   (8.22)
γ̃ w2 = ei(kv +φo ) (f0 + a10 w2 f1 ) = eiφ (f0 + a10 w2 f1 )

This is a more balanced set suitable for inversion; that is, for estimation of the
four unknown parameters φ0 , kv , a10 (w1 ) and a10 (w2 ) from two observations
of complex coherence. The following strategy then follows immediately from
equation (8.22). First we estimate the surface phase term (noting that f1 is
imaginary), not by line fitting as in equation (8.5), but by differencing the
complex coherences as shown in equation (8.23):

φ = kv + φo = arg(−i(γ̃ (w1 ) − γ̃ (w2 ))) (8.23)

 8.2 Estimation of height hv 297

We can then calculate kv from the real part of the phase-shifted coherence, as
shown in equation (8.24):
sin kv
Re(γ̃ (w1 )e−iφ ) = Re(γ̃ (w2 )e−iφ ) = 0 ≤ kv ≤ π (8.24)
kv
Note that to invert this relation to estimate kv we can use the following con-
venient invertible approximation for the SINC function, valid over the range
0 to π:
 
sin(x) (π − x) 1.25
y= ≈ sin 0 ≤ x ≤ π, 0 ≤ y ≤ 1
x 2
⇒ x ≈ π − 2 sin−1 (y0.8 ) (8.25)
⇒ k̂v ≈ π − 2 sin−1 (Re(γ̃ (w2 )e−iφ )0.8 )

Note that this approach, when combined with equation (8.23), allows us to
calculate the surface phase φ0 without the need for a separate straight-line
coherence region assumption. We can then calculate the structure parameter for
arbitrary polarization w from the imaginary part of the phase-shifted coherence,
as shown in equation (8.26):

Im(γ̃ w e−iφ ) Im(γ̃ e−i(kv +φ0 ) )kv2
â10 (w) = = (8.26)
|f1 | sin kv − kv cos kv
Finally, we can reconstruct the vertical profile by knowing the interferometric
wavenumber βz to calculate the height from kv and using the Legendre coef-
ficient to reconstruct the profile with unit integral over this height range, as
shown in equation (8.27):
 
2k̂v 1 2â10
ĥv = ⇒ fˆL1 (z) = (1 − â10 ) + z 0 ≤ z ≤ ĥv (8.27)
βz ĥv ĥv
Figure 8.7 shows a schematic summary of the types of structure function we can
construct from this simple first-order truncation. We note that the maximum and
minimum of this first-order structure function are given simply in terms of the
parameter a10 , as shown in Figure 8.7. Note that for a10 > 1 this has a negative
minimum on the surface. This may seem to violate the important physical
requirement that f (z) be non-negative (since physically it represents scattered
power as a function of depth). However, such a restriction is not necessary when

^ max
fL1
z = hv
^ ^ ^ min
fL1(0) = (1– a10) = fL1
^ ^ ^ ^
fL1(hv) = (1+ a10) = fL1max
a10 > 1 a10 < 0
a10 = 0

a10 < 1

^ min Fig. 8.7 Summary of first-order reconstruc-

fL1 f (z) tion of the vertical structure function
298 Parameter estimation using polarimetric interferometry

Structure function and Legendre approximation

1
6th order
0.9 4th order
2nd order
0.8 1st order
original
0.7

Normalized height
0.6

0.5

0.4

0.3

0.2

0.1

0
Fig. 8.8 Examples of bipolar Legendre struc- –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
ture function estimates of a non-negative step
structure function Relative scattering density

we realize that fL (z) is only a band-limited approximation to the true structure

function. While the true function is always non-negative, its approximation can
go negative, indicating more concentrated scattering from the top of the layer.
To illustrate this we show, in Figure 8.8, various Legendre approximations (up
to sixth order) for a vertical step function at 50% of layer depth; that is, the
true structure function involves uniform scattering, but only from the top half
of the layer. We see that all approximations (including the first-order as used
in fL (z)) go negative at low z values, and this is a direct consequence of the
true physical structure. Hence it is useful to allow such negative profiles in the
estimation, on the understanding that we only ever obtain an approximation to
the true profile. If it is important to maintain positivity in the approximation,
then the negative parts of the profile can be set to zero and the estimate still
have some physical correspondence with the true profile. In this case, for the
linear approximation of this model, the scattering is non-zero only for elevated
heights above (below) a critical height, zc , expressed simply in terms of a10 , as
shown in equation (8.28):
 
(a10 − 1) hv 1
|a10 | > 1 ⇒ zc = hv = 1− (8.28)
2a10 2 a10

For positive a10 , zc lies between the surface and half the height, while for
negative it lies between the top and the half height.
Before proceeding further it is important to consider the range of the three
unknown parameters φ0 , kv and a10 with a view to investigating the uniqueness
of this model inversion. The phase φ0 is defined in the range 0 to 2π, with
ambiguities arising for surface variations in excess of this. These correspond
to the classical phase unwrapping problem in radar interferometry. However,
here we are more concerned with phase shifts relative to φ0 , and can therefore
ignore the phase unwrapping problem—at least initially. If we consider scenar-
ios where βz is always positive (by appropriate selection of master and slave
 8.2 Estimation of height hv 299

tracks for the baseline generation) and hv is also positive, then a good working
range for kv is 0 ≤ kv ≤ π. Although, mathematically, kv can go to infinity, in
practice we would wish to restrict it to avoid phase ambiguities in the layer; that
is, we design the interferometer to ensure that the depth of the layer is always
less than the 2π ambiguity height of the interferometer. This then restricts kv
to the range specified.
The Legendre coefficient a10 , on the other hand, can be bipolar—negative
or positive—but must be constrained so that the magnitude of the right-hand
side in equation (8.21) is always less than or equal to 1 (to match the limits of
coherence on the left-hand side). This requires the following inequality to hold:
" &
'
1 − f02 ' kv2 − sin2 kv
|γ̃ | ≤ 1 ⇒ |a10 | ≤ =( (8.29)
|f1 |2
2
sin kv
kv2 − 2 sin kvkcos
v
kv
+ cos2 kv

The range of a10 allowed under this constraint varies as a function of kv , as

shown in Figure 8.9. Note that the range is limited to ±1.732 for low values,
and up to ±π for large values of kv . Note also that the variation of coherence
with structure (kv fixed and a10 variable) is a straight line in the complex plane,
intersecting the unit circle at two points corresponding to the limits given in
equation (8.29).
It is interesting to look at the variation of this line with kv . We show such a
loci in Figure 8.10, where we have removed any topographic phase so that the
surface phase is zero and lies at the point O. The set of lines are for variation
of kv from 0◦ to 180◦ in 5-degree steps. The loci of points of the uniform SINC
model (a10 = 0) are shown as black stars, which we see constitute a spiral.
However, passing through each SINC value is now a straight line (solid for
positive a10 values, and dashed for negative). Clearly, if φ0 is not zero then
the whole diagram is just rotated clockwise for negative and anticlockwise for

Variation of the first Legendre coefficient with kv

4

minimum
3 maximum

2
Maximum bounds on a 10

–1

–2

–3

–4 Fig. 8.9 Maximum (dash) and minimum

0 0.5 1 1.5 2 2.5 3
(solid) bounds on first Legendre coefficient
kv as a function of kv
300 Parameter estimation using polarimetric interferometry

90
1
120 60
0.8

0.6
150 30
0.4

0.2

180 0

210 330

Fig. 8.10 Coherence loci for first-order Leg-

endre approximation: positive a10 (solid) and 240 300
negative a10 (dash) for different kv values
(black stars) 270

positive phase shifts. In any case there are two important observations from
this result:
1. The first-order model does not cover the whole coherence diagram, and
severely limits the possible set of valid coherences. In fact, for a fixed
kv the valid coherences are constrained to lie along a single straight line
going through the appropriate SINC point.
2. It follows from this model that the coherence variation with polarisation
should also lie along a line in the complex plane. However, this line
does not intersect the unit circle at the topographic phase point, and
hence is not the same line as used in other coherence models, such as
the two-layer random-volume-over-ground or RVOG model.
We also note from Figure 8.10 that there is also some ambiguity for phase shifts
less than π/2, where we see the intersection of different lines. However, these
ambiguities can be explained physically as the equivalence of a thick layer with
all scattering from the surface region having the same phase centre as a thin layer
with all scattering coming from the top. In order to enable a unique inversion
we can restrict the model to positive values of a10 ; that is, to scenarios where
the scattering increases rather than decreases with height into layer 1. These
positive loci are shown in Figure 8.11, where we also superimpose the linear
coherence loci of RVOG and its distortion for temporal effects on volume-
only scattering. Clearly, from this the first-order Legendre series cannot be
used to fully represent coherences with large µ values in the RVOG model.
Nor can it even be used for all temporal decorrelations in the volume-only
scattering case.
From these observations we conclude that the conditions are very restrictive
for the inversion scheme of equations (8.23)–(8.27) to apply. Consequently,
we can consider the first-order Legendre model to be inappropriate for general
physical applications. At the very least, for fixed kv , we require better coverage
 8.2 Estimation of height hv 301

90 1
eif2 120 60
0.8

0.6
150 30
0.4

0.2

180 0
eif

210 330

240 300
Fig. 8.11 Overlay of RVOG model on posi-
270 tive first-order Legendre model

of the complex plane so that we can model a wider range of physical scenarios
(including the RVOG model). To do this we must extend the model to second
order, as we now consider. However, we shall see that in doing so we can still
make use of some of the inversion ideas from this first-order truncation.

8.2.2 Second-order Legendre model

We now consider truncation of the Legendre series to second order, as shown
in equation (8.30). Again we assume that we can isolate all polarisation depen-
dence in the Legendre spectrum, which amounts to the assumption that height,
topographic phase and baseline are all invariant to polarisation changes.
/      
sin kv  sin kv cos kv 3 cos kv 6 − 3kv2 1
γ̃ (w) = ei(kv +φo ) + a10 w i − + a20 (w) − + sin kv + R2
kv kv2 kv kv2 2kv3 2kv
(8.30)

Here R2 is again the truncation error. This error is of the order of the absolute
value of the next term in the coherence series normalized by |f0 |; in this case,
R2 ≈ |f3 |/|f0 |. This is generally small. From Figure 5.15 we see that for 0 ≤
kv ≤ π this is much smaller than the first-order truncation error R1 . Hence
the second-order truncation offers a much more accurate model; but it has an
increased number of parameters, and so we now turn to consider its suitability
for inversion.
We see that we now have two polarisation dependent coefficients, a10 and
a20 , which together with kv and φ0 constitute a set of four unknowns to be
determined from the two observations (amplitude and phase of coherence).
Here, polarisation diversity does not seem to help us, as each additional w adds
two observations but also adds two new unknowns. Instead we need to develop
a different strategy to invert equation (8.30). First, however, we investigate the
expanded coverage in the complex plane of this second-order model.
302 Parameter estimation using polarimetric interferometry

To visualize coverage of the second-order Legendre coherence model (equa-

tion (8.30)) inside the unit circle, we rewrite the coherence in the following
second-order form:

γ̃ (w) = ei(kv +φo ) (f0 + a10 (w)f1 + a20 (w)f2 )

= z(a20 ) + a10 d
 
ikv +φ sin kv cos kv
⇒d =e i − (8.31)
kv2 kv
 /   
ikv +φ sin kv 3 cos kv 6 − 3kv2 1
z(a20 ) = e + a20 − + sin kv
kv kv2 2kv3 2kv

In this form we see that a10 again generates a line in the complex plane passing
through a fixed-point z with direction d , but that now the fixed point is itself
determined by the second-order coefficient a20 . Since the function f2 is always
negative (see Figure 5.15), it follows from equation (8.31) that for negative a20
the fixed point moves radially outwards towards the unit circle, and for positive
a20 it moves inwards towards the origin of the coherence diagram. Thus, for
a fixed kv the result is a family of straight lines, all with the same slope and
defined by a set of fixed points generated by a radial line passing through the
SINC point, as shown in Figure 8.12. Here we show in thick black line the
spiralling SINC locus for φ0 = 0. For each point on this curve there is now
a family of lines generated for positive and negative a10 (shown as solid and
dashed lines respectively) and moving up and down the radial line through the
SINC point according to positive or negative a20 . As we move along the SINC
locus this pattern of lines is rotated and shifted accordingly. The bounds of a10
for fixed a20 can then be derived by setting the coherence to unity, as shown in

a10 > 0
90 1
120 60 a10 < 0
0.8
a20 < 0
0.6
150 30
0.4

0.2 a20 > 0

180 0

210 330

240 300
Fig. 8.12 Coherence loci for the second-
order Legendre approximation 270
 8.2 Estimation of height hv 303

equation (8.32):

|γ̃ |2 = (f0 + a10 f1 + a20 f2 ) (f0 + a10 f1 + a20 f2 )∗

= f02 + 2a20 f0 f2 + a10 f1 + a20
2 2 2 2
f2 = 1
" "
1 − (f0 + a20 f2 )2 1 − (f0 + a20 f2 )2
⇒− ≤ a10 ≤ (8.32)
|f1 | 2 |f1 |2

If we then set a10 = 0 we obtain the corresponding bounds of a20 , as follows:

1 − f0 (1 + f0 )
(f0 + a20 f2 )2 = 1 ⇒ ≤ a20 ≤ − (8.33)
f2 f2

We note that for a fixed kv value these bounds now lead to full coverage of
the unit circle. Hence, if we know the kv value then we can use the position of
any sample coherence to estimate the two parameters a10 and a20 , as shown in
equation (8.34):
 
Im(γ̃ w e−iφ ) Re(γ̃ w e−iφ ) − f0
a10 (w) = a20 (w) = (8.34)
|f1 | f2

The structure function itself can then be expressed as in equation (8.35):

 
1 2z 6z 2
fˆL2 (w, z) = 1 − â10 + â20 (w) + (â10 (w) − 3â20 (w)) + â20 (w) 2 0 ≤ z ≤ hv
hv hv hv
(8.35)

One interesting property of the second-order Legendre approximation fˆ2L (z)

is that its extreme value (maximum or minimum) no longer has to fall at the
boundaries of the layer. This provides us with more flexibility in representing
variations in the structure function itself, which is important in complex media
such as scattering from forest canopies (Woodhouse, 2006). The stationary
point of the estimated profile can be simply related to the Legendre coefficients
as shown in equation (8.36):
 
dfL2 (z) 1 a10
= 0 ⇒ zm = hv − (8.36)
dz 2 6a20

Here we see, for example, if a10 = 0 then the minimum (maximum) of the
profile occurs at half the layer depth for positive (negative) a20 . In general,
if the ratio a10 /a20 is positive then the extreme point will occur in the lower
half of the layer, while for a negative ratio the extreme point will occur in
the upper half. In this way we can represent a much wider variety of structure
functions than is possible using the classical RVOG model (which assumes
that the maximum always occurs at the top of the volume). In particular we
note that since a20 can be positive or negative, we can now represent functions
with a maximum response below the top of the layer (negative a20 ) or with an
enhanced response from the surface position at z = 0 (a20 positive). The former
is useful for representing non-exponential volume scattering profiles, while the
latter can be used to represent changes in µ—the effective surface-to-volume
scattering ratio.
304 Parameter estimation using polarimetric interferometry

To see an example of the flexibility of this second-order approximation we

consider its application to the two-layer RVOG model (see Section 7.4.2). The
special form of this model that we use is summarized in equation (8.37):

 iφ0 γ̃v + µ w sin kv
γ̃ w = e γ̃v = eikv (8.37)
1+µ w kv

For simplicity we show a case where the volume-only coherence (µ = 0)

is given by a simple zero extinction medium (a uniform structure function).
The factor µ then corresponds physically to the ratio of surface-to-volume
scattering. We also assume that the surface phase φ0 = 0 in this example.
Figure 8.13 shows how this model maps onto the Legendre coordinates (defined
by kv ) for µ in the range −30 dB to +30 dB in 1-dB steps.
Here we have used a specific example when hv = 10 m, βz = 0.2, and so
kv = 1. Each point of the RVOG model now has a set of Legendre coordinates
a10 , a20 , as shown in Figure 8.14. We see that when µ is small the coordinates are
both zero, corresponding to the assumption of a uniform structure. However, as
µ increases we see that a10 increases in a negative direction while a20 increases

90 1
120 60
0.8

0.6
150 30
0.4

0.2
Fig. 8.13 The RVOG model (stars) superim-
posed on the second-order Legendre coordi-
nate system 180 0

RYOG model: a10 (solid) and a20 (dash)

5

3
Legendre coefficients

–1

–2

Fig. 8.14 Variation of Legendre coefficients –3

a10 (solid) and a20 (dash), with µ for the –30 –20 –10 0 10 20 30
RVOG model mu (dB)
 8.2 Estimation of height hv 305

RVOG : Legendre structure function estimates

0.8
10
0.7

0.6
8
0.5

6 0.4
Height (m)

0.3

4 0.2

0.1
2 0

–0.1
Fig. 8.15 Variation of second-order Legen-
0
–30 –20 –10 0 10 20 30 dre structure function approximation of the
RVOG model with surface-to-volume scatter-
Mu (dB) ing ratio µ

in the positive direction. This reflects changes in the structure function itself.
Figure 8.15 shows an image of how the second-order approximation to the
RVOG structure function varies with µ. Each vertical profile extends over
10 m, and is normalized so that its integral is unity (as in equation (8.35)). On
the left we see the uniform volume scattering profile obtained when µ = 0. As
µ increases we see a shift in the structure function to more localized surface
scattering, as physically expected in the RVOG model. Again we note that as the
surface contribution increases, the second-order approximation is forced to go
negative at some points in the volume. This again reflects the approximate nature
of the truncation rather than any physical interpretation of negative scattering
amplitudes.
In conclusion, we have seen that a second-order Legendre expansion is the
lowest-order truncation capable of providing full unit circle coverage. We have
demonstrated its application to the widely used RVOG coherence model to
demonstrate its ability to reflect changes in the underlying structure function.
There remains, however, one issue with this model. It has too many unknowns
to be inverted and hence to be directly applied for height estimation. In the next
section we turn to consider methods for resolving this limitation.

8.2.3 Approximate height estimation from

the second-order Legendre series
In the previous section we showed that a second-order truncation of the Leg-
endre series is useful for characterising a wide range of different structure
functions. The problem with this model is that we have more unknowns than
observations. For a single polarization channel, single wavelength and single
baseline we have only two observations (one complex coherence), while the
model has four unknowns (the two Legendre coefficients and two structural
parameters φ0 and kv ). The only observation in our favour is that φ0 and kv
306 Parameter estimation using polarimetric interferometry

are invariant to changes in polarisation. In order to progress, therefore, we

need somehow to estimate two of these parameters so as to obtain a balance of
two unknowns and two observations. In this section we consider approximate
methods for achieving this. In particular we look at ways of estimating φ0 and
kv , with layer depth then following from the latter. To do this we will need to
impose some further constraints on properties of the unknown structure func-
tion, but we will see that these can still be rather lax, allowing some flexibility
(and critically more so than in the fixed structure approaches like RVOG) in
determining variations in structure.
We start by noting that if we adopt the slightly more general SVOG model
for our second-order Legendre series (see Section 7.4.3) we can still use a
line-fit technique to obtain estimates of φ0 ; that is, the surface phase can again
be estimated from equation (8.5) or (8.16). In this scheme we maintain the
assumption that the upper layer comprises random volume scatterers; the only
difference with RVOG is that the volume contribution can now have arbitrary
structure function and not just an exponential.
In order to estimate height we first use this φ0 estimate to obtain a phase-based
estimate, exactly as proposed in equation (8.19). However, as we noted earlier,
this phase centre separation, according to SVOG, can lie anywhere between
halfway and the top of the layer, and hence in general underestimates the true
layer depth.
To progress, one key idea is that this error can be at least partly compensated
by employing a coherence amplitude correction term. The idea is that as the
phase centre separation increases due to changes in structure function so, at
the same time, the effective volume depth decreases (as the structure function
becomes more localized near the top of the layer), and hence the level of volume
decorrelation will decrease. A convenient invertible model for this coherence
amplitude process is just the f0 or SINC coherence function, as discussed in
equation (8.25). Just as required, when coherence amplitude decreases, so this
height (or kv ) estimate will decrease at the same time as the phase estimate
increases. Finally, by combining these two terms with a scaling parameter η
we then obtain an approximate algorithm that can compensate variations in
structure, as shown in equation (8.38):
1@  0.8 A
k̂v = arg(γ̃wv e−iφ̂0 ) + η(π − 2 sin−1 (γ̃wv  )
2
2k̂v
⇒ ĥv = (8.38)
βz
The first term represents the phase component (using the estimated surface
phase together with our estimate of ‘volume-only’ coherence channel). The
second—the coherence amplitude correction—is weighted by η, to be selected
so as to make the full expression as robust as possible to changes in the structure
function.
This expression has the right kind of behaviour in two important special
cases. If the medium has a uniform structure function then the first term will
give half the height or βz hv /2, but the second will then also obtain half the
true height and yield βz hv /2 (if we set η = 1), and so half the sum gives the
correct kv estimate. At the other extreme, if the structure function in the volume
channel is localized near the top of the layer, then the phase height will give
 8.2 Estimation of height hv 307

the true height βz hv , and the second term will approach zero. Half the sum
then still produces the correct kv estimate. The idea is that equation (8.38) will
provide a reasonable estimate for arbitrary structure functions between these
two extremes. It requires estimates of only two parameters: the truesurface
topographic phase φ0 , and the volume-only complex coherence γ̃ wv .
We can extend this idea further and estimate an optimum value of weighting
factor η by using the second-order Legendre structure model for the volume
coherence channel, as shown in equation (8.39):
 
γ̃ (wv ) = ei(kv +φo ) (f0 + a10 wv f1 + a20 wv f2 ) (8.39)

Now, if we allow a10 and a20 their full range we can fit this coherence to
any kv value, and hence seem to undermine the approximation proposed in
equation (8.38). However, by making some reasonable physical assumptions
about volume scattering we can reduce the working range of a10 and a20 as
follows.
The basic idea is that in the selected special volume-only channel we assume
there is zero surface scattering component, and hence its structure function
should have a local minimum at the surface (at z = 0). This in turn requires
that in this polarization channel wv we restrict a10 ≥ 0 and a20 ≤ 0. When
combined with the limits derived in equations (8.32) and (8.33) we will see that
this constrains the kv values satisfying equation (8.39). Figure 8.16 illustrates
the results. Here we show along the abscissa a set of true kv values in steps of 0.1
over the range 0–2. The corresponding ordinate shows the spread of estimated
kv values obtained using equation (8.38) with η = 0.8. This value is selected to
fit the a20 = 0 variation, and the underestimates we see are then entirely due to
the presence of non-zero a20 structure. It also ensures that equation (8.38) will
always estimate the minimum height consistent with the data. While the trend
in Figure 8.16 is encouraging, some of the errors can apparently be quite large.
For example, if the estimate yields a value of 1 (in the ordinate of Figure 8.16)

Comparison of estimated vs. true kv values

2

1.8

1.6

1.4
Estimated kv value

1.2

0.8

0.6

0.4

0.2

0 Fig. 8.16 Estimated versus true kv values for

0 0.5 1 1.5 2
the full range of volume structure functions
True kv value a10 , a20 , and using η = 0.8 in equation (8.39)
308 Parameter estimation using polarimetric interferometry

Fractional error in kv estimate

0.2
0.1
0
–0.1
–0.2
–0.3
–0.4
–0.5
0
–1 2
1.5
–2 1
Fig. 8.17 Fractional error for full range of 0.5
–3 0
volume structure functions for kv = 2 a 20 a 10

then we see that the true value could actually lie anywhere between 1 and 1.6,
depending on the volume structure. However, this simple interpretation masks
an important issue. The underestimation errors for each abscissa point in Figure
8.16 increase in proportion to a20 , as shown for the kv = 2 case in Figure 8.17.
Here we plot the fractional error in kv estimate for all valid values of a10 and a20
(but note that this behaviour is typical of all other values, and so our conclusions
will apply equally for all values of kv ).
The key observation from Figure 8.17 is that the largest errors always occur
for large a20 values. This makes physical sense, as this occurs for a quadratic
profile, which is a second-order approximation to a Dirac delta function located
halfway up the volume. In this case we can achieve a combined high-coherence
and low-phase centre, in contradiction to the assumptions behind equation
(8.38). This is a problem faced by all height estimation techniques based
on interferometry. If the volume has a top height hv but the bulk of the
scattering comes from halfway up the volume, then interferometry ‘sees’ a
smaller effective height. In order to resolve such ambiguities we need to add
extra information beyond a single baseline single wavelength interferometer.
Other possibilities include adding more baselines or using a second frequency
(Treuhaft, 2000b, 2004; Reigber, 2000, 2001; Neumann, 2008).
If we accept that such errors can occur for a single baseline, but only for rather
extreme (and unlikely) cases of the structure function, then we can proceed to
employ equation (3.38) as a reasonable approximation, with around 10–15%
estimation errors for a wide range of structure function variations.
We now turn to consider depth retrieval under the more extreme assumption
of a fixed structure function: namely, the exponential assumed by the RVOG
class of models.

8.2.4 Height estimation using RVOG

The above algorithm assumes that the volume-only structure function fv (z)
has the property that its Fourier–Legendre coherence contributions satisfy the
 8.2 Estimation of height hv 309

bounds a10 ≥ 0, a20 ≤ 0. To avoid such assumptions about the Fourier–

Legendre expansion we can instead construct a similar approach, but based
directly on the two-parameter RVOG model, where mean wave extinction
replaces the Legendre structure parameter a10 . This automatically generates
a structure function that decreases with depth. The layer depth is now obtained
by minimising the following function G(λ), being the norm of the difference
between the volume coherence (itself obtained from the observed coherence
γ̃wv by a line shift through the parameter 0 ≤ λ ≤ 1) and the RVOG model
prediction for volume-only scattering (µ = 0). The phase φ2 is the second unit
circle intersection point of the straight line fit (see Figure 8.2).


 p=
2σe
? ? 

?  p1 hv − 1 ?

 cos θ
? iφ̂2 iφ̂ p e ?
min G (λ) = ?γ̃wV + λ e − γ̃wV − e ? where p1 = p + iβz
hv ,σ ? p1 ephv − 1 ? 


 4π θ

 βz =
λ sin θ
? ?
? p ep1 hv − 1 ?
? ?
min G(λ = 0) = ?γ̃wV e−iφ̂ − ? (8.40)
hv ,σ ? p1 ephv − 1 ?

Shown in the lower portion of equation (8.40) is the simplified version of this
algorithm obtained in the case λ = 0; that is, when we can assume that the
observed coherence γ̃wv itself has µ = 0—that it contains volume-only scatter-
ing. We have also shifted the topographic phase estimate onto the observable
coherence. Again it is interesting to investigate the range of the two parameters
hv and σe with a view to determining coverage and uniqueness of the solution in
the complex plane. The range of hv is again capped by the requirement to avoid
phase ambiguities in the layer, so that 0 ≤ βz hv ≤ 2π . The extinction can also
have non-negative infinite extent; but in practice, as extinction becomes large
enough so the coherence amplitude becomes insensitive to changes in height,
and varies only the phase of the top of the layer. These trends are apparent in
the loci shown in Figure 8.18. Here we see the variation of coherence over

90 1
120 60
–if 0.8
g^ w e
v

0.6
150 30
0.4

0.2

180 0

210 330

240 300
Fig. 8.18 Overlay of SVOG linear region on
270 RVOG coherence loci
310 Parameter estimation using polarimetric interferometry

the full range of height for varying extinction. The inside spiral corresponds to
the reference zero extinction or SINC loci. We see that for zero extinction the
coherence falls to zero at the 2π height. However, as the extinction increases,
the curvature of the spiral reduces until for large extinction the loci is almost
circular, maintaining high coherence amplitude because of the small effective
volume contributing to decorrelation.
Although the loci are no longer straight lines we note that they provide a set of
non-intersecting curves, and so again if we overlay a sample volume coherence
(shown as the point in Figure 8.18) then we can find a unique solution to
equation (8.40) (for fixed λ), and thus secure an estimate of hv . Note again that
here the extinction parameter is acting as a structure compensation parameter,
allowing height estimation for a wide range of structure functions approximated
by exponentials with varying extinction rates. Note also that the coverage is
not complete. If we draw a line through our sample coherence (shown as the
point in Figure 8.18) we see that if µ is above a certain level the coherence can
still fall outside coverage of the simple extinction model. In this case we have
to employ the free parameter λ in equation (8.40) to move the coherence back
into the valid region, but generally we have no idea which value of λ to use and
hence are likely to make errors in the height retrieval. In this sense we see that
the SINC curve defines a boundary for coverage, and any volume scattering
candidate coherence must lie above the SINC curve to enable a clear solution
in RVOG inversion.
The above algorithm uses the RVOG model to match the observed coherence
in both amplitude and phase. However, to do this we require estimates of the
topographic phase φ0 . We saw in Section 8.1 how to devise algorithms based
on RVOG and OVOG to estimate this parameter. However, sometimes the
coherence can be so low that this phase estimate is too noisy to use. In this
case we would like to employ an algorithm that does not make use of phase
information and relies only on coherence amplitude. We can devise such a
model based on the RVOG approach, as long as we assume that the structure
function has a known form (known extinction, in this case). We then have
only one unknown (the depth of the layer) and one observable, and can solve a
minimization problem as shown in equation (8.41):
?   ? 
?    p ep1 hv − 1  ?  p = 2σe
?     ?
min G = ? γ̃wv −   ? where cos θ (8.41)
hv ?  p1 ephv − 1  ? 
p1 = p + iβz

In the RVOG case, a known structure function implies knowledge of the mean
extinction coefficient in the medium σ e , as shown in equation (8.41). This
can sometimes be estimated from physical models of the environment or from
measurements and inversions from previous datasets (see Section 3.5.3). Note,
however, that matching coherence amplitude calls for good calibration and
compensation for SNR and temporal effects.
In conclusion, we have seen three important algorithms for estimating layer
depth (or height) from single baseline polarimetric interferometric data. Equa-
tion (8.38) represents an approximate method that makes minimal assumptions
about the layer structure function. Equation (8.40) assumes an exponential
structure via the RVOG model, and consequent matching of the coherence
in both amplitude and phase provides an estimate of both height and mean
 8.2 Estimation of height hv 311

extinction. Finally, if surface phase estimates are not available, a coherence

amplitude-only approach is shown in equation (8.41). This, however, requires
an a priori estimate of the mean extinction in the medium.

8.2.5 Depth estimation using OVOG

We turn now to consider problems faced in estimating layer depth hv when
layer 1 behaves as an oriented volume, with polarisation-dependent propagation
through the layer. In this case we can no longer consider having just a single
volume coherence point for inversion but a triplet of such points, as discussed
in Section 7.4.5. Nonetheless, we can still use the techniques developed for the
SVOG and RVOG models to derive simple algorithms for depth estimation in
oriented volumes, as follows (Treuhaft, 1999; Cloude, 2000a; Lopez-Sanchez,
2006, 2007).
In the same way as for topography estimation, we consider two special forms
of the oriented volume-over-surface model: the first with assumed small sur-
face contributions in all channels (to be called the finite OV problem), and
the secondly with high µ dispersion, when only one channel (with the high-
est extinction) has a volume-dominated response while the other (the lowest
extinction) has a large surface-to-volume ratio µ. We now consider each of
these cases in turn.

8.2.5.1 Finite OV height estimation algorithm

In this case we deal with backscattering by a cloud of volume scatterers, such
that there is no surface scattering from top or bottom of the layer, as shown
schematically in Figure 8.19. We are then interested in an algorithm for deriving
the layer depth and bottom phase from the triplet of observed coherences. As
the medium is not infinite we cannot use the OV solution of equation (8.17) to
find bottom or top phase by a circle fit. We therefore need to consider a more
integrated parameter estimation problem, whereby bottom phase is included as
an unknown at the same time as layer depth.

8.2.5.2 First-order Legendre OV inversion

We start by noting that despite the complications of anisotropic propagation, all
three coherences are characterized by the same depth hv (and hence the same
kv value) and surface phase φ, and differ only in their structure functions. So,
by assuming µ = 0 for all channels we obtain the ordered triplet of volume-
only coherences γ̃xx = γ̃1 , γ̃xy = γ̃2 , γ̃yy = γ̃3 from the eigenpolarisations x
and y, to set up the following Legendre cost function, based on that derived in


z = zo + hv

Oriented volume
Fig. 8.19 Geometry of scattering from an
z = zo
oriented volume with finite depth
312 Parameter estimation using polarimetric interferometry

90 1
120 60
0.8

0.6
150 30
0.4

0.2

180 0

slope = cot(kv + )

210 330

Fig. 8.20 Example oriented volume coher- 240 300

ence triplet superimposed on coherence loci
for a first-order Legendre structure function 270

equation (8.21).
3 ?
!   ?
? iφ ikv sin kv j sin kv cos kv ?
min G = ? γ̃j − e e + ia − ? (8.42)
kv ,a1 ,a2 ,a3 ? kv kv2 kv ?
j=1

Note that here there are five unknowns and six observables. The parameter
bounds are the same as those derived earlier (0 ≤ kv ≤ π , 0 ≤ aj ≤ π , 0 ≤ φ <
2π), and the only difference now is that for each fixed pair of values kv , φ we
seek a triplet of Legendre coefficients aj that minimize the above function. The
global minimum of such searches will then give us estimates of the parameters
of the layer. There is, however, a simple geometrical interpretation of this
inversion, as we now consider.
The first point to make is that the model of equation (8.42) implies that the
triplet of coherences lie along a line in the complex plane (see Figure 8.20).
These lines are shown in Figure 8.11, and derive from the assumption of a trun-
cated Legendre series. Hence a starting point for the suitability of the inversion
of (8.42) is to test whether or not the components of the triplet are collinear. If
they are, then equation (8.42) will have a good match with the data, otherwise
the assumptions of the truncated Legendre series may be invalid. Note, inci-
dentally, that this line does not intersect the unit circle at the surface topography
point. This is in contrast to the RVOG line fit employed in Section 8.1 to find sur-
face topography from this intersection. It is therefore important in applications
to be able to differentiate between RVOG and finite OV before applying the
appropriate parameter estimation. This can be accomplished in several ways—
for example, by checking the orthogonality of the optimum coherence states:
for single layer OV they will be orthogonal, while for two-layer RVOG they
will not be orthogonal.
The second point to note is that the slope of the line joining the three coher-
ences is given simply in terms of the two fixed parameters kv and φ, as shown in
equation (8.43). Therefore, by fitting a line through the three volume coherences
 8.2 Estimation of height hv 313

and measuring its slope, we can obtain directly an estimate of the parameter
kv + φ. We can then employ a single channel model fit—as in the random
volume case—to any one of the three coherences (for example, the maximum
coherence γ1 ) to obtain an estimate of kv from minimization of the function
shown in equation (8.43).
  
sin kv sin kv cos kv
γ̃ = eiφ eikv + iaj − = z + aj d
kv kv2 kv
 
ikv +φ sin kv cos kv
⇒d =e i −
kv2 kv
1
⇒m=− (8.43)
tan(kv + φ)
 
1
⇒ (kv + φ) = − tan−1
m
?   ?
? iφ+kv sin kv 1 sin kv cos kv ?
?
⇒ min G = ?γ̃1 − e + ia − ?
kv ,a 1 kv 2
kv kv ?

This value of kz will then, by definition, satisfy the other two polarisation
channels, and when combined with the slope estimate will provide an estimate
for the surface topography φ.
Having determined a simple algorithm based on the Legendre coherence
expansion, we now consider a solution based on exponential structure functions.

8.2.5.3 Exponential OV inversion

The differences in structure function with polarisation may be ascribed to expo-
nentials, as in the OVOG model. In this case we can derive a new cost function
as shown in equation (8.44):

 2σ1

 p =
? ?  1
 cos θ
3 ? (pj +iβz )hv − 1 ?


! ? pj e ? σ1 + σ2
min G = ?γ̃j − eiφ ? where p2 =
hv ,κ1 ,κ2 ? (pj + iβz ) e p j hv −1 ? 
 cos θ
j=1 



 p3 = 2σ2
cos θ
(8.44)
This has the advantage of having only four unknowns and six observables, as
it assumes a relationship between the co- and cross-eigenpolarisation channels.
This does, however, lead to an assumed rank ordering of the three polarisations
that the Legendre approach of equation (8.43) does not require. Again, hv and
φ are common to all three channels, and the differences between polarisations
are modelled by variation of extinction. There are no straight lines embedded
in this equation, and solution is best tackled by a brute-force iterative search
technique for the four-dimensional minimization.

8.2.6 OVOG model height estimation

In the second case to be considered we redirect attention to oriented volume
problems where the influence of the underlying surface cannot be ignored, as
314 Parameter estimation using polarimetric interferometry


z = zo + hv

Oriented volume

Fig. 8.21 Geometry of the oriented-volume- z = zo

over-ground (OVOG) model

shown schematically in Figure 8.21. Here the underlying surface contributes a

µ value in one or more polarisation channels, and so we cannot use the simple
volume decorrelation models of the previous section. Again, by restricting
attention to the co- and crosspolarised combinations of eigenpolarisations for
layer 1, we can formally set up the following OVOG cost function:

? ?
Govog = ?γ̃xx − eiφ (γ̃vo
xx xx ?
+ Fxx (1 − γ̃vo ))
? xy ?
+ ?γ̃xy − eiφ (γ̃vo + Fxy (1 − γ̃vo ))?
xy

? yy ?
+ ?γ̃yy − eiφ (γ̃vo + Fyy (1 − γ̃vo ))?
yy
(8.45)

This has six observables but seven unknowns, and hence is not well suited to
solution as it stands. To be able to make inversion tractable we need to make
some further assumptions about one or more of the parameters in the model.
The simplest of these is the high µ dispersion assumption—also used in Section
8.1.4 for topography estimation. In this case we assume that the extinction in
the x polarisation is so high as to reduce µ in this channel to zero, while the
low extinction channel y maintains a high surface-to-volume scattering ratio.
In this case we can simplify the cost function as shown in equation (8.46):

? ?
Govog = ?γ̃xx − eiφ γ̃vo
xx ?
? xy ?
+ ?γ̃xy − eiφ (γ̃vo + Fxy (1 − γ̃vo ))?
xy

? yy ?
+ ?γ̃yy − eiφ (γ̃vo + Fyy (1 − γ̃vo ))?
yy
(8.46)

This now has six unknowns and six observables. This can be further simplified
if we use the xx and yy channels to estimate surface topography using the line
fit technique, as described in Section 8.1.2. In this way we can reduce the
balance of equation (8.46) to five unknowns and six observables. Problems
arise with this approach if µ remains high in all channels. This can occur, for
example, in applications involving thin layers with low to moderate extinctions
at high angles of incidence. This causes several problems, as the coherences
then all migrate down their lines towards the surface topography point, and it
becomes more difficult to fit lines to the topography point or find the true volume
scattering channel for depth estimation. The only approach then is to constrain
the range of extinctions expected in the problem (by physical modelling or
external measurements), and then use these as known parameters in equation
 8.3 Hidden surface/target imaging 315

(8.46) to leave five unknowns for the six observables. Of course, if additional
further information is available (for example, the phase φ of the surface from
external measurements) then this can be added to further reduce the parameter
imbalance. Whichever course is taken, the result is a set of inversions across
the parameter range of extinctions. These provide us with a mean solution and
error bars associated with the spread of solutions.

8.3 Hidden surface/target imaging

In this section we consider methods for using polarimetric interferometry to
separate volume and surface contributions to the total backscattering cross-
section of two-layer problems of the form shown in Figure 8.22. Here we
show, on the left, the combined surface and volume scattering geometry, and
on the right the main objective of isolating the (effective) surface components.
Here there are two primary motives. The first is to be able to study the surface
properties such as surface roughness and moisture, even in the presence of a
vegetation or snow layer, for example (Cloude, 2005a). The second objective
is to be able to image the surface beneath the volume (using synthetic aperture
radar, for example) in order to detect objects located on the surface and obscured
by the top layer (Sagues, 2001). This application includes foliage penetration,
or FOPEN, in military detection as well as in search-and-rescue in forested or
avalanche conditions (Cloude, 2004).

8.3.1 RVOG estimation of µ

The first step in this process is to identify a polarisation channel, which if
possible contains only volume scattering and no surface component at all. We
can again find the best approximation to such a channel by using physical
arguments or the coherence optimization techniques of Section 6.2, where we
identify the polarisation channel with smallest µ as the one with the largest
phase bias. In either case this allows us to find a reference complex coherence
for the volume scattering component. If we now assume that layer 1 is a random
volume, then according to the SVOG model this point lies on a line joining the
volume coherence to the surface phase point on the unit circle. If we now
calculate the complex coherence in any other polarisation channel w, then we
find F(w)—the fraction of surface scattering in this channel by a line fit between
two complex values. This parameter can then be directly estimated from the

o(w) = ov(w) + oes(w)

 oes(w) = o(w) + ov(w)

Random volume  Fig. 8.22 Schematic representation of the

hidden surface imaging problem or 2-to-1
layer conversion
316 Parameter estimation using polarimetric interferometry

two complex coherence values, as shown in equation (8.47):

√ 
−B − B2 − 4AC 

F(w) = 0 ≤ F(w) ≤ 1  F(w)
2A ⇒ µ(w) =
 2  2 
 1 − F(w)
A = γ̃wv  − 1 B = 2Re((γ̃ (w) − γ̃wv ).γ̃w∗v ) C = γ̃ (w) − γ̃wv  
(8.47)

Note that this a just a special case of the general phase bias removal algorithm
of equation (8.5), the main difference being that now the reference point is
assumed to have µ = 0 and F is the desired parameter, rather than just being
an intermediate step towards phase estimation.
This approach works acceptably well for large µ, but for small µ the two
coherences are close in the complex plane, and so any small errors in coherence
estimation can lead to large errors in line fit. A more robust approach is therefore
to find the estimated surface phase φ using large µ separations or a least squares
line fit (as in equation (8.16)), and to then project all coherences onto the best-
fit line joining the volume coherence to the estimated unit circle point before
estimating F for a given polarisation w, as shown in equation (8.48):

γ̃p w − γ̃v
F(w) = (8.48)
eiφ̂ − γ̃v
Here the projected value of the coherence is given in terms of the line fit
parameters m and c, as shown in equation (8.49):
  
Re γ̃ (w) + m̂.Im γ̃ (w) − m̂.ĉ 
xi = 
1 + m̂2 γ̃p w = xi + iyi (8.49)


yi = m̂.xi + ĉ

This relation is derived by minimising the distance between the coherence γ̃ (w)
and the line with known slope and estimating y = m̂x + ĉ, posed as shown in
equation (8.50):
? 2 ?
? ?
min ?(Re(γ̃ ) − xi )2 + m̂xi + c − Im(γ̃ ) ? (8.50)
xi

F(w) g~p(w) This algorithm is summarized schematically in Figure 8.23, where we show the
end points of the linear coherence region for the SVOG model assumption. The
~ parameter F is then just the fractional distance along the line from the volume-
g w ~ eif
v g(w) only coherence point passing through the projected coherence towards the unit
circle. Clearly, if the coherence approaches the unit circle F = 1 we have 100%
Fig. 8.23 Projection of general coherence
onto the line model surface scattering. This fractional parameter can then be directly used with an
estimate of the total scattering cross-section σ to isolate the effective surface
component, as shown in equation (8.51):
    
σ w = σv w + σes w = (1 − F)σ w + Fσ w
 
⇒ σes w = Gs (w) = Fσ w (8.51)

Note the following important points about this decomposition:

1) The effective surface component is attenuated by extinction through the
volume, and hence is not the same as the bare surface return. However,
 8.3 Hidden surface/target imaging 317

under the SVOG assumption this attenuation is equal in all polarisation

channels, and so polarisation ratios will be preserved. This is important,
as several surface parameter retrieval algorithms employ ratios rather
than absolute values. For example, surface moisture and roughness under
the X-Bragg model (see Section 3.2.1) can be found from functions R
and M as ratios of the Pauli scattering components. In the two-layer
context we can now replace these formulae for bare surfaces with the
following ratios, to be used to estimate parameters for a surface hidden
beneath a random scattering layer:
   
Gs wHH −VV − Gs wHV Gs wHH −VV + Gs wHV
R=   M = 
Gs wHH −VV + Gs wHV Gs wHH +VV
(8.52)

However, this result masks a problem with this approach: the confusion
of direct and specular surface scattering in the polarimetric response, as
we now consider.
2) The surface component in polarimetric interferometry is ‘effective’ in
that it includes everything with a phase centre located on the surface.
As we have seen in Section 7.3, this includes not just the direct surface
return but also the specular second order scattering. Note that this acts
to enhance F(wHH−VV ) at the expense of F(wHH+VV ) because of the
π polarimetric phase change on specular reflection, and so will distort
the moisture and roughness estimates of equation (8.52). A method of
correcting for this is to estimate the full polarimetric coherency matrix
for the surface components and use an incoherent decomposition to sep-
arate the direct and dihedral components before applying appropriate
surface parameter techniques. One way to do this is to model the surface
components as a rank-3 reflection symmetric coherency matrix, which
can then be reconstructed from seven separate Gs estimates, as shown
in equation (8.53):
 
1
 Gs (w1 ) (Gs (w4 ) − Gs (w5 ) − i(Gs (w6 ) − Gs (w7 ))) 0 
1 2 
[Ts ] = m 
 (Gs (w4 ) − Gs (w5 ) + i(Gs (w6 ) − Gs (w7 ))) Gs (w2 ) 0  
2 
0 0 Gs (w3 )

where
         
1 0 0 1 1
         
      1   1  
w1 = 0
 
w2 = 1
 
w2 = 0
 
w4 = √ 1
 
w5 = √ −1
 
(8.53)
2 2
0 0 1 0 0
  

1 1
1 
 
 
1   
w6 = √  i  w7 = √ −i
2  2 
0 0

We can then apply any of the incoherent decomposition theorems of

Section 4.2 to separate the dihedral and direct components of this
matrix.
We now turn to consider a sensitivity analysis of the various surface/volume
separation algorithms. We have already seen that one way of estimating µ from
318 Parameter estimation using polarimetric interferometry

polarimetry alone (without the need for interferometry) is given by incoherent

decomposition (see Section 4.2.3), the model-based form of which is sum-
marized again in equation (8.54), where tii are the elements of the combined
surface and volume coherency matrix and we set Fp = 2 (see equation (4.62)).

(t11 + t22 − 3t33 ) ± (t11 − t22 − t33 )2 + 4 |t12 |2

mv = t33 md ,s =
2
) * 1
 2 −2
 t12 
αd ,s = cos −1 
1 +  


t 22 − t33 − md ,s

mmax = max(md ,ms ) ⇒ αmax

mmax 1
⇒ µmax = (sin2 αmax + cos2 αmax ) (8.54)
mv 2
Here we show how to estimate the maximum µ ratio from coherency matrix
data. Key to this is an assumption about the depolarisation caused by the vol-
ume component, in that it has the characteristic diagonal 2:1:1 structure of
dipole scattering. The problem with this approach is that the volume term has
a high scattering entropy and hence requires a large number of data samples to
reduce the variance of the estimated coherency matrix to sufficiently low levels
to be able to isolate any small surface contributions (Lopez-Martinez, 2005).
Furthermore, this speckle fluctuation can also lead to negative estimates for
scattered powers md and ms , unless explict attempts are made to enforce the
positive semi-definite nature of [T ].
As a measure of this lack of sensitivity we consider, as an example, a
simplified problem with a volume-only contribution—µmax = 0 and T =
diag(2,1,1)—and then use the Monte Carlo technique of Appendix 3 to gener-
ate random samples taken from a normal distribution with the same underlying
coherency matrix. We then use these samples to estimate the mean µmax
as a function of increasing number of samples. In the limit of an infinite
number of samples we will of course obtain µmax = 0, but we see in
Figure 8.24 that the convergence is rather slow, with more than 100 sam-
ples required to be able to identify −10 dB of surface contribution in a
volume scattering background. Again, this can be traced to the high scat-
tering entropy of the volume. On the other hand, coherent methods based
on polarimetric interferometry are potentially more sensitive, as the volume
decorrelation is now a function of the baseline/height product, which can be
designed to optimize performance. In addition, they involve relaxed assump-
tions about the nature of the volume scattering (as long as it maintains azimuthal
symmetry).

8.3.2 Optimum baseline for hidden surface detection

In contrast, we can express the sensitivity of interferometric coherence to sur-
face effects caused by µ by calculating the fractional change in the length
of the coherence line due to the presence of a surface component, as shown in
Figure 8.25. Here we can see that even with µ = −10 dB, the shift in coherence
is around 10% of the total line length. Therefore, in order to be able to detect
a small change in µ we need to choose the baseline so that the corresponding
 8.3 Hidden surface/target imaging 319

Estimated surface component as a a function of number of samples

0

–2

–4
Mu max (dB)

–6

–8

–10

–12

Fig. 8.24 Estimation of apparent surface

–14 scattering contribution in volume-only scat-
101 102 10 3
tering for the Freeman-eigenvalue model ver-
Number of samples sus number of looks

Fractional line length vs. surface component

0

–5

–10
mu (dB)

–15

–20

–25

–30 Fig. 8.25 Fractional line length in SVOG

0 0.1 0.2 0.3 0.4 0.5 model (F) versus relative level of surface
F scattering (µ)

change is detectable. To do this we therefore design the interferometer to have a

long line length in the complex plane, so we can be sensitive to the presence of
small surface components. The total line length itself is just given by |1 − γ̃vo |,
and hence determination of line length involves assumptions about the volume-
only coherence γ̃VO . On the other hand, we also wish to minimize the number
of data samples (L) required to ensure accurate estimates. These two require-
ments are in conflict, and require some compromize through correct baseline
selection, as follows. We start by calculating the derivative of coherence with
320 Parameter estimation using polarimetric interferometry

respect to the surface-to-volume ratio µ, as shown in equation (8.55):

∂ γ̂ (1 − γ̃vo )
= = f (µ)g (γ̃vo ) (8.55)
∂µ (1 + µ)2

Maximising sensitivity would then seem to require that γvo = −1; that is, the
baseline is chosen so that the phase centre for volume scattering lies at the π
height of the interferometer. To be realized, however, this phase must also occur
with a coherence magnitude of unity. This is not a realistic scenario, requiring
as it does infinite extinction in RVOG or a very localized structure function in
the Legendre approximation. Adopting the latter, we can express the line length
more realistically as a function of three parameters, kv a20 and a10 , as shown in
equation (8.56):
 
 
|1 − γ̃vo | ≈ 1 − eikv (f0 + a10 f1 + a20 f2 ) (8.56)

Before considering this in more detail, we first include the change of minimum
coherence along the line. This is important, as it impacts on the number of sam-
ples required to estimate coherence and hence on the accuracy and resolution
of any estimation. The µ for minimum coherence was found in equation (7.49).
Inserting this into the line coherence model we obtain the following expression
for the minimum coherence:

γ̃ (w) = eiφ(zo ) (γ̃vo + F(w) (1 − γ̃vo ))

 2 dL
⇒ L = γ̃ (w) = a + bF + cF 2 ⇒ = b + 2cF
dF
b |γ̃vo |2 − Re(γ̃vo )
⇒ Fmin = − = 
2c ∗
(1 − γ̃vo ) 1 − γ̃vo
  (8.57)
 |γ̃vo |2 − Re(γ̃vo ) 
 
⇒ |γ̃min | = γ̃vo +  (1 − γ̃vo )
 ∗
(1 − γ̃vo ) 1 − γ̃vo 
 
 Im(γ̃ ) 
 vo 
= 
 1 − γ̃vo ∗ 

The requirement of keeping this minimum as high as possible is in conflict

with the simultaneous desire to maximize the line length of equation (8.56).
As a compromise we choose to select a kv value that maximizes the product of
equations (8.56) and (8.57); that is, that maximizes the expression in equation
(8.58):
 
 Im(eikv (f + a f + a f ))   
 0 10 1 20 2   
|1 − γ̃vo | . |γ̃min | =   . 1 − eikv (f0 + a10 f1 + a20 f2 )
 1 − e v (f0 − a10 f1 + a20 f2 ) 
−ik

(8.58)

We can now investigate the upper and lower bounds of this function as we
change structure for a given baseline/height product. Note that a10 and a20
have a limited range, as we are considering the volume-only component of the
structure function (and so only structure functions that increase with height).
 8.3 Hidden surface/target imaging 321

Bound on line length vs. changes in structure for kv

2

1.8
maximum value
1.6 minimum value
Bounds on line length

1.4

1.2

0.8

0.6

0.4

0.2

0 Fig. 8.26 Bounds on line length versus kv for

0 0.5 1 1.5 2 2.5 3
all structure parameters in the range a10 ≥
kv 0, a20 ≤ 0

Bound on minimum coherence vs. changes in structure for kv

maximum value
Bounds on minimum coherence

minimum value
0.8

0.6

0.4

0.2

0 Fig. 8.27 Bounds on minimum coherence as

0 0.5 1 1.5 2 2.5 3
a function of kv for all structure parameters in
kv the range a10 ≥ 0, a20 ≤ 0

This limits the Legendre spectrum, so that a10 is non-negative and a20 is non-
positive. In Figures 8.26 and 8.27 we show how the two components of this
function vary with kv . The line length in equation (8.26) starts at zero, and
then increases to a maximum of 2 before falling again for high kv . However, at
the same time we see the minimum coherence value start at 1 for low kv , and
decrease to zero when the line goes through the origin. Indeed, we see that the
minimum can be zero for all structure configurations beyond kv = 2.
Finally, in Figure 8.28 we show the corresponding variation of the product
of these two components, and see a clear optimum range 1 ≤ kv ≤ 1.5. This
range then represents the best compromise between line-length (sensitivity) and
322 Parameter estimation using polarimetric interferometry

Optimum baseline selection vs. changes in structure for kv

0.8

Bounds on cost function

0.6

maximum value
0.4 minimum value

0.2

Fig. 8.28 Bounds on product of line length/ 0

0 0.5 1 1.5 2 2.5 3
minimum coherence versus kv for all structure
parameters in the range a10 ≥ 0, a20 ≤ 0 kv

number of samples required for estimation (resolution). We see that a design

value around kv = 1.25 (centre of the range) represents a good choice. We can
then calculate the optimum baseline to be used by first specifying a target layer
height hdesign . This can then be used with the baseline geometry (see Section
5.1) to calculate the spatial baseline required, B, for wavelength λ, as shown in
equation (8.59):

βz hdesign 2π Bn hdesign 1.25λR sin θ

kv = = = 1.25 ⇒ B =
2 λR sin θ 2π hdesign cos (θ − δd )
(8.59)

Note that the coherence so obtained can be much higher than the polarimetric
only coherence, and so by adopting interferometry we obtain a better situ-
ation (effectively a lower entropy) than the polarimetric approach based on
equation (8.54).

8.4 Structure estimation: extinction and

Legendre parameters
In this section we consder methods for estimating the vertical structure function
f (z) itself; that is, of estimating the vertical variation of scattering through layer
1 using polarimetric interferometry. This information is useful for classification
of different layer types (in forestry and vegetation, for example, where canopy
depth can be an indicator of species or plant stress), and for the estimation of
propagation parameters such as the mean total and differential wave extinction,
from which we can then indirectly obtain information about water content and
density (Ballester-Berman, 2005; Lopez-Sanchez, 2006, 2007; Cloude, 2006b,
2007a).
 8.4 Structure estimation: extinction and Legendre parameters 323

In the RVOG and OVOG models, polarisations other than volume-only are
related by a scale factor: the surface-to-volume scattering ratio, µ. This parame-
ter can be estimated for an arbitrary polarisation using the techniques described
in equation (8.48). The corresponding vertical profile is then obtained as a
weighted sum of an exponential and delta function, as shown in equation (8.60):
2σ̂e
cos θe− cos θ z 
fˆrvog (w, z) = 2σ̂e
+ µ̂ w δ (z) 0 ≤ z ≤ ĥv (8.60)
2σ̂e (e cos θ
ĥv
− 1)

In the Legendre approach, more interesting possibilities arise and lead to a gen-
eralization of the structure estimation problem, termed coherence tomography
(Cloude, 2006b), as we now consider.

8.4.1 Coherence tomography (CT)

Our starting point is to adopt the second-order Legendre expansion of coherence
as shown in equation (8.30). This allows us to include quadratic as well as linear
and constant scattering profiles. Here we have four unknowns on the right: kv ,
φ0 , and the two normalized Legendre coefficients a10 and a20 . On the left we
have only two observables (the complex coherence). We note that only two of
the parameters depend on polarisation. The next stage is therefore to isolate the
polarisation-dependent terms, as shown in equation (8.61), where a00 has unit
value, and where the real functions f0 , f2 and imaginary function f1 as defined
in equation (5.55), are independent of polarisation and depend only on kv .
 
γ̃ (w)e−i(kv +φo ) = γ̃k ≈ a00 f0 + a10 w f1 + a20 w f2 (8.61)

This can then be written in matrix form in terms of the real and imaginary parts
of the phase normalized coherence, as shown in equation (8.62):
     
1 0 0 a00 1
 0 −if1 0  .  a01  =  Im(γ̃k )  ⇒ [L]a = b (8.62)
0 0 f2 a02 Re(γ̃k ) − f0

The next important idea is that we can now invert this relationship to obtain
estimates of the polarisation-dependent Legendre parameters from coherence
and knowledge of the matrix [L], as shown in equation (8.63):

â = [L]−1 b̂ (8.63)

From the vector â we can then estimate the normalized vertical structure
function for a known layer depth hv , as shown in equation (8.64):
/ 
ˆ 1 2z 6z 2
0 ≤ z ≤ hv ⇒ fL2 (z) = 1 − â10 + â20 + (â10 − 3â20 ) + â20 2
hv hv hv
(8.64)

Note that when the quadratic term (a20 ) is zero this reverts to the linear approx-
imation fˆL1 (z), as developed in equation (8.27). Equations (8.62) and (8.63)
constitute a method for reconstructing the function f (z) from coherence, and
324 Parameter estimation using polarimetric interferometry

are therefore termed coherence tomography, or CT. We shall see that the matrix
formulation can be extended to arbitrary order of Legendre polynomial, and
hence to higher and higher resolution reconstructions by adding multiple base-
lines to the interferometer. However, there remains the important issue of how
to obtain estimates of the polarisation-independent terms kv and φ0 . We now
turn to consider this in more detail.
In equation (8.61) we separated coherence into polarisation-dependent and
independent components. The latter set comprises three parameters of interest:
the layer depth hv and interferometric wavenumber βz (which are then used
to calculate kv ), and the phase of the bottom of the layer, φ0 . The wavenum-
ber can be estimated from knowledge of the baseline geometry and operating
wavelength of the interferometer (see Section 5.1), but the other two parame-
ters require special attention. There are two principle ways of obtaining these
parameters.

8.4.1.1 CT using external data

In the first approach we can use separate external measurement of the layer
depth hv and surface phase (the latter by measuring the z coordinate z0 of the
bottom of the layer above the zero datum of the interferometer, and using βz
to obtain φ0 = βz z0 ). These can be obtained, for example, for laboratory-
based experiments (Cloude, 2007a) by direct measurement, and then directly
used in equation (8.60) to investigate the variation of structure function with
polarisation. In field experiments such estimation can be more difficult, but
can still be accomplished with the aid of global positioning technology such as
GPS, or depth profiling technologies such as laser sounding using LIDAR or
high-resolution microwave altimeters or scatterometers.
In this case we can estimate the structure function for arbitrary polarisations
w by first forming the interferogram, estimating complex coherence, phase
shifting the coherence using βz and φ0 , and then calculating the profile estimate
as summarized in equation (8.65):

βz 
β z hv
hv → kv = , γ̃k (w) = γ̃ (w)e−i(kv +φ0 )
 2
φ0
   
sin kv sin kv cos kv 3 cos kv 6 − 3kv2 1
fo = f1 = i − f2 = − + sin kv
kv kv2 kv kv2 2kv3 2kv
     
1 0 0 a00 1
0 −if1 0  . a01 (w) =  Im(γ̃k (w)) 
0 0 f2 a02 (w) Re(γ̃k (w)) − f0

⇒ â(w) = [L]−1 b̂
/ 
ˆ 1 2z 6z 2
⇒ fL2 (w, z) = 1 − â10 (+â20 (w) + (â10 (w) − 3â20 (w)) + â20 (w) 2
hv hv hv
(8.65)

This approach makes no assumptions about the shape of the coherence region,
and so can be used to investigate the most general profiles. We will consider
an example of such an approach based on laboratory anechoic chamber mea-
surements of maize plants in Chapter 9. Often, however, especially in remote
 8.4 Structure estimation: extinction and Legendre parameters 325

sensing applications, we have no access to the layer depth or supporting mea-

surements, and must therefore develop alternative techniques for estimating
the parameters directly from the data itself. There are two approaches to be
considered: dual baseline inversion, when we add new baselines to increase
the number of observables, and single baseline bootstrap techniques, where we
use the height and surface phase estimators of Sections 8.1 and 8.2 to enable
tomography with a single baseline. We now turn to consider such methods in
more detail.

8.4.1.2 Dual baseline inversion

The challenge is now to develop parameter estimation algorithms for the surface
phase φ0 and layer depth hv that involve minimal assumptions about the shape
of the structure function f (z). In this way we can use these estimates in the
second-order Legendre algorithm directly, and maintain its flexibility to deal
with general scattering scenarios.
The algorithms presented in Sections 8.1 and 8.2 for φ0 and hv estimation
can be proposed, but all of them made some further restrictive assumptions
about f (z) in the volume-only channel—that it is exponential for the RVOG
and OVOG models, or that the volume-only scattering coefficients satisfy
a20 ≤ 0 and a20 ≥ 0 for the Legendre approach. These assumptions strictly
only have to be valid for the volume-only polarisation channel, but because of
the SVOG assumption used to estimate topography they impact on the assumed
volume component of the response in arbitrary polarisation channels. Hence
such assumptions force the reconstructions to conform to a subset of structure
functions satisfying the requirements of the models. Only if these assump-
tions are a good match to the physical structure of the problem will they yield
good results. It is therefore of interest to see if we can avoid such restrictive
assumptions at all.
Here we consider one important way to achieve this, by using a dual-baseline
interferometer, with a second baseline, different from the first, used to obtain
four observables (the amplitude and phase of two coherences) with four model
unknowns (hv , φ0 , a10 , and a20 ). We start with the simpler case when φ0 is
known for both baselines, and so we can set φ0 = 0 without loss of general-
ity. We shall consider the full four-dimensional case later. This then reduces
the problem to three unknowns (hv , a10 , and a20 ). Even so, as we have seen,
problems arise with CT if we do not know the kv value in advance. In this case
we obtain multiple solutions for a whole range of kv a10 , a20 coordinates. As
shown in Figure 8.29, a single coherence point (in grey) can fit the model over

90 1
120 60
0.8

0.6
150 30
0.4

0.2 Fig. 8.29 Superimposed second-order Leg-

endre approximation for two different kv
values, showing structural ambiguity for a
180 0 single-baseline coherence (in grey)
326 Parameter estimation using polarimetric interferometry

a20 a20

a10 a10

Fig. 8.30 Schematic representation of well-

conditioned (left) and ill-conditioned (right)
solutions for dual-baseline inversion

a wide range of kv values. Here we show the set of coordinate lines for two kv
values, with the two origins (a10 = a20 = 0) shown as black points, and see
that the sample coherence point, although it has different a10 a20 values, can be
made to fit either. Hence a single baseline cannot be used to estimate the two
Legendre structure parameters uniquely. We can, however, estimate a family
of solutions using single baseline data. As we move around the SINC spiral in
Figure 8.29 we obtain a set of solution pairs a10 , a20 for the given coherence
(in grey). We can represent this family geometrically as a set of solution points
in the a10 , a20 plane, generated as kv varies from 0 to π . These generate a curve
in the plane, as shown schematically in Figure 8.30. We can then use this idea
to propose a method for estimating the true kv value by combining data from a
second baseline.
If we now consider that coherence data is available for a second additional
baseline, related to the first by a baseline ratio Br , then we can generate a second
set of a10 , a20 points by simultaneously solving the matrix equation for kv for
the first baseline and Br kv for the second. The correct kv value then occurs
when these two curves intersect, as shown schematically in Figure 8.30. Here
we show two possible scenarios. On the left, the solid curve is the curve of
solutions obtained for the first baseline, and the dashed curve is the solutions
for the second. This is a well-conditioned case, when the intersection point
occurs for nearly orthogonal curves. Such a scenario will be robust to errors in
the two coherence estimates. On the right of the figure is shown the opposite
case of a poorly conditioned solution where the intersection point occurs for
nearly parallel curves. For example, in the limiting case, if we take the two
baselines equal (Br = 1) then the loci will exactly overlap and a solution is
not possible. When nearly overlapping, any small perturbation of the curves
will lead to a large change in the solution. This ill-conditioning can undermine
the uniqueness of a solution using dual baselines by making the algorithm so
sensitive to noise that it cannot be used in practical applications (Hopcraft,
1992). The level of such ill-conditioning will be a function of the baseline
ratio Br . We shall examine the conditioning of coherence tomography and its
dependence on Br in more detail in Section 8.4.3. First, however, we consider
a formalization of this approach and how to generalize it for unknown surface
topography φ0 .
We can formally write the solution to the dual baseline kv estimate as mini-
mization of the coherence error as defined in equation (8.66), where subscripts
 8.4 Structure estimation: extinction and Legendre parameters 327

1 and 2 refer to the baselines used:

γ̃1 = eiφ1 eikv1 (f0 (kv1 ) + a10 f1 (kv1 ) + a20 f2 (kv1 )) ⇒ γ̃k1 = γ̃1 e−iφ1 e−ikv1
   −1  
â00 1 0 0 1
â01  = 0 −if1 (kv1 ) 0  . Im(γ̃k1 ) 
â20 0 0 f2 (kv1 ) Re(γ̃k1 ) − f0 (kv1 )
iBr φ1 ikv2

⇒ γ̃2 = e
est
e f0 (kv2 ) + â10 f1 (kv2 ) + â20 f2 (kv2 ) where kv2 = Br kv1
? ?
⇒ Coherence error = ?γ̃2 − γ̃2est ? (8.66)

The procedure (for φ1 = 0) is then to vary kv1 from 0 to π , calculating for each
value the Legendre spectrum a10 , a20 . We then use these values to estimate the
second baseline coherence and select the triplet kv1 , a10 , a20 that minimizes the
difference between this complex estimate and the true second baseline coher-
ence γ̃2 , as shown in equation (8.66). When topographic phase is also unknown,
the only difference we face is to search for the intersection of a10 , a20 loci in
a two-dimensional space of φ0 and kv1 rather than just kv1 . The estimate of
the second baseline coherence is then phase shifted by a scaled topography, as
shown in equation (8.66). (Note that we are also assuming that there are no
residual phase errors between baselines, as can occur, for example, in repeat
pass sensors. If not true, then we must add an extra unknown phase parameter
to equation (8.66).)
As a typical example of dual baseline performance, consider the choice (φ1 =
φ2 = 0) Br = 0.5 and kv = π/2, with a profile defined by a10 = 0.5 and a20 =
1. These lead to very distinct coherences of 0.87 for the smaller baseline and 0.54
for the larger. These two points can then be used to estimate a pair of solution
curves for a10 , a20 , as shown in Figure 8.31. Here we see a scenario that seems
poorly conditioned, despite the fact that the coherences from the two baselines
are very different, with the intersection point of the two curves occurring for

Legendre solution loci for two baselines

3

2.5

1.5
a 20

0.5

–0.5

–1 Fig. 8.31 Example solution loci for dual-

–1.5 –1 –0.5 0 0.5 1 1.5
baseline inversion, showing ill-conditioned
a 10 nature of solution
328 Parameter estimation using polarimetric interferometry

Kv estimate from dual baseline data

0

–5

–10

Coherence error (dB)

–15

–20

–25

–30

–35

Fig. 8.32 Variation of coherence error

for dual-baseline inversion example of –40
0 0.5 1 1.5 2 2.5 3
Figure 8.31, showing a minimum at correct
value (π /2) kv for baseline 1

Coherence error for unkonwn surface topography

0
350

300 –5
Topography phase (degrees)

250
–10

200
–15
150
–20
100

–25
50

Fig. 8.33 Coherence error for two-

dimensional search in baseline/height 0 –30
0.5 1 1.5 2 2.5 3
product kv1 and surface phase f1 (for true
values of kv = π/2 and f1 = π ) Kv for baseline 1

nearly parallel sections. In Figure 8.32 we show the corresponding coherence

error of equation (8.66), which correctly shows a unique minimum for kv =
π/2. Turning now to the general case when topography is also unknown, we
consider a situation where φ1 = π , φ2 = π/2. Figure 8.33 shows the two-
dimensional variation of coherence error. Again, formally, the correct solution
is located with a minimum at kv = π/2, φ1 = π , but we see a long ‘valley’ of
potential local minima stretching across the solution space.
The level of ill-conditioning is especially important when we consider that the
longer baseline coherence is quite low (around 0.54), and so will require a large
 8.4 Structure estimation: extinction and Legendre parameters 329

number of samples to minimize residual estimation noise. Such noise could lead
to large errors in the estimate of kv , and we need to balance the amplification
of error caused by the ill-conditioning against the number of samples required
in order to correctly assess these errors. This example nicely illustrates how
uniqueness is not the only criterion required for an assessment of algorithm
performance, and that the level of numerical stability or ill-conditioning must
also be quantified. We will consider such an analysis in Section 8.4.3.
Equation (8.66) represents one method of estimating kv , but it requires data
for two baselines. But this is often not available, and so it is of interest to consider
alternative single baseline strategies for estimation of kv that still enable use of
the second-order Legendre approach to CT.

8.4.2 Bootstrap polarisation coherence

tomography (PCT)
In this approach we try to use single baseline data itself to approximate the
two parameters φ0 and kv . The easiest case to deal with is the estimation of
topography φ0 . Here, by assuming only validity of the SVOG model—that
layer 1 shows azimuthal scattering symmetry and is random so that polarisation
dependence of coherence comes only from the variations of surface-to-volume
scattering ratio—we again obtain a linear coherence loci, and can therefore use
any of the line fit techniques developed in Section 8.1 to estimate topographic
phase. Note that this symmetry makes no assumptions about the shape of the
volume-only structure function, and assumes only that it is invariant in shape
(but not necessarily in amplitude) to polarisation.
For kv estimation using a single baseline configuration we must employ
an algorithm for layer depth estimation that is robust to changes in structure
and hence robust to changes in the a10 , a20 coefficients. We have already seen
an example of such an algorithm in equation (8.38), where we used separate
phase and coherence estimates to balance the errors across a range of structure
parameters.
With this established, we can provide a direct estimate of kv by first identify-
ing a volume-dominated polarisation channel and then calculating kv directly,
as shown in equation (8.67):

1@  0.8 A
kv = arg(γ̃wv e−iφ0 ) + 0.8(π − 2 sin−1 (γ̃wv  ) (8.67)
2

The estimated φ0 , kv values then establish a unique coordinate system for the
whole coherence diagram, so that the structure function for all other polari-
sations can be reconstructed up to second order using the matrix inversion of
coherence tomography.
This approximation works with only a single baseline, but requires the com-
bination of at least two interferograms formed for different polarisations: one
volume and the other surface dominated. It therefore requires some polarisation
diversity in measurements, and hence we term it polarisation coherence tomog-
raphy, or PCT, to distinguish it from standard CT. The basic steps involved in
PCT are summarized in Figure 8.34. Here we show how, starting from a pair
of polarisations, ws and wv , we can use the SVOG assumption to fit a line
330 Parameter estimation using polarimetric interferometry

Stage 1: Height and phase estimation

f = arg(g~wv – g~ws (1 – Lws)) 0 <

– Lws <– 1
2 –B– B2 –4 AC
ALws + BLws + C = 0 ⇒ Lws =
2A
A = g–ws2 –1 B = 2Re((g–wv – g–ws)·g–ws) C = g–wv – g–ws2

^ ^
– –
kv ≈ 12 arg(gwv e–if) + 0.8(p – 2sin–1(gwv0.8)

Stage 2: Coherence normalization

Select arbitrary polarisation w
g~k(w) = g(w)e
~ –ikv e –if

Stage 3: Legendre estimate

–1
a00 1 0 0 1
a10 = 0 f1i 0 Im(g–k)
a20 0 0 f2 Re (g– k) – f0
sin kv
fo =
kv
sin kv cos kv
f1i = ( – )
k2v kv
1 sin kv
f2 = (3 cos kv – (3 – kv2 ) )
kv2 kv

Stage 4: Profile estimate

2kv
hv = 0 < z < hv
kz
1 2z 6z2
Fig. 8.34 Summary of algorithm for single- fL2 (w, z) = {1 – a10 (+a10 (w) + (a10 (w) – 3a20 (w)) +a20 (w) 2 }
hv hv hv
baseline polarization coherence tomography
(PCT)

and find the topographic phase. (The two-state line fit is shown in Figure 8.34,
although the total least squares or other approach could equally be applied if
more polarisation channels were to be made available.) We must then select
one of the polarisations as a volume-dominated channel to obtain an estimate
of kv . We then use the kv and φ0 estimates to scale the coherence for arbitrary
polarisation w to generate a second order estimate of the structure function, as
shown. This can then be repeated for arbitrary polarisations as desired.
The procedure has a simple geometric interpretation as shown in Figure 8.35.
The process begins by identifying the volume coherence (grey point) and cor-
recting for surface topography (white point) using the SVOG linear coherence
assumption (black line). From the grey volume coherence we then estimate kv ,
which locates the origin for the a10 , a20 coordinate system (the black point in
Figure 8.35). This origin then defines a grid of a10 a20 values that can be used
to estimate the structure function for arbitrary location in the complex unit cir-
cle. However, under the SVOG assumption, coherence varies with polarisation
 8.4 Structure estimation: extinction and Legendre parameters 331

90 1
120 60
0.8

0.6
150 30
0.4

0.2 a20 > 0

180 0

210 330

240 300 Fig. 8.35 Superimposed SVOG coherence

region on coherence loci for second-order
270 Legendre approximation

along the black line and so we can access only a limited portion of the a10 /a20
space. Note, for example, that while a10 can be positive or negative, the line
permits only positive a20 values in the reconstruction. This is the price to pay
for using such a bootstrap approach.
We now turn to consider more quantitatively the ill-conditioning of the matrix
inversion embedded in both CT and PCT algorithms.

8.4.3 Condition number and error analysis

We have seen that estimation of the structure function using CT and PCT
involves the following key matrix inversion step:
     
1 0 0 a00 1
0 −if1 0  . a01  =  Im(γ̃k )  ⇒ [L]a = b ⇒ â = [L]−1 b
0 0 f2 a02 Re(γ̃k ) − f0
(8.68)

where the functions f0 , f1 and f2 are given in Figure 8.34. Some care must
be taken with this inversion, as it can lead to an amplification of any errors
in b so that the resulting a may represent a very poor estimate of structure.
An alternative way of thinking about this is to estimate the level of noise we
can tolerate in the coherence vector b so as to keep the fractional error in
a below a prescribed value. This will then allow us to estimate the number of
coherence samples required for good estimation. In this section we develop such
an algorithm for analysing the stability of matrix inversions like equation (8.68).
Key to quantifying this amplification process is the condition number (CN)
of the matrix [L] (see Appendix 1). The larger CN, the larger any amplification
of errors in b. As [L] is diagonal in equation (8.68), we can obtain an explicit
expression for the condition number of the matrix [F] as a ratio of the functions
332 Parameter estimation using polarimetric interferometry

Single baseline condition number

105

104

103

CN
102

101

Fig. 8.36 Variation of matrix condition num- 100

0 0.5 1 1.5 2 2.5 3
ber versus kv for single-baseline coherence
tomography kv

fi , as shown in equation (8.69):

1 kv2
CN = − =− (8.69)
f2 3 cos kv − (3 − kv2 ) sin kv kv

Figure 8.36 plots this function versus normalized wavenumber kv . Note that for
small baseline/height products the inversion is very poorly conditioned (a large
CN). For baseline/height products around unity, the condition number is around
10–20. Since [F] is diagonal we can also identify the worst-case scenario, when
the system becomes most sensitive to errors in b. From equation (8.68) this
arises for perturbations of a true solution of the form
   
1 1
b = 0 ⇒ b + δb = 0 (8.70)
0 δ

This physically corresponds to small radial coherence amplitude perturbations

about uniform zero-extinction volume scattering. In this worst case, the error
in the Legendre coefficient vector is amplified by the matrix inversion to the
order of CN.β. The coefficient β can now be related to the coherence and effec-
tive number of looks L by using the Cramer–Rao bound (see Appendix 3) and
considering the limiting case of zero-extinction volume scattering. Considering
the worst case from equation (8.70), it follows that the largest error contribu-
tion is from the real part of the phase corrected coherence γ̃k . For a uniform
zero-extinction volume the real part error is then dominated by the Cramer–
Rao variance on coherence rather than phase estimation. Taking the standard
deviation as a measure of the coherence error we can then write:
)   *
(1 − γv2 ) 1 sin kv 2
β≈ √ ≈√ 1− (8.71)
2L 2L kv
 8.4 Structure estimation: extinction and Legendre parameters 333

Fractional error in Legendre spectrum vs. number of looks

1

0.9 L = 25
L = 50
0.8 L = 100
Maximum fractional error

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 Fig. 8.37 Maximum bound on fractional
error in Legendre estimates versus kv for
kv different number of looks

where L is the number of looks. In this way we can estimate an upper bound on
the fractional error in the estimate of the Legendre coefficients as a function of
just two parameters, kv and the number of looks L, as shown in equation (8.72):

)? ?*
?δa? sin2 kv − kv2
max ? ? = CN .β = √  (8.72)
?a?
2L 3 cos kv − (3 − kv2 ) sinkvkv

Figure 8.37 shows how this bound varies as a function of kv for various number
of looks L. We should note that this represents a worst-case scenario, and
generally the errors will be better than this. This approach assumes that the
layer is almost a uniform volume scatterer (and so the volume coherence lies
along the bounding SINC curve in Figure 8.35). If this is not true, and the
volume channel has some other structure, then the errors will be less than this
bound.
This conditioning error is due to amplified noise in coherence estimation.
However, we have two other main sources of system noise to consider: the
effects of SNR, and temporal decorrelation in the interferometer. These can
now be incorporated into the CT and PCT formulations as follows.

8.4.4 SNR and temporal decorrelation in CT

Signal-to-noise ratio and temporal decorrelation effects can be included in
the CT formalism by noting that they act as scalar multiplying factors of the
observed coherence (see Section 5.2.5). Hence they do not distort the mean
phase of the complex coherence but reduce the coherence amplitude (increase
phase variance). This then scales the real and imaginary parts of the b vector
as shown in equation (8.73). Note that they do not influence f0 , which has now
334 Parameter estimation using polarimetric interferometry

90 1
120 60
0.8

0.6
150 30
0.4

0.2

180 0

210 330

Fig. 8.38 Effect of temporal decorrelation on 240 300

the RVOG/second-order Legendre combina-
tion 270

been separated in the formulation.

   −1  −1    −1  
a00 1 0 0 1 0 0 1 1 0 0 0
           
a =
 10   0 −if1 0   0 γ γ
snr t 0   Im( γ̃ )
k  −  0 −if 1 0  0
a20 0 0 f2 0 0 γsnr γt Re(γ̃k ) 0 0 f2 f0
(8.73)
These have the geometrical effect of shifting the coherence point along a radial
line towards the origin, as shown schematically for the volume coherence (grey
point) in Figure 8.38. In CT, the effect will be to amplify the quadratic compo-
nent of the structure function (with an increased positive a20 coordinate) without
influencing the [L] matrix elements. For bootstrap PCT, however, we use the
volume coherence itself to estimate kv , and so temporal/SNR decorrelation will
impact on the [L] matrix as well as the b vector. We see from Figure 8.38 that
a radial shift will initially cause an overestimation of kv . This will continue
until the radial line intersects the SINC locus (shown as the light grey point in
Figure 8.38).
For larger SNR/temporal decorrelation the volume coherence moves below
the SINC boundary, and the Legendre approximation no longer provides a
solution for kv . We see from the curvature of the SINC locus that such SNR
and temporal effects (which are independent of baseline) will be more serious
for small kv , where the locus approaches the unit circle. Therefore, one way to
provide increased robustness to such effects is to work at larger spatial baselines.
The best way to minimize such effects, however, is to avoid temporal effects by
employing a single-pass interferometer and to ensure high SNR in the selected
polarisation channels.

8.4.5 Multiple baseline CT

In the previous sections we saw how we can employ first- and second-order Leg-
endre approximations of the structure function to model complex coherence.
 8.4 Structure estimation: extinction and Legendre parameters 335

These relationships, under certain assumptions, can be inverted to provide esti-

mates of the structure function from coherence measurements in a technique
called coherence tomography (CT). However, such reconstructions are lim-
ited, as we have seen, to second-order polynomial variation. It is natural to
ask if we can further improve the resolution of the reconstruction by estimat-
ing higher-order terms of the Fourier–Legendre expansion. In this section we
consider such an extension, and show how, with knowledge of layer depth and
surface position, we can employ multiple baseline interferometry to reconstruct
the structure function to higher and higher resolutions—albeit at the price of
increasing condition number with increasing resolution (Cloude, 2007a).
In single-baseline CT we have two observables (one complex coherence)
and two unknowns (a10 and a20 ), assuming we have knowledge of layer depth
and surface position. Hence the addition of a second baseline adds two new
observables and allows us to further extend the Legendre series by a further
two orders to fourth order, as shown in equation (8.74). The new functions f3
and f4 are given in equation (8.75), where we note that f3 is pure imaginary and
f4 is real:

γ̃ e−ikv e−iφo = γ̃k = f0 + a10 f1 + a20 f2 + a30 f3 + a40 f4 (8.74)

sin kv
fo =
kv
 
sin kv cos kv
f1 = i −
kv2 kv
 
3 cos kv 6 − 3kv2 1
f2 = − + sin kv
kv2 2kv3 2kv
    
30 − 5kv2 3 30 − 15kv2 3
f3 = i + cos kv − + sin kv
2kv3 2kv 2kv4 2kv2
   
35(kv2 − 6) 15 35(kv4 − 12kv2 + 24) 30(2 − kv2 ) 3
f4 = − cos kv + + + sin kv
2kv4 2kv2 8kv5 8kv3 8kv
 
−2kv4 + 210kv2 − 1890 30kv4 − 840kv2 + 1890
f5 = i cos kv + sin kv
kv5 kv6
   
42kv4 − 2520kv2 + 20790) 2kv6 − 420kv4 + 9450kv2 − 20790)
f6 = cos kv + sin kv
kv6 kv7
(8.75)

We now see that there is a natural extension of this idea to multiple baselines,
adding two new structure parameters per baseline, so that in general N baselines
yields 2N + 1 terms of the Fourier–Legendre series. Returning to the N = 2
case, CT inversion then takes the form of a matrix equation based on the use of
equation (8.74) for two baselines ‘x’ and ‘y’, as shown in equation (8.76):
     
1 0 0 0 0 a00 1 
0 −if x 0 −if x 0  a10   Im γ̃ x 
 1 3      k x
0 f4  . a20  = Re γ̃k − f0 
    −1
 ⇒ â = [L] b
0 f2x 0 x x

0 −if y
0 −if3 0  a30   Im γ̃k
y y 
1
y y y y
0 0 f2 0 f4 a40 Re γ̃k − f0
(8.76)
336 Parameter estimation using polarimetric interferometry

Note that the real matrix [F] is now 5 × 5 (for N baselines it is (2N + 1) ×
(2N + 1)), and is no longer diagonal in structure. From the estimated vector of
Legendre coefficients we can determine the shape of the corresponding structure
function up to fourth order, as shown in equation (8.77):

fˆ (z ) = 1 + â10 P1 (z ) + â20 P2 (z ) + â30 P4 (z ) + â40 P4 (z ) −1≤z ≤1

(8.77)

Extending this to N = 3 (to three baselines) leads to the following model for
coherence, where again the new functions f5 and f6 are given in equation (8.75).

γ̃ e−ikv e−iφo = γ̃k

= f0 + a10 f1 + a20 f2 + a30 f3 + a40 f4 + a50 f5 + a60 f6 (8.78)

When applied across the three baselines ‘x’, ‘y’ and ‘z’, this leads to a cor-
responding 7 × 7 matrix inversion for coherence tomography, as shown in
equation (8.79). Note again that [F] is a real non-diagonal matrix with elements
a function of kv .
    
1 0 0 0 0 0 0 a00 1 
0 −if x 0    
−if3x −if5x  a10   Im x γ̃k x 
x
 1 0 0
0 0 f2x 0 f4x 0 f6  a20  Re γ̃k − f0 
x    
 
0 −if y 0 −if3
y
0 −if5
y
0 a30  =  Im γ̃ y  (8.79)
 1
y     k 
0 0 f2
y
0 f4
y
0 f6  a40  Re γ̃k − f0 
    x y
 
0 −if z 0 −if3z 0 −if5z 0  a50   Im γ̃kz 
1
0 0 f2x 0 f4x 0 f6x a60 Re γ̃k − f0
z z

This then permits an even higher-resolution reconstruction of the structure

function, as shown in equation (8.80):

fˆ (z ) = 1 + â10 P1 (z ) + â20 P2 (z ) + â30 P4 (z ) + â40 P4 (z )

+ â50 P5 (z ) + â60 P6 (z ) (8.80)

Figure 8.39 summarizes the differences between single, dual and triple base-
line reconstructions by plotting the polynomials employed in the corresponding
reconstructions. We can clearly see the improvement in resolution with increas-
ing baseline. However, while formulation of CT in this way is straightforward,
note from Figure 5.15 that the functions fi tend to zero with increasing order
and hence anticipate problems with the conditioning of the inversion. This will
provide a practical limit to the achievable resolution, as eventually we will
demand impossible limits on the control of error in coherence estimation for
the vector b. In order to assess this, we now turn to quantify the condition
number of multi-baseline CT.
In general we can analyse the conditioning of multi-baseline CT using a
singular value decomposition of the matrix [L]. This allows us to represent
the inversion in terms of a (2N + 1) × (2N + 1) diagonal matrix [], just as
we did for the single-baseline case in equation (8.68), but now with different
 8.4 Structure estimation: extinction and Legendre parameters 337

Single baseline Dual baseline Triple baseline

1 1 1

0.9 0.9 0.9

0.8 0.8 0.8

0.7 0.7 0.7

0.6 0.6 0.6

Height

Height

Height
0.5 0.5 0.5

0.4 0.4 0.4

0.3 0.3 0.3

0.2 0.2 0.2

0.1 0.1 0.1

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1 Fig. 8.39 Legendre functions for single,
Relative density Relative density Relative density dual, and triple baseline inversions

orthogonal frames [U ] and [V ] for the vectors a and b. Formally we can write
the matrix [L] in the form shown in equation (8.81) (see Appendix 1):

[L] = [U ] · [] · [V ]∗T s1 ≥ s2 ≥ · · · ≥ s2N +1 (8.81)

where the (2N + 1) real parameters si are the singular values. The formal solu-
tion to the CT estimation problem can be written in terms of these matrix
components, as shown in equation (8.82):

â = [V ].[]−1 [U ]∗T b,
 −1 
s1 0 0 .. 0
 0 s−1 0 .. 0 
 2  (8.82)
 0 s3−1 
[]−1 =  0 .. 0 
 .. 
 : : : . : 
−1
0 0 0 0 s2N +1

The condition number of the inversion is then defined as the ratio of maximum
to minimum singular values, as shown in equation (8.83):
s1
CN = (8.83)
s2N +1

As an example, we show in Figure 8.40 the variation of CN on a dB scale for

baseline pairs kv1 .kv2 in dual-baseline CT. Note that along the diagonal the CN
goes to infinity, since the rows of [F] no longer provide independent information
about the structure. We see that in useful portions of the kv space (around 1)
the CN is very high, of the order of 1000 or more.
338 Parameter estimation using polarimetric interferometry

CN(dB) of dual baseline (F) matrix

50
3
45

2.5 40

35
2
30

Kvy
25
1.5
20

1 15

10
0.5
5

Fig. 8.40 Variation of condition number for 0

0.5 1 1.5 2 2.5 3
dual-baseline inversions with baselines kvx
and kvy Kvx

CN(dB) of dual baseline (F) matrix

50
3
45

2.5 40

35
2
30
Kvy

25
1.5
20

1 15

10
0.5
5

Fig. 8.41 Variation of condition number for 0

0.5 1 1.5 2 2.5 3
singular value filtered dual-baseline inver-
sions with baselines kvx and kvy Kvx

One way to deal with this high condition number is to filter the [L] matrix,
which can be achieved by removing the smallest singular value to reconstruct
a profile with a matrix [f ], as shown in equation (8.84). Here we obtain
a matrix with a lower condition number, given by s1 /s2N . Note that in this
case we lose some resolution (given by one pair of singular vectors), but still
gain resolution over the reduced baseline case. Figure 8.41 shows the condition
number of this filtered matrix. Here we see an order of magnitude improvement
 8.4 Structure estimation: extinction and Legendre parameters 339

in conditioning, with condition numbers around 100 for the useful part of the kv
range.

âf = [V ].[f ]−1 [U ]∗T b,

 −1 
s1 0 .. .. 0
 0 s−1 .. .. 0
 2  (8.84)
 
[f ]−1 =  : :
..
. : :
 
 0 0 0 s−1 0
2N
0 0 0 0 0

We will provide an example of dual and triple baseline CT processing of

anechoic chamber data in Chapter 9.
 Applications of
9 polarimetry and
interferometry

In this chapter we turn to consider some applications of polarimetry and polari-

metric interferometry in remote sensing. A comprehensive survey would be
impossible, and so instead we select a few representative examples taken from
different areas. We do this firstly to reinforce the theoretical ideas introduced in
earlier chapters, but also to present an idea of the wide range of topics in which
these concepts can be applied. We start with a general introduction to synthetic
aperture radar (SAR) (Curlander, 1991; Mensa, 1991), as it is with radar imag-
ing that most applications currently occur. In particular we outline a hierarchy
of polarimetric modes in radar imaging, starting with single-channel SAR and
then interferometric SAR, or InSAR (Bamler, 1998), before developing into
both compact and quad polarimetric, or POLSAR (Lee, 2008; Kong, 1990;
Mott, 2007; Ulaby, 1990), and finally to imaging polarimetric interferometry.
or POLInSAR (Cloude, 1998, 2001b; Krieger, 2005).
We then turn to consider several application themes, starting with bare surface
scattering and then considering the effects of vegetation cover, first through agri-
culture or short vegetation and then considering the important case of forestry.
We finally turn to consider applications centred around the study of isolated
point scatterers, such as occur in urban areas and in ship detection and moni-
toring. In this way we cover a broad range of topics that illustrate many of the
concepts introduced in earlier chapters.

9.1 Radar imaging

We begin by considering the basic principles of radar imaging. More details
can be found in the specialist monographs by Curlander (1991), Mensa (1991),
and Franceschetti (1999). Consider a static transmitter/receiver configuration as
shown schematically in Figure 9.1. When we employ a transmitter and receiver
separated by bistatic angle  and operating at a single wavelength λ, then
scattering by the environment around the transmitter leads to a total received
signal in amplitude and phase, represented by a complex number. This complex
number in fact represents the amplitude of a Fourier component located at point
P in a wavenumber space, as shown on the right-hand side of Figure 9.1. The
polar coordinates of the point P in this space are then defined by the geometry
of the transmitter and receiver configuration and the propagation phase delay
between the two, defined as exp iβ(r1 + r2 )), leading to the triangular construc-
tion shown in Figure 9.1. Clearly such a single static configuration does not
lead to an image of the environment. The signal obtained in the receiver is the
coherent summation from many points in the scene, depending on many factors
 9.1 Radar imaging 341

2
=

Transmitter Receiver =
p 2
z r=2 cos
r 2
FT

e 2 cos
2

Fig. 9.1 Wave-space geometry for a single

transmitter/receiver combination

A O x-axis A

Radar flight trajectory
z-axis


z = zo

T Fig. 9.2 Synthetic aperture geometry

including the beamwidth of the transmitter and receiver antennas. To obtain an

image requires diversity over one or more of the three parameters λ, θ and  in
order to fill a sector of wave space. When such a sector has been filled, then an
inverse two-dimensional Fourier transform can be used to reconstruct an image
of the environment, as we now formally demonstrate.
By far the most common radar configuration is to employ a monostatic sen-
sor ( = 0) working in backscatter, with a finite bandwidth W representing
wavelength diversity, and then linear motion of the radar system to generate
θ diversity. The latter generates a finite line segment along the radar flight
trajectory, as shown by AA’ in Figure 9.2. This segment can be considered
a synthetic antenna aperture (which is much larger than the actual physical
antenna aperture)—hence the term synthetic aperture radar, or SAR.
A two-dimensional image of the environment in the z–x plane can then be
obtained using an ω–β SAR processor as follows. (Note that in many texts this
is called an ω − k processor, where k is the symbol for wavenumber. However,
here we use k for scattering vector, and so to avoid confusion we refer to the
ω − β processor.) In Figure 9.2, T is a general reflecting point in the scene,
O is the (monostatic) radar observation point and AA’ the linear ‘aperture’ of
the radar flight path. If we denote the wave field caused by an apparent source
at T as d (x, z, t) then we know that d everywhere obeys the following wave
equation (Gazdag, 1984; Cafforio, 1991):

∂ 2d ∂ 2d 1 ∂ 2d
+ = (9.1)
∂x2 ∂z 2 v2 ∂t 2
342 Applications of polarimetry and interferometry

If we now take the Fourier transform of this equation with respect to time and
to x, we obtain the following (ordinary) differential equation for the transform
quantity D(βx , ω, z):

d 2D ω2 d 2D ω2 β 2 v2
βx2 D − = 2D⇒ = − 2 (1 − x 2 )D (9.2)
dz 2 v dz 2 v ω
This ‘ODE’ can then be factored in terms of upward and downward (±z) prop-
agating waves. The latter we can then use to ‘migrate’ the field from the line
AA back to the source line at z = zo . This will render the wave field sensed
along AA as an ‘image’ of the apparent ‘sources’ along z = zo ; that is, it will
focus the radar image. We shall see that the larger the aperture AA , the higher
the resolution in this image. For the downward propagating waves we have the
following factorization:
"
 
dD ω βx v 2 ω
= f .D = i 1− D=i 1 − η2 D (9.3)
dz v ω v

This equation has a simple plane wave solution. Therefore, if we know D across
a surface AA then we can propagate or ‘migrate’ the data to any other z value,
as shown in equation (9.4):
ω
√
1−η2 z
D(βx , ω, z + z) = D(βx , ω, z)ei v (9.4)

Finally, we can obtain the image of the sources by an inverse Fourier transform
(FT) with respect to βx and a summation w.r.t. ω as follows:
!!
d (x, t = 0, z) = D(βx , ω, z)eiβx x (9.5)
kx ω

If the velocity is constant (v = c/2 to account for the two-way propagation) then
this summation can be performed very efficiently by using a two-dimensional
Fourier transform as follows. We first propagate the data from AA to z = zo in
one single step by phase rotation, as shown in equation (9.6):
ω
D(βx , ω, zo ) = D(βx , ω, z = 0)eiβz zo βz = 1 − η2 (9.6)
v
We then perform the summation and inverse transform to obtain the following
integral:

d (x, t = 0, zo ) = D(βx , ω, z = 0).ei(βz (ω)z+βx x) d βx d ω (9.7)

This is almost a two-dimensional FT operation, and we can complete the process

by a change of variable ω = 2c βx2 + βz2 and integration with respect to βz
instead of ω to obtain the following Fourier transform relationship between the
measured spectral function D and original source distribution:

cβz
d (x, t = 0, zo ) = D(βx , ω, z = 0).ei(βz z+βx x) d βx d βz (9.8)
βx2 + βz2
 9.1 Radar imaging 343

This approach of wave migration to generate a SAR image is called the

wavenumber processor. It is only one of several approaches to SAR processing
(Bamler, 1992, 1998; Curlander, 1991), but for our purposes provides a direct
link to the wave equation and propagating polarised EM waves. As shown
above, there are three major steps involved in the ω–β processor.
• Collect raw signal data d (x, t, z = 0) and perform a two-dimensional
Fourier transform (FT) to obtain D(βx , ω, z = 0). In practice this stage
can be very efficiently implemented by using coherent IQ sampling of
the signal and a digital signal processor.
• Evaluate the complex function D over a regular grid in βz , βx (called Stolt
interpolation).
• Multiply by the Jacobian and inverse two-dimensional FT to obtain a
two-dimensional image.
Note that only two parameters are important for correct focusing of the image:
• The platform velocity νp (βx depends on νp ).
• The parameter z0 —the distance to the front of the range gate used.
Hence both of these need to be known accurately in order to focus the image
correctly. This basic ω − β SAR processor (we have ignored, for example,
important practical issues such as motion compensation, by assuming that
the sensor moves in a perfect straight line) is summarized geometrically in
Figure 9.3. The resulting image is complex, as at each pixel we obtain an
estimate of the scattering in both amplitude and phase from that point. The
resolution we obtain depends on the angular and radial extent of the measured
sector in wave space. It is common to process a narrow sector θ centred
around θ = 0◦ (by pointing the real antenna axis at right angles to the flight
vector). In this case the resolution in z (range) and cross-range or azimuth (x)
can be simply related to two system parameters: the transmitter bandwidth W ,
and the real antenna dimension in the along-track or x direction. To see this
we use the relationship between Fourier transform variables and estimates of
the bandwidth in wave space in both the z and x directions, as shown in Figure
9.4. Here we see that the resolution in the range direction depends only on
the bandwidth W of the transmitted signal, while the cross-range resolution

SAR processing
pixel

2-D IFT
z
bx Image
Space

x
2dFT
d(x,t) D(bx,v)
Stolt Interpolation c 2 2 o2
D(bx,v) U(bx,bz) bx , +( + – o
2 c
2dIFT Fig. 9.3 Schematic geometry and key steps
U(bx,bz) u(x,z) in SAR processing
344 Applications of polarimetry and interferometry

bx
x
bz
Image
2p
= bz
4pW
⇒ z =
c
Pixel z
z c 2W
2p 2v l
= bx ≈ bu = u ⇒ x =
x c 2u u

Fig. 9.4 Resolution in SAR imaging

depends on the angular width of the sector. At first sight this may seem to
depend on several system variables, but there is a simple relationship between
the maximum angular width (and hence best resolution) and real antenna size,
as shown schematically in Figure 9.5. The beamwidth (in radians) of the real
antenna is approximately equal to the size of the aperture in wavelengths. When
dr we substitute this result into Figure 9.4 we obtain the well-known result, from
SAR theory, that the cross-range resolution is given by half the real antenna
= aperture size:
dr

λ dr
θ ≈ ⇒ x ≈ (9.9)
dr 2
This result has important implications for the exploitation of polarisation effects
in radar imaging, as we now show.

9.1.1 PRF, antenna size and Doppler bandwidth

From equation (9.9) we see that the smaller the antenna, the wider its
beamwidth, the wider the measured sector in wavespace, and hence the better
the resolution. However, since SAR involves a sampled measurement system
dr on a moving platform, we must be careful that the sampling of phase across
the aperture is performed fast enough so as to avoid any sampling errors due
Fig. 9.5 Approximate expression for antenna to aliasing. The pulses are transmitted at a rate called the PRF or pulse repe-
beamwidth tition frequency and, to avoid sampling errors, this PRF must be greater than
or equal to the maximum rate of change of phase. The time rate of change of
phase across the aperture is just the Doppler frequency of the received signal
due to the relative motion between radar and sample point. Doppler shift is
zero when the velocity vector is perpendicular to the line of sight vector to the
point—at the centre of the antenna pattern in side-looking geometry. In general,
the Doppler shift of the signal from a point with angular position θ inside the
beam is proportional to sinθ , as shown on the left-hand side of equation (9.10):

4v sin θ 4v θ 2v λ 2v
fd = ⇒ fd max ≈ · = · = (9.10)
λ λ 2 λ dr dr
The maximum Doppler shift therefore occurs at the outer edges of the real beam
(positive for approaching points, and negative for receding). By again using our
approximation for the beamwidth of the antenna in terms of the real aperture
size (Figure 9.5), we can obtain a simple expression for the maximum Doppler
 9.2 Imaging interferometry: InSAR 345

shift as a function of the ratio of platform speed to real aperture size, as shown
on the right-hand side of equation (9.10). One important constraint required
for undistorted SAR imaging is therefore that the PRF be greater than or equal
to this maximum shift. For small antennas on fast-moving platforms this can
require a very high PRF. However, there are two consequences of operating
at high PRF. The first is the requirement for a transmitter with higher mean
power (given by the peak power times the duty cycle of the radar or τ *PRF,
where τ is the pulse width), which may be expensive or difficult to obtain at the
desired operating frequency. The second issue, however, is more important for
imaging, in that the PRF also impacts on the range extent of the image or range
swath in pulsed systems. The problem here arises from range ambiguities. If
the PRF is too high and the range variation across the image too large, then
there can be an ambiguity as to which transmitter pulse any particular received
pulse actually belongs. A quantitative analysis (Curlander, 1991) shows that the
PRF must be bounded by the following inequality in order to avoid such range
ambiguities:

c
PRF ≤ (9.11)
2Ws

where Ws is the width of the image in the range direction (the range swath size).
This tends to demand low PRF for wide image coverage of the system, which,
as we have seen, is in direct contrast with the requirements for high resolution
in the cross-range direction. The compromise between such conflicting PRF
requirements is one of the central engineering steps in imaging radar design.
Polarisation switching places further constraints on this relationship, as we
show in Section 9.3; but first we turn to consider radar interferometry.

9.2 Imaging interferometry: InSAR

The above ideas can be extended to imaging interferometry by combining two
SAR images generated by linear trajectories separated by a baseline vector b.
Figure 9.6 shows a schematic representation of this process. The two tracks
will in general fill different sectors of wavenumber space, shown as θ1 and θ2
in the figure. By applying a two-dimensional inverse Fourier transform (IFT)
to the separate images we obtain two complex images. However, for successful
interferometry we require good coherence between the two images, and so the
same regions should be processed to generate the two images. In general, the
two regions will overlap only over part of their wave space coverage, and this
will reduce the resolution. This is called common-band filtering, and we see
that in the imaging context it is a two-dimensional process. In the azimuth or
x-direction we require there to be an angular sector of the same width and with
the same mean. This implies that the same squint of the real antenna be used. A
common approach is to employ zero squint; that is, to process to zero Doppler in
both images. This then maintains coherence, and maximizes overlap and hence
resolution. In the z or range direction the radars should have the same carrier
frequency and bandwidth; but we also note that in this direction range, spectral
filtering will be required to shift the pulse bandwidth of the second track so as
to remove baseline decorrelation according to the discussion in Section 5.1.1.1.
346 Applications of polarimetry and interferometry

2-D
IFT

wave-space P
P

T
IF
2-D
Image space

Fig. 9.6 Schematic of wavenumber interfer-

ometric SAR processing

N
2-D
IFT

wave-space P
P P M

T
IF
2-D Image space
Fig. 9.7 Boxcar estimation of complex
coherence in radar imaging interferometry

By co-registering images so that the point P has exactly overlapping pix-

els in the two image spaces, we obtain an image with resolution given by the
SAR process, at each pixel of which we can generate a phase difference. In
this way we generate a high-resolution interferometer whereby we can track
spatial changes in interferometric phase and coherence across a scene. Under
assumptions of stationarity and ergodicity we can then estimate the mean inter-
ferometric phase and coherence using a rectangular window in image space,
centred on the pixel of interest, as shown in Figure 9.7 (Touzi, 1999).
5MN ∗ /
i=1 s1i s2i 0 ≤ |γ̃ | ≤ 1
γ̃ = 5 5 (9.12)
MN ∗ MN ∗ 0 ≤ arg(γ̃ ) < 2π
i=1 s1i s1i i=1 s2i s2i

If the window size is M × N pixels, we have MN samples available for coher-

ence estimation. Clearly, by using large windows we can secure more accurate
estimates of coherence, but at the same time are reducing the effective res-
olution of the image. This idea of multiple channel imaging and combining
channels coherently is also employed in polarimetric SAR, as we consider in
the next section.
 9.3 Polarimetric synthetic aperture radar (POLSAR) 347

9.3 Polarimetric synthetic aperture radar

(POLSAR)
The extension of the SAR concept to the polarimetric case is in principle
straightforward. In place of a single complex number at each location in wave
space, we require a set of four complex numbers representing the scattering vec-
tor at that wave space coordinate. This is shown schematically on the left-hand
side of Figure 9.8. Repeating the SAR imaging process (the ω–β algorithm of
equation 9.8) for each of the four channels separately leads to four images—one
for each component of the scattering matrix, as shown on the right of Figure 9.8.
We can then take linear combinations of the (complex) elements using the w
vector concept to form an image of scattering mechanism w. We can also study
local variations in depolarisation by estimating the coherency matrix from a
weighted sum about the pixel of interest, as shown schematically in Figure 9.9.
Note that this assumes stationarity and ergodicity in that the spatial locality of
the pixel is assumed to consist of random samples from an underlying stochas-
tic process with the same coherency matrix. Under this assumption we can
then estimate the coherency (or Mueller) matrix locally and apply eigenvalue
decomposition or any of the other processing techniques discussed in Chapter 4.
There are two important points to note:
1. The coherency matrix obtained in this way is only ever an estimate,
usually obtained from a relatively small number of samples (depending on

x
HH
HV
VH
VV P
P
P
P
z

Image space
Fig. 9.8 Wave-space interpretation of POL-
SAR imaging

1 M N *T
T = k-ij k-ij
NM i =1 j =1
N
Image space
k
ij
P

M pixel

Fig. 9.9 Pixel averaging of POLSAR data

348 Applications of polarimetry and interferometry

window size M × N ). Hence it contains estimation errors due to speckle fluc-

tuations (Lee, 1994a), and these must be accounted for using, for example,
the multivariate Wishart distribution in any quantitative assessment of the ele-
ments of [T ] (see Appendix 3) (Conradson, 2003; Schou, 2003; Ferro-Famil,
2008).
2. The weights for each pixel inside the window need not be unity (which
corresponds to the standard so called ‘boxcar’ filter). One reason for varying the
elements is that such a window degrades the effective resolution of the image.
The convolution of the rectangular window shape with the image is equivalent
to multiplication of the corresponding Fourier spectra and hence to a low-pass
filtering of the image with a SINC reference spectrum with a width inversely
proportional to window size. A better approach is often to adaptively estimate
the weights over the image, using an estimate of local statistics (around the
pixel of interest). The most popular form of such locally adaptive filtering in
radar imaging is the Lee filter (Lee, 1999, 2008). This filter forms an estimate
of coherency matrix from local samples according to the following weighted
contributions:

T̂ = T  + f (k i k ∗T
i − T ) (9.13)

where f is to be determined from the local statistics. In homogeneous areas

(areas with fully developed speckle), f = 0 and the average matrix is taken
as the estimate. On the other hand, for inhomogeneous areas (isolated point
scatterers, for example) f = 1, and the estimate is obtained using only the
central pixel itself, so preserving spatial resolution. Note that in order to preserve
the correct polarimetric information in the coherency matrix, the same f should
be used on all elements of the T matrix. Details of the expression for f in terms
of local statistics can be found in Appendix 3.
As a popular extension of the Lee filter, the window shape itself is modified to
account for edges at 0◦ and 45◦ to the image boundaries. This leads to a family
of eight Lee filters for each pixel, with the best matched to the local scene being
selected (Lee, 2008). This additional complexity is used in an attempt to further
improve the balance between preservation of spatial resolution in heterogeneous
parts of the image (such as at edges, and for point scatterers) while maintaining
good radiometric resolution (reducing estimation bias of coherency matrices)
in homogeneous regions.
All of this forms a natural extension of single-channel imaging, but we have
so far assumed that all four elements of the scattering matrix can be measured
simultaneously for all points in wave space. In practice this is not possible, and
the coding schemes employed have important implications for the PRF of the
imaging radar, as we now consider.

9.3.1 Pulse switching requirements for

POLSAR imaging
Measurement of the four complex matrix elements of [S] requires transmission
of two orthogonal polarisations x and y, represented by end points of a diameter
of the Poincaré sphere. In principle, x and y can be any orthogonal pair, but
the most common selections are horizontal and vertical linear (H and V) or
 9.3 Polarimetric synthetic aperture radar (POLSAR) 349

SHH SVH SHH SVH SHH SVH

[S] = [S] = [S] =
H SHV SVV V SHV SVV H SHV SVV V
Fig. 9.10 Pulse switching in quadpol
systems

left and right circular (L and R). In order to measure the first column of [S]
we illuminate the scatterer with x polarisation and measure, simultaneously in
amplitude and phase, the scattered field components in the orthogonal x and
y channels. Simultaneous dual reception can be achieved using a two-channel
receiver preceded by an orthogonal polarisation splitter, although, as we shall
see, this complicates the calibration, as multiple channels need to be balanced
in both amplitude and phase (Freeman, 1992).
The second column of [S] can be similarly measured by illuminating the
scatterer with orthogonal y polarisation and again measuring coherently the
x and y components of the scattered radiation. In this way, all four complex
matrix elements are obtained. Ideally, as SAR involves a moving platform we
should transmit x and y polarisations simultaneously. This could in princi-
ple be achieved using suitable orthogonal coding. However, by far the most
common method is to employ a single carrier frequency and time multiplex
the two orthogonal states on a pulse-by-pulse basis, as shown schematically in
Figure 9.10. Here we first transmit a horizontal or H polarised pulse, and receive
in the co- and crosspolarised channels the first column of [S], as shown. The
next pulse is then transmitted with V polarisation, and we measure the second
column of [S]. In this switching scheme there is an inherent time delay of one
PRI (pulse repetition interval) between the first and second column, and so the
bandwidth of the transmitter switch needs to be much faster than any decorre-
lation time of the scattering process, so as to maintain coherence between the
columns of [S]. Bandwidths in the kHz region are typical for imaging radar
applications.
We see, however, that this interleaved switching arrangement also inter-
feres with the sampling requirements for SAR processing. There are two main
options. The first is to keep the same PRF (and hence mean power) as for a
single polarisation system. However, this means that the effective PRF for each
column measurement is halved, as H is transmitted only on every second pulse
in Figure 9.10. This in turn means that the azimuth resolution is also halved,
and a larger antenna (with twice the size in the azimuth direction) is required
to avoid Doppler aliasing. Both of these are unattractive options for imaging
radar systems.
In the second scenario we can instead double the PRF of the system so as
to maintain the column sampling at the same rate as before, and so maintain
azimuth resolution and keep the same antenna size. However, in this case the
mean power of the transmitter is doubled (unless the pulse length is also halved
to maintain the same duty cycle, which then has further implications for the
range resolution, which may have to be halved). This option also leads to a halv-
ing of the range swath and hence a smaller image, due to the possibility of range
ambiguities between columns of [S] (especially in the crosspolarised channels).
In equation (9.11), therefore, we need to use the full PRF in determining the
range swath, leading to reduction by a factor of 2. All these considerations are
worsened by the fact that systems always have finite isolation between orthog-
onal channels; that is, there will inevitably be some y component radiated when
350 Applications of polarimetry and interferometry

Transmitter path [T] Scattering matrix [S] Receiver path [R]

x t11 r11

t21 t12 r21 r12

Fig. 9.11 Calibration diagram for transmitter y t22 r22

and receiver distortions in radar polarimetry

x is selected, and vice versa. Methods for dealing with such practical issues via
calibration are dealt with in the next section.

9.3.2 Polarimetric calibration

Practical devices and systems are never perfect, and there will inevitably be
some corruption of the measured scattering matrix elements by system imper-
fections due, for example, to undesired cross-talk between channels, and
amplitude or phase imbalance in transmitter and receiver systems (van Zyl,
1990; Freeman, 1992; Sarabandi, 1992b; Quegan, 1994). To quantify such dis-
tortions we can employ a cascade of matrices in a composite product, as shown
in equation (9.14). Figure 9.11 shows how this distortion chain originates. First
the ‘ideal’ orthogonal states x and y are passed through the transmitter chain
(including antenna), which incurs some distortions via the channel imbalance
terms t11 and t22 , as well as undesired cross-talk via t12 and t21 .
This transmitted wave then interacts with the scatterer, and the desired
changes in amplitude and phase caused by scattering are imprinted on the sig-
nal. On return to the receiver there is yet another series of distortions and the
addition of thermal noise in the receiver before the observed matrix elements
Oij are obtained. We can formulate an expression for all these processes based
on matrix multiplication, as shown in equation (9.14). This is a standard model
widely used for calibration of polarimetric radar systems (Papathanassiou,
1998a; Kimura, 2004).
     
r11 r12 S11 S12 t11 t12 n n12
[O] = + 11
r21 r22 S21 S22 t21 t22 n21 n22
= [R] [S] [T ] + [N ] (9.14)

There are two strategies for dealing with these distortions. In design, every
effort can be made to reduce cross-talk by good antenna design and careful
system layout. As a second strategy we can employ the process of calibration
to estimate the distortion matrices [R] and [T ], and remove them by matrix
inversions, so that in the absence of noise, for example, we can obtain an
estimate of the true scattering matrix, as shown in equation (9.15):

[S] = [R]−1 [O] [T ]−1 (9.15)

There are various methods available for the estimation of the elements of [R] and
[T ], involving a combination of internal and external calibration techniques.
Internal methods involve monitoring of signals by test channels inside the radar
to estimate imbalances, while external methods (which have the advantage
 9.3 Polarimetric synthetic aperture radar (POLSAR) 351

that they include the full system, including antenna and propagation effects)
involve measuring signals from external active and passive reflectors, which
send signals back through the radar system from an object with known polari-
metric behaviour. By arranging for a set of four such reflectors with orthogonal
scattering vectors (see Section 4.1.4), a set of sixteen equations in the sixteen
unknowns of [R] and [T ] can be obtained. In practice, simpler deployments
are favoured, often of only one or two types of reflector with additional con-
straints (such as symmetry assumptions in the scattering from random media;
see Section 2.4.2.1) used to solve the remaining calibration equations. To see
how these arise we now turn to a vector formulation of the system calibration
equations.

9.3.3 Scattering vector formulation of

polarimetric calibration
One important application of the scattering vector formulation is in the treat-
ment of polarimetric system calibration. We showed in equation (9.14) how the
distortions due to system imperfections can be represented as a triple matrix
product. Ignoring noise, and using the expansion of such a product into a single
matrix equation, we obtain the scattering vector distortion matrix [Z], as shown
in equation (9.16):

[O] = [R] [S] [T ] ⇒

     
OHH r11 t11 r11 t21 r12 t11 r12 t21 SHH
 OHV  r11 t12 r11 t22 r12 t12 r12 t22   
k obs =     .  SHV 
 OVH  = [Z] k s = r21 t11 r21 t21 r22 t11 r22 t21   SVH 
OVV r21 t12 r21 t22 r22 t12 r21 t22 SVV
⇒ k s = [Z]−1 k obs (9.16)

The two key features of this formulation are the presence of quadratic products
of the distortion matrices appearing in [Z], and the simple mathematical form
of the correction process. If we can estimate the elements of [R] and [T ], then
their distortions can be offset by a single matrix inversion as shown. There are
two important special forms of this calibration matrix to consider, both of which
stem from the important case of reciprocal backscatter when SHV = SVH . In
this case the matrix [Z] is no longer square and has dimension 4 × 3, as shown
in equation (9.17):

   
OHH r11 t11 r11 t21 + r12 t11 r12 t21  
 OHV  r11 t12  SHH
r11 t22 + r12 t12 r12 t22  
k obs =   
 OVH  = [Z] k s = r21 t11 . SHV 
r21 t21 + r22 t11 r22 t21 
SVV
OVV r21 t12 r21 t22 + r22 t12 r22 t22
−1 ∗T
⇒ k s = Z ∗T Z Z k obs (9.17)

Note that the observed scattering vector k obs violates reciprocity, though this is
due entirely to the effect of system distortions. This observation can be used,
352 Applications of polarimetry and interferometry

for example, to test the quality of system calibration. If the Pauli channel image
OHV − OVH is formed, it should, for properly calibrated backscatter data,
behave like noise (have zero coherence, and so on). If it contains structure,
then the calibration is not perfect. Note that calibration of the data does not
now involve matrix inversion directly, but instead a pseudo-inverse based on
a least squares solution can be employed, as shown in equation (9.17). This
arises because we are using reciprocity to reduce the number of unknowns
below the number of observations, and hence have an overdetermined system of
equations.
In practice, a further simplification can often be made by assuming that the
cross-talk terms (the off-diagonal elements of [R] and [T ]) are small compared
to the copolar distortion terms. In this case we can set to zero elements of the
[Z] matrix that involve products of the small crosspolar terms. In this ‘small-
coupling’ assumption the [Z] matrix takes on the simplified form shown in
equation (9.18):
 
r11 t11 r11 t21 + r12 t11 0
r11 t12 r11 t22 r12 t22 
[Z] = 
r21 t11
 (9.18)
r22 t11 r22 t21 
0 r21 t22 + r22 t12 r22 t22

A special case of this matrix occurs in the limit of zero cross-talk. If the design
isolation of the system is very good (typically better than –30 dB), then we
can set all off-diagonal terms of [R] and [T ] to zero and establish a simplified
calibration matrix as shown in equation (9.19):
 
1 0 0
0 t22 /t11 0 
[Z] = r11 t11 
0

 (9.19)
r22 /r11 0
0 0 r22 t22 /r11 t11

Note that this matrix still causes lack of reciprocity in the observed vector due
to differences in the receiver and transmitter copolar distortion channels.
To illustrate how this matrix vector formulation can be used to derive practical
calibration algorithms, we summarize here the main steps involved in a widely
used POLSAR calibration algorithm, first derived in Quegan (1994), and then
further modified by Papathanassiou (1998a) and Kimura (2004). It employs
the small cross-talk hypothesis of equation (9.18) to express the relationship
between calibrated and observed (uncalibrated) scattering four-element vectors
in the form shown in equation (9.20).

   −1  
hh k2 0 0 0 hh
hv  0 0  
  =Y  k 0  [Z]−1 . hv  (9.20)
vh 0 0 k 0 vh
vv calibrated 0 0 0 1 vv uncalibrated

Here two scalar factors Y = r22 t22 k = rr2211 have been factored from the matrix
inverse. These are found by imaging a single point target with a known scattering
 9.3 Polarimetric synthetic aperture radar (POLSAR) 353

matrix. Often a trihedral corner reflector (Figure 1.21) is used for this purpose.
It has the identity matrix as a true scattering matrix, and hence by measuring
the ratio of apparent copolarised elements for the pixel we can establish a direct
estimate of ‘k’. The radiometric factor Y can then be determined from the
known radar cross-section (RCS) of the trihedral. The 4 × 4 matrix [Z]−1 can
be written in terms of a set of four cross-talk ratios u, v, w and z, and a factor
‘a’ as shown in equation (9.21):

 
1 −v −w vw
1 −az a azw −aw
[Z]−1 =  
a  −u uv 1 −v 
azu −au −az a (9.21)

r21 t21 r12 t12 r22 t11

u= v= w= z= a=
r11 t22 r22 t11 r11 t22

Importantly, these five parameters can be estimated from observations of the

covariance matrix for a distributed or depolarising region of the image. Solving
for these components requires two key assumptions:

1. The first is reciprocity, so that for the true matrix HV = VH. Hence any
departure from this in the measured data is attributed to the effects of
the system distortions u, v, w, z, and a.
2. In addition, we also need to assume that the depolariser has reflection
symmetry, so that the true covariance matrix has zero elements whenever
co- and crosspolarised channels are multiplied (see Section 2.4.2).

The second is a more restrictive assumption, as it may not be true in the presence
of surface slopes or in heterogeneous regions, where discrete point scatterers
exist (in urban areas, for example). It also forces the use of a large number of
looks to reduce bias in the estimation of zero cross-products. For this reason,
calibration must be applied over very flat homogeneous regions containing
strong depolarising effects (volume scattering). Flat, forested areas (such as the
Amazon basin for spaceborne sensors) are typical of regions of choice. On the
other hand, urban and mountainous regions (even if vegetated) must be avoided
in the calibration process, as they are likely to violate reflection symmetry. In
practice these are masked out of the calibration by first estimating the co-/cross-
correlations and rejecting pixels where this is high (Papathanassiou, 1998a;
Kimura, 2004).
The detailed algorithm is shown in equation (9.22). Starting on the right
we form the Quadpol scattering vector for each pixel, and then average over
azimuth to obtain a 4 × 4 covariance matrix [C] (having also masked out pixels
with low SNR as well as those which violate reflection symmetry). These
elements can then be used to solve for all the unknowns:

   
hh c11 c12 c13 c14
hv  + , c c22 c23 c24 
k clutter =  ∗T  21
vh ⇒ [C] = k clutter k clutter = c31

c32 c33 c34 
vv c41 c42 c43 c44
354 Applications of polarimetry and interferometry

 c44 c31 − c41 c34


 u=

 c11 c44 − c14 c41





 c11 c34 − c31 c14

v=

 c 11 c44 − c14 c41



 c11 c24 − c21 c14


w = c11 c44 − c14 c41

⇒ (9.22)

 c44 c21 − c41 c24

z=

 c11 c44 − c14 c41



 c33 − uc13 − vc43



a1 =

 c23 − zc13 − wc43



 (c − zc13 − wc43 )∗

a2 = 23
c22 − z ∗ c21 − w∗ c24
These calibration parameters are then used to correct the observed single look
complex (SLC) data for each pixel by the matrix multiplication shown in equa-
tions (9.20) and (9.21). Note that the two parameters a1 and a2 are combined to
estimate ‘a’as shown in equation (9.23) (assuming equal noise in all polarisation
channels):

|a1 a2 | − 1 + (|a1 a2 | − 1)2 + 4 |a2 |2

a= (9.23)
2 |a2 |
From this we can symmetrize the matrix; that is, we can estimate the
true crosspolarised component as a linear combination of the measured
crosspolarised signals, as shown in equation (9.24):
(a∗ hv + vh)
xx = (9.24)
1 + aa∗
Two recent modifications of this basic algorithm have been proposed. In the
first—by Kimura (2004)—the assumptions about equal noise in all channels can
be relaxed. This can in principle allow the treatment of low SNR regions, as
can occur in power-limited applications such as spaceborne sensors. However,
given the additional multi-looking requirements in noisy regions and the higher
SNR achievable with airborne systems, it is often easier just to mask out those
few areas of low SNR using, for example, the fourth eigenvalue of [C], or more
simply the HV/VH coherence. If the HV/VH coherence is less than, say, 0.9
(around 10 dB SNR), then mask out the pixels from the calibration algorithm.
The second recent development (Ainsworth, 2006) has been the relaxing of
the requirement for zero correlation between co- and crosspolarisation channels
(the reflection symmetry assumption). This allows application of the technique
over a much wider range of terrain types at the expense of computational
complexity. (An iterative algorithm is now required where the parameters are
first estimated from the data and then fed back into the model to improve the
estimation.)

9.3.4 Compact polarimetry

Sometimes the complexity, bandwidth and range swath coverage restrictions of
switching the transmitter polarisation are undesirable, and so-called compact
 9.3 Polarimetric synthetic aperture radar (POLSAR) 355

polarimetry systems have been developed as a compromise (Souyris, 2005;

Raney, 2006, 2007). In these systems the transmitter polarisation state is fixed,
but the dual channel coherent receiver configuration is maintained. This yields
measurement of only part of the complex scattering matrix, although interesting
permutations arise by allowing the transmitter state to be a different polarisa-
tion to that of the receiver; for example, transmitting circular polarisation and
receiving horizontal and vertical linear components. In this section we outline a
general approach to such compact designs and highlight some of their strengths
and weaknesses.
In general we start by considering the S matrix represented in an arbitrary
orthogonal basis xy used in the receiver. The fixed transmitter polarisation
is then represented by complex components px and py in this basis. The key
constraint of compact polarimetry is that px and py are fixed and form a uni-
tary vector (with unit amplitude). The two orthogonal receiver channels then
measure complex signals s1 and s2 , as shown in equation (9.25):

    
s1 S SXY px
= XX (9.25)
s2 SYX SYY py

Each of these received signals is a linear combination of the elements of [S].

The real utility of compact polarimeters lies, however, not in coherent analyses
but in the characterization of depolarisers. In this case, interest centres not so
much on the complex signals s1 and s2 but on their 2 × 2 coherency matrix [J ],
as shown in equation (9.26):

 + ∗ , + ∗ ,  
s1 = SXX px + SXY py |px |2 +|py | =1
2
s s +s1 s2∗ , = JXX JXY
−→ [J ] = + 1 1∗ ,
s2 = SYX px + SYY py s2 s1 s2 s2 JYX JYY
(9.26)

This matrix has only four parameters, while the full scattering coherency matrix
has up to sixteen. However, by assuming two symmetries in the scattering
process we can reduce this discrepancy. The first—reciprocity in backscatter
forces Sxy = Syx and the full coherency matrix—then has rank 3 with nine
parameters. The second—reflection symmetry with an axis aligned parallel
to x or y—forces cross-products involving mixed co- and crosspolar terms to
zero; that is, SXX SXY  = SYY SYX  = 0. This reduces the scattering coherency
matrix [T ] and covariance matrix [C] in the xy basis to the reduced 3 × 3 forms
shown in equation (9.27), both of which have only five unknowns.

   
t11 t12 0 c11 0 c13
[T ] = t12
∗ t22 0  ⇔ [C] =  0 c22 0 (9.27)
0 0 t33 ∗
c31 0 c33

Now we wish to relate the four observations of [J ] obtained in compact

polarimetry to the five unknowns of the full scattering covariance matrix under
reciprocal reflection symmetry. This we can do by expanding equation (9.26)
and using reciprocity and reflection symmetry relations to obtain the following
356 Applications of polarimetry and interferometry

set of linear equations:

 
  c11  
px p ∗ py py∗ 0 0 0   JXX
 x   c22   
 0 px px∗ py py∗ 0 0    JYY 
   c33 = 
 0 Re(px py∗ ) Re(px py∗ ) −Im(px py∗ )   Re(J ) ⇒ [P]c = j
 0    XY 
 Re(c13 ) 
0 −Im(px py∗ ) 0 ∗ ∗
Im(px py ) Re(px py ) Im(JXY )
Im(c13 )
(9.28)

Here we have four equations in five unknowns, and so cannot solve for all
five elements of [C], whatever the choice of px and py . There are then three
important special cases of compact polarimetry that arise in practice. They all
derive from the choice of xy = HV; that is, for linear horizontal and vertical on
receive. In the simplest case we can then choose px = 1, py = 0—fixed horizontal
transmit. In this case the [P] matrix takes the following form:
 
1 0 0 0 0
0 1 0 0 0
px = 1, py = 0 ⇒ [P] =  0 0 0 0 0
 (9.29)
0 0 0 0 0

Here we see that [J ] then only contains information about the scattered power in
co- and cross-channels. (Remember that we are assuming reflection symmetry
and so the HH and HV channels are uncorrelated.) In order to access information
related to the other elements a different choice of px and py are required. In the
π/4
B√ compact mode, for example, the transmitter is set to 45◦ linear—px = py =
1 2—and the matrix [P] takes the form shown in equation (9.30):
 
1 1 0 0 0
1 1 1 0 1 1 0 0
px = √ , py = √ ⇒ [P] = 

 (9.30)
2 2 2 0 1 0 1 0
0 0 0 0 1

We note two important aspects of this mode. The first is that we now have access
to linear combinations of all the elements of [C], and hence some sensitivity
to all the elements of the covariance matrix. The second is the factor of 1/2 in
front of the matrix [P]. This implies a 3-dB loss of signal compared to a full [S]
matrix system. Such signal loss is an inevitable consequence of mismatching the
transmitter and receiver bases. Finally, another√ mode that has√ been proposed
is to transmit circular polarisation:px = 1/ 2, py = ±i/ 2. This case is
very similar to the π /4 mode, but with a [P] matrix of the form shown in
equation (9.31):
 
1 1 0 0 0
1 ±i 1 0 1 1 0 0
px = √ , py = √ ⇒ [P] =   (9.31)
2 2 
2 0 0 0 0 ±1
0 ±1 0 ∓1 0

Again we see that there is a 3-dB loss of signal due to antenna mismatch, but
again there is information from all components of [C] present in the mixture.
 9.3 Polarimetric synthetic aperture radar (POLSAR) 357

Some authors have tried to extend this approach so as to be able to reconstruct

the reflection symmetric [C] matrix in full (Souyris, 2005). To do this we require
an extra constraint equation between the elements of [C] to reduce the number
of unknown to four, so matching the number of observations. Ideally we would
like to find an extra linear relationship so that we could make [P] a 5 × 5 square
matrix and then solve for the elements of c by matrix inversion. However,
so far no such linear relationship has been found, and instead a non-linear
constraint is widely used. We can motivate the development of this approach as
follows.
One way to reduce the number of unknowns in [C] is to assume a model
of scattering. In common with our discussions in Sections 4.2.3 and 7.4.1, we
adopt a random-volume-over-ground (RVOG) model for scattering by natural
terrain. In this case we assume the volume scattering component shows the
much stronger azimuthal symmetry, and it is only the presence of the direct
surface or dihedral returns that break this symmetry and leads to a reflection
symmetric composite. This model can now be used to relate the normalized
level of crosspolarisation to the copolar coherence, as shown in equation (9.32).
We start by considering the limiting case of zero surface component; that is,
pure volume scattering. In this case the coherency matrix is diagonal with two
degenerate eigenvalues. This leads to an additional relationship between the
crosspolarised power and the power in the second Pauli channel as shown.
By expanding and using the fact that for azimuthal symmetry the copolarised
powers in XX and YY are equal, we obtain a relationship between the HH/VV
coherence and normalized crosspolarised power. The key assumption we can
now make is that this relationship applies even when we add a non-zero surface
component.

[T ] = [Ts ] + [TV ]
 
t11 0 0 9 : 9 :
[TV ] =  0 t22 0  ⇒ 4 |SXY |2 = |SXX − SYY |2
0 0 t22
9 : 9 : 9 :
∗ ∗
= |SXX |2 + |SYY |2 − 2Re(SXX SYY ) = 2 |SXX |2 − 2Re(SXX SYY )
+ , ∗ )
4 |SXY |2 2Re(SXX SYY
⇒+ , + , = 1 − + , + ,
|SXX |2 + |SYY |2 |SXX |2 + |SYY |2
2Re(SXX SYY ∗ )
=1− + , = 1 − |γXXYY |
2 |SXX |2
+ ,
|SXY |2 1
⇒+ , + , = (1 − |γXXYY |) (9.32)
|SXX | + |SYY |
2 2 4

To check this we first ask what happens in the limit as the volume tends to
zero and we are left with bare surface scattering. In this case (according to the
RVOG model) the coherency matrix is rank-1 (or with very small secondary
eigenvalues), and is therefore represented by a symmetric scattering matrix,
which we also assume is diagonal in the XY basis (due to Bragg scattering
358 Applications of polarimetry and interferometry

from a flat surface, for example). Hence it has zero crosspolarisation combined
with a high polarimetric coherence equal to unity. We see that this combination
is still consistent with equation (9.32). For the general mixed case between
these two extremes we can adopt a simple two-component decomposition as
shown in equation (9.33):
   
cos2 α sin α cos αeiδ 0 0.5 0 0
   
[T ] = [Ts ] + [TV ] = ms sin α cos αe−iδ sin2 α 0 + mv  0 0.25 0 
0 0 0 0 0 0.25
   
1 + sin 2α cos δ 0 cos 2α + i sin 2α sin δ 3 0 1
ms   mv  
⇒ [C] =  0 0 0 + 0 2 0
2 8
cos 2α − i sin 2α sin δ 0 1 − sin 2α cos δ 1 0 3
(9.33)

Here we combine two components: one a rank-1 surface mechanism with mag-
nitude ms , and the second a random dipole cloud with scattering cross section
mv . This is very similar to the Freeman decomposition (see Section 4.2.4) or the
RVOG model (see Section 7.4.1). We can now express the cross-to-copolarised
scattering ratio and HH/VV coherence as functions of the surface-to-volume
scattering ratio µ = ms /mv and scattering mechanisms α and δ, as shown in
equation (9.34):
+ ,
4 |SHV |2 1
+ , + ,=
|SHH | + |SVV |
2 2
2µ + 32
 
  (9.34)
 1
+ µ(cos 2α + i sin 2α sin δ) 
 
1 − |γHHVV | = 1 −   4 
 9 
 µ2 (1 − sin2 2α cos2 δ) + 32 µ + 16 

We can now use this to check the equality of equation (9.32) for arbitrary
mixtures of surface and volume scattering mechanisms. Figure 9.12 shows
some example calculations. Here we plot, along the x axis, the cross-to-copolar
ratio, and along the y axis one minus the coherence amplitude. For equation
(9.32) to be valid, therefore, we require the points to lie along a line at 45◦ .
We show how the two parameters vary for µ ranging from –30 dB to +30
dB; that is, from the limiting cases of zero surface to zero volume scattering.
We show the results for steps of 15◦ in alpha (always with δ = 0 to simplify
the situation), starting from zero. We note that for α = 0 the equality holds
for all mixtures, and that the two limiting points of zero surface (the origin)
and zero volume (when both approach 2/3) are also satisfied for all scattering
mechanisms, as expected. However, for alpha angles greater than 30◦ we note
significant departures from the model. In particular, for alpha = 45◦ we see that
we have a situation where the coherence can be zero even when there is low
crosspolarisation. This arises because one of HH or VV scattering coefficients
goes to zero for this mechanism.
More significantly we see that for all alpha values greater than 45◦ —
for dihedral scattering of all types—there is always a µ value that leads to
zero coherence, and consequently to large deviations from the simple linear
 9.3 Polarimetric synthetic aperture radar (POLSAR) 359

XPOL ratio and HHVV coherence as a function of surface-to-volume ratio

1

0.9

0.8

0.7
1 - HHV Coherence

0.6

0.5 0 degrees
15 degrees
0.4 30 degrees
45 degrees
0.3 60 degrees
75 degrees
0.2 90 degrees
Compact Pol
0.1 Fig. 9.12 The variation of cross-to-
copolarised ratio versus 1-HHVV coherence
0 for varying mixture of surface and vol-
0 0.2 0.4 0.6 0.8 1
ume scattering and different scattering
Cross-to-Copolarised ratio mechanisms

relationship. We can see from equation (9.34) that this arises when cos2α is
negative, as it can then cancel the positive numerator contribution from the
volume scattering. As such this effect has its origin in the 180-degrees phase
shift caused by double reflection.
Despite these limitations, relations such as equation (9.32) are widely used
in the compact polarimetry community. The reason is that by combining this
result with [P] leads to a set of five (non-linear) equations in five unknowns
(the elements of [C]), as shown in equation (9.35), which can then be solved
by iteration for all the elements of [C] and therefore [T ] or the Mueller
matrix [M ].
 
  c11  
px px∗ py py∗ 0 0 0 JXX
 c22 
 0 px px∗ ∗
py py 0 0    
   c33  =  JYY 
 0 Re(px py∗ ) 0 Re(px py∗ ) −Im(px py∗ ) 
Re(c13 )
 Re(JXY )
0 −Im(px py∗ ) 0 Im(px py∗ ) Re(px py∗ ) Im(JXY )
Im(c13 )
 
4c22  c13 
+ √ −1=0
c11 + c33  c11 c33 
(9.35)

In this way, compact polarimetry can be used to provide estimates for the
full coherency or covariance matrix elements without compromising the PRF
requirements of a single channel SAR system. Note, however, that this is
only true for the class of depolarisers satisfying the combined assumptions
of reciprocity, reflection symmetry and especially the equality between copolar
coherence and cross-to-copolarised power (which, as we have shown in Figure
(9.12), is the weakest assumption). In addition to the case shown above, the
compact assumptions also do not apply to general point scatterers (such as
rotated dihedrals or dipoles, for example) or to scattering from sloped terrain—
both of which introduce correlation between co- and cross-polarisations and
360 Applications of polarimetry and interferometry

lead to a high coherence, even in the presence of significant crosspolarisation.

For these more general scenarios, full [S] matrix polarimetry is required. Hence
the user must be aware of exactly what kind of applications are in mind when
deciding on the best mode for use in imaging radar polarimetry.
Finally, we note that in [S] or compact POLSAR imaging it is important to
measure the phase as well as the amplitude of the scattered signal, and hence
coherent in-phase and quadrature (IQ) detection is required. If such detectors
are not available, an alternative indirect measurement of [S] can be made, based
entirely on incoherent (intensity) detectors only. However, in this case a com-
bination of four transmitter states (typically linear H, V, 45-degree, and circular
polarisations) and four receiving filters are used to measure the sixteen elements
of the Mueller matrix [M ] (see Section 2.2). From these elements, under some
restrictions (see Section 2.2), we can then estimate seven of the eight com-
ponents of [S]. (Absolute phase cannot be determined by this technique, and
hence interferometry cannot benefit from this approach.) Alternatively we can
use this to estimate the covariance or coherency matrix directly from [M ]. An
extreme form of this approach is to dispense with the transmitter completely
and rely on natural radiation of a scene. In this case the four-element inco-
herent receiver measures the Stokes vector of the scattered wave (see Section
1.3.4). Normally, the incident wave in such configurations is considered a ran-
domly polarised wave; that is, one with zero coherence, in which case only
the first column of the Mueller matrix is measured, providing access to limited
information about the scattering matrix (see Section 2.6). Such an approach is
widely used at optical wavelengths where direct phase measurements are not
possible. There have been some examples of this approach in radar applications
(Boerner, 1992; Sarabandi, 1992a), but primary interest has been in coherent
imaging system applications. This brings us to consider the most general case,
in which imaging polarimetry and interferometry are combined into the most
flexible sensor configuration: POLInSAR.

9.4 Polarimetric SAR interferometry

(POLInSAR)
The final stage in our radar imaging hierarchy is to consider imaging polarimet-
ric interferometry or POLInSAR (Cloude, 1998; Krieger, 2005). This involves
measurement of the full scattering matrix with wave space diversity for two
spatial tracks separated by a baseline vector b. This then enables SAR imaging
using Stolt interpolation and inverse Fourier transforms of each of the eight
complex channels, followed by interferometry between co-registered complex
images using arbitrary complex linear combinations based on weight vectors
w1 and w2 , as shown schematically in Figure 9.13.
This then provides maximum flexibility in terms of combined image-based
polarimetric and interferometric processing. The resolution and coverage issues
are the same as for polarimetry, and the same balance as regards compact or
quad polarimetry must be considered. The calibration and compact polarimetry
requirements do, however, deserve special attention, as they have some impor-
tant differences from standard polarimetry. We now turn to consider each of
these in turn.
 9.4 Polarimetric SAR interferometry (POLInSAR) 361

x
kz 2-D IFT

2-D IFT P
P
P
1 P
2-D IFT z
kx W1
2-D IFT
k-space P
Image space P
x
kz 2-D IFT

2-D IFT P
P Interferometry
P W2
1 P
2-D IFT z Fig. 9.13 Schematic of steps involved
kx in polarimetric SAR Interferometry, or
2-D IFT
POLInSAR

9.4.1 Calibration of POLINSAR systems

In this section we consider the effect of polarimetric calibration errors on coher-
ence estimation in polarimetric interferometry. We saw in equation (9.16) that
calibration errors can be represented by a distortion matrix [Z], which multi-
plies the true scattering vector k to yield the observed or measured vector. Any
uncorrected distortions will then impact on the estimate of the interferometric
coherency matrix and hence on estimation of coherence itself. In polarimetric
interferometry we must further allow for the possibility that we have different
distortion matrices Z1 and Z2 at the two different spatial/temporal positions
across the baseline. Hence we can generate the following general distorted
forms of the composite 6 × 6 matrix 2 .

 
Z1 T11 Z1∗T Z1 12 Z2∗T
[2 ] =
Z2 ∗T
12 Z1
∗T Z2 T22 Z2∗T

w∗T ∗T
1 Z1 12 Z2 w 2
⇒ γ̃ (w1 , w2 ) =   (9.36)
w∗T ∗T ∗T ∗T
1 Z1 T11 Z1 w 1 · w 2 Z2 T22 Z2 w 2

This impacts on the estimation of coherence for scattering vectors w1 and w2 ,

as shown on the right-hand side of equation (9.36). In simple terms it is clear,
therefore, that calibration errors do change coherence, and hence will act to
distort, for example, the coherence region. In particular, if [Z1 ] and [Z2 ] are
different we obtain an unknown mixture of polarimetric and interferometric
coherence, which will act to distort our interpretation of scattering behaviour
in the pixel by mixing interferometric and polarimetric coherences. This again
points to the need for good polarimetric calibration procedures, driving the
matrices Z1 and Z2 as close as possible to the identity matrix. However, it
is one advantage of the optimization approach to POLInSAR that it provides
some robustness to residual calibration errors, as we now demonstrate.
Using the distorted form of 2 including calibration effects, we can now
rewrite the optimization eigenvalue relations (see Section 6.2) for constrained
362 Applications of polarimetry and interferometry

and unconstrained optimization, as shown in equation (9.37):

w1 =w2 −1 
−→ Z1 T11 Z1∗T + Z2 T22 Z2∗T Z1 12 Z2∗T + Z2 ∗T ∗T
12 Z1 w = λr w
if Z1 = Z2
(Z ∗T )−1 T −1 Z ∗T w = λr w
⇒ T −1 w = λr w

 
 Z2∗T −1 T22
w1 =w2
−1 ∗T −1
12 T11 12 Z2∗T w2 = λopt w2
−→
 ∗T −1 −1 −1 ∗T ∗T
Z1 T11 12 T22 12 Z1 w1 = λopt w1


−1 ∗T −1
T22 12 T11 12 w2 = λopt w2
⇒
T −1  T −1 ∗T w = λ w
11 12 22 12 1 opt 1
(9.37)

In the upper part we show the effect on constrained optimization (w1 = w2 ),

and see that as long as Z1 = Z2 then it follows that the effects of calibration
distortion can be absorbed into the eigenvectors, and that the optimum coher-
ence values themselves (the eigenvalues) remain unchanged. In the lower part
of equation (9.37) we show a much stronger form of this result for the uncon-
strained optimization case. Here we see that even if Z1 and Z2 are different, the
effects of calibration can still be absorbed into the eigenvectors, and hence the
optimum coherences remain unchanged. This can be important in applications
where only the coherences themselves are used rather than the eigenvectors.
In such cases the calibration requirements can be relaxed compared to those
required for polarimetry alone.

9.4.2 Compact POLInSAR

By fixing the transmit polarisation, but receiving orthogonal components x and
y for two sampling positions separated by a baseline B, we obtain a 4 x 4
polarimetric interferometric matrix, as shown in equation (9.38):
 9 : 9 : 9 : 9 :
sx1 = SXX
1 p + S1 p
x XY y 
 sx1 sx1 ∗ sx1 sy1 ∗ sx1 sx2 ∗ sx1 sy2 ∗
 9 
1 p + S1 p 

  1 1 ∗ : 9 1 1 ∗ : 9 1 2 ∗ : 9 1 2 ∗ :
sy1 = SYX x y  |px | +|py | =1
2  s s s s s s s s 
9 y x : 9 y y : 9 y x : 9 y y :
YY 2
−→ [J ] =  
sx2 = SXX
2 p + S2 p
x 
XY y 
 s2 s1 ∗ s2 s1 ∗ s2 s2 ∗ s2 s2 ∗ 

  x x x y x x x y 
2 p + S2 p 
sy2 = SYX 

9 : 9 : 9 : 9 :
x YY y 2
sy sx 1 ∗ 2
sy sy 1 ∗ 2
sy sx 2 ∗ 2
sy sy 2 ∗

 C 
T11 C 12
= (9.38)
C ∗T T C
12 22

This matrix can be partitioned into 2 × 2 sub-matrices for polarimetry and

interferometry, as shown by the C superscript for ‘Compact’, and then the same
optimization procedures as used for quadpol interferometry can be applied.
Both constrained and unconstrained optimization algorithms can be developed.
 9.5 Applications of polarimetry and interferometry 363

The unconstrained optimization, for example, can be implemented as shown in

equation (9.39):
  2
w1 =w2
C −1 C ∗T T C −1 C w = λ λ∗ w
T22 12 11 12 2 1 2 2 K1C w1 = νw1 = γopt  w1
−→ λ1 = λ2 = γ̃opt ⇒  2
T C −1 C T C −1 C ∗T w = λ λ∗ w
11 12 22 12 1 1 2 1 K2C w2 = νw2 = γopt  w2
(9.39)

We can also derive the coherence region for compact polarimetry using con-
strained optimization, as we did in Section 6.2. The key difference here is
that we are now working in a two-dimensional rather than three-dimensional
complex space, and hence the coherence region will always be an ellipse (see
Section 6.2.4). It is interesting, from an applications point of view, to deter-
mine if we can use these ideas to approximate the true region shape by using
such a compact system. We return to this point in Section 9.5.3, when we
consider the coherence region for vegetation scattering obtained from chamber
measurements.
We now turn to consider some illustrative applications of polarimetry and
polarimetric interferometry.

9.5 Applications of polarimetry and

interferometry
In this section we illustrate application of the theory developed in previous chap-
ters. To do this we employ data from four sources—the first being polarimetric
interferometry measurements made inside a large 10-m anechoic chamber: the
European Microwave Scattering Laboratory (EMSL), located at the European
Joint Research Centre (JRC), at Ispra, in Italy: http://www-emsl.jrc.it/ (Cloude,
1999; Sagues, 2000, 2001; Lopez, 2000).
The second is data from advanced computational electromagnetic simula-
tions made by the NASA Goddard Space Center in New York: http://www.
giss.nasa.gov/∼crmim/ (Mishchenko, 2007). These represent an example of the
very latest developments in the computer-based solution of Maxwell’s equa-
tions for a complicated system of interacting particles—in this case a cluster
of dielectric spheres. This approach provides a full solution free of many of
the simplifying assumptions usually employed in multiple scattering calcula-
tions. One advantage of this approach is that it provides a full vector solution,
including depolarisation effects, allowing us to explore full parameterization
of a complex scattering problem.
The third is data from an airborne imaging POLInSAR system: the E-SAR,
operated by the German Aerospace Centre (DLR) at Oberpfaffenhofen, near
Munich, Germany: http://www/dlr.de/hr/en/desktopdefault.aspx. This was one
of the first systems to successfully demonstrate repeat-pass polarimetric inter-
ferometry at low radar frequencies (L and P bands) (Papathanassiou, 1998;
Reigber, 2000, 2001), and since then has been a major source of such data to
the wider radar sciences community (Hajnsek, 2009).
Finally we employ data from the ALOS-PALSAR sensor—an L-band POL-
SAR satellite system operated by the Japanese space agency JAXA, and
364 Applications of polarimetry and interferometry

– z +

 < 0º  > 0º

m
.56
inc

=9
sct

RO
2.0 m Y
0.3 m
Turntable

Fig. 9.14 Geometry of EMSL anechoic

chamber in Ispra, Italy

launched in January 2006 (Rosenqvist, 2007). This spaceborne imaging radar

provides a global Earth observation and monitoring role using full [S] matrix
and dual polarimetry imaging modes.

9.5.1 Application 1: depolarisation by surface scattering

To illustrate the nature of depolarisation caused by rough surface scattering we
employ data from the European Microwave Scattering Laboratory (EMSL):
http://www-emsl.jrc.it/EMSLdata/nvt04-07-11/. The EMSL is located in a large
anechoic chamber, enabling environmentally controlled broadband fully polari-
metric measurements of surface and volume scattering. Figure 9.14 shows the
geometry of the measurement chamber used. The transmitter is fixed, and can
be used for scattering measurements at various angles of incidence θ. A separate
receiver can be used for making monostatic or bistatic measurements on the
same surface.
Computer-generated surface profiles with isotropic Gaussian statistics were
machined for use in the surface scattering experiments. The two surfaces used
are both a composite of sand + ethanediol + water, with rms heights of s = 2.5
cm (rough) and 0.4 cm (smooth), and with the same correlation length l = 6
cm. The surfaces are contained in a cylinder of 2 m diameter and 0.4 m depth,
as shown in Figure 9.15. The bottom of the cylinder was lined with absorbing
material to minimize boundary effects on the measurements.
The complex dielectric constant of the surface was measured experimentally,
and shows some decrease with increasing frequency. To provide an idea of the
values obtained, at 5 GHz the dielectric constant has a value ε = 7 – i3, rising
to ε = 9 – i4 at 2 GHz, and falling to ε = 5.5 – i2 at 10 GHz.

Surface backscatter
Wideband scattering matrix measurements were made in monostatic mode
(backscatter) over the frequency range 1–19 GHz and incidence angles θ of
10–50◦ (in 5 or 10-degree steps). For each angle of incidence the turntable is
rotated through 360◦ in 5-degrees steps. In this non-imaging case, averaging
 9.5 Applications of polarimetry and interferometry 365

Fig. 9.15 Image of the computer manufac-

tured rough surface located in the EMSL
chamber

Backscatter (HH, HV, VH, VV) at 30 degrees

5

–5
Sattering coefficient (dB)

–10

–15

–20
HH
HV
–25 VH
VV
–30

–35

–40
2 4 6 8 10 12
Fig. 9.16 Rough surface backscatter in linear
Frequency (GHz) basis as a function of frequency

is made over 360◦ of azimuth coverage (72 samples) combined with some fre-
quency smoothing over a 160-MHz bandwidth (sixteen frequency steps). By
averaging over such combined azimuth/frequency variations we ensure that
the surface has a scattering coherency matrix of the reflection symmetric form
shown in equation (2.75). Starting with the rough surface, Figure 9.16 shows
the backscatter cross-section as a function of frequency for 30-degree angle of
incidence. Here we show the cross-section in the linear basis, HH, HV VH and
VV, being the diagonal elements of the covariance matrix [C]. We see that there
is little change with frequency over the band, and that the copolarised chan-
nels HH and VV are some 10 dB greater than the crosspolarisation channels.
As expected, we see that HV = VH (due to reciprocity and good polarimetric
calibration of the data). Although the copolarised channels are also equal in
amplitude (HH = VV), this is due to another reason: the particular scattering
366 Applications of polarimetry and interferometry

Backscatter eigenvalue spectrum at 30 degrees

5

–5

Sattering coefficient (dB)

–10

–15

–20
lambda 1
lambda 2
–25 lambda 3
lambda 4
–30

–35

–40
2 4 6 8 10 12
Fig. 9.17 Coherency eigenvalue spectra for
rough surface scattering Frequency (GHz)

symmetries of this surface. By generating the coherency matrix [T ] at each

frequency and calculating its eigenvalues, we obtain the variations shown in
Figure 9.17. Here we see a maximum eigenvalue around 3 dB larger than the
linear HH or VV channels (due to the eigenvector, which in this case is close
to the coherent sum SHH + SVV ). The minimum eigenvalue is around –40 dB,
and this represents an eigenvector of the form SHV –SVH . By reciprocity this
should be exactly zero, but noise and residual calibration errors in the data give
us around 40 dB of dynamic range in this dataset.
One interesting feature of Figure 9.17 is the presence of two small eigen-
values around 10 dB below the maximum. These represent the depolarisation
subspace. The fact they are roughly equal illustrates the noise-like behaviour
of this subspace—itself due to the symmetry of the rough surface. Secondly,
the fact that it is a two-dimensional subspace means that not only crosspolari-
sation HV gives a small, depolarised return, but some coherent combination of
copolarised channels is also depolarised (in this case SHH –SVV ).
Figure 9.18 shows how this eigenvalue spectrum varies with angle of inci-
dence (now for a fixed frequency of 10 GHz). Note that because of the roughness
of the surface, the dominant eigenvalue does not vary much with angle of
incidence. However, the depolarisation subspace shows significantly more
variation, with the depolarised eigenvalues decreasing with increasing angle
of incidence. One way to demonstrate the balance of this polarised/depolarised
decomposition is to normalize the eigenvalues at each frequency by their sum
and display them as probabilities, as shown in Figure 9.19. Here we see that
at small angles the depolarised signal can be around 20% of the total, while
at larger angles it reduces to around 2%. Hence, despite the roughness of this
surface (and in a normalized sense it represents a wavenumber/rms rough-
ness product βs = 5.236) the signal actually remains strongly polarised. This
means that polarimetric phase and amplitude ratios remain coherent and can
 9.5 Applications of polarimetry and interferometry 367

Eigenvalues at 10 GHz (bs = 5.236)

0

–5

–10
Scattering amplitude (dB)

–15

–20

–25

–30

–35

–40

–45
10 15 20 25 30 35 40 45 50
Fig. 9.18 Variation of coherency eigenvalue
Angle of incidence spectra with angle of incidence

Normalised eigenvalues at 10 GHz (bs = 5.236)

100

10–1

10–2

10–3 Fig. 9.19 Variation of normalized eigenvalue

10 15 20 25 30 35 40 45 50
spectra with angle of incidence for rough
Angle of incidence surface scattering

be estimated from the data for the purposes, for example, of surface parameter
estimation.
Another way to represent this information is to use the entropy/alpha
approach. Figure 9.20 shows the results when applied to this rough surface
scattering data. In the upper graph we show the entropy, which reduces with
increasing angle of incidence. This is just another way of representing the angle
of incidence dependence of the depolarised eigenvalues in Figure 9.19. We see
that at small angles of incidence the entropy is over 0.6, while at larger angles
it falls to around 0.1. However, by comparing Figures 9.19 and 9.20 we realize
368 Applications of polarimetry and interferometry

Scattering entropy at 10 GHz (bs = 5.236)

1
0.8
0.6 *

Entropy
0.4
* *
0.2 *
*
0
10 15 20 25 30 35 40 45 50
Angle of incidence

Alpha angle at 10 GHz (bs = 5.236)

15

Alpha angle (degrees)

10

5
*
* * *
Fig. 9.20 Entropy (upper) and alpha angle 0 *
(lower) values for rough surface scattering as 10 15 20 25 30 35 40 45 50
a function of angle of incidence Angle of incidence

that an entropy of 0.6 still represents a strongly polarised signal. In the lower
portion of Figure 9.20 we see the corresponding alpha angle. If the eigenvector
were truly of the form (1,0,0), this should be zero. We see that experimentally
it lies around 3◦ .
In conclusion, we have seen that rough surface backscattering represents
only a weak depolariser, with an isotropic noise-like depolarisation subspace
and a dominant eigenvector with scattering entropies below 0.6. However, the
polarised eigenvector itself seems rather trivial: just the coherent sum of the HH
and VV channels. The question is, do we ever obtain more interesting variation
of eigenvectors, allowing us to use variation of the polarised ratios for parameter
estimation? To answer this we turn attention to the smooth surface scattering
behaviour.
The smooth surface is characterized by a smaller rms roughness s (although
the same correlation length), and hence at a given frequency the product βs
will be smaller and the surface electrically smoother. With this in mind we
show, in Figure 9.21, the variation of linear basis backscatter as a function of
frequency (at a 30-degree angle of incidence). Here we see two features of
interest. The first is a lower backscatter level compared to the rough surface,
with the crosspolarised response now around 20 dB below the copolarised.
The second feature of interest is the separation of copolar coefficients at low
frequencies. Here we see that VV is a few dB above HH, in qualitative agreement
at least with the predictions of the small perturbation or Bragg scattering model.
Figure 9.22 shows the corresponding eigenvalue spectra of the coherency
matrix. Again we note the small fourth eigenvalue, as expected from reci-
procity, and also the presence again of a dominant eigenvalue, showing that
again the backscatter is strongly polarised. However, now we see that the depo-
larised subspace is anisotropic; that is, that the second and third eigenvalues
are not equal, demonstrating that the depolarisation caused by the surface is
 9.5 Applications of polarimetry and interferometry 369

Backscatter (HH, HV, VH, VV) at 30 degrees

10

HH
0 HV
VH
VV
–10
Scattering coefficient (dB)

–20

–30

–40

–50

–60
2 3 4 5 6 7 8 9 10
Fig. 9.21 Smooth surface backscatter in lin-
Frequency (GHz) ear basis as a function of frequency

Backscatter eigenvalue spectrum at 30 degrees

0

lambda 1
–10 lambda 2
lambda 3
lambda 4
Scatterinng coefficient (dB)

–20

–30

–40

–50

–60
2 3 4 5 6 7 8 9 10
Fig. 9.22 Coherency eigenvalue spectra for
Frequency (GHz) smooth surface scattering

now polarisation dependent. Again we can expose the polarised nature of the
scattering by normalizing the eigenvalue spectra to have unit sum. Figure 9.23
shows the results as a function of angle of incidence (for a fixed low frequency
of 2 GHz). Here we see a strongly polarised response, with the depolarised
power making up less than a few percent of the scattered signal. The aver-
age phase angle between HH and VV is small—typically a few degrees, as
shown in Figure 9.24—again agreeing qualitatively with predictions from the
Bragg surface scattering model. However, the main new feature of interest is
the change in the eigenvector parameters with incidence angle. Figure 9.25
shows how the amplitude of the Pauli components of the dominant eigenvector
370 Applications of polarimetry and interferometry

Normalized eigenvalues at 2 GHz (bs = 0.16755)

100

10–1

10–2

10–3
Fig. 9.23 Variation of normalized eigenvalue 10 15 20 25 30 35 40
spectra with angle of incidence for smooth
surface scattering Angle of incidence

HH/VV Phase
20

15

10

5
Phase (degrees)

-5

-10

-15

-20
Fig. 9.24 HH/VV phase for the smooth sur- 2 3 4 5 6 7 8 9 10
face at 30-degree angle of incidence Frequency [GHz]

vary across the spectrum. Here we see that while the first component is still
dominant, the second Pauli component is non-zero and increases with angle of
incidence. This is reflected in the corresponding entropy/alpha variation shown
in Figure 9.26. Here we see low scattering entropy at all angles, but combined
with an alpha parameter that steadily increases with angle of incidence. This
variation is directly related to the dielectric constant of the surface, as we now
demonstrate.
We begin by noting that the simple first-order Bragg surface scattering model
can be used to estimate an alpha parameter, which is a function of the ratio of sum
 9.5 Applications of polarimetry and interferometry 371

Principal eigenvector components

0.8

0.6

0.4

0.2

Fig. 9.25 Variation of the amplitude of the

0 Pauli components of the dominant eigenvec-
10 15 20 25 30 35 40
tor of [T ] as a function of angle of incidence
Angle of incidence for the smooth surface

Scattering entropy at 2 GHz (bs = 0.16755)

0.5
0.4
0.3

0.2

0.1 * *
* * *
*
0*
10 15 20 25 30 35 40
Angle of incidence

Alpha angle at 2 GHz (bs = 0.16755)

20

15
*
10 *
*
5 *
* *
*
0 Fig. 9.26 Entropy (upper) and alpha angle
10 15 20 25 30 35 40
(lower) values at 2 GHz for smooth surface
Angle of incidence scattering as a function of angle of incidence

and difference of copolarised scattering coefficients, and hence for a given θ

depends only on the dielectric constant and not surface roughness (see Section
3.1.3). We can therefore use the Bragg model with the measured dielectric
constant of the surface to predict an alpha angle variation, and compare this
directly with the estimates obtained from the coherency matrix. Before doing
this, however, we note that the simple Bragg model ignores any influence
of wave depolarisation, which while weak, still occurs even for this smooth
372 Applications of polarimetry and interferometry

Alpha angle at 2 GHz (bs = 0.16722)

20

18

16

14

Alpha angle (degrees)

12

10

Fig. 9.27 Alpha angle variation at 2 GHz for

4
smooth surface scattering as a function of
angle of incidence. Solid line is the eigenvec- 2
tor estimate, dotted line is the prediction of
the Bragg scattering model, and dashed line 0
10 15 20 25 30 35 40
is the corrected alpha value using the X-Bragg
scattering model of depolarisation Angle of incidence

surface backscatter. We saw in Section 3.2 how the extended or X-Bragg model
provides a method for parameterising the effects of depolarisation on the Bragg
model.
According to this approach (see equation (3.40)), the alpha parameter must be
corrected for depolarisation by estimating it—not directly from one eigenvector,
but from a ratio of diagonal terms of the full coherency matrix, as shown in
equation (9.40):
 
t22 + t33
αb = tan−1 (9.40)
t11

Figure 9.27 shows how these three estimates compare for low-frequency (2-
GHz scattering). In the solid line we again show the eigenvector estimates from
the lower part of Figure 9.26. We show too the reference Bragg values obtained
from the dielectric constant. Finally, in dash we show the corrected estimated
alpha values using the X-Bragg model. We see that the correction is of the
order of 2◦ , and for angles of incidence greater than 20◦ the correction brings
the estimates into close agreement with the Bragg predictions. However, for
angles less than 20◦ , both methods seem to overestimate alpha. This can be
traced to the small separation of HH and VV scattering coefficients for angles
near normal incidence, which makes estimation of small alpha values from
experimental data more difficult.
The importance of this depolarisation correction can be made even more
apparent by considering a higher frequency. At 10 GHz, for example, the sur-
face roughness leads to βs = 0.84, which is well outside the usual bounds for
validity of the simple Bragg model. Figure 9.28 shows the various alpha esti-
mates for this high-frequency case. Here we see that the maximum eigenvector
dramatically underestimates the alpha parameter, especially at high incidence
angles, whereas when we add the correction for depolarisation (dashed line)
 9.5 Applications of polarimetry and interferometry 373

Alpha angle at 10 GHz (bs = 0.16722)

20

18

16

14
Alpha angle (degrees)

12

10

4 Fig. 9.28 Alpha angle variation at 10 GHz

for smooth surface scattering as a function of
2 angle of incidence. Solid line is the eigenvec-
tor estimate, dotted line is the prediction of
0 the Bragg scattering model, and dashed line
10 15 20 25 30 35 40
is the corrected alpha value using the X-Bragg
Angle of incidence scattering model of depolarisation

we see much better agreement. This result can, for example, be used to esti-
mate surface dielectric constant by first using the X-Bragg model to estimate
alpha corrected for surface depolarisation. From alpha, and knowing the angle
of incidence, we can then obtain a direct estimate of the dielectric constant by
using the standard Bragg relations.
In this section we have seen that surface backscattering provides a polarised
return over a wide range of roughness scales and angles of incidence. We now
turn to consider the generalization of these ideas to bistatic surface scattering—
the situation in which transmitter and receiver are separated.

Bistatic surface scattering

The geometry we now consider is shown in Figure 9.29. The rough surface
is illuminated by the transmitter at a fixed angle of –40◦ . The receiver is then
moved in 10-degree steps from near backscatter to beyond specular reflection
(at +40◦ ). In this case we can no longer assume that HV = VH by reciprocity,
and hence must employ a full 4 × 4 coherency matrix formulation. It is one
of our objectives here to see how important this new fourth eigenvalue and
associated eigenvector is for bistatic surface scattering. Figure 9.30 shows an
example: the normalized eigenvalue spectra variation with scattering angle for
the rough surface at 10 GHz (βs = 5). Here we see that there is again a dominant
eigenvalue and hence a strongly polarised scattering component for all angles.
The second and third eigenvalues are again less than 10% of the signal, while
the new fourth eigenvector, although larger than in the backscatter case, remains
two orders of magnitude below the main polarised response. Note the dip in
the level of depolarisation around the specular direction. This is expected, as
for this scattering angle the return is more coherent due to specular reflection
at the rough surface. However, we see that even away from these angles the
surface return remains strongly polarised.
374 Applications of polarimetry and interferometry

EMSL chamber geometry

10

Metres
0 + + +
+ + +
–2

–4

–6

–8

Fig. 9.29 Geometry of bistatic surface scat- –10

–10 –5 0 5 10
tering measurements in the EMSL chamber
(* = TX, and O = RX) Metres

Eigenvalues of coherency matrix

100

10–1
Eigenvalue level

10–2

10–3
Fig. 9.30 Normalized eigenvalue spectra for –30 –20 –10 0 10 20 30 40 50
bistatic scattering from rough surface at
10 GHz Scattering angle

The Pauli components of the dominant eigenvector are shown in Figure 9.31.
For angles close to backscatter we again see a dominant first component, corre-
sponding to alpha around zero. However, as the scattering angle increases so the
second Pauli component becomes more important. We note that the third and
fourth remain equal and close to zero over the full range of angles. Figure 9.32
shows the corresponding entropy and alpha variations. Note that the entropy
is now defined in base 4, to account for the fourth eigenvalue. We see a dip in
 9.5 Applications of polarimetry and interferometry 375

Principal eigenvector components

1
Eigen vector component level

0.8

0.6

0.4

0.2

0 Fig. 9.31 Pauli components of dominant

–30 –20 –10 0 10 20 30 40 50
eigenvector of bistatic scattering coherency
Scattering angle matrix for rough surface at 10 GHz

Bistatic entropy
0.5
0.4
Entropy

0.3
0.2
0.1
0
–30 –20 –10 0 10 20 30 40 50
Scattering angle

Principal eigenvector of T
Alpha angle (degrees)

40

30

20

10

0
–30 –20 –10 0 10 20 30 40 50
Fig. 9.32 Bistatic entropy (upper) and alpha
Scattering angle angle (lower) for bistatic

the entropy for specular scattering. The alpha angle is taken from the dominant
eigenvector, and shows a steady increase with scattering angle, with a trend
similar to that found in Figure 9.28. However, in this case the physical origin
of this variation is quite different. Here we must use the bistatic geometry and
dielectric constant in the BRDF model of Section 3.2.2 (Priest, 2000). This
model can be used to predict the variation of alpha for the bistatic scattering
matrix. In Figure 9.33 we show the amplitude of the Pauli components of this
matrix as a function of scattering angle. Superimposed, we again show the Pauli
376 Applications of polarimetry and interferometry

Principal eigenvector components

Eigen vector component level

0.8

0.6

0.4

0.2

Fig. 9.33 Variation of Pauli coefficients for

the BRDF model for the rough surface (solid 0
–30 –20 –10 0 10 20 30 40 50
lines) versus the variation of the dominant
eigenvector components from Figure 9.31 Scattering angle

components for the dominant eigenvector. We see that the simple BRDF model
is quite good at explaining the trend of the alpha variation.
In conclusion, we have seen from our analysis of rough and smooth surface
scattering data that bare surface scattering is characterized by a strong polarised
return (the entropy is seldom above 0.5), maintained over a wide range of
angles of incidence and frequencies. The corresponding dominant eigenvector
is characterized by an alpha parameter in the range 0 ≤ α ≤ π/4, and which
shows dependence on local scattering geometry (angle of incidence and bistatic
scattering angle) and surface dielectric constant. We now turn to consider a
different class of interactions: volume scattering, where scattering entropies
and hence depolarisation can be much higher.

9.5.2 Application 2: depolarisation by volume scattering

Depolarisation by volumes can be caused by two basic physical processes:
particle anisotropy (in shape or material composition), or multiple scattering
between particles. We begin by looking at the latter and by considering a vol-
ume made up of wavelength-sized spherical particles. Spheres have a strong
symmetry, which means that in single scattering they have zero depolarisation
and hence zero scattering entropy. Regardless of size and dielectric constant,
in backscatter they always yield a scattering matrix proportional to the iden-
tity (in the BSA coordinate system). However, when several such particles
are brought together in a volume their mutual interactions destroy this simple
picture and lead to depolarisation. To illustrate this phenomenon we consider
scattering by a random cloud of dielectric spheres. We make use of some recent
‘exact’ calculations of scattering by a cloud of particles using the superposition
T-matrix method (Mishchenko, 2000). These simulations were provided cour-
tesy of Michael Mishchenko and his group at the NASA Goddard Institute in
New York.
 9.5 Applications of polarimetry and interferometry 377

The three-dimensional simulations are based on numerical solution of

Maxwell’s equations for the cloud of particles considered as a whole
(Mishchenko, 2007). In this regard there are no approximations involved, and
the technique can be considered ‘exact’ and to include all effects of multiple
and single scattering such as coherent speckle, wave depolarisation, enhanced
backscattering, and so on (van Albada, 1988; Mishchenko, 1992; Macintosh,
1999). The output from the simulator is all sixteen elements of the Stokes phase
matrix (which is related to the Mueller matrix (Hovenier, 2004) in the FSA or
wave-based coordinate system. These can then be converted in a 1–1 mapping
into the elements of the 4 × 4 Hermitian coherency matrix (see Section 2.3),
which can then be expressed in terms of its eigenvalue decomposition. These
matrix elements can be determined in the simulator for arbitrary incident and
scattered wave directions. From the symmetry of the problem it is sufficient for
us to consider some fixed but arbitrary incident vector and variation of the scat-
tering angle ψ in a plane formed by the incident and scattered vectors, varying
from 0◦ (forward scatter) to 180◦ (backscatter).
As an example we choose to consider scattering by lossless particles with
a low dielectric constant, so that the refractive index n = 1.32 (εr = 1.74) R
(representing the properties of water and ice particles at visible wavelengths)
and size βr = 4, where β is the wavenumber in the surrounding medium and r is
the particle radius. The N particles, where N = 1…240, are randomly placed in a
spherical volume of size βR = 40 (see Figure 9.34), corresponding to variations
from 0.1%–24% in particle concentrations.
In this way the effect of multiple scattering can be investigated by looking at
the transition from single to multiple particle configurations. By symmetry, the
Mueller and coherency matrices for arbitrary scattering angle must then have 2r
the form shown in equation (9.41) (see equation 2.108):

    Fig. 9.34 Geometry for calculation of scatter-

m11 (ψ) m12 (ψ) 0 0 t11 t12 0 0 ing by random cloud of spheres
m12 (ψ) m22 (ψ) 0 0  t ∗ t22 0 0
[M ] = 
 0
 ⇔ [T ] =  12 
0 m33 (ψ) m34 (ψ) 0 0 t33 0
0 0 −m34 (ψ) m44 (ψ) 0 0 0 t44
1
t11 = (m11 + m22 + m33 + m44 )
2
t12 = m12 − im34 (9.41)
1
t22 = (m11 + m22 − m33 − m44 )
2
1
t33 = (m11 − m22 + m33 − m44 )
2
1
t44 = (m11 − m22 − m33 + m44 )
2

Hence we can in general have four non-zero eigenvalues, but the eigenvectors
are limited in their structure to C2—a two-dimensional complex subspace.
Hence we can use the scattering sphere concept (see Section 2.4.3.2) to represent
variations in the eigenvector information. We choose to consider three special
cases: a single particle for reference, a low concentration of 0.5%, and a high
concentration of 16%. The first parameter of interest is the phase function. This
378 Applications of polarimetry and interferometry

Phase function
103

1 particle
5 particle
102 160 particle

101

M11
100

10–1

Fig. 9.35 Phase function for three special 10–2

0 20 40 60 80 100 120 140 160 180
cases: single particle, five particles, and 160
particles Scattering angle

is just the m11 element of the Stokes scattering matrix, normalized so that
π
1
m11 (ψ) sin ψd ψ = 1 (9.42)
2
0

Equivalently, m11 represents one half the trace of the coherency matrix (or sum
of the eigenvalues of [T ]), which physically represents the total scattered power
of the signal. Figure 9.35 shows how this function varies for three different par-
ticle concentrations (scattering angle = 0 is forward scatter in this notation). We
immediately see that the particles scatter more in the forward than the back-
ward direction (a typical feature of wavelength size particles), and that as more
particles are added there is coherent addition in the forward direction, and also
the appearance of a small enhanced backscatter peak in the backscatter direc-
tion (scattering angle equal to 180◦ ). We see that the phase function becomes
smoother, compared to the single particle case, as we add more particles to the
volume. This is the effect of multiple scattering.
We are now particularly interested in the depolarisation properties of this
process. The single particle case is trivial to consider, as it yields a single non-
zero eigenvalue for all scattering angles—the eigenvector corresponding to
which is obtained from the scattering matrix [S] for the particle (and which can
be obtained exactly using standard Mie scattering theory; see Section 3.3.2). Of
more interest are the low and high concentration results. Figures 9.36 and 9.37
show the normalized coherency spectra as a function of scattering angle for
these two cases.
In the low concentration example we see a dominant eigenvalue, with the
secondary eigenvalues less than 1% of the scattered power, except for a couple
of peaks around 60◦ and 100◦ . In forward scatter the entropy is zero (a coherent
wave propagating through the volume), and in backscatter one of the minor
eigenvalues goes to zero (because of reciprocity), while the other two are equal
 9.5 Applications of polarimetry and interferometry 379

Coherency eigenvalues for 0.5% by volume

100

lambda 1
lambda 2
lambda 3
10–1 lambda 4
Relative amplitude

10–2

10–3

10–4
0 20 40 60 80 100 120 140 160 180
Fig. 9.36 Normalized coherency spectra for
Scattering angle low concentration of particles

Coherency eigenvalues for 16% by volume

100

10–1
Relative amplitude

10–2

lambda 1
lambda 2
lambda 3
10–3 lambda 4

10–4
0 20 40 60 80 100 120 140 160 180
Fig. 9.37 Normalized coherency spectra for
Scattering angle high concentration of particles

and lie around the 1% mark. The denser concentration starts to lose structure,
as multiple scattering dominates. Here we see a smoother variation of eigen-
values, still with zero entropy in forward scatter, but much higher entropy in
the backscatter hemisphere. Note again that for exact backscatter one of the
eigenvalues still falls to zero as a result of reciprocity. These results can be
expressed in terms of the scattering entropy, as shown in Figure 9.38. Here we
show the variation of entropy for the three cases. The single particle case yields
zero entropy for all scattering angles, while the denser concentration leads
to more depolarisation and increased entropy, especially in the backscattering
hemisphere.
380 Applications of polarimetry and interferometry

Bistatic entropy

0.9

0.8

0.7

0.6
160 particle

Entropy
0.5

0.4

0.3

0.2 5 particle
0.1

0 1 particle
–0.1
0 20 40 60 80 100 120 140 160 180
Fig. 9.38 Scattering entropy of a cloud of
spheres versus scattering angle for three par-
ticle concentrations Scattering angle

We see that the entropy peaks around 0.8 for the denser concentration but
then reduces slightly for backscatter. However, if we ignore the fourth eigen-
value in the backscatter direction (because reciprocity, rather than scattering
symmetries, forces it to zero) and calculate the monostatic entropy, we find that
the backscatter entropy rises from 0.75 to 0.94. This is close to the maximum
entropy obtained from single scattering by a cloud of prolate particle shapes,
even though we are dealing with a cloud of spheres. In this way we see that
multiple scattering can be an important source of depolarisation.
Turning now to the dominant eigenvector, we start by noting that it has only
two non-zero elements: the first and second Pauli components. Figure 9.39
shows how these two vary in amplitude for each of the three cases considered:
single particle, five particles, and 160 particles. We see that in forward scattering
we have the identity matrix (first Pauli component) as consistent with forward
scattering in the FSA system. In backscatter, by contrast, we have the second
Pauli component with the first falling to zero. This corresponds to a backscat-
tering matrix with equal amplitudes but 180-degree phase shift between HH
and VV, as expected in the FSA system. Between these two extremes we see
that the eigenvectors change significantly. Indeed, these changes not only occur
in amplitude but in phase. As the eigenvector is an element of C2—the two-
dimensional complex space—we can map it in both amplitude and phase on the
surface of the scattering sphere. This allows us to visualize changes in the com-
plex nature of the eigenvector between forward and back scattering. Figure 9.40
shows the dominant eigenvector variation on the scattering sphere for the three
cases considered. In forward scattering they all begin on the equator at zero
phase. As scattering angle changes we then see combined amplitude and phase
variations trace a complicated loci across the sphere, ending for backscatter
on the equator again at 180◦ . For the single-particle case (the black line) these
are just the polarimetric phase and amplitude variations due to Mie scattering
by a lossless dielectric sphere. In the low concentration case we see that the
 9.5 Applications of polarimetry and interferometry 381

Dominant eigenvector:Pauli 1 Dominant eigenvector:Pauli 2

1 1

0.9 0.9

0.8 0.8

0.7 0.7
Component magnitude

Component magnitude
0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 1 patticle 0.2

5 particles
0.1 0.1
160 particles
0 0 Fig. 9.39 Variation of dominant eigenvector
0 50 100 150 0 50 100 150 200
Pauli components with scattering angle for 1,
Scattering angle Scattering angle 5, and 160 particle clouds

Alpha sphere for dominant eigenvector

90 1
120 60
0.8 N=1

0.6 N=5
150 30
0.4
N = 160
0.2

180 0

210 330

Fig. 9.40 Alpha sphere representation of the

240 300 eigenvector amplitude and phase informa-
tion for N = 1 particle, N = 5 particles, and
270 N = 160 particles

dominant eigenvector closely follows this loci, maintaining information about

the particle. In the high concentration case, however, we see that the dominant
eigenvector is smoothed out by the multiple scattering effects.
Another way to present these ideas is to use the alpha angle. The alpha
variation corresponding to the eigenvector fluctuations in Figure 9.39 are shown
in Figure 9.41. Here we see that for forward scatter α = 0 (in the FSA coordinate
system), and for backscatter this rises to π/2. The single particle and low
382 Applications of polarimetry and interferometry

Dominant eigenvector alpha

90
1 particle
80
5 particle
70 160 particles

Alpha angle (degrees)

60

50

40

30

20

10

Fig. 9.41 Alpha angle variation of dominant 0

eigenvector as a function of scattering angle 0 20 40 60 80 100 120 140 160 180
for 1, 5, and 160 particle clouds Scattering angle

Front view Antennas

m
56
Vegetation 9.
0=
sample
R


EMSL focal point

38 cm Y

Turntable

Fig. 9.42 Schematic of the maize scattering

measurements in the EMSL Ground

concentration cases move between these two boundaries in a series of steps,

while the high concentration smoothes out these details.
In this example we have seen how multiple scattering can lead to high
entropies, even for spherical particles. In the next we give a second exam-
ple of volume scattering—this time from EMSL chamber measurements of
a vegetated surface, where both multiple scattering and anisotropic particle
shape contribute to wave depolarisation. We will then use this example of a
complex depolarising problem as a test case for the application of polarimetric
interferometry and polarisation coherence tomography techniques.
 9.5 Applications of polarimetry and interferometry 383

Fig. 9.43 Image of maize sample used for

EMSL measurements

9.5.3 Application 3: coherent scattering from vegetation

In this section we consider analysis of a measurement dataset that combines
surface and anisotropic volume scattering. The backscatter measurement geom-
etry inside the EMSL anechoic chamber is shown in Figure 9.42. The target, in
this case, is a sample of 6 × 6 maize plants of 1.8 m height, uniformly planted
in a square container of sides 2 m (http://www-emsl.jrc.it/EMSLdata/). Sep-
aration between plants is around 30–35 cm. The plants are characterized by
green vertical stems of around 4 cm diameter—each carrying wide leaves from
a height of 40 cm to the top. The leaves themselves are 30–40 cm long and 7–8
cm wide. The leaves are oriented in a range of angles from 20–45◦ , as shown
in Figure 9.43.
The vegetation sample is placed on a rotating turntable so that measurements
can be made over 360◦ of azimuth for a given angle of incidence. The antenna
beamwidth is such that the sample is uniformly illuminated by the transmitter at
all times. In this experiment there were 72 azimuth steps of 5◦ . At each position
the frequency is stepped across the frequency range 1.5–9.5 GHz (in 10-MHz
steps), and the elevation angle is incremented in 0.25-degree steps from 44◦ to
45◦ . The complete scattering matrix [S] is measured at each frequency across
the full band. In this way, multi-baseline polarimetric interferometric analysis,
and even coherence tomography, can be performed, with a minimum angular
baseline of θ = 0.25◦ and a maximum of 5◦ . Note, finally, that the focus for the
chamber (zero phase position for interferometry) is located around 38 cm above
the soil surface of the sample. Hence the interferometric phase of the true surface
position φ o will not be zero, and will change with frequency and baseline. This
dataset provides an interesting testbed for polarimetric, interferometric, and
tomographic processing (Sagues, 2000; Lopez-Sanchez, 2006, 2007; Cloude,
2007a).

Depolarisation in vegetation scattering

We begin with an assessment of the level of depolarisation caused by this scat-
tering environment. Figure 9.44 shows the variation of normalized eigenvalues
with frequency. We note that the maximum eigenvalue is now much reduced
384 Applications of polarimetry and interferometry

Normalized eigenvalues
1

0.9

0.8

Relative eigenvalue amplitude

0.7

0.6
lambda 3
0.5 lambda 2
lambda 1
0.4

0.3

0.2

0.1

0
2 3 4 5 6 7 8 9
Fig. 9.44 Variation of normalized eigenvalue
spectra with frequency for the maize sample Frequency (GHz)

Principal eigenvector components

1
Pauli components of eigenvector

Pauli 1
0.8
Pauli 2
Pauli 3
0.6

0.4

0.2

0
2 3 4 5 6 7 8 9
Fig. 9.45 Variation of Pauli components of
the maximum eigenvector (maize sample) Frequency (GHz)

to around 60–70% of the signal energy. The depolarisation is quite large across
the spectrum, and shows some anisotropy at low frequencies, which reduces as
the frequency increases.
Note that here we have increased the averaging window to 320 MHz (in
addition to averaging over the 72 azimuth samples) in order to reduce the speckle
noise on the estimates. The Pauli components of the maximum eigenvector vary
with frequency, as shown in Figure 9.45. Here we see that at low frequencies
the second Pauli component is significant, but that as frequency increases the
 9.5 Applications of polarimetry and interferometry 385

Scattering entropy
1

0.9
Entropy

0.8

0.7

0.6
2 3 4 5 6 7 8 9
Frequency (GHz)

Alpha parameter
40
Alpha (degrees)

30

20

10

0
2 3 4 5 6 7 8 9 Fig. 9.46 Entropy (upper) and alpha (lower)
Frequency (GHz) variation with frequency for maize sample

Entropy/Alpha Diagram
90

80

70

60

50 Ap = 20dB
Alpha

40
Ap = 6dB
30

20

10

0 Fig. 9.47 Entropy/alpha plane representa-

0 0.2 0.4 0.6 0.8 1
tion of maize data points superimposed on
Entropy prolate and oblate H /α loci (see Figure 3.29)

eigenvector tends to the first Pauli component. These results all indicate a high
level of wave depolarisation by the vegetation sample. This is confirmed in
Figure 9.46, which shows the entropy/alpha variations corresponding to Figures
9.44 and 9.45. Here we see high entropy, around 0.8 at low frequency and falling
slightly to 0.7 at higher frequencies. The dominant eigenvector alpha falls from
around 20◦ at low frequencies to nearly zero at high frequencies. However,
386 Applications of polarimetry and interferometry

Entropy/Alpha diagram

42
Ap = 10dB
40

38

36
Ap = 8dB
34

Alpha
32

30

28

26

24

Fig. 9.48 Zoom on entropy/alpha plane rep-

0.65 0.7 0.75 0.8 0.85
resentation of maize data points superimposed
on prolate H /α loci Entropy

given the relatively high entropy we can expect the maximum eigenvector to
be corrupted by depolarisation (as we saw in the cloud of spheres in Figure 9.38,
for example). Hence we again need some way of compensating the eigenvector
information for the depolarisation that is occurring. One way to do this is
to employ the average alpha, formed by the sum of the alpha values for each
eigenvector weighted by the normalized eigenvalues (see Section 2.4.2.4). This
leads to a representation on the entropy/alpha plane as shown in Figure 9.47,
with a zoom on the data points shown in Figure 9.48. Also superimposed on
these diagrams are the entropy/alpha loci for a cloud of anisotropic particles,
as developed in Section 3.4.1. We show results for prolate particle clouds of
varying orientation structure, and the same for oblate particle clouds. One way
to interpret the maize data points is in terms of the response for an equivalent
cloud of such particles. In this regard we see that at low frequencies the response
is equivalent to a cloud of strongly prolate particles (with a shape anisotropy
factor Ap around 10 dB), but with a non-random orientation distribution; while
as frequency increases two things happen: initially (up to 6 GHz) the cloud
becomes more random, while maintaining the same particle anisotropy. From 6–
9 GHz, however, the behaviour changes, the effective particle shape anisotropy
(Ap ) reduces (which causes a drop in the entropy), and the distribution stays
random (the lower curve of the entropy/alpha diagram represents azimuthal
symmetry in the volume).
However, this interpretation—in terms of the response of an equivalent cloud
of anisotropic particles—involves many simplifying assumptions, one of the
most important being that we see only volume scattering from the maize and
that the underlying surface plays no part in the scattered return. This is not
necessarily a good assumption, especially at low frequencies. We can see,
for example, that there is a change in scattering behaviour around 6 GHz, and
cannot dismiss the idea that for frequencies lower than this the surface response
is playing an important role.
 9.5 Applications of polarimetry and interferometry 387

To resolve these issues we could try using one of the many model-based
decomposition methods based on mixed surface and volume scattering, but
these require us to make even more stringent assumptions about the nature of
the volume scattering. (The Freeman–Durden, for example, assumes a cloud
of dipole particles.) These assumptions are not supported for this example,
and so would not necessarily help us resolve issues about the true ratio of
surface-to-volume scattering. Instead we choose to look at the role that radar
interferometry can play in helping resolve this by adding additional information
about the structure of such complex scattering problems.

9.5.4 Application 4: tomographic imaging of

vegetated surfaces
Interferometric processing of wide-band signals from the EMSL chamber starts
with the two calibrated complex signals s1,2 from angular positions θ and θ +θ
at frequency f and with transmit/receive polarisation ‘pq’, as shown in equation
(9.43):

pq pq
s1 = Eθ (f ) s2 = Eθ+θ (f ) (9.43)

The wide-band interferogram is then formed from the product of common spec-
tral band filtered and phase offset signals, as shown in equation (9.44), where
the phase offset is required because the chamber focus lies below the surface
of the sample container:

  
f θ 4π f θ
s1 .s2∗ e−i c sin θ zo
pq pq
= Eθ (f ) · conj Eθ+θ f − (9.44)
tan θ

Finally, the complex coherence for polarisation combination ‘pq’and frequency

f is calculated as shown in equation (9.45):

+ ∗,
s1 s2
γ̃pq (f ) = + ,+ , (9.45)
s1 s1∗ s2 s2∗

Figure 9.49 shows how this coherence amplitude and phase vary for three
polarisations as a function of frequency for the smallest baseline (0.25◦ ). We
see that the coherence falls as frequency increases. This is caused by volume
388 Applications of polarimetry and interferometry

decorrelation, as discussed in Section 5.2.4. To confirm this, in dash we show the

coherence values expected from a simple uniform structure function (the SINC
model). We also show coherence results for three polarisation channels, HH,
HV and VV. We note that the SINC gives the same trend as seen in the data (in
both amplitude and phase); but there are some important differences, relating,
as we shall see, to departures of the vertical structure function from uniform. To
further emphasize these differences we show, in Figure 9.50, coherence results
for a larger baseline (0.5◦ ). We note the following features:
1. The phase centre generally lies below the SINC level (which is set
at half the top height of the vegetation). This implies that the phase
centre is closer to the surface of the sample than expected from volume-
only scattering models, which again implies the presence of a mixed
surface/volume scattering scenario.
2. There is separation of the polarisation channels, with VV closer to the
SINC phase level (higher in the volume) than HH. However, the phase
separation between polarisations is not maximized by this special selec-
tion of states. To see how a larger separation can be achieved, we
show, in Figure 9.51, the phase and amplitude variation for the 0.5-
degree baseline but for two of the Pauli channels. In particular we again
show the crosspolar HV channel, but now compare it with the coherent
combination of a difference of copolar channels HH–VV. Here we see
much better phase separation across the band. This fits a physical inter-
pretation based on a mixed surface/volume scenario with the effective
surface component, in fact caused by double bounce or dihedral scat-
tering, which introduces a 180-degree polarimetric phase shift between
copolar channels.

Interferometric coherence amplitude (0.1 degree baseline)

1

0.8
Coherence

0.6
VV
0.4
VH
0.2 SINC

2 3 4 5 6 7 8 9
Frequency (GHz)

Interferometric phase (0.1 degree baseline)

3
2.5
Phase (rads)

2
1.5
1
0.5
Fig. 9.49 Variation of interferometric coher-
ence amplitude (upper) and phase (lower) for 0
2 3 4 5 6 7 8 9
0.25-degree baseline and for polarisations VV
and VH Frequency (GHz)
 9.5 Applications of polarimetry and interferometry 389

Interferometric coherence amplitude (0.2 degree baseline)

1
VV
0.8
VH
Coherence

0.6 SINC

0.4

0.2

0
2 3 4 5 6 7 8 9
Frequency (GHz)

Interferometric phase (0.2 degree baseline)

3
2
Phase (rads)

1
0
–1
–2
–3 Fig. 9.50 Variation of interferometric coher-
2 3 4 5 6 7 8 9 ence amplitude (upper) and phase (lower) for
0.5-degree baseline, and for polarisations VV
Frequency (GHz) and VH

To confirm this we can use these complex coherences to estimate the vertical
structure function using coherence tomography (Cloude, 2007a; Zhou, 2006).
In this case we know from the chamber geometry the true phase of the sur-
face position and the height of the vegetation, and so can use these directly to
reconstruct the profile using the baseline geometry. Figure 9.52 shows how the
interferometric wavenumber β z varies across the spectrum for the two baselines.
This can then be used with the known vegetation height (1.8 m) to calculate
the normalization parameter kv , as shown in Figure 9.53. We see that it varies
from around 0.4 for the small baseline at low frequencies to more than 4.0 for
the larger baseline at high frequency. These values provide good sensitivity to
changes in structure function in the vegetation layer (see Section 8.4). These
can then be used with the known surface phase to reconstruct a truncated Leg-
endre series expansion of the structure function. With one baseline we obtain
a second-order expansion, while by combining data from two baselines we
obtain a fourth-order reconstruction. Figures 9.54 and 9.55 show the recon-
structed vertical profiles through the vegetation as a function of frequency for
the single and dual baseline data. We note the following important features:

1. The dual baseline data has higher vertical resolution than the single.
This is due to the presence of higher-order terms in the Legendre series
reconstruction.
2. Both single and dual baseline datasets show a strong surface component
at low frequencies (below 6 GHz), with a more dominant volume con-
tribution for frequencies above this. However, the volume component
is not a simple exponential function. It shows a small response from the
top with a peak near the centre of the layer. This therefore represents a
390 Applications of polarimetry and interferometry

Interferometric coherence amplitude (0.2 degree baseline)

1
HH-VV
0.8
HV

Coherence
0.6 SINC

0.4

0.2

0
2 3 4 5 6 7 8 9
Frequency (GHz)

Interferometric phase (0.2 degree baseline)

3
2
Phase (rads)

1
0
–1
–2
Fig. 9.51 Variation of interferometric coher- –3
ence amplitude (upper) and phase (lower) for 2 3 4 5 6 7 8 9
0.5-degree baseline, and for two polarisations
HH-VV and HV Frequency (GHz)

Interferometric wavenumber
5

4.5
0.1 degree baseline
4 0.2 degree baseline

3.5
Vertical wavenumber

2.5

1.5

0.5

Fig. 9.52 Interferometric wavenumber as a 0

1 2 3 4 5 6 7 8 9 10
function of frequency for the two baselines
used in the EMSL chamber Frequency (GHz)

simple example where the RVOG model assumptions do not seem to be

valid (see Section 7.4).
3. We note important differences between polarisations. In particular we
see that VV has a much lower surface response at higher frequencies
than HH. The idea that this arises from a double bounce contribution
can be confirmed by imaging in the HH–VV Pauli channel, as shown in
 9.5 Applications of polarimetry and interferometry 391

Wavenumber/height product
4.5

4
0.1 degree baseline
3.5 0.2 degree baseline

2.5
Kv

1.5

0.5
Fig. 9.53 Normalized wavenumber/height
0 product factor kv as a function of frequency
1 2 3 4 5 6 7 8 9 10
for the two baselines used in the EMSL
Frequency (GHz) chamber

HH Channel
2 1
Height (m)

1 0.5

0 0
2 3 4 5 6 7 8 9
Frequency (GHz)
HV Channel
2 1
Height (m)

1 0.5

0 0
2 3 4 5 6 7 8 9
Frequency (GHz)
VV Channel
2 1
Height (m)

1 0.5
Fig. 9.54 Reconstructed single-baseline ver-
0 0 tical profiles for the 1.8-m vegetation layer
2 3 4 5 6 7 8 9
in HH (upper), HV (centre), and VV (lower)
Frequency (GHz) polarisations

Figure 9.56 for the single (upper) and dual (lower) baseline data. Here we
see a strong surface component across most of the spectrum, confirming
that the structure function in this Pauli channel remains localized around
the surface.
In conclusion, we have seen that the presence of vegetation on a surface causes
complexity in backscatter, with an increase in the scattering entropy and hence
in the level of wave depolarisation. This causes a drop in polarimetric coherence
and hence an erosion of our ability to exploit polarimetric phase information.
392 Applications of polarimetry and interferometry

HH Channel
2 1

Height (m)
1 0.5

0 0
2 3 4 5 6 7 8 9
Frequency (GHz)
HV Channel
2 1

Height (m)
1 0.5

0 0
2 3 4 5 6 7 8 9
Frequency (GHz)
VV Channel
2 1
Height (m)

1 0.5
Fig. 9.55 Reconstructed dual-baseline verti-
0 0
cal profiles for the 1.8-m vegetation layer in
2 3 4 5 6 7 8 9
HH (upper), HV (centre), and VV (lower)
polarisations Frequency (GHz)

HH-VV channel
2 1

0.8
1.5
0.6
1
0.4
0.5
0.2
0 0
2 3 4 5 6 7 8 9

HH-VV channel
1
2
0.8
Height (m)

1.5
0.6
1 0.4
0.5 0.2
Fig. 9.56 Reconstructed vertical profiles for
the 1.8-m vegetation layer in the HH-VV 0 0
channel for single (upper), and dual (lower) 2 3 4 5 6 7 8 9
baselines Frequency (GHz)

However, by combining polarimetry with interferometry we can control the

variation of interferometric coherence with polarisation and use this to quantify
more carefully the balance of surface and volume scattering.
We now turn to an example that combines all these ideas, but in a more
challenging environment more typical of remote sensing applications: airborne
radar imaging.
 9.5 Applications of polarimetry and interferometry 393

9.5.5 Application 5: forest height estimation using

POLInSAR
E-SAR is an airborne multi-frequency SAR system operated by the German
Aerospace centre (DLR) (Papathanassiou, 1998b; Hajnsek, 2009). It operates
in four frequency bands—X (9.6 GHz), C (5.5 GHz), L (1.3 GHz), and P
(350 MHz)—with a repeat-pass quad polarimetric interferometry mode at the
two lower bands of L and P. It operates with a range swath width of 3–5 km,
with range and azimuth resolutions of the order of 1.5 × 1.5 m, so providing
multichannel images for data analysis. In this section we consider use of the
lowest-frequency P-band in POLInSAR mode for forest height estimation in
tropical forests.
We employ P-band repeat-pass polarimetric data collected by the DLR E-
SAR airborne system as part of the ESA sponsored INDREX-II campaign
(November 2004) (Hajnsek, 2009). We concentrate on tracks collected over
the Mawas forest reserve in central Kalimantan (114◦ 36 E, 2◦ 12 S). This site
is a typical example of a tropical peat swamp forest environment. The forest
covers a large area (540,000 ha), and has a wide biomass range: 50–400 ton/ha,
with a corresponding height range of 5–30 m. One key objective of this study
was to investigate the potential of POLInSAR to retrieve forest biomass from
height using allometric relationships (Mette, 2004, 2007; Woodhouse, 2006).
The more traditional route to biomass follows directly from radar backscatter
(Imhoff, 1995), but this approach is plagued by two key issues: backscatter
saturation—especially at high frequencies—and high variability due to sensi-
tivity to changes in forest structure (Woodhouse, 2006). Height estimates, on
the other hand, overcome such saturation effects and, by using a robust height
estimation algorithm, can be made tolerant to changes in structure (Mette, 2004;
Treuhaft, 2004; Papathanassiou, 2005; Praks, 2007).
Here we make use of two tracks at P-band (λ = 0.86m), with a nominal
horizontal baseline of 30 m (and 75 minutes time baseline). The single-look
complex (SLC) SAR images are provided with high resolution, operating with a
slant range/azimuth pixel size of 1.4985 m/0.72 m at P-band. One key advantage
of this site is the exceptionally flat topography, with slopes less than 0.1%. A
further reason for employing data for the Mawas-E site is the availability of in
situ tree-height measurements for validation. In all, 1,049 trees were measured
in two parallel transects. Figure 9.57 shows a radar backscatter image of the
scene with the in situ transects marked as black lines. In the upper picture
we show an aerial photograph of the scene. We note the access road and non-
forested area to the right, with the main forest to the left. In the lower part
we show a P-band polarised power (maximum eigenvalue of [T ]) image. To
visualize the polarimetric information we show an entropy image in Figure 9.58.
Here we see high entropy over the forest, and lower entropy (more polarised
signal) over the non-forested region to the right.
Turning now to interferometry, Figure 9.59 shows a raw interferogram
(before flat Earth removal) of the scene (in HH polarisation). Note that the
presence of the forest can be clearly seen as an increase in phase variance.
This can be confirmed by calculating the corresponding coherence. Figure 9.60
shows the interferometric coherence as a greyscale image with white = 1 and
black = 0. Note the high (white) coherence over the non-forested region, with
the forest showing a range-dependent variation increasing from near to far
 Polarised power (lambda 1)
46
45
Angle of incidence 44
43
42
41
40
Fig. 9.57 Aerial photograph of Mawas-E test 39
area (top), and P band POLSAR Image 500 1000 1500 2000 2500
(lower), with two in situ data transects marked
as black lines Azimuth (m)

Scattering entropy
1
46
0.95
45 0.9
Angle of incidence

44 0.85
43 0.8
0.75
42
0.7
41 0.65
40 0.6
39 0.55
0.5
500 1000 1500 2000 2500
Fig. 9.58 Entropy image of MAWAS P-band
data Azimuth (m)

P band interferogram

Fig. 9.59 P-band HH repeat pass interfero-

metric phase for forest/non-forest boundary
 9.5 Applications of polarimetry and interferometry 395

HH interferometric coherence
1

50 0.9

0.8
Angle of incidence (degrees)

45 0.7

0.6
40
0.5

0.4
35
0.3

30 0.2

0.1

25 0
0 1000 2000 3000 4000 5000
Fig. 9.60 Interferometric coherence of the
Azimuth (m) MAWAS-E test area for P-band HH channel

range. This is characteristic of volume decorrelation, with the 30-m horizon-

tal baseline having a higher effective β z value in the near range (see equation
(5.15)). This leads to more decorrelation for a given tree height at near range
than far. We can now combine the coherence phase and amplitude information
to estimate surface topography and top height, using the coherence separation
optimization of equation (6.29).
In Figure 9.61 we show an image of the highest P-band phase centre (upper)
and a transect through each of the two in situ datasets (lower). Here we show
the location of the two optimum phase centres, noting around 5 m of separation
between the optima. In solid, the surface topography estimate from a line fit;
and in dash, the top height estimates from equation (8.38). We also show data
points corresponding to in situ tree-height measurements (located around 800
m in azimuth). We note good general agreement with the POLInSAR height
estimates, with around 20 m tree-height in the upper, and only 10 m in the lower
transect.
These results indicate how the variation of interferometric coherence with
polarisation can be used to estimate important biophysical parameters of inter-
est, such as forest height, with key implications for the estimate of biomass.
One key question remains as to whether we can now scale these ideas up to
continental or global scale using satellite remote sensing technology. We now
turn to consider this final topic in more detail.

9.5.6 Application 6: spaceborne satellite

radar polarimetry
Finally, we turn to consider data from the first fully polarimetric radar satellite
system to be launched: the Phased Array L-band SAR (PALSAR) on board
the Advanced Land Observation Satellite (ALOS), launched by the Japanese
Space Agency (JAXA) in January 2006 (Rosenqvist, 2007). Figure 9.62 shows
an image of the satellite, with the large solar panels on the right and the SAR
396 Applications of polarimetry and interferometry

Fig. 9.61 P-band phase centre separations

along upper and lower azimuth transects for
optimum polarizations, showing in situ height
measurements as stars

Parameter Value
Wavelength 0.236 m
Launch date Jan 24th 2006
Orbit height 691 km
Orbit Repeat 46 days

Chirp bandwidth 14 MHz

Peak transmit power 2 kW
Duty cycle 3,5% (7 %/2)
Noise figure 4 dB
PRF (Quadpol) 3.8 kHz
Antenna size (Tx, Rx) 8.9 m  3.1 m
Fig. 9.62 Schematic image and technical
details of the ALOS-PALSAR satellite radar Quadpol mode 21.5 degrees
incidence angle
system

antenna located on the left. Some key system parameters are also shown on the
right-hand side.
We note four key points in particular. The first is that the transmitted band-
width in polarimetric mode is only 14 MHz, which is much smaller than the
airborne E-SAR system, and hence the range resolution is poorer (around 11 m
 9.5 Applications of polarimetry and interferometry 397

HH image HH image
0 0 0 0

10 10
–5 –5

20 20
–10 –10
30 30
–15 –15

Range (m)
Range (m)

40 40
–20 –20
50 50

–25 60 –25
60

70 –30 70 –30

80 80
–35 –35

90 90
–40 –40 Fig. 9.63 Satellite SAR Images (HH left,
0 20 40 0 20 40
and HV right) for a trihedral corner reflector
Azimuth (m) Azimuth (m) before Faraday rotation correction

slant range for PALSAR compared to 1.5 m for E-SAR). The second point is
the relatively large PRF in polarimetric mode (around 4 kHz), which as men-
tioned in Section 9.3, limits the range coverage of the sensor. In this case the
range swath width is reduced to around 15 km in slant range, which for 21.5-
degree incidence (the default mode for the satellite) translates to around 40 km
of ground range. Thirdly, we note that the satellite repeats its orbit only every
46 days, so that repeat-pass interferometry can be implemented only with this
minimum temporal baseline. However, such long baselines lead to large levels
of temporal decorrelation, and hence are limiting for quantitative applications
such as POLInSAR height retrieval and coherence tomography. Finally, we
note the high orbit of the satellite, by which propagation effects through the
ionosphere cannot be ignored.
To illustrate this we show data for calibration trihedral corner reflectors (see
Figure 1.21) deployed in Adelaide, South Australia. The data was collected on
10 June 2006 at 13:50 UT. Figure 9.63 shows SAR images in HH (left) and
HV (right) channels. This data has been calibrated using the scheme of Section
9.3.2, but has not been corrected for Faraday effects. From scattering theory
we expect the HV backscatter to be zero for the trihedral, but we see it is only
around 20 dB below the copolar signal. This is due to a small rotation of the
plane of polarisation through the ionosphere. The level of Faraday rotation can
be estimated using equation (4.83). The estimated level is around +3-degree
one-way rotation. When this is removed from each pixel of the image we
obtain the corrected imagery shown in Figure 9.64. Here we see much better
isolation and residual cross-talk levels below –30 dB. This nicely illustrates
how vector-wave propagation models can be used to improve data quality and
system calibration for spaceborne radar systems.
One important application of such radars is in the identification and char-
acterization of coherent (polarised, in our context) scattering points in radar
imagery (Ferro-Famil, 2003; Schneider, 2006). Figure 9.65 is an example, in
398 Applications of polarimetry and interferometry

HH image HV image
0 0 0 0

10 10
–5 –5

20 20
–10 –10
30 30
–15 –15

Range (m)
Range (m)
40 40
–20 –20
50 50

60 –25 60 –25

70 –30 70 –30

80 80
–35 –35

90 90
Fig. 9.64 Satellite SAR images (HH left, and –40 –40
0 20 40 0 20 40
HV right) for a trihedral corner reflector after
Faraday rotation correction Azimuth (m) Azimuth (m)

HH RCS image

50
100
150
200
Slant range (m)

250
300
350
400
450
500
550
Fig. 9.65 HH image of ALOS-PALSAR cor- 500 1000 1500 2000 2500 3000
ner reflector scene (Adelaide, South Aus-
tralia) Azimuth (m)

which is shown an expanded view of the corner reflector scene. (The three cor-
ner reflectors are seen—that at far left being used in Figures 9.63 and 9.64.) We
then use a 3 × 5 (range x azimuth) window centred on each pixel to estimate the
local entropy from the eigenvalues of the coherency matrix. Figure 9.66 shows
the distribution of entropy/alpha values for the whole scene. We note that the
bulk of background pixels lie with high entropy, but there are several distinct
bands of low entropy distributions with various alpha values. The trihedrals are
seen lying close to the origin (small alpha and low entropy). However, there
 9.5 Applications of polarimetry and interferometry 399

Entropy/alpha diagram
90

80

70

60

50
Alpha

40

30

20

10

0
0 0.2 0.4 0.6 0.8 1 Fig. 9.66 Distribution of entropy/alpha val-
Entropy ues for all pixels in Figure 9.69

are clearly coherent points (low entropy) with alpha values greater than 45◦ ,
indicating second-order dihedral scattering.
The differences in alpha then relate directly to the boundary conditions for that
scattering process. In particular they relate primarily to the dielectric constant
of the reflection. Furthermore, given the small angle of incidence of the data
(typically a 24-degree angle of incidence in mid swath) and its small variation
across the limited range swath, this dependence is primarily determined by one
of the two surfaces involved, and not both (see Figure 3.15). For dry materials
(low εr ) we can even approach the Brewster angle for reflection on one of the
surfaces, so that we can observe either 0- or 180-degree phase shift between
HH and VV, depending on which side of the Brewster point we operate. A phase
shift of 180◦ will be seen for dielectric constants greater than 5 (the Brewster
angle corresponding to which is 66◦ , matching the 24-degree local incidence
angle of ALOS). Hence the alpha angle for dihedral scattering at small angles
of incidence can be less than π /4. We can quantify this for the geometry of
PALSAR by calculating the variation of the apparent alpha of the boundary
conditions as a function of the dielectric constant.
Figure 9.67 shows an example. Here we show the variation of alpha with
dielectric constant of the vertical surface (for εr = 10 for the horizontal). We
show results for three angles of incidence—23◦ , 24◦ , and 25◦ (covering the
swath ofALOS-PALSAR)—and see only a small change. We see that increasing
alpha is directly linked to increasing dielectric constant of the vertical surface,
and that we have high sensitivity to changes in dielectric constant. This provides
us with a way to estimate εr of such dihedrals directly from spaceborne radar
data. For example, we can use Figure 9.67 to relate alpha directly to dielectric
constant using a suitable curve fit. One key feature of the fit must be that when
α tends to π/2 so εr tends to infinity (a metallic dihedral has an alpha of π/2),
and hence the function must have a pole at π/2. We then obtain the following
400 Applications of polarimetry and interferometry

fit from Figure 9.67 (fitting θ = 24 as the middle of the swath):

3.2299
εr ≈ 2.5 − 0.8522 (9.46)
π
2 − αs

This relation then enables us to estimate the dielectric constant of the vertical
surface reflection in dihedral scattering directly from the alpha angle of the
dominant eigenvector.
These few examples help illustrate the potential of spaceborne radar
polarimetry. The successful fusion of interferometry with polarimetry from
space will have to await the development of single-pass spaceborne systems,
or at least short temporal baseline low-frequency repeat pass sensors. At the
time of writing, no such system is yet operational, but plans for the launch of
Tandem-X in 2009 will see deployment of a spaceborne single-pass POLInSAR
system at X-Band (Krieger, 2007). This will not only open new applications in
the fusion of interferometry with polarimetry, but also provide new stimulus to
continued study of our understanding and exploitation of the ‘memory’ effect
imprinted on polarised electromagnetic waves scattered by random media.

Dihedral alpha: solid = fresnel, dash = pole model fit

80

70

60
Alpha parameter

23 degrees

50

25 degrees
40

30

Fig. 9.67 Relation between alpha and dielec-

tric constant of vertical reflecting surface of 20
dihedral for ALOS-PALSAR angles of inci- 100 101 102
dence (23–25◦ ) Dielectric constant
Introduction to
matrix algebra A1
In this Appendix we gather together some basic definitions and concepts from
matrix linear algebra. (For more details see Gershenfeld (1999), Strang (2004),
and Press (2007).) We concentrate on those of particular importance to the
subject matter of this book: namely, the matrix description of polarised wave
scattering and propagation.

A1.1 Matrix definition, diagonal, upper and

lower triangular forms
A matrix [A] (sometimes written in this book as a bold capital without brackets,
A) is a rectangular array of numbers arranged with m rows and n columns. The
dimensions of the matrix are then m x n. The general element located in row
i and column j is termed aij . If all elements are zero expect when i = j, the
matrix is termed diagonal. Two important variations of this idea are upper and
lower triangular matrices—the former having zeros below the main diagonal,
and the latter above. [A] can be a real or complex matrix, depending on whether
one or more of the elements aij are real or complex numbers.

A1.2 Matrix multiplication, Kronecker sums

and products
Two matrices can be multiplied only if they are size compatible; that is, two
matrices [A] and [B] can be multiplied to generate [C] = [A] · [B] only when
[A] is m × n and [B] is n × q, so the product [C] will have dimensions m × q.
The elements of [C] cij are then formed from the inner (scalar) product of the
ith row of [A] with the jth column of [B].
There are two other important ways in which matrix elements can be com-
bined to generate a new matrix. The Kronecker product [C] is an mp-by-nq
matrix derived from two matrices [A] which is m × n and [B] which is p × q,
as shown in equation (A1.1):
 
a11 B · · · a1n B
 .. 
[C] = [A] ⊗ [B] =  ... ..
. .  (A1.1)
am1 B · · · amn B

As such, this operation replaces every element of [A] by its ij element multi-
plied by the whole matrix [B]. This form is particularly useful for converting
402 Introduction to matrix algebra

composite matrix products into a single matrix vector operation, so if we have a

matrix [X ] and two product matrices [A] and [B], as shown in equation (A1.2),
then the vectorization of [X ] is transformed by a Kronecker product matrix as
shown:

[A][X ][B] ≡ [A] ⊗ [B]T x (A1.2)

where x is a lexicographic row vectorization of the elements of [X ]:

   
x11 x12 . . . x1n x11
x21 x22 . . . x2n   
  x
⇒x=
12 
[X ] =  .   (A1.3)
 : : . . :  :
xn1 xn2 . . . xnn xnn

The Kronecker sum of two square matrices A (n x n) and B (m x m) is likewise

defined as shown in equation (A1.4):

[C] = [A] ⊕ [B] = [A] ⊗ [Im ] + [In ] ⊗ [B] (A1.4)

where [I ] is the m x m identity matrix, defined such that [I ]·[A] = [A]·[I ] = [A].

A1.3 Square matrices, powers and

the exponential
An important special class of matrices arises when m = n; that is, [A] has
an equal number of rows and columns. In this case the matrix [A] is square,
and can be multiplied with itself to generate power series; so, for example,
[A]2 = [A] · [A] is well defined, and also has dimension m × m. One key
consequence of this idea is that we can define matrix functions from their series
representations. The most important example of this is the matrix exponential
function, defined formally by the following series:

[A]2 [A]n
exp([A]) = [I ] + [A] + + ··· + + ··· (A1.5)
2! n!
This function is widely used in polarisation algebra (see Appendix 2 for more
details). Finally, we note that for two different square matrices of the same
dimension [A] and [B] (both m × m), the products [A] · [B] and [B] · [A] are
both well defined but are not necessarily equal. Hence matrix multiplication
(unlike ‘ordinary’ multiplication) is non-commutative. The ability of matrices
to represent differences in the order of multiplications is a key attraction for
their use in problems involving rotations and transformations as encountered
in polarimetry.

A1.4 Inverse matrix, minors and determinants

The inverse of a square matrix [A], denoted [A]−1 , is defined by the relationship
[A]−1 · [A] = [A] · [A]−1 = [I ]. Although this relationship applies for arbitrary
 A1.4 Inverse matrix, minors and determinants 403

matrix dimension, special attention is paid to the case m = 2—2×2 matrices—

for which the following simple expression can be used to directly calculate the
inverse:
   
a a12 −1 1 a22 −a12
[A] = 11 ⇒ [A] = (A1.6)
a21 a22 a11 a22 − a12 a21 −a21 a11

The multiplicative scale factor in front of the inverse matrix a11 a22 –a12 a21 is
called the determinant of the matrix |A| or det([A]), and evidently problems
arise with the existence of the inverse when this determinant goes to zero. The
concept of determinant is important, and can be extended to arbitrary matrix
dimensions. In this case we define the determinant as a scalar obtained from an
n × n matrix [A], as shown in equation (A1.7):
 
a11 a12 . . . a1n 

a21 a22 · · · a2n  ! n !n

|A| =  . .. .. ..  = a C
ij ij = aij (−1)i+j Mij (A1.7)
 .. . . 
.  j=1
 j=1
an1 an2 . . . ann 

Here Cij is called the cofactor associated with the element aij , and is in turn
related to a signed version of Mij , called the minor, associated with the element
aij . The minor is itself a determinant, but importantly for a matrix of reduced
dimension (n−1)—the matrix formed by eliminating the ith row and jth column
of [A]. In this way we can reduce every determinant by reduction of dimension
to calculation of a series of 2 x 2 sub-matrices. The summation on the right-
hand side of equation (A1.7) is called the Laplace expansion of determinant
by the ith row (where i can be chosen arbitrarily). Similar expressions can be
used for a Laplace expansion by the jth column, again involving minors and
cofactors.
For arbitrary dimensions the inverse can be now be formally written as
follows:

adj([A])
[A]−1 = adj([A]) = [Cij ]T (A1.8)
det([A])

where adj([A] is called the adjoint matrix—itself formed as the transpose of

the matrix of cofactors of [A]. The transpose of a matrix is an operation that
exchanges rows and columns, formally replacing the ij element by the ji ele-
ment. If a matrix [A] equals its transpose or [A] = [A]T , then [A] is termed
symmetric, having the property that aij = aji . Another special case is when
[A] = −[A]T , in which case the matrix is called skew or anti-symmetric. In
this case it is easy to see that the diagonal elements of an antisymmetric matrix
must all be zero.
Another important scalar function of matrix elements is the trace5 or Tr([A]).
This is defined as the sum of the diagonal elements Tr([A]) = ni=1 aii . Two
very important properties of trace used extensively in analytic studies fol-
low from the ability to change the order of matrix products inside the trace
operation, so that for dual products, for example, Tr(AB) = Tr(BA), and
404 Introduction to matrix algebra

also the cyclic property of the trace for triple products so that Tr(ABC) =
Tr(BCA) = Tr(CAB). One important consequence of this result is that the
trace is invariant to unitary similarity transformations of a matrix [A], since
Tr(U∗TAU) = Tr(AUU∗T ) = Tr(A). Hence the trace is equal to the sum of
the eigenvalues of [A], and represents, in polarimetry, an important invariant
quantity identified as the total scattered power by an object.

A1.5 Hermitian and anti-Hermitian matrices

For the case of complex matrices, another important operation can be defined by
combining transpose with conjugation. Firstly we define the conjugate matrix
[A]* as the matrix obtained by forming the complex conjugate of each element
of [A]. By then combining this with the transpose we obtain the conjugate-
transpose or Hermitian adjoint operation:

[B] = [A]∗T ⇒ bij = aji∗ (A1.9)

Note that sometimes in the literature the transpose symbol ‘T’ is omitted for
notational convenience, and that only the conjugate sign is used to indicate
both operations implicitly. There is then, of course, the possibility of ambiguity
of notation. To counter this, the ‘dagger’ symbol † is often used to represent
the combined transpose and conjugate operations, so [A]† = [A]∗T . Again, an
important class of matrices arises when [A]† = [A]. These are termed Hermitian
matrices, and arise a great deal in the development of polarised wave scatter-
ing and propagation. Cleary, such matrices must have real elements along the
diagonal and complex conjugate elements on matching off-diagonal locations.
If [A]† = −[A], then [A] is termed skew or anti-Hermitian, and again must
have zeros along the diagonal.

A1.6 Orthogonal and unitary matrices

Two important classes of matrix can be defined by relating transpose and
Hermitian adjoint operations to the matrix inverse. In the first case we define
a matrix as orthogonal if [A]T · [A] = [I ] that is, if the inverse of the matrix is
just its transpose. In this case the n × n matrix [A] can be decomposed into a
set of n mutually orthogonal n-element column vectors a, where, for example,
aT1 a2 = 0, and so on, as shown in equation (A1.10):
 
[A] = a1 a2 . . . an (A1.10)

In a related sense, if a matrix satisfies the relation [A]∗T [A] = [I ], then it

is termed unitary, and again it can be decomposed into a set of n column
vectors—this time with orthogonality defined in the complex Hermitian sense:
a∗T
1 a 2 = 0. Since orthogonality is a key physical concept in polarimetry, such
matrices often arise in applications, particularly in the context of similarity
transformations.
 A.8 Rayleigh’s quotient, positive definite, positive semidefinite matrices 405

A1.7 Similarity and congruent transformations

Two n × n matrices [A] and [B] are called similar if there exists an invertible
n × n matrix [P], so that [A] and [B] are related as follows:

[P]−1 [A][P] = [B] (A1.11)

Another way to think about this relationship is that the matrix [A] is transformed
into the matrix [B] by operation of [P]. This is termed a similarity transfor-
mation of [A] by [P]. Often the matrix [P] is unitary, in which case equation
(A1.11) becomes a unitary similarity transformation and the matrices [A] and
[B] are unitarily similar. Similar matrices share many properties in common.
For example, the determinants are equal, so det([A]) = det([B]). More impor-
tantly, their eigenvalues are equal (although their eigenvectors are different).
An important variation of this scheme is the congruent transformation. In this
case, [A] and [B] are related again by a third matrix [P], but now in the form:

[P]T [A][P] = [B] (A1.12)

This is to be clearly distinguished from the more common similarity transfor-

mation of equation (A1.11), and arises in the description of the backscatter of
polarised waves.

A1.8 Rayleigh’s quotient, positive definite,

positive semidefinite matrices
Another important class of composite products is the embedding of a matrix
between two vectors to generate a scalar, as shown in equation (A1.13). Further-
more, if the matrix is real orthogonal [O], or complex Hermitian [H ], then this
scalar is always real. To see this, just take the transpose (for [O]) or conjugate
transpose (for [H ]), and note the invariance of the scalar to these operations,
implying that the scalar must be real. When this scalar is greater than zero for all
vectors x, the matrix is termed positive definite, or PD. When the scalar is just
non-negative (so including zero), the matrix is termed positive semi-definite, or
PSD. These conditions are summarized in equation (A1.13):

xT [O]x > 0 x∗T [H ]x > 0 PD

(A1.13)
xT [O]x ≥ 0 x∗T [H ]x ≥ 0 PSD

A classical problem in matrix algebra is to find the x vector that maximizes the
scalar. To do this we need to first normalize by the magnitude of the vector to
obtain Rayleigh’s quotient R, as shown in equation (A1.14):

 x∗T [H ]x
R x = (A1.14)
x∗T x

Such ratios arise a great deal in polarimetry applications. It is therefore of

importance to be able to test if a matrix is PD or PSD, and to then to find a
systematic way of maximising or minimising R. One of the best ways to do this
is to use an eigenvalue approach, as we now show.
406 Introduction to matrix algebra

A1.9 Subspaces, eigenvectors and eigenvalues

We are often interested in a subset of the full set of vectors x in relations such
as equation (A1.14). One particularly important set form is the nullspace of the
matrix [A]. These are by definition the set of vectors xN satisfying the following
relation:

[A]xN = 0 (A1.15)

A second important set are the eigenvectors e. A square matrix [A] of dimension
n × n has n such eigenvectors, defined by the following equation:

[A] e = λe
⇒ [A] e − λe = 0 (A1.16)
⇒ det([A] − λ[I ]) = 0

where λ is a scalar called the eigenvalue. In this sense we can consider [A]
to operate on a vector x, and equation (A1.16) states that for some special
vectors, e, this operation will leave the vector x unchanged, apart from multi-
plication by a scalar. This physical interpretation is useful in many applications
in polarimetry, although equation (A1.16) is also an important general mathe-
matical idea. The eigenvalues can be found from the zeros of the determinant
(generally an nth order polynomial with n complex roots), as shown in equation
(A1.16). It is of special interest to be able to express a general matrix [A] as
a function of its eigenvectors and eigenvalues. An important theorem, called
Schur’s theorem, gives us a systematic way to achive this. Schur’s theorem
states that for any square matrix [A] there exists a unitary matrix [U ], such
that [U ]−1 [A][U ] = [B] is upper triangular. The eigenvalues of [A] must be
shared by the similar matrix [B] and appear along its main diagonal. The unitary
matrix [U ] is then composed of the eigenvectors of [A] as its columns. Hence
we can write a general matrix [A] in terms of its eigenvalues and eigenvectors,
as shown in equation (A1.17):
 
λ1 δ12 · · · δ1n
0 λ2 · · · δ2n   
  ∗T
[A] = [U ] .  . .. .. ..  · [U ] [U ] = e1 e2 . . . en
 .. . . . 
0 0 · · · λn
(A1.17)

An important class of matrices has an even simpler form, when the off-diagonal
δ terms are all zero and the matrix [B] is diagonal. Symmetric and Hermitian
matrices fall into this category. The proof for Hermitian matrices is very sim-
ple, since if [A] is Hermitian then so is [B], and the only way [B] can be
Hermitian (equal to its conjugate transpose) is if it is diagonal, with real eigen-
values. Hence we have proved that Hermitian matrices have real eigenvalues
and orthogonal eigenvectors—a result of great importance in polarimetry. In
general, the required condition that [B] is diagonal is that the matrix [A] be nor-
mal, which by definition means that it commutes with its conjugate transpose;
that is, [A][A]∗T − [A]∗T [A] = 0. Unitary matrices, for example, also satisfy
 A1.10 Norms, condition number, least squares and the SVD 407

this condition (when the conjugate transpose equals the inverse), and so can
also be expressed in diagonal form.
Eigenvalue decompositions are also of interest for solving optimization prob-
lems. Taking, for example, Rayleigh’s quotient again (equation (A1.14)), we
return to the problem of how to maximize this ratio. To solve this we can employ
the Lagrange multiplier method of constrained optimization. We first set up a
Lagrangian function L as follows:

L = x∗T [H ]x − λ(x∗T x − 1) (A1.18)

and then differentiate with respect to the variable x to find the stationary points
as zeros of the derivative, as shown in equation (A1.19):
dL
= [H ]x − λx = 0 (A1.19)
dx
This implies that the optimum solution for x is the eigenvector of [H ] corre-
sponding to the maximum eigenvalue λmax . This simple example illustrates how
eigenvalue decompositions and optimization theory can be formally linked.
Finally, we consider an extension of this concept to the singular value decom-
position, or SVD. The Schur decomposition employs a single unitary matrix
[U ]. If, on the other hand, we employ two unitary matrices [U ] and [V ], then
any matrix (even those that are not square) can be expressed as a function of a
purely diagonal matrix [D], as follows:

[A] = [U ] . [D] . [V ]∗T

 
s1
 .. 
 . 
 
 sp   
[D] = 


 |s1 | ≥ |s2 | · · · ≥ sp  (A1.20)
 0 
 .. 
 . 
0

The diagonal elements of [D] are termed the singular values of the matrix [A]
and [U ], the left singular vectors, while [V ] is the set of right singular vectors.
In the general case when [A] has dimensions m × n, [U ] is m × m, [D] is m × n,
and [V ] is n×n. However, only p elements of [D] are non-zero. The dimensions
p and m−p define two important subspaces that have important applications,
as we now consider.

A1.10 Norms, condition number,

least squares and the SVD
Very often we wish to consider matrix equations of the following form:

[A] x = b (A1.21)

There are two important subspaces to this problem. The range of the matrix
[A] is the space of all possible vectors b for which the equation is solvable.
408 Introduction to matrix algebra

The dimension of the range is called the rank. Secondly, the set of vectors x
which satisfy the homogeneous equation (b = 0) define the null space of [A].
If there is no null space, then [A] is of full rank. The SVD provides a useful
perspective on these subspaces. The columns of [U ] in equation (A1.20) that
are associated with non-zero singular values form an orthonormal basis for the
range of [A], while the columns of [V ] associated with the zero singular values
form an orthonormal basis for the null space.
These concepts are particularly important in the solution of least squares
problems. In these cases we are usually interested in finding solution vector x
that solves an overdetermined set of equations. Given the possibility that an
exact solution may not exist, we ask instead to find x that has the minimum
residual; that is, one that minimizes the function L, defined as:

 2
L(x) = [A] x − b (A1.22)

The classical way to solve this is to expand L and differentiate with respect to
x and set to zero. This yields the so-called normal equations, as follows:

   
Ax − b2 = Ax − b T Ax − b

= xT AT Ax − 2xT AT b + bT b
∂ (A1.23)
= 0 ⇒ AT Ax − AT b = 0
∂x
−1 T
⇒ x = AT A A b

There are two issues to be faced with such a solution. Firstly, does the inverse
matrix exist at all; and secondly, how sensitive is the solution x to perturbations
or small errors in the vector b? If there is any serious amplification of such errors
we say the equations are ill-conditioned. The SVD provides an important insight
into both these situations, as we now consider.
The generalized or pseudo-inverse of [A], designated [A]+ , can be defined
from the SVD as follows. Writing the function L in terms of the SVD we obtain
the following simplified expansion:

 2  2
L = UDV T x − b = DV T x − U T b
.
UTb = z   ! p
!m (A1.24)
 
⇒ L = Dy − z  = (s y
i i − z i )2
+ zi2
VTx = y i=1 i=p+1

Minimising L is now seen to occur for the following choice:

zi
yi = , x = V y ⇒ x = VD+ U T b = A+ b (A1.25)
si

where D+ is a rather strange kind of inverse matrix, formed as follows. When

the diagonal elements of D are non-zero then we take the reciprocal, but when
 A1.10 Norms, condition number, least squares and the SVD 409

they are zero we leave then as zero:

 
1
 s1 
 
 .. 
 . 
 
 1 
[D]+ =  

 (A1.26)
 sp 
 0 
 
 .. 
 . 
0

Equation (A1.25) then provides a formal solution to linear least squares prob-
lems, even in the case when the inverse of ATA in the normal equations does
not exist. The generalized or pseudo-inverse of [A] is then defined as shown
in equation (A1.25). When the singular values of [A] have a clear cut-off the
above formulation is clear, but more often the situation is that the singular val-
ues fall off slowly and never actually go to zero. In this case we face another
challenge. Here the matrix is technically of full rank, but practically we can
obtain an ill-conditioned system. Again we can use the SVD to quantify this
via the concept of condition number of the matrix [A] or CN(A) as follows.
We are interested in how any fractional errors in the input vector b are mapped
into errors in the solution vector x. To obtain this we first define the norm of a
matrixA, which is a scalar defined in equation (A1.27):
 
Ax ? ? ? ?
A = max   ⇒ ?Ax? ≤ A ?x? (A1.27)
x =0 x 

This norm has one very important property: the norm of a product is less than
or equal to the product of the norms. We can find an expression for the norm
of [A] by using the SVD and first generating the squared norm, as shown in
equation (A1.28):

xT AT Ax
A2 = (A1.28)
xT x

Now the matrix ATA is always symmetric, and so this Raleigh quotient is
maximized by the maximum eigenvalue of ATA (see equation (A1.19)), λmax .
However, ATA = VD2 VT , and so this eigenvalue is equal to the squared modulus
of the maximum singular value of [A]. Hence we finally have an expression for
the norm of matrix [A] as shown in equation (A1.29):

A = |s1 | (A1.29)

Returning to the issue of error amplification, we have b = Ax, and so any

error δx will satisfy δx = A−1 δb. Using norms on both these equations, and
re-ordering, we can then write an expression for the condition number CN, as
shown in equation (A1.30):
? ? ? ? . ? ? ? ?  
?b? ≤ A . ?x? ?δx? ?δb? ? ?  smax 
? ? ? −1 ? ? ? ⇒ ? ? ≤ CN ? ? → CN = ? ?A −1 ?
? . A =  
?δx? ≤ ?A ? . ?δb? ?x? ?b? smin 
(A1.30)
410 Introduction to matrix algebra

Here we see that the amplification factor depends on the ratio of maximum
to minimum singular values of [A]. It is easy to show that if [A] is unitary,
for example, then CN = 1, and the errors are not amplified at all. In general,
however, we find that CN > 1, and some care must be taken to ensure that
errors are not amplified too much. Note from equation (A1.30) that we have a
way of controlling this amplification by just setting any small singular values
to zero in the pseudo-inverse, so increasing smin and reducing the CN until
an acceptable ratio is achieved. The price to pay for this, however, is a rank
reduction of [A].
Unitary and rotation
groups A2
In Appendix 1 we defined some very basic properties of matrices and matrix
algebra as used in a description of polarised waves. Here we extend the discus-
sion to examine several more advanced ideas, largely based on group theory,
which are widely used in a description of the propagation and scattering of
polarised electromagnetic waves, forming the subject of polarisation algebra.
There are three classical routes by which the transition from scalar to higher
dimensional forms can be achieved mathematically (Murnaghan, 1962; Misner,
1973; Goldstein, 1980; Cornwell, 1984; Penrose, 1984; Rosen, 1995; Georgi,
1999). These are:
• Scalars_Vectors_Tensors
• Scalars_Complex Numbers_Quaternions_Bi-Quaternions
• Scalars_Spinors_Twistors
Each of these has its strengths and weaknesses in terms of ease of formula-
tion, potential for quantitative analysis, and physical insight. However, two
important general themes arise from all approaches: firstly, the relationship
(sometimes conflict, as in relativistic quantum theory) between real and com-
plex formulations of a problem (which in our context relate to the role of phase
in multi-channel systems); and secondly, the unifying role played by group the-
ory, which provides not only a convenient unifying framework, but also aids
physical insight by exposing deep symmetries that can then be exploited to aid
analysis of complicated problems.
Among the many concepts used in abstract algebra, one of the most useful
is that of a group. We can introduce these ideas by identifying a hierarchy of
algebraic concepts as follows, each one building on the properties of the simpler
concepts to its left.
SET ⇒ GROUP ⇒ FIELD ⇒ VECTOR SPACE ⇒ ALGEBRA
A group G is then defined as a set of elements together with a composition
(generalized concept of a product) xy with x,y ∈ G, such that the following four
conditions hold:
a) Closure, xy ∈ G
b) Associative, x(yz) = (xy)z
c) Existence of the identity, xI = Ix = x
d) Existence of the inverse, xx−1 = I
If in addition we have the property shown in e), then the group is called Abelian
and such groups, as we shall see, have a simpler form and generate the basic
building blocks of a general classification theory.
e) Commutative for group multiplication, xy = yx
412 Unitary and rotation groups

We then specify G as finite or infinite if the number of elements is finite or

infinite (called the order of the group). Again, finite and infinite groups have
very different properties. Finally, if the elements of a group are functions of a
continuous parameter—that is, rotation through an angle θ—then the group is
called continuous.
It is remarkable that armed only with a set of such simple rules we can for-
malize many complex transformation problems and expose new and important
underlying patterns in the description of polarised wave scattering.
Continuing the hierarchy, a field is a group with the extra concept of addition
of elements, under which the field is commutative. There are three main fields
of interest in physics and engineering: real numbers R, complex numbers C, and
quaternions Q, all of which are important in polarisation algebra. An algebra
itself is then a still more complicated structure, and consists of a group, a
field, and three additional concepts: addition, scalar and vector multiplications.
These, then, are the basic building blocks for polarisation algebra, as we now
demonstrate.
We have seen that in the development of polarisation geometry, the mathe-
matics of mapping from complex to real domains is of central significance. Two
important examples are the SU(2)–SO(3) homomorphism, which underlies the
geometry of the Poincaré sphere and the SL(2,C)–SO(3,1,R) homomorphism
that leads to a Lorentz transformation of the Stokes vector and a real 4×4 matrix
representation of scattering. In this Appendix we develop a general approach
to parameterising complex unitary and real rotation groups of arbitrary dimen-
sion (Murnaghan, 1962; Cornwell, 1984; Cloude, 1995b; Rosen, 1995; Georgi,
1999). This formalism will clarify many of the features already discussed, and
also highlight a third important mapping between SU(4) and SO(6), which can
be used to provide a general physical interpretation of bistatic scattering in
random media.
One of the most important practical examples of a continuous group is the
general linear group formed from the set of n × n non-singular real GL(n,R)
and complex GL(n,C) matrices. There is also a set of important sub-groups of
GL such as SL(n), the special linear group of matrices with unit determinant;
U(n,C) and SU(n,C), the set of unitary and special unitary complex matrices;
and O(n,R)) and SO(n,R), the orthogonal and special orthogonal groups. A
complete classification for simple continuous groups such as these was first
developed independently by Sophus Lie in 1870 and Wilhelm Killing in 1880,
and was refined by Elie Cartan in 1894 (Cartan, 1966). This classification
leads to four infinite series of groups designated A, B, C, and D, together with
five exceptional groups. These can be used to identify general mappings from
complex into real domains, as we now show.
We begin with the concept of a Lie algebra L, named after the nineteenth-
century Norwegian mathematician Sophus Lie (1842–1899). This is an n-
dimensional linear vector space equipped with a Lie product or commutator
defined between elements a and b, as shown in equation (A2.1):

[a, b] = ab − ba (A2.1)

where group (matrix) multiplication is implicit in terms such as ‘ab’. If [a,b] = 0

then the group is called Abelian. Sophus Lie was the first to show that the
 A2.1 The real Lie algebra L = su(n) and the group SU(n)n ≥ 2 413

r , defined
properties of the algebra are embodied in a set of structure constants cpq
from the commutation by

  !n
ap , aq = r
cpq ar (A2.2)
r=1

The connection between Lie algebras and groups is often provided by the matrix
exponential function, so that we can define a general matrix element A as shown
in equation (A2.3):

a2 a3 an
A = exp(a) = I + a + + + ··· + ··· (A2.3)
2! 3! n!

Consider the following two important examples.

A2.1 The real Lie algebra L = su(n) and

the group SU(n)n ≥ 2
In this case A is unitary, and so L is the set of traceless anti-Hermitian n × n
matrices, since
 /
exp(a) exp(a)∗T = I a∗T = −a
⇒ (A2.4)
det(exp(a)) = 1 Tr(a) = 0

The dimensionality of su(N) is N2 − 1. For the algebra su(2) we then have

the following three-dimensional representation with commutation relations as
shown:
     
1 0 i 1 0 1 1 i 0
a1 = a2 = a3 =
2 i 0 2 −1 0 2 0 −i (A2.5)
[a1 , a2 ] = −a3 [a2 , a3 ] = −a1 [a3 , a1 ] = −a2

By exponentiation we then arrive at the Pauli spin matrices σi = −2iai . The Table A2.I Commutation
commutation properties of the Pauli matrices can be conveniently represented matrix of SU(2)
as a matrix, the pqth element of which is 1, 0 or –1 according to equation (A2.5),
σ 1 2 3
as shown in Table A.2.I.
Turning now to higher dimensions, the algebra su(3) is likewise constructed 1 0 1 −1
from eight basis matrices, as shown in equation (A2.6). Just as su(2) leads to 2 −1 0 1
3 1 −1 0
the Pauli spin matrices as so-called generators for SU(2), so the corresponding
set for the group SU(3) are the Hermitian Gell–Mann matrices, obtained as
λk = −iak . Note that the scale factor in a8 is used to ensure that for all products
Tr(ai aj ) = −2δij .
Finally, the algebra su(4) can be represented by the set of fifteen matrices
shown in equation (A2.7). The corresponding generators for the group SU(4) are
called Dirac matrices. The pattern is now clearly developed for representation of
higher-dimensional unitary groups, although we see that the number of elements
414 Unitary and rotation groups

quickly increases as we go to higher dimensions. We shall show later that there

is an important way of classifying these groups based on a smaller dimensional
set called the Cartan sub-algebra. However, we first consider a second important
set of algebras related to the rotation groups.

     
0 i 0 0 1 0 i 0 0
a1 =  i 0 0 a2 = −1 0 0 a3 = 0 −i 0
0 0 0 0 0 0 0 0 0

     
0 0 i 0 0 1 0 0 0
a4 = 0 0 0 a5 =  0 0 0 a6 = 0 0 i (A2.6)
i 0 0 −1 0 0 0 i 0

   
0 0 0 i 0 0
1 
a7 = 0 0 1 a8 = √ 0 i 0 
0 −1 0 3 0 0 −2i

     
0 i 0 0 0 0 i 0 0 0 0 i
i 0 0 1 0 −1 0
 0 0   0 0   0 
a1 =   a2 =   a3 =  
0 0 0 −1 i 0 0 0 0 1 0 0
0 0 1 0 26 0 −1 0 0 24 i 0 0 0 46

     
0 i 0 0 i 0 0 0 0 0 0 1
i 0   0 0
 0 0   0 i 0 0   0 i 
a4 =   a5 =   a6 =  
0 0 0 1  0 0 −i 0  0 i 0 0
0 0 −1 0 35 0 0 0 −i 14
−1 0 0 0 16

     
0 0 −1 0 0 0 i 0 0 0 0 −1
0 i 0 −1 0 0
 0 0   0 0   0 i 
a7 =   a8 =   a9 =  
1 0 0 0 i 0 0 0 0 i 0 0
0 i 0 0 12 0 1 0 0 15 1 0 0 0 34

     
i 0 0 0 0 1 0 0 0 0 0 i
0 −i 0 −1 0 0 0
 0   0 0   0 1 
a10 =  a11 =  a12 = 
0 0 i 0 0 0 0 i 0 −1 0 0
0 0 0 −i 35 0 0 i 0 23 i 0 0 0 13

     
0 0 1 0 0 −1 0 0 i 0 0 0
0 i 1 0 0 −i 0
 0 0   0 0   0 
a13 =  a14 =  a15 = 
−1 0 0 0 0 0 0 i 0 0 −i 0
0 i 0 0 45 0 0 i 0 56 0 0 0 i 25
(A2.7)
 A2.2 The real Lie algebra L = so(n) and the group SO(n) 415

A2.2 The real Lie algebra L = so(n) and

the group SO(n)
In this case A is orthogonal, and so L is the set of traceless anti-symmetric n × n
matrices, since
 / T
exp(a) exp(a)T = I a = −a
⇒ (A2.8)
det(exp(a)) = 1 Tr(a) = 0
Table A2.II shows a comparison of the dimensionality of this algebra compared
to su(n). Also shown for completeness are the dimensions of other impor-
tant classical matrix groups. An important example already encountered in
polarimetry is so(3), which has a three-dimensional algebra formed from the
following matrices:
     
0 0 0 0 0 −1 0 1 0
a1 = 0 0 1 a2 = 0 0 0  a3 = −1 0 0 (A2.9)
0 −1 0 1 0 0 0 0 0
Also shown in Table A2.II is the generalized Lorentz group SO(n,1,R), which
again we have encountered in an interpretation of the geometry of the scattering
matrix (see Section 1.5.3). This group combines n-dimensional rotations with
a boost in one direction, and is formed from the set of real n+1 dimensional
matrices satisfying the following equation involving the Lorentz metric:
 
1 0 ... 0
0 −1 ... 0
 
[L] =   = [M ]T [L] [M ] (A2.10)
0 0 . . . 0 
0 0 ... −1
For so(3,1) there is a homomorphism with sl(2,c), and the former can be
represented by six matrices of the form shown in equation (A2.11):
   
0 0 0 0 0 0 −1 0 0 1 0 0
0 0 1 0 0 0 0 0 −1 0 0 0
   
0 −1 0 0 1 0 0 0  0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0
    (A2.11)
0 0 0 −1 0 0 0 0 0 0 0 0
 0 0 0 0  0 0 0 −1 0 0 0 0
   
 0 0 0 0  0 0 0 0  0 0 0 −1
−1 0 0 0 0 −1 0 0 0 0 −1 0

Table A2.II Dimensionality of important matrix groups

n SO(n,R) SU(n,C) SL(n,C) SO(n,1,R)

0.5n(n − 1) n2 − 1 2n(n2 − 1) 0.5n(n2 + 1)
2 1 3 6 3
3 3 8 16 6
4 6 15 30 10
5 10 24 48 15
6 15 35 70 21
416 Unitary and rotation groups

A2.3 The Killing form and Cartan matrix

To proceed towards a more general approach, we now consider construction of
the Killing form for a Lie algebra L, named after Wilhelm Killing (1847–1923),
and defined as shown in equation (A2.12):

B(x, y) = Tr {ad (x).ad (y)} (A2.12)

where x and y are elements of the algebra, and

 
ad (x) = x, xj (A2.13)

is an n×n matrix, the jth row of which consists of the structure constants for the
commutation of x with the jth element of L. The structure constants themselves
form an n-dimensional representation of L called the adjoint representation. For
example, for su(2) we have the following 3 × 3 matrix:
     
0 0 0 0 0 −1 0 1 0
ad (a1 ) = 0 0 1 ad (a2 ) = 0 0 0  ad (a3 ) = −1 0 0
0 −1 0 1 0 0 0 0 0
(A2.14)

The Killing form is then formed from the trace of matrix products, as shown in
equation (A2.15):
 
−2 0 0
Bsu(2) =  0 −2 0  = −2δpq (A2.15)
0 0 −2

The Killing form may initially seem a rather contrived concept, but it is used
as the basic distinguishing feature of the Lie algebra, and is central to the
general classification scheme. As mentioned earlier, it is possible to classify
these algebras by identifying an important sub-algebra: the Cartan algebra H .
This is Abelian by definition, and hence has very simple structure, being asso-
ciated with a set of commuting or simultaneously diagonal matrices in the
representation. Physically it can be considered a generalization of absolute
phase in coherent signal analysis. When binary (two channel) operations are
applied between signal channels this phase tends to disappear (in interferom-
etry, for example), and we shall see that similar properties hold for the phase
transformations generated by matrices obtained from the Cartan sub-algebra.
The dimension of H is called the rank of L and, importantly, this is always
much smaller than the dimension of L itself. For su(n), for example, there are
n–1 independent diagonal matrices, and hence the rank of su(n) is n–1. Hence
su(2), su(3), and su(4) have rank 1, 2, and 3 respectively, while their dimensions
(from Table A2.II) are 3, 8, and 15.
The algebra L can now be expressed as the direct sum of H , with rank k, and
a remaining root subspace R, so that

L=H ⊕R (A2.16)

Conveniently, the roots that span the space R can themselves always be written
in terms of a subset of k simple roots rs (where k equals the dimension of H ),
 A2.3 The Killing form and Cartan matrix 417

so that

!
k
r= αs rs (A2.17)
s=1

where the coefficients αs are generally complex. The classification can then
be generated by constructing an r × r matrix from the simple roots, called the
Cartan matrix A. The jkth element of A is obtained from the simple roots aj and
ak as

2B aj , ak
Ajk =  (A2.18)
B aj , aj

where B(..) is the Killing form. Clearly, the diagonal elements of A are equal
to 2, but less obvious is that the off-diagonal elements are limited in value
to 0, −1, −2, or −3. The Cartan matrix can be constructed from knowledge
of the roots and vice versa; that is, we can construct the whole algebra from
knowledge of A. Hence this matrix is the ‘signature’ of the algebra, and allows
us to compactly describe higher dimensional groups.
Two examples will illustrate the method. We start with su(2), which has basis
a1 , a2 , and a3 (equation (A2.5)), and the Cartan sub-algebra, which is one-
dimensional, with h1 = a3 . From equation (A2.5) we then have the following
(quasi-)eigenvalue conditions for the non-zero roots:

[h1 , (a1 + ia2 )] = i (a1 + ia2 )

(A2.19)
[h1 , (a1 − ia2 )] = −i (a1 − ia2 )

There are consequently two roots α1 and −α1 with α(h1 ) = i. The one-
dimensional root subspaces are then defined from complex linear combinations
of a1 and a2 as λ (a1 + ia2 ) and µ (a1 − ia2 ), where λ and µ are arbitrary
complex numbers. The Cartan matrix for su(2) has only one element: A = 2.
Turning to the more complicated case of su(3), from equation (A2.6) the
Killing form is now

B ap , aq = −12δpq p, q = 1, 2, ...8 (A2.20)

and with rank 2, the Cartan sub-algebra is spanned by h1 = a3 and h2 = a8

(see equation (A2.6)). It follows that we can write the following equations:

[h1 , a2 − ia1 ] = 2(a2 − ia1 ) [h2 , a2 − ia1 ] = 0

√
[h1 , a7 − ia6 ] = −1(a7 − ia6 ) [h2 , a7 − ia6 ] = 3 (a7 − ia6 )
√
[h1 , a5 − ia4 ] = 1(a5 − ia4 ) [h2 , a5 − ia4 ] = 3 (a5 − ia4 )
[h1 , −a2 − ia1 ] = −2(−a2 − ia1 ) [h2 , −a2 − ia1 ] = 0
√
[h1 , −a7 − ia6 ] = 1(−a7 − ia6 ) [h2 , −a7 − ia6 ] = − 3(−a7 − ia6 )
√
[h1 , −a5 − ia4 ] = −1(−a5 − ia4 ) [h2 , −a5 − ia4 ] = − 3(−a5 − ia4 )
(A2.21)
418 Unitary and rotation groups

There are therefore six roots, α1 , α2 , α3 , −α1 , α2 , and −α3 , which can all
be expressed as linear combinations of two simple roots: r1 = ah1 + bh2 .
The coefficients a and b can be obtained from the commutation properties of
the roots combined with the Killing form to generate a pair of simultaneous
equations of the following form:

aB(h1 , hj ) + bB(h2 , hj ) = α(hj ) j = 1, 2

α1 (h1 ) = 2 α1 (h2 ) = 0
√ (A2.22)
α2 (h1 ) = −1 α2 (h2 ) = 3
√
α3 (h1 ) = 1 α3 (h2 ) = 3

The coefficients a and b can then be calculated as

 
1 0 0
1 1
α1 = h1 = 0 −1 0
6 6 0 0 0
√  
0 0 0
1 3 1
α2 = − h1 + h2 = 0 1 0  (A2.23)
12 12 6 0 0 −1
√  
1 0 0
1 3 1
α3 = h1 + h2 = 0 0 0
12 12 6 0 0 −1

where we have used the fact that B(h1 , h1 ) = B(h2 , h2 ) = 12 and B(h1 , h2 ) = 0.
From these we can calculate the following Killing forms and Cartan matrix:

1 1 1 1
B(α1 , α1 ) = B(α1 , α2 ) = − B(α1 , α2 ) = − B(α2 , α2 ) =
3 6 6 3
  (A2.24)
2 −1
A=
−1 2

One important point is that we can now reverse this procedure and use the Cartan
matrix to generate the roots and hence the whole algebra. This is facilitated
through use of a geometrical construction called Dynkin diagrams. These will
finally lead us to the complete classification scheme.

A2.4 Dynkin diagrams: classification of unitary

and rotation groups
The construction of Dynkin diagrams provides a convenient geometrical
method for classifying Lie algebras (Cornwell, 1984; Georgi, 1999). This pro-
cedure is only part of a more general geometrical approach to the study of
Lie algebras that involves the association of roots with vectors in a Euclidean
space, with the Killing form employed as a scalar product. The commutation
properties of the roots, together with the restrictive integer range for the Killing
 A2.4 Dynkin diagrams: classification of unitary and rotation groups 419

form, mean that the set of non-zero vectors in this space is very restricted. In
this way we can construct the full set of rank N spaces by employing purely
geometrical methods.
A Dynkin diagram is constructed for each algebra L by associating a node
with each simple root and connecting nodes corresponding to roots aj and ak by
a number of lines given by Ajk Akj , where Ajk is the jkth element of the Cartan
matrix. Each node is also given a weight ωj = ωαj , αj , where ω is a constant
chosen such that the minimum weight is unity. These diagrams are used to
generate the whole classification scheme by starting with a root space of rank
1 and using an iterative scheme to generate root spaces of higher rank.
This scheme leads to classification of the four infinite sets of algebras, denoted
Ai , Bi , Ci , and Di with Dynkin diagrams shown in Figure A2.1. There are
also five exceptional algebras E6 , E7 , E8 , F4 , and G2 , with irregular Dynkin
diagrams as shown in Figure A2.2. The classical continuous groups associated
with the four infinite sets of Lie algebras through the matrix exponential function

1 2 i–1 i
1 1 1 1
Ai

2 2 2 1
Bi

1 1 1 2
Ci

1
1 1 1
i>2
Di
Fig. A2.1 Dynkin diagrams for the algebras
1 A, B, C and D

1 1 1 1 E6

1 1 1 1
E7

1
1 1 1 1 1 1
E8

F4

2 2 1

G2 Fig. A2.2 Dynkin diagrams for the 5 excep-

3 1 tional algebras
420 Unitary and rotation groups

can then be identified as follows:

AN −1 ⇒ SU (N )
BN ⇒ SO(2N + 1)
(A2.25)
CN ⇒ USp(2N )
DN ⇒ SO(2N )
where USp(N ) is a unitary group but with a symplectic inner product in N
dimensions. (These symplectic groups are related to quaternions, and involve
skew-symmetric bilinear forms; see Cornwell (1984) for more details.) The key
concept is that for a homomorphism to exist between the various groups they
must have the same Dynkin diagram, and their algebras are then isomorphic. In
this way the constructs in FigureA2.1 can be used to identify higher-dimensional
mappings from complex to real groups. We have seen in this book that such
mappings are central to polarisation algebra.
Consider the following important examples. If L = A1 , the Dynkin diagram
is a single node with unit weight. In this case we can construct the Cartan
matrix as A = 2, and assign a one-dimensional root space. However, we also
note that B1 and C1 have the same Dynkin diagram. Hence the three algebras are
isomorphic. This result leads to the SU(2)–SO(3) homomorphism underpinning
the geometry of the Poincaré sphere. Extending this, if L = A3 then the Cartan
matrix is of the following form:
 
2 −1 0
A = −1 2 −1 (A2.26)
0 −1 2
which is the same as that for D3 . Hence there is also a homomorphism between
the groups associated with A3 and D3 . Finally, we see that B2 = C2 , and
with this we have exhausted all possible isomorphisms between the algebras.
These important results are summarized in Table A2.III. The SU(2)–SO(3) and
SU(4)–SO(6) homomorphisms are of particular interest in polarimetry studies.
The SU(2) example is well known from studies of the Poincaré sphere, and so
here we summarize the main details of the less well known SU(4) mapping; that
is, given an element U4 of SU(4,C) generate an equivalent element O6 of SO(6).
The algebra su(4) has rank 3, and a suitable basis for the Cartan sub-algebra
can be obtained from equation (A2.7):
     
1 0 0 0 1 0 0 0 1 0 0 0
     
0 1 0 0 0 −1 0 0 0 −1 0 0
h1 =   h2 =   h3 =  
0 0 −1 0 0 0 1 0 0 0 −1 0
0 0 0 −1 0 0 0 −1 0 0 0 1
(A2.27)

Table A2.III Important homomorphic (isomorphic)

relationships between Lie groups (algebras)

Algebras Group Homomorphism Dimension

A1 = B1 SU(2)−SO(3) 3
B2 = C2 Sp(2)−SO(5) 10
A3 = D3 SU(4)−SO(6) 15
 A2.4 Dynkin diagrams: classification of unitary and rotation groups 421

The Dynkin diagram and corresponding Cartan matrix are then of the form
shown in equation (A2.28):
 
1 1 1 2 −1 0
A = −1 2 −1 (A2.28)
1 2 3 0 −1 2

A suitable set of generators for the group SU(4) are given by ηk = −iak , where
ak are defined in equation (A2.7). These fifteen matrices have the following
commutation properties:

!
15
[ηa , ηb ] = 2iεabc ηc (A2.29)
c=1

where the permutation symbol may be represented by a 15 × 15 matrix, as

shown in Table A2.IV. The ijth element of this matrix is zero if ηi and ηj
commute, and ±1 depending on the sign of the non-commuting elements. For
SU(2) the corresponding matrix is shown in Table A2.I. To illustrate the power
of this theory we consider the detailed mapping from SU(4) to SO(6), which is
performed in three distinct stages, as follows.

SU(4)–SO(6) homomorphism stage 1

We begin by considering two vector spaces U and V. The tensor product U ⊗ V
consists of a new vector space with basis ui ⊗ vj where i = 1, 2, 3 . . . N, j =
1, 2, 3 . . . M, and N and M are the dimensions (dim) of U and V. Consequently,
dim(U ⊗ V) = dim(U) dim(V) = MN. Typically we take repeated rth-order ten-
sor products of a space with itself: U ⊗ U ⊗ U. . . = Lr (UN ). An element in this
space is known as an rth-order tensor, and Lr (UN ) is called the carrier space
for an rth-order tensor representation.
Special consideration is given to second-order tensor space L2 (UN ), were
it is possible to form new tensors as linear combinations of the basis vectors
ei ,ej , which are antisymmetric under an interchange of subscripts. We do this

Table A2.IV Commutation matrix for SU(4)

η 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 0 −1 1 0 0 −1 1 0 0 −1 1 0 0 −1 1
2 1 0 −1 0 1 0 −1 0 1 0 −1 0 1 0 −1
3 −1 1 0 0 −1 1 0 0 −1 1 0 0 −1 1 0
4 0 0 0 0 0 0 0 1 1 1 1 −1 −1 −1 −1
5 0 −1 1 0 0 −1 1 1 1 0 0 −1 −1 0 0
6 1 0 −1 0 1 0 −1 1 0 1 0 −1 0 −1 0
7 −1 1 0 0 −1 1 0 1 0 0 1 −1 0 0 −1
8 0 0 0 −1 −1 −1 −1 0 0 0 0 1 1 1 1
9 0 −1 1 −1 −1 0 0 0 0 −1 1 1 −1 0 0
10 1 0 −1 −1 0 −1 0 0 1 0 −1 1 0 1 0
11 −1 1 0 −1 0 0 −1 0 −1 1 0 1 0 0 1
12 0 0 0 1 1 1 1 −1 −1 −1 −1 0 0 0 0
13 0 −1 1 1 1 0 0 −1 1 0 0 0 0 −1 1
14 1 0 −1 1 0 1 0 −1 0 −1 0 0 1 0 −1
15 −1 1 0 1 0 0 1 −1 0 0 −1 0 −1 1 0
422 Unitary and rotation groups

by forming a wedge product or bivector defined as ei ∧ ej = ei ⊗ ej − ej ⊗ ei =

−ej ∧ ei . Note that this implies that ei ∧ ei = 0. In general, the number of basis
vectors existing in a fully antisymmetric subspace of Lr (UN ) is N!/(r!(N–r)!).
For a second-order space there are therefore N(N–1)/2 basis vectors. Signifi-
cantly for us, if we start with U = C4–that is, N = 4, a four-dimensional complex
space—then this results in six dimensions. This allows us the first part of our
objective by mapping from C4 to C6 as follows.
From the basis vectors in C4, e1 , e2 , e3 , and e4 , we first generate a six-
dimensional space with basis vectors ti generated from all possible wedge
products given by equation (A2.30):

t1 = e1 ∧ e2 t2 = e2 ∧ e3 t3 = e3 ∧ e1
(A2.30)
t4 = e3 ∧ e4 t5 = e1 ∧ e4 t6 = e2 ∧ e4

SU(4)–SO(6) homomorphism stage 2

We now consider a general 4 × 4 complex matrix A in C4, and note how it
maps into this new six-dimensional space C6. This can be obtained explicitly
as shown in equation (A2.31):
! ! !!
ei ∧ ej ⇒ [A] ei ∧ [A] ej = ari ei ∧ asj ej = ari asj er ∧ es
(A2.31)

This mapping corresponds to a 6 × 6 matrix W, the 36 elements of which are

derived from the 36 2 × 2 minors of A. In general, the ijth element of W is then
the minor formed from the ith and jth rows with the ith and jth columns, so, for
example, w11 = a11 a22 − a12 a21 , and so on.
Given any 4×4 unitary matrix U 4 , therefore, we can generate a 6×6 unitary
matrix by calculating the 36 2 × 2 determinants of U 4 ; that is, the 1,1 element
of U 6 is the 2 × 2 determinant from columns 1,2 and rows 1,2 of U 4 . Similarly,
the 1,2 element is formed from the determinant from columns 1,2 and rows 2,3,
and so on. In this way we achieve a mapping from C4 into C6.

SU(4)–SO(6) homomorphism stage 3

In general, r × r minors are involved in the exterior algebra Lr (UN ). A special
case arises when r = N , when there is only one basis vector, called the volume
element associated with the basis set e1 , e2 . . . eN . In our case we have, for
L4 (U4 ), the following result;

[A] e1 ∧ [A] e2 ∧ [A] e3 ∧ [A] e4 = det([A])e1 ∧ e2 ∧ e3 ∧ e4 (A2.32)

Hence when [A] ∈ SU(4) it has unit determinant and therefore when all four
terms in the volume element are distinct their coefficient is 1, whereas the
coefficient is 0 for all combinations with repeated indices.
This result is important, because we can now consider generation of a scalar,
the inner product of two bivectors x and y, as shown in equation (A2.33):
5 
x = 5 xi ti
⇒ x ∧ y = f (x, y)e1 ∧ e2 ∧ e3 ∧ e4
y = yj tj (A2.33)
f (x, y) = x1 y4 + x2 y5 + x3 y6 + x4 y1 + x5 y2 + x6 y3
 A2.4 Dynkin diagrams: classification of unitary and rotation groups 423

This can be shown quite easily by explicit expansion using the basis vectors
defined in equation (A2.30). Note that there is a cyclic permutation of ordering
in y, and so to obtain a scalar product in C6 we consider not matrix products such
as WT W but must also include a permutation matrix P, as shown in equation
(A2.34), where I 3 is the 3 × 3 identity matrix:
 
0 I3
[P] = ⇒ [W ]T [P][W ] (A2.34)
I3 0

Since we are considering SU(4), the matrix product will then have 1 in positions
14,25 and so on, where four distinct basis vectors occur and 0 elsewhere. In
other words, the following matrix identity holds:

[W ]T [P][W ] = [P] (A2.35)

This result introduces the permutation matrix P, which has the useful property
that P2 = I 6 and is central to the next and final stage of our mapping procedure.
Recall that so far we can go from a 4 × 4 unitary matrix A to a 6 × 6 matrix W,
which is also unitary. (To prove this, use the fact that A maps to W, A∗T maps to
W∗T , and then finally, I 4 maps to I 6 , to show that W∗T W = I 6 .) However, we
seek a mapping from SU(4) to SO(6), which we know exists from the associated
Dynkin diagrams. We now therefore seek a similarity transformation to convert
W into a real orthogonal matrix; that is, we seek a 6 × 6 matrix Q such that
R = QWQ−1 is real orthogonal. We can construct such a matrix using P, as
follows.
We start by defining a symmetric matrix Q = QT , as shown in equation
(A2.36):

[Q]2 = [P] ⇒ [Q]4 = [P]2 = [I6 ] ⇒ [Q]−1 = [Q]3 (A2.36)

The reason for this choice becomes clear when we consider testing whether R
is orthogonal, as shown in equation (A2.37). We see that by using the properties
in equation (A2.36) we guarantee that R is orthogonal, as required.
 T 
[R]T [R] = QWQ−1 QWQ−1 = Q3 W T PWQ3 = Q3 PQ3 = I6 (A2.37)

Technically, R could still be complex orthogonal, but we can show that it is real
orthogonal by demonstrating that it is unitary as well as orthogonal; that is, that
R∗T R = I 6 . This follows from a similar expansion to that shown in equation
(A2.37). Hence R is real orthogonal as required.
The matrix Q can then be defined explicitly, as shown in equation (A2.38):
       
(1 − i) I3 iI3 I iI3 I iI3 0 I3
Q= ⇒ Q2 = −i 3 · 3 = =P
2 iI3 I3 iI3 I3 iI3 I3 I3 0
(A2.38)

This finally brings us to a general algorithm for mapping elements of SU(4)

into corresponding elements of SO(6). We start by generating the 6 × 6 unitary
424 Unitary and rotation groups

matrix W = U 6 from stage 2. If U 6 is then partitioned as shown in equation

(A2.39),
 
A B
U6 = (A2.39)
C D

where A, B, C, and D are 3 × 3 sub-matrices, then we can always generate a

6 × 6 real orthogonal matrix by the following transformation:
 
1 − i I3 iI3
Q=
2 iI3 I3
 
−1 i B − C + i(A + D) A − D + i(B + C)
⇒ O6 = QU6 Q =
2 D − A + i(B + C) C − B + i(A + D)
(A2.40)

For example, consider the element of U 4 given by a simple phase shift between
elements in C4, as shown on the left of equation (A2.41). This maps into an
‘equivalent’ 6 × 6 real orthogonal matrix O6 , as shown. The generator for this
U 4 matrix is –ia15 (see equation (A2.7)), and it maps into a rotation in the 2,5
plane. This is shown in equation (A2.7) as a subscript of the form [. . . ]25 .
Equation (A2.42) shows a second example: mapping a (real) 1,6 plane rota-
tion into a (complex) unitary matrix with generator –ia6 . This procedure can
be extended to all fifteen of the generators of SU(4). The plane rotations cor-
responding to each of the fifteen elements of su(4) are shown as subscripts in
equation (A2.7).
 
 iφ  1 0 0 0 0 0
e 0 0 0 0 cos 2φ 0 0 − sin 2φ 0
 
 0 e−iφ 0 0  0 0 1 0 0 0
U4 =    ⇒ O6 =   
0 0 e−iφ 0  0 0 0 1 0 0
0 0 0 eiφ 0 sin 2φ 0 0 cos 2φ 0
0 0 0 0 0 1
(A2.41)
 
cos φ 0 0 0 0 − sin φ
 0 1 0 0 0 0 
 
 0 0 1 0 0 0 
O6 =   0

 0 0 1 0 0 
 0 0 0 0 1 0 
sin φ 0 0 0 0 cos φ
 
cos φ2 0 0 sin φ2
 
 0 cos φ2 i sin φ2 0 
⇒ U4 =    (A2.42)

 0 i sin φ2 cos φ2 0 
φ
− sin 2 0 0 cos φ2
Coherent stochastic
signal analysis A3
In any discussion of noise processes, prime consideration is usually given to
Gaussian random variables. Their importance stems from the Central Limit
Theorem, which briefly states that the sum of a large number of independent
and identically distributed random variables will be normally distributed. Due
to the generic nature of this theorem, Gaussian statistics are often encoun-
tered in random wave propagation and scattering problems. (Some aspects of
non-Gaussian statistics have been explored for polarised waves; see, for exam-
ple, Bates (1998).) In this Appendix we briefly summarize the main impact of
such stochastic models on coherent signal analysis, particularly on phase and
coherence statistics (Touzi, 1999; Lee, 1994b, 2008; Lopez-Martinez, 2005;
Ferro-Famil, 2008).
Signals generated by Gaussian random processes are characterized by sta-
tistical and not deterministic measures. Consequently, any single sample of
such a process essentially contains zero information, and it is only by obtaining
multiple samples and forming sums or integrals that the signal can be charac-
terized by its statistical moments. The value of x for any particular sample is
then independent of any previous values, and is taken from a normal distribu-
tion such that it is characterized by a probability density function (pdf) p(x), as
summarized in equation (A3.1):
 ∞



 x = E(x) = x.p (x) dx = m




1 (x−m)2
 −∞
p (x) = G(m, σ ) = √ e− 2σ 2 ⇒ ∞
σ 2π 


x =
2 x2 .p (x) dx = σ 2 + m2



 −∞

σx2 = E((x − m)2 ) = σ 2
(A3.1)

For Gaussian signals the process is fully characterized by the mean m and
standard deviation σ . Pure noise signals are often characterized by a zero mean
process, written in shorthand as G(0,σ ). Hence Gaussian noise has only one
free parameter, σ . Figure A3.1 shows an example of a signal composed of
256 samples taken from a G(0,1) random number generator. Such generators
provide the basis for modelling multi-channel polarimetric and interferometric
signals, and are a useful way to investigate the statistical properties of wave
depolarisation, as we show later in this Appendix. First, however, we need to
extend the simple scalar model of equation (A3.1) to complex signals, and then
to the multidimensional complex signal vectors encountered in polarimetry.
426 Coherent stochastic signal analysis

2.5

1.5

0.5

Signal level
0

–0.5

–1

–1.5

–2

–2.5
50 100 150 200 250
Fig. A3.1 256 noise signal samples gener-
ated from a G(0,1) process Sample number

For complex signals, with amplitude and phase, as encountered in wave prop-
agation and scattering, we must extend these ideas to account for both real and
imaginary components of the signal. This is summarized in equation (A3.2).
One of the most important relations in equation (A3.2) is that the expectation of
the product between real and imaginary parts is zero. This is just a consequence
of the fact that noise carries zero phase information, and hence the real and
imaginary parts are independent Gaussian random variables. The phase, there-
fore, has a uniform distribution in the range 0 to 2π , while the intensity, defined
as shown in equation (A3.2), has an exponential distribution. The amplitude
A or square root of the intensity has a Rayleigh distribution as shown. Both
have large variances, and so some care is required to minimize errors when
estimating parameters from real data.

 σ
n = nI + inQ ∈ GC (0, σ ) ⇒ nI ,Q = G 0,
2


 E(nI ) = E(n Q) = 0

E nI nQ = 0
⇒ 

 σ2
E(n2 ) = E(n22 ) = ⇒ E n2 = σ 2
I
2
nI ⇓
/
1 − I2 E(I) = σ 2
p(I = nI + nQ ) = 2 e σ ⇒
2 2
nR σ var(I ) = σ4
 √
 π
√ 2A −A22  E(A) = σ
p(A = I ) = 2 e σ ⇒ 2
var(I ) = (4 − π )σ
 2
σ
4
(A3.2)
 Coherent stochastic signal analysis 427

As a measure of the fluctuations in such data, the coefficient of variation (CV)

can be defined as shown in equation (A3.3):

1 standard deviation
CV = = (A3.3)
ENL mean

This coefficient is also related—as shown in equation (A3.3)—to the ‘effective

number of looks’, or ENL, which is widely used in SAR image analysis. We
see, for example, that for exponentially distributed data as in equation (A3.2),
the ratio CV = 1. This emphasises that these fluctuations are not due to thermal
noise. As the signal strength increases, so its variance also increases to keep
CV = 1. Such fluctuations therefore cannot be reduced by increasing signal
power. These fluctuations are common to all types of coherent imaging, where
they are termed speckle noise (Lee, 1994a, 2008).
The simplest way to reduce speckle—the variance of the estimate, and there-
fore reduce the CV (increase ENL)—is to employ multi-look averaging. Here L
independent samples are summed to obtain an estimate of the mean intensity. In
this case the intensity distribution has a chi-square distribution with 2L degrees
of freedom, as shown in equation (A3.4):

/
LL I L−1 −LI E(I ) = σ 2
p(I ) = e σ2 I ≥0⇒ (A3.4)
var(I ) = σL
4
(L − 1)!σ 2L

√
The ratio of the standard deviation to the mean is then reduced to 1/ L. This
observation forms the basis for the design of speckle filters in SAR imaging.
For example, the Lee filter (Lee, 1994a) proposes that we estimate the intensity
of a pixel Î using a mixture of the pixel value itself, I , and its local mean, I
(usually estimated locally using a small M × N window centred on the pixel),
based on a local Taylor expansion for the intensity of a pixel of the form shown
in equation (A3.5):

 CV 2 − 1
I = I + k(I − I ) ⇒ k = L
(A3.5)
CV 2

If the area is homogeneous (called fully developed speckle in the SAR context)
then from equation (A3.5), k = 0, and the mean value is taken. However, if
the region is very heterogeneous (a point target, for example) then CV will be
much greater than 1, and so k = 1, and the local value is kept. In this way,
local statistics can be used to strike a balance between spatial and radiometric
resolution. Note that application of this approach to image data relies on two
key assumptions: ergodicity in the mean, and local wide-sense stationarity—
that space and time averages converge to the same mean value so that the spatial
averaging locally around a SAR pixel can be considered equivalent to obtaining
multiple samples of the same random process.
We have seen in Chapter 1 that polarimetry involves a two-dimensional com-
plex space C2. Hence we need to take one further step in our characterization of
noise by considering the statistical properties of signals in C2: pairs of complex
428 Coherent stochastic signal analysis

signals of the form shown in equation (A3.6):

s1 = s1I + is1Q ∈ Gc (0, σ )

(A3.6)
s2 = s2I + is2Q ∈ Gc (0, σ )

We have seen, for example, that the product s1 · s2 * arises in many applica-
tions. This product is important, because the conjugate sign implies the phase
difference between signals 1 and 2. Hence, while s1 and s2 individually have
random phase, the phase difference can still be deterministic. Noise processes
in C2 must therefore be further specified by the following added constraints:
 
E s1 s1∗ = σ 2 E(s1 s2∗ ) = 0 E s2 s2∗ = σ 2 (A3.7)

where again the zero expectation of the cross terms forces the phase difference
to be uniformly random. We can summarize these properties of noise signals in
C2 by generating a 2 × 2 covariance matrix from the expectation of the outer
product of a vector in C2 with its conjugate, as shown in equation (A3.8):
  
s1  ∗
[C]noise =E . s1 s2
s2
   2   
E(s1 s1∗ ) E(s1 s2∗ ) σ 0 2 1 0
= = = σ (A3.8)
E(s2 s1∗ ) E(s2 s2∗ ) 0 σ2 0 1

To generalize the above discussion we must consider signals where the cross
expectation is not zero. In this regard, one of the most useful relationships is
the Schwarz inequality, which can be formulated as shown in equation (A3.9):
 2
 b  b b
 
 s1 (x) .s2 (x) dx ≤
∗
|s1 (x)| dx
2
|s2 (x)|2 dx (A3.9)

 
a a a

where the equality only-holds if s1 (x) = k ·s , k ∈ C. Using more compact

+ 2 (x)
notation we can write s1 (x) s2∗ (x)dx = s1 s2∗ , and with this, equation (A3.9)
can be rewritten as shown in equation (A3.10):
+ ∗ ,
 s1 s 
2
0 ≤ γ = + , + ∗, ≤ 1 (A3.10)
∗
s1 s1 s2 s2

This ratio of integrals is called the coherence γ between signals s1 and s2 . From
equation (A3.8) we see that the coherence is always zero for noise signals, while
for polarised EM waves it follows that we can always write Ex = kEy for some
complex constant k, and the coherence of polarised waves is always unity.
Furthermore, as the mean phase of s1 s2 * may not be zero, it is convenient to
define the complex coherence as shown in equation (A3.11):
+ ∗,
s1 s2
γ̃ = γ e iφ
= + ,+ , (A3.11)
s1 s1∗ s2 s2∗
 Coherent stochastic signal analysis 429

90 1
120 60
0.8

0.6
150 P
30
0.4

0.2


180 0

210 330

240 300
Fig. A3.2 Unit circle in the complex coher-
270 ence plane

This has a magnitude between 0 and 1 and a phase from 0 to 2π, hence we
can represent the coherence as a point P inside the unit circle of the complex
coherence plane as shown in FigureA3.2. Noise sits at the origin of this diagram,
while coherent signals lie around the outer unit circle.
Coherence is a ratio of random variables and hence is a stochastic quantity,
so attention must be paid to its statistics. In the most general case where corre-
lation is allowed between the signals s1 and s2 , the probability density function
becomes a multivariate Gaussian of the form shown in equation (A3.12):
 
s 1 ∗T −1
u = 1 ⇒ p(u) = 2 e−u [C] u (A3.12)
s2 π det([C])

where [C] = E(u · uT ∗ ) is the 2 × 2 Hermitian covariance matrix defined as

=  > +s s∗ , +s s∗ ,
s1  ∗ ∗  1 1 1 2
[C] = · s1 s2 = + , + ∗, det([C]) ≥ 0 (A3.13)
s2 ∗
s2 s1 s2 s2

Note that if s1 = ks2 then det([C]) = 0, and the density function must be
replaced by a delta function at u = u0 . The single-look density function for the
phase of s1 s2 * now has the following form:
0 1
1 − γ2 1 − ψ 2 + ψ π − cos−1 ψ
P(φ) = 1.5
2π 1 − ψ 2
ψ = γ cos φ −π ≤φ ≤π (A3.14)

which we note is a function of the underlying coherence γ . As the coher-

ence reduces so the width of this distribution increases and the noise variance
increases. Hence coherence and phase variance are closely related: low coher-
ence leads to high variance, and in the limit of unit coherence the phase variance
falls to zero and the phase becomes a deterministic parameter. Again, multi-look
430 Coherent stochastic signal analysis

averaging can be used to reduce the variance of the estimates for any given γ .
We can therefore define the maximum likelihood estimate of [C], denoted [Z],
as shown in equation (A3.15):

1!
L
[Z] = uk u∗T
k (A3.15)
L
j=1

The matrix [Z] is then itself a random matrix with a probability distribution;
the complex Wishart distribution—a function of the number of samples L and
of the general form shown in equation (A3.16) (Lee, 1994b, 2008; Conradsen,
2003):

LLq det([Z])L−q exp(−L.Trace([C]−1 [Z]))

pL ([Z]) =
K(L, q) det([C])L (A3.16)
K(L, q) = π 0.5q(q−1)
(L).. (L − q + 1)

Here q is the dimension of the complex vector u (2 in this case, but 3 for
monostatic polarimetry, 4 for bistatic, and 6 for single baseline polarimetric
interferometry and so on). Equation (A3.4) is a special case of this distribution
for q = 1. This distribution then leads to the following pdf for the phase as a
function of the number of looks L:
  
 L + 12 (1 − γ 2 )L ψ (1 − γ 2 )L 1 2
P(φ) = √ L+ 1 + F L, 1 : ; ψ
2 π  (L) 1 − ψ 2 2 2π 2

ψ = γ cos (φ − φm ) (A3.17)

where F is a Gauss hypergeometric function, and φm is the mean phase. Com-

bining these ideas, the maximum likelihood sample complex coherence is often
directly used as shown in equation (A3.18):

5
L
∗
s1i s2i /
i=1 0 ≤ |γ̃ | ≤ 1
γ̃ = " " (A3.18)
5
L 5
L 0 ≤ arg(γ̃ ) = φ̂m < 2π
∗
s1i s1i ∗
s2i s2i
i=1 i=1

The pdf of the sample coherence magnitude g = |γ̃ | for jointly Gaussian signals
can be derived analytically, and is a function of the coherence magnitude γ , the
number of integrated independent samples, L, and the hypergeometric function
F, as shown in equation (A3.19):

p (g, γ ) = 2(L − 1)(1 − γ 2 )L g(1 − g 2 )L−2 F(L, L; 1; g 2 γ 2 ) (A3.19)

from which the moments of order k can be deduced as shown in equation

(A3.20):

(L) 1 + k2  
k k N
mk =  3 F2 1 + , L, L; L + , 1; γ
2
1 − γ2 (A3.20)
 L + k2 2 2
 Coherent stochastic signal analysis 431

where p Fq is the generalized hypergeometric function. Of particular interest is

the expression for the first moment of g, shown in equation (A3.21):
  
(L) 1 + 12 3 1 N
E(g) =  3 F2 , L, L; L + , 1; γ 2
1 − γ2 (A3.21)
 L+ 2 1 2 2

This shows a bias towards higher coherence values, especially for low coherence
with a small number of samples L. The variance of the estimate can also be
derived from equation (A3.22) using equations (A3.20) (k = 2) and (A3.21).

var(g) = E g 2 − E(g)2 (A3.22)

Useful as these expressions are, they are difficult to interpret without detailed
calculations. For this reason the Cramer–Rao (CR) lower bounds on variance
of phase and coherence have also been derived (see Seymour (1994) and yield
simpler expressions, as shown in equation (A3.23):
2
1 − γ2 1 − γ2
var(φ) > var(g) > (A3.23)
2Lγ 2 2L
Figure A3.3 shows examples of these CR bounds as a function of coherence
and increasing number of looks. Generally, the higher the coherence, the lower
the number of looks required to obtain a specified variance.
Our representation of coherence inside the unit circle of Figure A3.2 is there-
fore rather misleading. In fact, each point P has a minimum cloud of uncertainty
around it representing the Cramer–Rao bounds in radius (coherence) and polar
angle (phase) fluctuations. Note that this cloud will be elliptical in shape. For a
given number of looks L the phase variance is larger than the radial coherence
variance. Figure A3.4 shows a schematic representation of this concept. This
result underpins our distinction in Chapters 7 and 8 between coherence loci and
associated coherence regions.

Phase variance Coherence variance

90 0.14

L=4
80 L = 16
0.12
L = 64
70
0.1
60
Variance (degrees)

50 0.08
Variance

40 0.06

30
0.04
20

0.02
10

0 0 Fig. A3.3 Cramer–Rao bounds on fluctua-

0 0.5 1 0 0.5 1 tions in complex coherence estimates (phase
Coherence Coherence (left) and coherence amplitude (right)
432 Coherent stochastic signal analysis

90 1
120 60
0.8

0.6
150 30
0.4

0.2


180 0

210 330

240 300
Fig. A3.4 Complex coherence as a stochastic
variable inside the unit circle 270

Having established the general properties of coherence and stochastic signals

in C2, we now turn to consider the special case of depolarisation effects in
polarimetry. Equation (A3.16) represents a general expression for the analysis
of fluctuation statistics for coherency matrices of arbitrary dimension q. For
example, q = 1 represents a scalar channel, when the distribution reduces to
the gamma distribution for L > 1, with the special case of the exponential
for L = 1. Dual polarised systems based on the wave coherency matrix [J ]
and single polarisation radar interferometry are both examples represented by
q = 2. Polarimetry based on the full scattering matrix requires either q = 3
for reciprocal backscatter, or q = 4 for bistatic scattering. When we consider
extension to polarimetric interferometry, then the dimension increases to q =
MN, where M −1 is the number of baselines and N the number of polarisations.
In these higher-dimensional cases, analytical manipulation to find marginal
distributions for phase parameters is difficult (Lopez Martinez 2005). It is then
useful to employ numerical investigations based on Monte Carlo simulations
using Gaussian random number generators. To see this, we start with a reference
or desired MN dimensional coherency matrix MN . This positive semi-definite
Hermitian matrix can always be expressed in terms of its eigenvalue/eigenvector
decomposition, as shown in equation (A3.25):
 
λ1 0 0 0
 0 λ2 0 0 
 
[MN ] = [UMN ] [DMN ] [UMN ]∗T [DMN ] =  
 0 0 ... 0 
0 0 0 λMN
(A3.25)

This can then be used to generate a sequence of random N -dimensional complex

sample vectors u, all of which have a coherency matrix equal to  (in the limit
of an infinite number of samples). We can generate such a numerical sequence
using MN sets of pairs of G(0, σ ) random number generators, as shown in
 Coherent stochastic signal analysis 433

equation (A3.26):
 
e1 0 0
 ..  2  3
u = [UMN ] [E] =  0 . 0  ei = λi Ga 0, 12 + iGb (0, 12 )
0 0 eMN
0 1 !L
L→∞
ˆ MN = uu∗T −−−−→ [MN ] (A3.26)
i=1

We start by generating two (independent) real random sequences Ga and Gb as

shown, then combine them into a complex series before scaling by the square
root of the appropriate eigenvalue of . This process is then repeated MN times
for each eigenvalue, to obtain a set of MN complex series. Finally, we introduce
the complex correlations between samples by multiplying by the matrix of
eigenvectors [UMN ]. The vector u then has the property that its coherency matrix
u · u∗T  converges to . This provides us with a practical way to generate test
sequences in polarisation statistics and depolarisation studies.
Very often in applications we make a measurement of a scattering matrix
(or vector k) and wish to determine which class it belongs to from a set of
preselected reference states. This comparison process is made complicated by
the stochastic nature of such measurements. For example, if k is complex normal
distributed then an individual sample may not correspond exactly to the correct
class mean, and there will be some natural fluctuation. One way to deal with this
is to employ a maximum likelihood (ML) approach. According to this we assign
a sample to the class with the maximum probability. To do this we first need to
assume a distribution (multivariate normal, for example), and then characterize
each reference state by the parameters of this distribution. In the normal case
this is just the covariance matrix [C], as shown in equation (A3.27):
 
s1
 ..  1 ∗T
[C]−1 u
u =  .  ⇒ p(u) = e−u (A3.27)
π q det([C])
sq

Each class is then characterized by a q × q class covariance matrix [Ci ], which

we must calculate or measure before the comparison takes place. We then take
the measured vector k and compare it to all the class matrices. Geometrically this
reduces to a distance measure between the sample vector and class covariance.
As the normal distribution involves the exponential function, it is common to
consider distances based on the so-called log-likelihood function, obtained from
the normal distribution by taking the natural logarithm, as shown in equation
(A3.28), where we have used the cyclic property of the trace operation to
simplify the centre term.

− ln |Ci | − Tr(Ci−1 kk ∗T ) − q ln π (A3.28)

From this we can define a non-negative distance measure such that we assign
k to the class with the shortest distance d , defined from equation (A3.28) by
434 Coherent stochastic signal analysis

ignoring elements that do not depend on the class, as shown in equation (A3.29):

d k, Ci = ln |Ci | + Tr(Ci−1 kk ∗T ) (A3.29)

This is formally a measure of the ‘closeness’ of k to class i. It forms the

basis for image classification and hypothesis testing in radar polarimetry and
interferometry (Lee, 2008).
In the depolarising case we may wish to compare not a single k vector but an
average coherency or covariance matrix C itself. A distance measure for this
case can be obtained in a similar fashion to equation (A3.29), but starting from
the complex Wishart distribution of equation (A3.16), and again forming the
log-likelihood function and ignoring constant terms to obtain equation (A3.30):

d (C, Ci ) = ln |Ci | + tr(Ci−1 C) (A3.30)

Note that since the Mueller matrix [M ] can be mapped 1–1 with the scattering
coherency matrix [T ], which itself is unitarily similar to [C], the metric in
equation (A3.30) is invariant to use of [C] or [T ]. In this way we can also
provide a statistical distance measure between experimental Mueller matrices.
Bibliography

Abhyankar, K. D. and Fymat, A. L. (1969). Relations between the elements of

the phase matrix for scattering. Journal of Math. Phys., 10, 1935–1938.
Ablitt, B. P., Hopcraft, K. I., Turpin, K. D., Chang, P. C. Y. and Walker,
J. G. (1999). Imaging and multiple scattering through media containing
optically active particles. Waves in Random Media, 9, 561–572.
Ablitt, B. P. (2000). Characterisation of Particles and their Scattering Effects
on Polarized Light. PhD thesis, University of Nottingham, UK.
Ainsworth, T. L., Ferro-Famil, L. and Lee, J. S. (2006). Orientation angle
preserving a posteriori polarimetric SAR calibration. IEEE Transactions
on Geoscience and Remote Sensing, 44 (4), 994–1003.
Ainsworth, T. L., Preiss, M., Stacy, N., Nord, M. and Lee, J. S.
(2007). Analysis of compact polarimetric SAR imaging modes. Pro-
ceedings of the Third ESA Workshop on Polarimetry and Polarimet-
ric Interferometry, POLInSAR 2007, Frascati, Italy, January 2007.
http://earth.eas/int/workshops/polinsar2007/
Allain, S. (2003). Caractérisation d’un Sol nu à partir de données SAR
Polarimétriques: Etude Multi-fréquentielle et Multi-résolutions. PhD
thesis, University of Rennes, France.
Anderson, D. G. M. and Barakat, R. (1994). Necessary and sufficient conditions
for a Mueller matrix to be derivable from a Jones matrix. J. Opt. Soc. Am.
A, 11 (8), 2305–2319.
Askne, J., Dammert, P. B., Ulander, L. M. and Smith, G. (1997). C-band repeat
pass interferometric SAR observations of the forest. IEEE Transactions
on Geoscience and Remote Sensing, 35, 25–35.
Askne, J., Santoro, M., Smith, G. and Fransson, J. E. S. (2003). Multitemporal
repeat-pass SAR interferometry of boreal forests. IEEE Transactions on
Geoscience and Remote Sensing, 41, 1540–1550.
Askne, J. and Santoro, M. (2007). Selection of forest stands for stem volume
retrieval from stable ERS tandem InSAR observations. IEEE Geoscience
and Remote Sensing Letters, 4, 46–50.
Attema, E. P. and Ulaby, F. T. (1978). Vegetation modeled as a water cloud.
Radio Science, 13, 357–364.
Azzam, R. M. (1978). Propagation of partially polarised light. J. Opt. Soc. Am.,
68, 1756–1767.
Azzam, R. M. A. and Bashara, N. M. (1987). Ellipsometry and Polarized Light.
North–Holland.
436 Bibliography

Ballester-Berman, J. D., Lopez-Sanchez, J. M. and Fortuny-Guasch, J. (2005).

Retrieval of biophysical parameters of agricultural crops using polarimet-
ric SAR interferometry. IEEE Transactions on Geoscience and Remote
Sensing, 43 (4), 683–694.
Ballester-Berman, J. D. and Lopez-Sanchez, J. M. (2007). Coherence loci
for a homogeneous volume over a double-bounce ground return. IEEE
Geoscience and Remote Sensing Letters, 4 (2), 317–321.
Bamler, R. (1992). A comparison of range-Doppler and wavenumber domain
SAR focusing algorithm. IEEE Transactions on Geoscience and Remote
Sensing, 30, 706–713.
Bamler, R. and Hartl, P. (1998). Synthetic aperture radar interferometry. Inverse
Problems, 14, R1–R54.
Barakat, R. (1981). Bilinear constraints between the elements of the 4 × 4
Mueller–Jones matrix of polarization theory. Opt. Comms., 38, 159–161.
Barakat, R. (1987). Conditions for the physical realisability of polarization
matrices characterising passive systems. J. Mod. Optics, 34, 1535–
1544.
Bates, A. P., Hopcraft, K. I. and Jakeman, E. (1998). Non-Gaussian fluctuations
of Stokes parameters in scattering by small particles. Waves in Random
Media, 8, 235–253.
Baum, C. and Kritikos H. N. (eds.) (1995). Electromagnetic Symmetry. Taylor
and Francis, Washington.
Bessette, L. A. and Ayasli, S. (2001). Ultra wide band P-3 and Carabas II
foliage attenuation and backscatter analysis. Proceedings of IEEE Radar
Conference, 357–362.
Bickel, S. H. and Bates, R. H. T. (1965). Effects of magneto-ionic propagation
on the scattering matrix. Proc. IEEE, 53 (8), 1089–1091.
Bickel, W. S. and Bailey W. M. (1985). Stokes vectors, Mueller matrices and
polarized scattered light. American Journal of Physics, 53, 468–478.
Bicout, D. and Brosseau C. (1992). Multiply scattered waves through a spa-
tially random medium: entropy production and depolarization. J. Phys.
I. France, 2, 2047–2063.
Boerner, W. M. (1981). Polarization dependence in electromagnetic inverse
problems. IEEE Trans. Antennas and Propagation, AP-29, 262–274.
Boerner, W. M. (ed.) (1992). Direct and Inverse Methods in Radar Polarimetry,
Parts 1 and 2. NATO ASI Series C: Mathematical and Physical Sciences,
Vol. 350, Kluwer.
Borgeaud, M. and Noll, J. (1994). Analysis of theoretical surface scatter-
ing models for polarimetric microwave remote sensing of bare soils.
International Journal of Remote Sensing, 15 (14), 2931–2942.
Born, M. and Wolf E. (1998). Principles of Optics, Chapters 1 and 10. Pergamon
Press, sixth edition.
Brosseau, C. (1990). Analysis of experimental data for Mueller polarization
matrices. OPTIK, 85, 83–86.
Brosseau, C. and Bicout, D. (1994). Entropy production in multiple scattering
of light by a spatially random medium. Phys. Rev. E, 50, 4997–5005.
Brosseau, C. (1998). Fundamentals of Polarized Light: a Statistical Approach.
Wiley.
 Bibliography 437

Byrne, J. (1971). Classification of electron and optical polarization transfer

matrices. J. Phys. B, 4, 940–953.
Cafforio, C., Pratti, C. and Rocca, F. (1991). SAR data focusing using seismic
migration techniques. IEEE Trans. Aerospace and Electronic Systems,
27, 194–205.
Cameron, W. L., Youssef, N. N. and Leung, L. K. (1996). Simulated polari-
metric signatures of primitive geometrical shapes. IEEE Transactions on
Geoscience and Remote Sensing, 34 (3), 793–803.
Cartan, E. (1966). The Theory of Spinors. Dover Press.
Chandrasekhar, S. (1960). Radiative Transfer. Dover Press.
Chen, H. C. (1985). Theory of Electromagnetic Waves: a Coordinate-Free
Approach. McGraw-Hill.
Cloude, S. R. (1985). Radar target decomposition theorems. IEEE Letters, 21
(1), 22–24.
Cloude, S. R. (1986). Group theory and polarization algebra. OPTIK, 75 (1),
26–36.
Cloude, S. R. (1989). Physical realisability of matrix operators in polarime-
try. SPIE, 1166, Polarization Considerations for Optical Systems II,
pp. 177–185.
Cloude, S. R. (1995a). An Introduction to Electromagnetic Wave Propagation
and Antennas. UCL Press.
Cloude, S. R. (1995b). Lie groups in EM wave propagation and scattering.
In Baum, C. and Kritikos, H. N. (eds.), Electromagnetic Symmetry,
Chapter 2. Taylor and Francis, Washington.
Cloude, S. R. and Pottier, E. (1995c). The concept of polarisation entropy in
optical scattering. Optical Engineering, 34 (6), 1599–1610.
Cloude, S. R. and Pottier, E. (1996). A review of target decomposition theo-
rems in radar polarimetry. IEEE Transactions on Geoscience and Remote
Sensing, 34 (2), 498–518.
Cloude, S. R. and Pottier, E. (1997a). An entropy based classification scheme for
land applications of polarimetric SAR. IEEE Transactions on Geoscience
and Remote Sensing, 35 (1), 68–78.
Cloude, S. R. and Papathanassiou, K. P. (1997b). Polarimetric optimisation in
radar interferometry. Electronics Letters, 33 (13), 1176–1178.
Cloude, S. R. and Papathanassiou, K. P. (1998). Polarimetric SAR interferome-
try. IEEE Transactions on Geoscience and Remote Sensing, GRS-36 (5),
1551–1565.
Cloude, S. R., Fortuny, J., Lopez, J. M. and Sieber, A. J. (1999). Wide
band polarimetric radar inversion studies for vegetation layers. IEEE
Transactions on Geoscience and Remote Sensing, 37/2 (5), 2430–2442.
Cloude, S. R., Papathanassiou, K. P. and Boerner, W. M. (2000a). The remote
sensing of oriented volume scattering using polarimetric radar inter-
ferometry. Proceedings of International Symposium on Antennas and
Propagation, ISAP 2000, Fukuoka, Japan, 549–552.
Cloude, S. R., Papathanssiou, K. P. and Reigber, A. (2000b). Polarimetric SAR
interferometry at P band for vegetation structure extraction. Proceedings
of the Third European SAR Conference, EUSAR 2000, Munich, Germany,
249–252.
438 Bibliography

Cloude, S. R., Papathanssiou, K. P. and Boerner, W. M. (2000c). A fast method

for vegetation correction in topographic mapping using polarimetric radar
interferometry. Proceedings of the Third European SAR Conference,
EUSAR 2000, Munich, Germany, 261–264.
Cloude, S. R. (2001a). A new method for characterising depolarisation effects
in radar and optical remote sensing. Proceedings of the IEEE Inter-
national Geoscience and Remote Sensing Symposium, IGARSS 2001,
Sydney, Australia, 2, 910–912.
Cloude, S. R., Papathanassiou, K. P. and Pottier, E. (2001b). Radar polarime-
try and polarimetric interferometry. IEICE Transactions on Electronics,
E84-C (12), 1814–1822.
Cloude, S. R., Woodhouse, I. H., Hope, J., Suarez Minguez, J. C., Osborne, P.
and Wright G. (2001c). The Glen Affric Radar Project: forest mapping
using dual baseline polarimetric radar interferometry. ESA Symposium
on Retrieval of Bio and Geophysical Parameters from SAR for Land
Applications, University of Sheffield, England, 333–338.
Cloude, S. R. and Corr, D. G. (2002a). A new parameter for soil moisture esti-
mation. Proceedings of the IEEE International Geoscience and Remote
Sensing Symposium, IGARSS 2002, Toronto, Canada, 1, 641–643.
Cloude, S. R. (2002b). Helicity in radar remote sensing. Proceedings of the
IEEE International Geoscience and Remote Sensing Symposium, IGARSS
2002, Toronto, Canada, 1, 411–413.
Cloude, S. R. and Papathanassiou, K. P. (2003). A 3-stage inversion process for
polarimetric SAR interferometry. IEEE Proceedings, Radar, Sonar and
Navigation, 150 (03), 125–134.
Cloude, S. R., Corr, D. G. and Williams, M. L. (2004). Target detection beneath
foliage using polarimetric SAR interferometry. Waves in Random Media,
14 (2), S393–S414.
Cloude, S. R. and Williams, M. L. (2005a). The negative alpha filter: a new pro-
cessing technique for polarimetric SAR interferometry. IEEE Geoscience
and Remote Sensing Letters, 2, 187–191.
Cloude, S. R. (2005b). On the status of bistatic polarimetry theory. Proceedings
of IEEE Geoscience and Remote Sensing Symposium, IGARSS 2005,
Seoul, South Korea, 3, 2003–2006.
Cloude, S. R. (2006a). Information extraction in bistatic polarimetry. Pro-
ceedings of the Sixth European SAR Conference, EUSAR 06, Dresden,
Germany.
Cloude, S. R. (2006b). Polarization coherence tomography, Radio Science, 41,
RS4017.
Cloude, S. R. (2007a). Dual baseline coherence tomography. IEEE Geoscience
and Remote Sensing Letters, 4 (1), 127–131.
Cloude, S. R. (2007b). The dual polarization H/alpha decomposition. Pro-
ceedings of the Third ESA Workshop on Polarimetry and Polarimetric
Interferometry, POLInSAR 2007, Frascati, Italy.
Colin, E., Titin-Schnaider, C. and Tabbara, W. (2006). An interferometric
coherence optimization method in radar polarimetry for high-resolution
imagery. IEEE Transactions on Geoscience and Remote Sensing, 44 (1),
167–175.
 Bibliography 439

Colin, E. (2005). Apport de la Polarimétrie à l’Interférométrie Radar pour

l’Estimation des Hauteur’s de Cibles et de Paramétres de Forêt, PhD
thesis, Université de Paris.
Collet, E. (1993). Polarized Light. Marcel Dekker, New York.
Collin, R. E. (1985). Antennas and Radiowave Propagation. McGraw-Hill.
Conradsen, K., Nielsen, A. A., Schou, J. and Skriver, H. (2003). A test statistic
in the complex Wishart distribution and its application to change detection
in polarimetric SAR data. IEEE Transactions on Geoscience and Remote
Sensing, 41 (1). 4–19.
Cornwell, J. F. (1984). Group Theory in Physics, Vol. 1, ‘Techniques in physics’,
7. Academic Press.
Curlander, J. C. and McDonough, R. N. (1991). Synthetic Aperture Radar:
Systems and Signal Processing. Wiley Series in Remote Sensing.
Dall, J., Papathanassiou, K. P. and Skriver, H. (2003). Polarimetric SAR inter-
ferometry applied to land ice: first results. Proceedings of the IEEE
Geoscience and Remote Sensing Symposium, IGARSS ’03, Toulouse,
France, 3, 1432–1434.
Deschamps, G. A. (1951). Geometrical representation of plane polarized waves.
Proc. IRE, 39, 540.
Dobson, M. C., Ulaby, F. T., Hallikainen, M. and El-Rayes, M. A. (1085).
Microwave dielectric behaviour of wet soil: II Four component dielectric
mixing models. IEEE Transactions on Geoscience and Remote Sensing,
23, 35–46.
Dong, Y., Forster, B. C. and Ticehurst, C. (1998). A new decomposition of radar
polarization signatures. IEEE Transactions on Geoscience and Remote
Sensing, GRS-36, 933–939.
Dubois, P. C., van Zyl, J. J. and Engman, T. (1005). Measuring soil mois-
ture with imaging radars. IEEE Transactions on Geoscience and Remote
Sensing, GE-33, 916–926.
Ferro-Famil, L. and Pottier, E. (2000). Description of dual frequency polari-
metric data using Gell–Mann parameter set. Electronics Letters, 36 (19),
1646–1647.
Ferro-Famil, L., Pottier, E. and Lee, J. S. (2001). Unsupervised classifica-
tion of multifrequency and fully polarimetric SAR images based on the
H/A/Alpha–Wishart classifier. IEEE Transactions on Geoscience and
Remote Sensing, 39 (11), 2332–2342.
Ferro-Famil, L., Reigber, A., Pottier, E. and Boerner, W. M. (2003). Scene char-
acterization using subaperture polarimetric SAR data. IEEE Transactions
on Geoscience and Remote Sensing, 41 (10), Part 1, 2264–2276.
Ferro-Famil, L. and Neumann, M. (2008). Recent advances in the deriva-
tion of POLInSAR statistics: study and applications. Proceedings of the
Seventh European Conference on Synthetic Aperture Radar (EUSAR),
Friedrichshafen, Germany, 2, 143–146.
Flynn T., Tabb, M. and Carande, R. (2002). Coherence region shape estima-
tion for vegetation parameter estimation in POLINSAR. Proceedings of
IGARSS 2002, Toronto, Canada, V 2596–2598.
Franceschetti, G. and Linari, R. (1999). Synthetic Aperture Radar Processing.
Chapter 4. CRC Press.
440 Bibliography

Freeman, A. (1992). SAR calibration: a review. IEEE Transactions on

Geoscience and Remote Sensing, GE-30(6), 1107–1121.
Freeman, A. and Durden, S. L. (1998). A three component model for polari-
metric SAR data. IEEE Transactions on Geoscience and Remote Sensing,
GE-36, 963–973.
Freeman, A. (2004). Calibration of linearly polarized polarimetric SAR
data subject to Faraday rotation. IEEE Trans., GRS-42 (8), 1617–
1624.
Freeman, A. (2007). Fitting a two component scattering model to polarimetric
SAR data. IEEE Trans. GRS-42 (8), 2583–2592.
Fry, E. S. and Kattawar, G. W. (1981). Relationships between elements of the
Stokes matrix. Applied Optics, 20, 2811–2814.
Fung, A. K., Li, Z. and Chen, K. S. (199). Backscattering from a randomly
rough dielectric surface. IEEE Transactions on Geoscience and Remote
Sensing, 30, 356–369.
Gatelli, F., Monti Guarnieri, A., Parizzi, F., Pasquali, P., Prati, C. and Rocca,
F. (1994). The wavenumber shift in SAR interferometry. IEEE Trans.,
GRS-32, 855–865.
Gazdag, J. and Sguazzero, P. (1984). Migration of seismic data. Proceedings
of the IEEE, 72, 1302–1315.
Georgi, H. (1999). Lie Algebras in Particle Physics. Perseus Books.
Gershenfeld, N. (1999). The Nature of Mathematical Modeling. Cambridge
University Press.
Gil, J. J. and Bernabeu, E. (1985). A depolarization criterion in Mueller
matrices. Optica Acta, 32, 259–261.
Gil, J. J. and Bernabeu, E. (1986). Depolarization and polarization indices of
an optical system. Optica Acta, 33, 185–189.
Girgel, S. S., (1991). Structure of the Mueller matrices of depolarised optical
systems. Sov. Phys. Crystallogr., 36, 890–891.
Giuli. D. (1986). Polarization diversity in radars. Proceedings of the IEEE, 74,
245–269.
Givens, C. R. and Kostinski, A. B. (1993). A simple necessary and sufficient
condition on physical realizable Mueller matrices. J. Mod. Opt., 40,
471–481.
Goldstein, H. (1980). Classical Mechanics, second edition. Addison–
Wesley.
Graham, R. (1974). Synthetic interferometric radar for topographic mapping.
Proceedings of the IEEE, 62, 763–768.
Graves, C. D. (1956). Radar polarization power scattering matrix. Proceedings
of the IRE, 44 (2), 248–252.
Hagberg, J. O., Ulander, L. and Askne, J. (1995). Repeat-pass SAR interferom-
etry over forested terrain. IEEE Transactions on Geoscience and Remote
Sensing, 33 (2), 331–340.
Hajnsek, I., Papathanassiou, K. P. and Cloude, S. R. (2001). Removal of additive
noise in polarimetric eigenvalue processing. Proceedings of the IEEE
Symposium on Geoscience and Remote Sensing, IGARSS ’01, 6, 2778–
2780.
 Bibliography 441

Hajnsek, I., Pottier, E. and Cloude, S. R. (2003). Inversion of surface parameters

from polarimetric SAR. IEEE Transactions on Geoscience and Remote
Sensing, 41, 727–744.
Hajnsek, I., Kugler, F., Lee, S. K. and Papathanassiou, K. P. (2008). Tropical
forest parameter estimation by means of POLInSAR: the INDREX-II
Campaign. IEEE Transactions on Geoscience and Remote Sensing, 47,
481–493.
He, C. and Watson, G. A. (1997). An algorithm for computing the numerical
radius. IMA J. Numer. Anal., 17, 329–342.
Hecht, E. and Zajac, A. (1997). Optics. Third edition. Addison–Wesley.
Hopcraft, K. I. and Smith, P. R. (1992). An introduction to electromagnetic
inverse scattering. Developments in EM Theory and Applications, 7.
Hovenier, J. W. (1994). Structure of a general pure Mueller matrix. Applied
Optics, 33, 8318–8324.
Hovenier, J. W. and van der Mee, C. V. M. (1996). Testing scattering matri-
ces: a compendium of recipes. Journal of Quantitative Spectroscopy and
Radiative Transfer, 55, 649–661.
Hovenier, J. W., van der Mee, C.V.M. and Domke, H. (2004). Transfer of
Polarized Light in Planetary Atmospheres: Basic Concepts and Practical
Methods. Kluwer Academic Publishers, Astrophysics and Space Science
Library, Vol. 318.
Hunt, B. J. (1991). The Maxwellians. Cornell University Press.
Huynen, J. R. (1970). Phenomenological Theory of Radar Targets, PhD thesis,
Technical University, Delft, Netherlands.
Huynen, J. R. (1987). Phenomenological theory of radar targets. In Uslenghi,
P. L. E. (ed.), Electromagnetic Scattering, Academic Press, New York.
Imhoff, M. L. (1995). Radar backscatter and biomass saturation: ramifica-
tions for global biomass inventory. IEEE Transactions on Geoscience
and Remote Sensing, 33, 511–518.
Iniesta, J. C. del Toro (2003). Introduction to Spectropolarimetry. Cambridge.
Ioannidis, G. A. and Hammers, D. E. (1979). Optimum antenna polarizations for
target discrimination in clutter. IEEE Trans. Antennas and Propagation,
AP-27, 357–363.
Ishimaru, A. (1991). Electromagnetic Wave Propagation, Radiation and
Scattering. Prentice Hall International.
Jackson, J. D. (1999). Classical Electrodynamics. Third edition. Wiley.
Jin, Y. Q. and Cloude, S. R. (1994a). Numerical eigenanalysis of the coherency
matrix for a layer of random non-spherical scatterers. IEEE Transactions
on Geoscience and Remote Sensing, 32., 1179–1185.
Jin, Y. Q. (1994b). Electromagnetic Scatering Modelling for Quantitative
Remote Sensing. World Scientific Publishing.
Jones, D. S. (1989). Acoustic and Electromagnetic Waves. Oxford Science
Publications.
Jones, R. C. (1941). New calculus for the treatment of optical systems. J. Opt.
Soc. Am., 31, 488–493.
Jones, R. C. (1948). New calculus for the treatment of optical systems, VII:
Properties of the N-matrices. J. Opt. Soc. Am., 38, 671–685.
442 Bibliography

Kampes, B. M. (2006). Radar Interferometry: Persistent Scatterer Technique.

Kluwer.
Kennaugh, E. M. (1952). Polarization Properties of Radar Reflections. MSc
thesis, Electro-Science Laboratory, Ohio State University.
Kim, K., Mandel, L. and Wolf, E. (1987). Relationship between Jones and
Mueller matrices for random media. Journal of Opt. Soc. Am. A, 4,
433–437.
Kimura, H., Mizuno, T., Papathanassiou, K. P. and Hajnsek, I. (2004). Improve-
ment of polarimetric SAR calibration based on the Quegan algorithm.
Proceedings of the IEEE IGARSS 04 Symposium, 1.
Kong, J. A. (1985). Electromagnetic Wave Theory. Wiley.
Kong, J. A. (ed.) (1990). Polarimetric remote sensing. Progress in Electromag-
netics Research, PIER 3. Elsevier.
Konnen, G. P. (1985). Polarized Light in Nature. Cambridge University Press.
Kostinski, A. B. and Boerner, W. M. (1986). On foundations of radar
polarimetry. IEEE Trans. Antennas and Propagation, AP-34, 1395–1404.
Krieger, G., Papathanassiou, K. P. and Cloude, S. R. (2005). Spaceborne polari-
metric SAR interferometry: performance analysis and mission concepts.
EURASIP Journal of Applied Signal Processing, 20, 3272–3292.
Krieger, G., Moreira, A., Fiedler, H., Hajnsek, I., Werner, M., Younis, M. and
Zink, M. (2007). TanDEM-X: a satellite formation for high-resolution
SAR interferometry. IEEE Transactions on Geoscience and Remote
Sensing, 45, 3317–3341.
Krogager, E. (1993). Aspects of Polarimetric Radar Imaging. PhD thesis,
Technical University of Denmark.
Krogager, E. (1992). Decomposition of the Sinclair matrix into fundamental
components with applications to high resolution radar imaging. In Boer-
mer, W. M. et al. (eds.), Direct and Inverse Methods in Radar Polarimetry,
2, 1459–1478. Kluwer Academic Publishers.
Lakhtakia, A., Varadan, V. V. and Varadan, V. K. (1989). Time Harmonic EM
Fields in Chiral Media. Springer.
Le, C. T. C., Ishimaru, A., Kuga, Y. and Hae Yea, J. (1998). Angular memory
and frequency interferometry for mean height profiling of a rough surface.
IEEE Transactions on Geoscience and Remote Sensing, 36, 61–67.
Lee, J. S. (1994a). Speckle filtering of SAR images: a review. Remote Sensing
Reviews, 8, 313–340.
Lee, J. S., Hoppel, K. W., Mango, S. A. and Miller, A. (1994b). Intensity
and phase statistics of multi-look polarimetric and interferometric SAR
imagery. IEEE Transactions on Geoscience and Remote Sensing, GE-32,
1017–1028.
Lee, J. S., Grunes, M. R., Ainsworth, T. L., Du, L. J., Schuler, D. L. and
Cloude, S. R. (1999). Unsupervised classification using polarimetric
decomposition and the complex Wishart distribution. IEEE Transactions
on Geoscience and Remote Sensing, 37/1 (5), 2249–2259.
Lee, J. S., Schuler, D. L. and Ainsworth, T. L. (2000). Polarimetric SAR sata
compensation for terrain azimuth slope variation. IEEE Transactions on
Geoscience and Remote Sensing, 38/5, 2153–2163.
Lee, J. S., Schuler, D. L., Ainsworth, T. L., Krogager, E., Kasilingam, D. and
Boerner, W.M. (2002). On the estimation of radar polarization orientation
 Bibliography 443

shifts induced by terrain slopes. IEEE Transactions on Geoscience and

Remote Sensing, 40, 30–41.
Lee, J. S. and Pottier, E. (2008). Polarimetric radar imaging: from basics to
applications. Optical Science and Engineering Series, 143, CRC Press.
Li, R. C. (1994). Relations between the Field of Values of a Matrix and those
of its Schur Complements. Report No. UCB//CSD-94-849, Computer
Science Division, University of California at Berkeley.
Lopez-Martinez, C., Pottier, E. and Cloude, S. R. (2005). Statistical assessment
of eigenvector-based target decomposition theorems in radar polarimetry.
IEEE Transactions on Geoscience and Remote Sensing, 43 (9), 2058–
2074.
Lopez, J. M., Fortuny, J., Cloude, S. R. and Sieber, A. J. (2000). Indoor polari-
metric radar measurements on vegetation samples at L, C, S and X bands.
Journal of Electromagnetic Waves and Applications, 14 (2), 205–231.
Lopez-Sanchez, J. M., Ballester-Berman, J. D. and Fortuny-Guasch, J. (2006).
Indoor wide-band polarimetric measurements on maize plants: a study of
the differential extinction coefficient. IEEE Transactions on Geoscience
and Remote Sensing, 44 (4), 758–767.
Lopez-Sanchez, J. M., Ballester-Berman, J. D. and Marquez-Moreno, Y.
(2007). Model limitations and parameter-estimation methods for agricul-
tural applications of polarimetric SAR interferometry. IEEE Transactions
on Geoscience and Remote Sensing, 45 (11), Part 1, 3481–3493.
Lu, S. Y. and Chipman, R. A. (1994). Homogeneous and inhomogeneous Jones
matrices. JOSA A, 11 (2), 766–773.
Lu, S. Y. and Chipman, R. A. (1996). Interpretation of Mueller matrices based
on polar decomposition. JOSA A, 13 (5), 1106–1113.
Ludwig, A. (1973). Definition of cross polarisation. IEEE Transactions on
Antennas and Propagation, AP-21, 116–119.
Luneburg, E. (1996). Polarimetry: a revision of basic concepts. In Cloude, S. R.,
Serbest, A. H. (eds.), Direct and Inverse Electromagnetic Scattering.
Pitman Research Notes in Mathematics, Vol. 361, 257–275. Longman
Scientific and Technical.
Luneburg, E. and Cloude, S. R. (1997). Optimisation procedures for bistatic
scattering. SPIE Proceedings on Wideband Interferometric Sensing and
Imaging Polarimetry, 3120.
Macintosh, F. C., Zhu, J. X., Pine, D. J. and Weitz, D. A. (1989). Polarization
memory of multiply scattered light. Phys. Rev. B, 40 (13), 9342–9345.
Mattia, F., Le Toan, T., Souyris, J. C., De Carolis, C., Floury, N., Posa, F. and
Pasquariello, N. G. (1997). The effect of surface roughness on multifre-
quency polarimetric SAR data. IEEE Transactions on Geoscience and
Remote Sensing, 35 (4), 954–966.
Mendez, E. R. and O’Donnell, K. A. (1987). Observation of depolarization
and backscattering enhancement in light scattering for Gaussian random
surfaces. Opt. Comm., 61, 91–95.
Mengi, E. and Overton, M. L. (2005). Algorithms for the computation of
the pseudospectral radius and the numerical radius of a matrix. IMA J.
Numerical Analysis, 25, 648–669.
Mensa, D. L. (1991). High Resolution Radar Cross-Section Imaging. Artech
House.
444 Bibliography

Mette, T., Papathanassiou, K. P. and Hajnsek, I. (2004). Biomass estima-

tion from POLInSAR over heterogeneous terrain. Proceedings of IEEE
Geoscience and Remote Sensing Symposium, IGARSS 2004, Anchorage,
Alaska, 20–24 September 2004.
Mette, T. (2007). Forest Biomass Estimation from Polarimetric SAR Interfer-
ometry. DLR Research Report 2007-10.
Mishchenko, M. I. (1992). Enhanced backscattering of polarized light from dis-
crete random media: calculations in exactly the backscattering direction.
J. Opt. Soc. Am. (A), 9, 978–982.
Mishchenko, M. I. and Hovenier, J. W. (1995). Depolarization of light backscat-
tered by randomly oriented nonspherical particles. Optics Letters, 20 (12),
1356–1359.
Mishchenko, M., Hovenier, J. W. and Travis, L.D. (2000). Light Scatter-
ing by Nonspherical Particles: Theory, Measurements and Applications.
Academic Press.
Mishchenko, M. I., Travis, L. D. and Lacis, A. A. (2006). Multiple Scattering
of Light by Particles: Radiative Transfer and Coherent Backscattering.
Cambridge.
Mishchenko, M. I., Liu, L., Mackowski, D. W., Cairns, B. and Videen, G.
(2007). Multiple scattering by random particulate media: exact 3D results.
Optics Express, 15 (6, 19), 2822–2836.
Misner, C. W., Thorne, K. S. and Wheeler, J. A. (1992). Gravitation. W. H.
Freeman and Co.
Mott, H. (1992). Antennas for Radar and Communications. Wiley.
Mott, H. (2007). Remote Sensing with Polarimetric Radar. Wiley Interscience.
Murnaghan, F. D. (1932). On the field of values of a square matrix. Proc. N. A.
S. Mathematics, 246–248.
Murnaghan, F. D. (1962). The Unitary and Rotation Groups. Spartan Books.
Nelander, A. (1995). Analysis of wide band polarimetric radar. Proceedings
of the third International Workshop on Radar Polarimetry (JIPR ‘95),
IRESTE, University of Nantes, France, 89–98.
Neumann, M., Ferro-Famil, L. and Reigber, A. (2008). Multibaseline polari-
metric SAR interferometry coherence optimization. IEEE Geoscience
and Remote Sensing Letters, 5, (1), 93–97.
Nghiem, S. V., Yueh, S. H., Kwok, R. and Li, F. K. (1992). Symmetry properties
in polarimetric remote sensing. Radio Science, 27 (5), 693–711.
Novak, L. M., Sechtin, M. B. and Cardullo, M. J. (1989). Studies of target detec-
tion algorithms which use polarimetric radar data. IEEE Trans. Aerospace
and Electronic Systems, AES-25, 15–165.
Nye, J. F. (1999). Natural Focusing and Fine Structure of Light. IoP Publishing.
Oh, Y., Sarabandi, K. and Ulaby, F. T. (1992). An empirical model and
an inversion technique for radar scattering from bare soil surfaces.
IEEE Transactions on Geoscience and Remote Sensing, GE-30(2),
370–381.
O’Neill, E. L. (1991). Introduction to Statistical Optics. Dover Press.
Pancharatnam, S. (1956). Generalised theory of interference and its applica-
tions. I: Coherent pencils. Proc. Indian Acad. Sci. A, 44, 247–262.
 Bibliography 445

Papathanassiou, K. P. and Cloude, S. R. (1997). Polarimetric effects in repeat-

pass interferometry. Proceedings IGARSS 97, Singapore, 3–8 August
1997, 1926–1928.
Papathanassiou, K. P. and Zink, M. (1998a). Polarimetric calibration of the
airborne experimental SAR system of DLR. Proceedings of European
SAR Conference, EUSAR 1998, Friedrichshafen, Germany.
Papathanassiou, K. P., Reigber, A., Scheiber, R., Horn, R., Moreira, A.
and Cloude S. R. (1998b). Airborne polarimetric SAR interferometry.
Proceedings of IEEE Symposium on Geoscience and Remote Sensing
(IGARSS), Seattle, USA, July 6–10.
Papathanassiou, K. P. and Cloude, S. R. (2001). Single baseline polarimet-
ric SAR interferometry. IEEE Transactions on Geoscience and Remote
Sensing, GRS-39/11, 2352–2363.
Papathanassiou, K. P. and Cloude, S. R. (2003). The effect of temporal decor-
relation on the inversion of forest parameters from POLInSAR data.
Proceedings of IEEE International Geoscience and Remote Sensing
Symposium (IGARSS 2003), Toulouse, France, July 21–25.
Papathanassiou, K. P., Cloude, S. R., Liseno, A., Mette, T. and Pret-
zsch, H. (2005). Forest height estimation by means of polarimetric
SAR interferometry: actual status and perspectives. Proceedings of
the Second ESA POLInSAR Workshop, Frascati, Italy, January 2005.
http://earth.esa.int/workshops/polinsar2005/
Pascual, C., Gimeno-Nieves, E. and Lopez-Sanchez, J. M. (2002). The equiv-
alence between the polarisation subspace method (PSM) and coherence
optimisation in polarimetric radar interferometry. Proceedings of the
Fourth European Synthetic Aperture Radar Conference, EUSAR 2002,
589–592.
Penrose, R. and Rindler, W. (1984). Spinors and Space-Time, Volume 1: Two
Spinor Calculus and Relativisitic Fields. Cambridge University Press.
Perrin, F. (1942). Polarization of light scattered by isotropic opalescent media.
Journal of Chemical Physics, 10, 415–427.
Poelman, A. J. and Hilgers, C. J. (1991). Effectiveness of multinotch logic-
product polarisation filters in radar for countering rain clutter. IEE
Proceedings F, Radar and Signal Processing, 138, 427–437.
Poincaré, H. (1997). Theorie Mathematique de la Lumiere II. Chapter 12. Paris,
1892.
Pottier, E. and Cloude, S. R. (1997). Application of the H-A-α polarimetric
decomposition theorem for land classification. SPIE International Sym-
posium on Optical Science Engineering and Instrumentation, Wideband
Interferometric Sensing and Imaging Polarimetry, San Diego, California,
USA, 27 July–1 August 1997.
Praks, J., Kugler, F., Papathanassiou, K. P., Hajnsek, I. and Hallikainen, M.
(2007) Height estimation of boreal forest: interferometric model based
inversion at L and X bands versus HUTSCAT profiling scatterometer.
IEEE Geoscience and Remote Sensing Letters, 4, 466–470.
Press, W. H., Teukolsky, S. A., Vetterling, W. T. and Flannery, B. P.
(2007). Numerical Recipes 3rd Edition: The Art of Scientific Computing.
Cambridge University Press.
446 Bibliography

Priest, R. G. and Germer, T. A. (2000). Polarimetric BRDF in the microfacet

model, theory and measurements. Proceedings of 2000 Military Sensing
Symposia, Speciality Group on Passive Sensors, Ann Arbor, Michigan,
August 2000, 1, 169–181,
Quegan, S. (1994). A unified algorithm for phase and cross-talk calibration
of polarimetric data-theory and observations. IEEE Trans., GRS-32,
89–99.
Raney, R. K. (2006). Dual-polarized SAR and Stokes parameters. IEEE
Geoscience and Remote Sensing Letters, 3 (3), 317–319.
Raney, R. K. (2007). Hybrid-polarity SAR architecture. IEEE Transactions on
Geoscience and Remote Sensing, 45 (11), 3397–3404.
Reigber, A. and Moreira, A. (2000). First demonstration of airborne SAR
tomography using multi-baseline L-band data. IEEE Transactions on
Geoscience and Remote Sensing, 38/5, 2142–2152.
Reigber, A., Papathanassiou, K. P., Cloude, S. R. and Moreira, A. (2001). SAR
tomography and interferometry for the remote sensing of forested terrain.
Frequenz, 55, 119–123.
Roman, P. (1959a). Generalized Stokes parameters for waves with arbitrary
form. Il Nuovo Cimento, 13, 2546–2554.
Roman, P. (1959b). Decomposition of 3 × 3 matrices. Proc. Phys. Soc., 74,
649–657.
Rosen, J. (1995). Symmetry in Science: An Introduction to the General Theory.
Springer.
Rosenqvist, A., Shimada, M., Ito, N. and Watanabe, M. (2007). ALOS
PALSAR: A pathfinder mission for global-scale monitoring of the
environment. IEEE Transactions on Geoscience and Remote Sensing,
GRS45(11), 3307–3316.
Sagues, L., Lopez-Sanchez, J. M., Fortuny, J., Fabregas, X., Broquetas, A. and
Sieber, A. J. (2000). Indoor experiments on polarimetric SAR interfer-
ometry. IEEE Transactions on Geoscience and Remote Sensing, GRS-38,
671–684.
Sagues, L., Lopez-Sanchez, J. M., Fortuny, J., Fabregas, X., Broquestas, A.
and Sieber, A. J. (2001). Polarimetric radar interferometry for improved
mine detection and surface clutter rejection. IEEE Trans. Geoscience and
Remote Sensing, GRS-39, 1271–1278.
Sarabandi, K. (1992a). Derivation of phase statistics from the Mueller matrix.
Radio Science, 27, 553–560.
Sarabandi, K., Pierce, L. E. and Ulaby, F.T. (1992b). Calibration of a polarimet-
ric imaging SAR. IEEE Transactions on Geoscience and Remote Sensing,
GRS-30 (3).
Saxon, D. S. (1955). Tensor scattering matrix for the electromagnetic field.
Phys. Rev., 100, 1771.
Schmeider, R. (1969). Stokes algebra formalism. Journal of the Optical Society
of America, 59, 297–302.
Schneider, R. Z., Papathanassiou, K. P., Hajnsek, I. and Moreira, A. (2006).
Polarimetric and interferometric characterization of coherent scatterers
in urban areas. IEEE Transactions on Geoscience and Remote Sensing,
GRS-44, 971–984.
 Bibliography 447

Schou, J., Skriver, H., Nielsen, A. A. and Conradsen, K. (2003). CFAR edge
detector for polarimetric SAR images. IEEE Transactions on Geoscience
and Remote Sensing, 41 (1), 20–32.
Schuler, D. L., Lee, J. S., Kasilingam, D. and Nesti, G. (2002). Sur-
face roughness and slope measurements using polarimetric SAR data.
IEEE Transactions on Geoscience and Remote Sensing, 40 (3),
687–698.
Seymour, S. and Cumming, I. G. (1994). Maximum likelihood estimation
for SAR interferometry. Proceedings of IEEE Geoscience and Remote
Sensing Symposium, IGARSS’94, Pasadena, USA.
Sharma, J. J., Hajnsek, I. and Papathanassiou, K. P. (2007). Vertical profile
reconstruction with POLInSAR of a subpolar glacier. Proceedings of
IEEE Geoscience and Remote Sensing Symposium, IGARSS’07, 1147–
1150.
Simon, R. (1987). Mueller matrices and depolarization criteria. Journal of
Modern Optics, 34, 569–575.
Souyris, J. C., Imbo, P., Fjortoft, R., Mingot, S. and Lee, J. S. (2005). Compact
polarimetry based on symmetry properties of geophysical media: the
pi/4 mode. IEEE Transactions on Geoscience and Remote Sensing, 43
(3), 634–646.
Stebler, O., Meier, E. and Nueesch, D. (2002). Multi-baseline polarimetric
SAR interferometry: first experimental spaceborne and airborne results.
ISPRS Journal of Photogrammetry and Remote Sensing, 56 (3).
Stokes, G. G. (1852). On the composition and resolution of streams of polar-
ized light from different sources. Cambridge Philosophical Society, 9,
399.
Strang, G. (2004). Linear Algebra and its Applications. Fourth edition. Brooks
Cole.
Tabb, M. and Carande, R. (2001). Robust inversion of vegetation struc-
ture parameters from low frequency polarimetric interferometric SAR.
Proceedings of IEEE International Geoscience and Remote Sensing
Symposium (IGARSS 2001), Sydney, Australia, July 2001.
Tabb, M., Orrey, J., Flynn, T. and Carande, R. (2002a). Phase diversity:
a decomposition for vegetation parameter estimation using polarimet-
ric SAR interferometry. Proceedings of the Fourth European Synthetic
Aperture Radar Conference, EUSAR 2002, 721–724.
Tabb, M., Flynn, T. and Carande, R. (002b). Direct estimation of vegetation
parameters from covariance data in POLINSAR. Proceedings of IGARSS
2002, Toronto, Canada, 1908–1910.
Topp, G. C., Davis, J. L. andAnnan, A. P. (1980). Electromagnetic determination
of soil water content: measurements in coaxial transmission lines. Water
Resources Research, 16, 574–582.
Touzi, R., Lopes, A., Bruniquel, J. and Vachon, P. W. (1999). Coherence esti-
mation for SAR imagery. IEEE Transactions on Geoscience and Remote
Sensing, 37/1, 135–149.
Touzi, R. (2007). Target scattering decomposition in terms of roll-invariant tar-
get parameters. IEEE Transactions on Geoscience and Remote Sensing,
45 (1), 73–84.
448 Bibliography

Tragl, K. (1990). Polarimetric radar backscattering from reciprocal random

media. IEEE Transactions on Geoscience and Remote Sensing, 28,
856–864.
Treuhaft, R. N., Madsen, S., Moghaddam, M. and van Zyl, J. J. (1996). Vegeta-
tion characteristics and underlying topography from interferometric data.
Radio Science, 31, 1449–1495.
Treuhaft, R. N. and Cloude S. R. (1999). The structure of oriented vegetation
from polarimetric interferometry. IEEE Transactions on Geoscience and
Remote Sensing, 37/2 (5), 2620.
Treuhaft, R. N. and Siqueria, P. (2000a). Vertical structure of vegetated land
surfaces from interferometric and polarimetric radar. Radio Science, 35
(1), 141–177.
Treuhaft, R. N., Law, B. E. and Asner, G. P. (2000b). Structural approaches
to biomass monitoring with multibaseline, multifrequency, polarimet-
ric interferometry. Proceedings of the Third European SAR Conference
(EUSAR), Munich, Germany, 253–255.
Treuhaft, R. N., Law, B. E. and Asner G. P. (2004). Forest attributes from
radar interferometric structure and its fusion with optical temote sensing.
BioScience, 56 (6), 561–571.
Tsang, L., Kong, J. A. and Shin, R. T. (1985). Theory of Microwave Remote
Sensing. Wiley Interscience.
Ulaby, F. T., Moore, R. K. and Fung, A. K. (1982). Microwave Remote Sensing:
Active and Passive. Vol. II: Radar Remote Sensing and Surface Scattering
and Emission Theory. Addison–Wesley.
Ulaby, F. T., Moore, R. K. and Fung, A. K. (1986). Microwave Remote Sensing:
Active and Passive. Vol. III: From Theory to Applications. Artech House.
Ulaby, F. T. and Elachi, C. (eds.) (1990). Radar Polarimetry for Geoscience
Applications. Artech House, Norwood, MA.
van Albada, M. P., van der Mark, M. B. and Lagendijk, A. (1988). Polarization
effects in weak localisation of light. Journal of Physics D, 21 (105),
28–31.
van de Hulst, H. C. (1981). Light Scattering by Small Particles. Dover Press.
van der Mee, C. V. M. and Hovenier, J.W. (1992). Structure of matrices
transforming Stokes parameters. J. Math. Phys., 33 (10), 3574–3584.
van der Mee, C. V. M. (1993). An eigenvalue criterion for matrices transforming
Stokes parameters. J. Math. Phys., 34, 5072–5088.
van Zyl, J. J., Zebker, H. A. and Elachi, C. (1987). Imaging radar polarization
signatures: theory and observations. Radio Science, 22, 529–543.
van Zyl, J. J. (1989). Unsupervised classification of scattering behaviour using
radar polarimetry data. IEEE Transactions on Geoscience and Remote
Sensing, GE-27(1), 36–45.
van Zyl, J. J. (1990). Calibration of polarimetric tadar images using only
image parameters and trihedral corner reflectors. IEEE Transactions on
Geoscience and Remote Sensing, GE-28, 337–348.
Wanielik, G. amd Stock, D. J. R. (1992). A proposed polarimetric CFAR-
detector and analysis of its operation. In Boerner, W. M et al. (eds.),
Direct and Inverse Methods in Radar Polarimetry, 2, 999–1010. Kluwer
Academic Publishers.
 Bibliography 449

Wiener, N. (1930). Generalized harmonic analysis. Acta Mathematica, 55,

118–258.
Woodhouse, I. H. and Turner, D. (2002). On the visualization of polarimetric
response. International Journal of Remote Sensing, 24 (6), 1377–1384.
Woodhouse, I. H. (2006). Predicting backscatter-biomass and height-biomass
trends using a macroecology model. IEEE Transactions on Geoscience
and Remote Sensing, GRS-44, 871–877.
Wright, P. A., Quegan, S., Wheadon, N. S. and David Hall, C. (2003). Faraday
rotation effects on L-band spaceborne SAR data. IEEE Transactions on
Geoscience and Remote Sensing, GRS-41 (12), 2735–2744.
Yamada, H., Yamaguchi, Y., Rodriguez, E., Kim, Y. and Boerner, W. M.
(2001). Polarimetric SAR interferometry for forest canopy analysis by
using the super-resolution method. IEICE Transactions on Electronics,
E84-C (12), 1917–1924.
Yamaguchi, Y., Moriyama, T., Ishido, M. and Yamada, H. (2005). Four-
component scattering model for polarimetric SAR image decomposi-
tion. IEEE Transactions on Geoscience and Remote Sensing, 43 (8),
1699–1706.
Zebker, H. A. and Villasenor, J. (1992). Decorrelation in interferometric radar
echoes. IEEE Transactions on Geoscience and Remote Sensing, 30 (5),
950–959.
Zhou, Z. S. and Cloude, S. R. (2006). Application of polarization coherence
tomography to GB-POLInSAR data. Proceedings of IEEE International
Symposium on Geoscience and Remote Sensing, IGARSS06, Denver,
Colorado, July 2006.
This page intentionally left blank
Index

along track interferometry (ATI), 219 for semi-infinite random volume, 258 contrast optimization, 193
ALOS-PALSAR, 396 SNR decorrelation, 221 COPOL nulls, 57
alpha parameter geometric decorrelation, 224 Cramer–Rao bounds, 431
and dihedral dielectric constant, 399 image co-registration errors, 223 critical baseline, 215
definition, 184 interferometric, 220 cross-polarisation, 8
for Bragg scattering, 128 polarimetric, 73
for dihedral scattering, 130 polarimetric interferometric, 235
mean alpha, 98 temporal decorrelation, 222
decomposition
anisotropy scattering parameter A, 97 coherence loci
eigenvalue decomposition of [J], 74
definition, 252 eigenvalue decomposition of [T], 87
for IWCM model, 280
of interferometric coherence, 233
baseline components, 213 for oriented volume scattering, 263 of N-dimensional coherency matrices, 92
for OVOG model, 282
baseline decorrelation, 224 of the Stokes vector, 78
for OVUG model, 283
biaxial material, 13 of the wave coherency matrix, 76
bidirectional reflectance distribution function for random volume scattering, 257 propagation distortions, 205
for RVOG, 274
(BRDF), 139 roll invariance, 179
for RVUG model, 283
birefringence the point reduction theorem, 185
for surface scattering, 255
for circular polarisations, 24 degree of polarisation, 76
coherence optimization
general definition, 14 depolarisation
bivariate Gaussian distribution, 429 estimation bias, 250 definition, 71
Bragg surface scattering, 126 constrained, 243 state vector, 92
for oriented volume scattering, 261
Brewster angle, 122 depolarisers
for random volume scattering, 255
azimuthal depolarisation, 96
for surface scattering, 254
characterisation of, 110
SVD interpretation, 242 isotropic, 91
C2 symmetry, 11 unconstrained, 241 reflection depolarisation, 97
calibration coherence region, 245
of POLInSAR systems, 361 dielectric constant
coherence tomography
of soil, 120
of POLSAR systems, 351 condition number for multi-baselines, 338 Polder–Van Santen/de Loor formula, 171
polarimetric, 350 dual baseline, 325
soil moisture, 121
Quegan algorithm, 352 multi-baseline reconstruction, 335
soil salinity, 120
Cameron decomposition, 182 single baseline, 323
Cartan matrix, 417 dielectric tensor, 12
single baseline condition number, 332
differential interferometry, 217
Cartan sub-algebra, 416 temporal decorrelation and SNR dihedral scattering
application to coherency matrix, 92 effects, 334 from dielectrics, 124
Chandrasekhar decomposition, 191 coherency matrix from metal structures, 54
chiral media and CFAR detection, 194
Dirac matrices, 413
admittance parameters, 23 and contrast optimization, 193
directional hemispherical reflectivity
D and L-rotatory materials, 24 for forward scattering, 165
spatial dispersion, 23 (DHR), 139
for general scattering, 86
specific rotatory power of, 25 discrete dipole approximation (DDA), 152
for waves, 73
wave equation for, 23 DLR E-SAR, 392
propagation effects, 202
Dynkin diagrams, 418
Cloude–Pottier decomposition, 192 compact polarimetry, 354
coefficient of variation (CV), 427 and Bragg scattering, 176
coherence and Rayleigh scattering by spheroids, 176
for two-layer surface/volume problems, 270 π/4 mode, 356 effective length of an antenna, 51
for circular polarisation, 136 compact POLInSAR, 362 effective number of looks (ENL), 427
for exponential profile, 231 condition number of a matrix, 409 eigenpropagation states, 14
452 Index

entropy ordinary wave, 20 Mie scattering, 152

backscattering, 97 plasma frequency, 18 Minkowski metric, 43
general scattering entropy, 108 isomorphism, 41 Mishchenko decomposition, 164
of a wave, 77 Monte Carlo polarisation simulations, 432
entropy-alpha decomposition, 194 Mueller matrix
entropy-alpha diagram Jones calculus backscatter form, 83
for backscatter, 99 definition, 25 definition, 79
for bistatic scattering, 109 diattenuator, 26 filtering, 90
for compact polarimetry, 101 homogeneous propagation channel, 28 for isotropic depolariser, 84
for dual polarisation, 101 inhomogeneous propagation channel, 29 formal connection to [T], 113
N matrix, 25 main properties, 80
retarder, 27 mapping to [T], 88
Faraday rotation pure matrices, 80
basic properties, 21 reciprocity theorem for backscatter, 85
estimation from data, 206 Kennaugh matrix, 83 sum of pure matrices, 86
observations from satellites, 397 Killing form, 416 test for a pure matrix, 90
vectorised form, 205 Krogager decomposition, 180 the phase function, 153
field of values of a matrix, 243 multivariate normal distribution, 433
flat earth phase component, 210
flat earth phase removal, 216 Lagrange multipliers
Foldy Lax Equations, 33 and coherence optimization, 241 N matrix, 14
forest propagation extinction models, 174 and Rayleigh quotient, 407 noise
Fourier–Legendre coherence basis and the [S] matrix, 48 estimation from crosspolar channels, 189
functions, 228 Lambertian surface, 139 in coherence tomography, 333
Fourier–Legendre series, 225 leaf-area-index (LAI), 173 in decomposition theorems, 189
Freeman–Durdan decomposition, 197 Lee filter, 348, 427 in interferometry, 221
Freeman-eigenvalue hybrid Legendre coherence model norm of a matrix, 409
decomposition, 198 first order, 296 nullspace
Fresnel equations, 117 second order, 301 and decomposition theorems, 191
Fry–Kattawar relations, 82 Lie algebra, 412 definition, 406
line fit algorithm
for two states, 287
Gell–Mann matrices, 414 total least squares (TLS), 290 orthogonal scattering mechanisms
Graves power matrix, 48 log-likelihood function, 433 definition, 186
gyrotropic media, 18 Lorentz force equation, 20 in natural media, 187
Lorentz spin matrix, 43 OVOG model, 281
Lorentz transformation, 43 OVUG model, 283
height estimation and [S] matrix, 67
structure free algorithm, 306 connection to special relativity, 66
using RVOG, 309 conservation of zero wave entropy, 82 Pancharatnam phase, 37
Hermitian matrices, 404 homomorphism, 64 Pauli spin matrices, 16
homomorphisms of Lie groups, 420 Lorentz boost, 66 penetration depth, 142
Huygens source, 8 Ludwig wave co-ordinates, 6 permanent scatterers (PS), 224
Huynen decomposition, 192 phase bias removal, 286
Huynen parameters, 63 plane of polarisation, 11
Malus’s law, 64 Poincaré sphere, 42
matrix exponential function, 15 polar decomposition of a matrix, 27
InSAR, 345 Maxwell’s equations polarimetric interferometry
integral equation model (IEM), 138 boundary conditions, 115 change of wave basis, 237
interferometric water cloud model coherent surface scattering, 129 forming vector interferograms, 235
(IWCM), 278 differential form, 3 optimum baseline, 318
interferometry dipole radiation, 5 phase normalization, 235
blind angles, 214 duality transformation, 6 SVD and Schur decompositions, 248
memory line, 212 Green’s function, 4 polarisation
vertical wavenumber, 213 Helmholtz equation, 4 C-lines, 37
π -height, 214 inhomogeneous plane wave, 119 conjugate semi-diameters, 36
ionosphere physical optics approximation, 123 definition of sense, 19
basic properties, 18 TE/TM Waves, 116 equation of ellipse, 35
cyclotron frequency, 19 TEM waves, 5 left and right circular, 19
extraordinary wave, 20 vector wave equation, 3 L-lines, 37
 Index 453

polarisation coherence tomography scattering sphere RVOG estimation, 316

(PCT), 329 application to particle scattering, 154 synthetic aperture radar (SAR), 341
polarisation fork, 62 definition, 108
polarisation frame or basis, 11 scattering symmetries
polarisation signatures, 60 backscatter azimuthal symmetry, 95 T-matrix method, 152
polarization of matter, 10 backscatter reflection symmetry, 94 trace of a matrix, 403
POLInSAR, 360 backscatter rotation symmetry, 95
POLSAR, 347 bistatic symmetries, 104
Poynting vector, 37 degree of symmetry, 183 Ulaby scattering model, 170
principal material axes, 13 symmetric scatterers, 181 uniaxial material, 13
scattering vector unitary matrix definition, 404
lexicographic order, 68 unitary transformations
Pauli expansion, 68 and double angles, 41
range spectral filtering, 212
Schur decomposition and coherence, 247 Cartan sub-algebra, 74
Rayleigh quotient
Schur’s theorem, 406 compound form for SU(2), 39
and optimum transmittance, 29
Schwarz inequality, 428 congruent, 56
definition, 405
similarity transformations, 405 homomorphic, 41
Rayleigh scattering, 145
SINC coherence model, 229 special unitary, 12
and the entropy-alpha plane, 159
singular value decomposition unitary reduction operator, 91
bistatic scattering, 160
and coherence optimization, 242
by a chiral spheroid, 150
and transformation of the [S] matrix, 55
by a cloud of chiral spheroids, 158
of the scattering amplitude matrix, 47 vector radiative transfer (VRT)
by a cloud of spheroids, 156
SVD, 407 cyclical components, 164
by a small spheroid, 148
skin depth, 119 ladder terms, 164
particle anisotropy Ap , 148
small perturbation model (SPM), 125 vegetation bias, 229
spheroidal particle functions, 149
Snell’s law, 116 vertical structure function
reciprocity theorem, 49
speckle filtering, 427 definition, 225
refractive index, 11
sphere-diplane-helix (SDH) estimation using CT, 323
retarder, 18
decomposition, 181
RVOG model, 271
spin matrix, 40
minimum coherence, 275
Stokes criterion, 84 water cloud model (WCM), 168
Stokes reflection matrix, 83 wave dichotomy, 75
Stokes vector, 40 Wishart distribution, 430
scattering amplitude matrix connection to w, 114
backscatter theorem, 50 definition, 43
bisectrix, 46 for three-dimensional Waves, 45 X-Bragg model
BSA co-ordinate system, 50 surface slope and entropy-alpha plane, 136
definition of [S], 47 azimuth slope, 131 definition, 134
FSA co-ordinate system, 50 range slope, 131 effect of surface slope, 136
scattering angle, 46 surface topography estimation moisture parameter M, 135
scattering plane, 46 for the OVOG model, 293 roughness parameter R, 135
singular matrices, 64 for the OVUG model, 294 XPOL nulls, 57
transformation invariants, 58 for the RVOG model, 287
scattering matrix group, 49 surface-to-volume ratio
scattering mechanism, 69 estimation from polarimetry, 200 Yamaguchi decomposition, 198