November 2011

Msc thesis in geomatics

IDENTIFICATION AND ANALYSIS
OF POINT SCATTERERS
IN TIME SERIES INSAR
Piers Titus van der Torren
Piers Titus van der Torren: Identification and analysis of point scatterers in
time series InSAR, Master of Science thesis in Geomatics, © November
2011

supervisors:
Ramon Hanssen
Sami Samiei Esfahany

Delft University of Technology

cover image:
Deformation speed change rate over Delft, as measured with Persistent
Scatterer InSAR (PSI) using TerraSAR-X radar images from April 2009
till November 2011. Featured in red is the construction of a train tunnel,
started in 2010. See section 6.2 for more details.
deformation speed change [mm/y²]
3 2 1 0 -1 -2 -3

Legend:
ABSTRACT

In this study methods are developed for improved analysis and pro-
cessing of PSI data. PSI, or radar interferometry, makes it possible to use
satellite radar images to measure deformation of the Earth’s surface
and objects on it with millimetre accuracy. However, interpretation of
the measurements and identifying the actually measured objects is still
a common problem.
There are no dedicated tools available for validation, for finding both
falsely detected and falsely rejected points, or for deeper analysis of PSI
results. Existing algorithms for automatic coherent scatterer selection
need a lot of acquisitions to obtain reliable results, which makes it
necessary to collect data for many months or years before processing
can be done.
In this study a suite of tools is developed that facilitate detailed analysis
of results and versatile processing of radar data. This suite consists of a
visual inspection tool, and a toolbox that handles metadata and can do
versatile processing of radar data.
Furthermore a method is developed for reliable point scatterer selection,
that works for a small number of acquisitions, among other improve-
ments.

iii
CONTENTS

1 i n t ro du c t i o n 1

2 r a da r s i g na l t h e o r y 5
2.1 Introduction 5
2.2 Radar 5
2.3 Scatterers 7
2.4 The impulse response 8
2.5 Interferometry 10
2.6 Signal, clutter and coherence 10

3 i n sa r i n v e s t i g at i o n t o o l s 15
3.1 Introduction 15
3.2 Methodology 16
3.3 InSAR Inspector 17
3.3.1 Zoomable background 17
3.3.2 Phase time series plot. 18
3.3.3 Point information table 20
3.3.4 Additional information plot 20
3.3.5 The GUI object 21
3.4 Object oriented InSAR toolbox 21
3.4.1 Slc and SlcStack 22
3.4.2 SlcImage 23
3.4.3 PointSet 24
3.4.4 Spectrum 25
3.4.5 Other classes 26
3.5 Outlook 26
3.5.1 GIS Integration 27
3.5.2 3D AHN2 viewer 27
3.5.3 Bayesian Point Selection 27

4 p o i n t s c at t e r e r s e l e c t i o n 29
4.1 Introduction 29
4.2 Algorithm 29
4.2.1 Implementation 31
4.2.2 Stack processing 31
4.2.3 Subpixel position 31
4.3 Results 32
4.3.1 Focus on point scatterers 32
4.3.2 Single observation per scatterer 32
4.3.3 More accurate quality measure for smaller stacks 32
4.3.4 Quality measure per acquisition 36
4.3.5 Sub pixel precision location 36
4.3.6 Amplitude calibration is not necessary 38
4.4 Relation to SCR 38
4.5 Conclusions 39

5 g e o c o d i n g i m p rov e m e n t s 41
5.1 Introduction 41
5.2 Geocoding 41
5.3 Topographic phase computation 42

v
vi contents

5.4 Conclusions 43

6 cases 45
6.1 Introduction 45
6.2 Delft train tunnel 45
6.3 Storage tank monitoring 49
6.4 Corner reflector experiments 53

7 c o n c l u s i o n s a n d r e c o m m e n dat i o n s 55

bibliography 57

i appendix 59

a t o o l b ox m a n ua l 61
ACRONYMS

irf impulse response function

sar synthetic aperture radar
insar Interferometric SAR
tsinsar Time Series InSAR
psi Persistent Scatterer InSAR
scr signal to clutter ratio
sscr signal to signal+clutter ratio
gui graphical user interface
islr integrated side lobe ratio
slc single look complex
mrm mean reflectivity map
ahn2 Actueel Hoogtebestand Nederland versie 2

vii
INTRODUCTION
1
On a lot of places on Earth the surface or objects on it move due to
natural phenomena or human activities, such as earthquakes, volcano
eruptions, mining, and water extraction. Often these movements are
very subtle, in the order of millimetres. Even so they may be precursors
to bigger events, or lead in accumulation to failure of infrastructure or
houses. Measuring them is important for better understanding of the
earth and can give early warnings for more serious events.
The traditional way of measuring these movements is using geodetic
surveying techniques such as levelling. While those techniques are very
accurate and reliable, it is very labour intensive per measured point.
Because of this resolution in both space and time is usually low.
In the last decades a new technology has been researched which over-
comes these problems, while allowing similar accuracy. This is achieved
using satellite radar. The type of radar that is used is an imaging side
looking radar, that can achieve a very high resolution using Synthetic
Aperture Radar (SAR).
Satellite radar collects complex scattering signals from the Earth’s
surface and objects on it, which results in a radar image. The phase
difference between two of such images is related to topography, defor-
mation and atmosphere. By processing time series those parameters
can be estimated. This technique is called radar interferometry (InSAR).
Radar interferometry makes it possible to use satellite radar images
to measure deformation of the Earth’s surface and objects on it with
millimetre accuracy [Hanssen, 2001]. Since new images are continuously
acquired all over the world, and have been acquired already for almost
two decades, the amount of available data gives a lot of opportunities.
It has been successfully applied for measuring deformations caused
by earthquakes, volcano eruptions, gas and salt mining, and for dike
monitoring [Hooper et al., 2004, Ketelaar and Hanssen, 2006, Hanssen
and van Leijen, 2008].
However, interpretation of the measurements and identifying the actu-
ally measured objects is still a common problem. This is caused by the
nature of radar scatterers, which are the ’objects’ that are measured by
the radar. Opposed to geodetic surveying techniques, where a measure-
ment is done on a known point, with imaging radar the geometry and
material properties of nearby objects define the measurement. The ge-
ometry of such a scatterer can be quite complex, and can include parts
which are not directly expected. For example the wall of a building
and the street in front of it can together form a scatterer, of which the
movement can be related to either one. This is because all radar signal
that travels the same distance arrives at the same time, and hence is in
the same measurement. Since this way of observing is so different from
human observation interpretation errors are very probable.

1
2 introduction

Furthermore the amount of data is very large, and is ambiguous, multi-

interpretable, relative, and multivariate.1 . Also only parts of the image
are usable for deformation measurement, because for the parts where
the composition of a scatterer has changed one cannot speak of a single
movement. When nothing in the composition of a scatterer has changed
it is called coherent.
The goal of this study is to develop a method to identify coherent
scatterers and analyse them in detail. For this analysis properties of the
scatterers can be used, such as precise location, deformation, amplitude
behaviour, and other characteristics derived from the radar signal. Those
properties have to be estimated and interpreted either automatically or
manually. Automatic processing is very useful for reducing the amount
of data and for focusing on interesting parts. But for validation, for
finding both falsely detected and falsely rejected points, or for deeper
analysis, manual interpretation is necessary.
There are no dedicated tools available for this detailed analysis, making
it a complicated and error prone task. Existing algorithms for automatic
coherent scatterer selection need a lot of acquisitions to obtain reliable
results, which makes it necessary to collect data for many months or
years before processing can be done.
The primary focus will lie on making it possible to query and visualise
all steps of the processing interactively and in detail. This way the raw
data can be analysed in detail, just as all steps of the processing algo-
rithms. Auxiliary data can be shown along to show possible relations.
An interactive way of handling data allows getting better understand-
ing of the problems that are accompanying time series InSAR, and
analysing the origin of peculiar results.
For this a graphical user interface (GUI) is developed, featuring an
overview plot and detailed information and plots of a selected point
with respect to a selected reference point. From the GUI parts of the
processing can be executed. To be able to show all point and radar
properties and run processing steps on the fly an object-oriented toolbox
will be developed, including existing and newly developed or improved
algorithms.
The secondary focus is on algorithmic parts: developing improved
methods for the selection of coherent point scatterers and accurate
geocoding. A method is developed for reliable point scatterer selec-
tion that will work for a small number of acquisitions, among other
improvements.

1 The phase of the radar signal is the part containing the deformation signal. It is ambiguous
because as it is only the phase it is impossible to distinguish signals one wavelength apart.
The phase is influenced by elevation, atmosphere and deformation, making multiple
interpretations possible. A phase value only has meaning if it can be related to the same
point at another observation, ánd another point in these observations, hence to another
phase value in space and time. At last multiple properties of each point can be extracted,
besides the position in the image also properties such as height, size, orientation, and
behaviour in time, resulting in a multivariate dataset.
 introduction 3

The structure of this thesis is as follows: Chapter 2 gives an overview

of the theory of interferometry, radar and signal processing. Chapter 3
describes the developed tools and GUI and their possibilities. Chapter
4 introduces a new scatterer selection method and reports the results of
a comparative study. Chapter 5 explains improvements in geocoding,
improving accuracy, computation time, and ease of use. Chapter 6
describes some cases which show the capabilities of the tools and algo-
rithms. The last chapter contains conclusions and recommendations.
R A D A R S I G N A L T H E O RY
2
2.1 introduction

This chapter is meant to give a background on the satellite radar image

signal. It tries to give a deeper understanding of radar images and how
to get the most information out of it. It is most useful as an introduction
to chapter 4, the chapter about the selection of point scatterers, but also
interesting for those who just want to understand radar images better.
For those interested in the history, synthetic aperture radar (SAR) image
formation, and broader information the book on this topic by Hanssen
[2001] and references in it are recommended. Ketelaar [2009] goes
deeper into using time series of radar images using PSI.
This chapter is built up as follows. In section 2.2 interpreting general
imaging radar is explained, section 2.3 handles the elements from which
the radar image is built: scatterers. Section 2.4 explains the exact shape
of those scatterers, and how that differs per satellite or processing chain.
In section 2.5 the step to the phase observations and the interesting
opportunities of interferometry is made, after which section 2.6 goes
more in detail about describing the quality of those interferometric
measurements.

2.2 radar

Radar observes in quite a different way than we are used to as humans.

In this chapter the focus will directly be on imaging radar, the type
of radar that is used for satellite based earth observation. Imaging
radar produces a 2D image, at the first glance not so different from
optical systems. But besides using a different wavelength which makes
it look through clouds, and using its own illumination which makes it
independent of the sun, there is a more structural difference. Instead
of observing angles, as an eye does, it observes range. For each pixel
in range the radar observes the amount of signal that is reflected back
to the satellite at that range. This causes the projection to be quite
different from what we are used to, and also shadows work the other
way around.
Because all reflections at the same range end up in the same pixel some
part are overlapping, while other parts are shadowed out, there just
void is observed, see Figure 1. It is also possible for the radar signal
to reflect multiple times. In that case the signal takes a longer way
and will end up in the image further away in range than the reflecting
objects are in reality. Imagine, on figure 1, where a ray bouncing on the
top of the high building and then on the small building in front will
end up in the radar image. In figure 2 a bigger part of a radar image is
shown, with some interesting features, and a high amount of detail.

5
Figure 1: The way radar observes the earth. High points get overlayed on low
points, and behind buildings a shadow is observed.

Figure 2: Radar image showing part of Delft University of Technology. The

radar looks from the top, so the bottom of the buildings is lower
in the image than the top, as you would expect in a optical picture.
Shadows can be clearly seen below a number of buildings. The most
prominent building, the 22 story EWI building in the centre of the
image, is overlayed on top of the Mekelpark and the civil engineering
building and its shadow. Note that you’re looking at the ’back’ side
of the buildings, the side that is pointing away or up, as the radar
observes from the top. Especially interesting in this view is the lack
of shadow of the EWI building. While it is by far the highest building
there is no shadow at all. This is because the building acts as a mirror,
and the terrain and building at the radar side of the building are
seen through this mirror. The signal has to travel a longer way, and
as such it ends up further away in range in the image. This image
is created by averaging 72 images taken between 2009 and 2011 by
the TerraSAR-X satellite. It is taken with a wavelength of 16 mm, a
ground resolution of 3x3 m, and an incidence angle of 24◦ .

6
 2.3 scatterers 7

3
scatterers
samples
signal 2
real part of signal

phase
0

−1

−2

−3
samples

Figure 3: Scatterers, sampled signal, and fully reconstructed signal. This gives
an idea of how a scattering surface could look and what signal that
would give.

2.3 scatterers

An object that reflects radar signal back to the satellite is called a

scatterer. What can be seen as one scatterer can have a size of anything
between a couple of centimetres and something very large, like a
skyscraper (when acting as double bounce corner reflector).
Essentially everything that causes the radar signal to return at the radar
at the same time can be part of the same scatterer.
The influence of a scatterer in a radar image extends over a resolution
cell, which has a size and shape dependent on the bandwidth and
spectrum of the radar signal. The shape of this influence area is the
impulse response. In other words: the radar signal is a convolution of
the scatterers with the impulse response.
A radar image is a sampled 2 dimensional signal, and since it is sampled
above the Nyquist frequency, it is a discrete, non-aliased, signal. The
signal can be perfectly reconstructed from the sampled signal. An
example of such a reconstructed signal is shown in Figure 3.
Every point of the signal is built up as the sum of all reflections in a
resolution cell.
When the energy is concentrated in a small area the signal represents a
point scatterer. This type of scatterer is the easiest to deal with since it
has defined properties: one location with one radar cross section.
This also means that a point scatterer isn’t influenced by geometric
decorrelation, which means that small incidence angle variations do
not influence the phase.
Point scatterers generally are corners or poles, which are triple respec-
tively double bounce scatterers. A single bounce pointscatterer will be
less strong, so can only be detected when there’s little signal in the
surroundings.
If the energy is more distributed over the resolution cell the signal
represents distributed scattering. This type of scattering is undergoing
8 radar signal theory

geometric decorrelation, and also is more prone to temporal decorre-

lation since the signal changes as soon as only a part of the reflecting
objects is changed.
The distribution of distributed scattering can be divided in three classes.
It can be a 1, 2, or 3 dimensional distribution. If it is a 1d or 2d
distribution the signal of observations from slightly different looking
directions can be filtered to match each other. If it is a 3d distribution
it can only be coherent if the looking direction is the same (it is very
prone to geometric decorrelation). A point scatterer in this sense is a 0d
distribution, hence a point.

2.4 the impulse response

An impulse response is the response of a dynamic system to a very

short input signal, a dirac pulse. For a point scatterer all reflected energy
is concentrated at one point, hence an infinitely short signal. So the
shape of the radar signal around a point scatterer equals the impulse
response function of the signal. The impulse response, and hence the
shape of a point scatterer, can be easily calculated if the spectrum of
the signal is known, by taking the Fourier transform of the spectrum.
The spectrum of the radar image is defined by the system and process-
ing characteristics. The most simple spectrum is a rectangular spectrum.
This spectrum gives a sinc as impulse response, which has very strong
sidelobes. To reduce the intensity of the sidelobes a window can be ap-
plied during processing. For SAR commonly applied windows are the
raised cosine (RC) or generalized Hamming window, and the Kaiser
window, which has generally better characteristics [Cumming and
Wong, 2005]. The sidelobe patterns of these windows can be seen in
figure 4
The spectrum of SAR systems is often described in the metadata. Since
a SAR image is a 2D signal it has two independent spectra, one in range
and one in azimuth. The spectrum in range is equal to the spectrum
of the chirp, and is mostly rectangular, and in azimuth it depends on
the antenna pattern. For some systems the antenna pattern is removed
during processing, such as TerraSAR-X and RADARSAT, otherwise
it is necessary to take it into account. The antenna pattern is a sinc2
function [Swart, 2000, Geudtner, 1995] , which is dependent on the size
of the antenna and the wavelength.
The spectra of several SAR systems can be found in table 1. These
are the spectra that come from the default SAR processing for those
satellites. The SAR images can be refiltered to a more optimised window
as the Kaiser window [Cumming and Wong, 2005], as that will either
reduce sidelobes or increase resolution, depending on the setting. This
window is designed to have most energy in the main lobe for a given
resolution related parameter. This parameter can be chosen to balance
resolution and sidelobe suppression. A Kaiser window with β = 3 very
much resembles a window of Envisat/ERS in azimuth direction. With
β = 3.55 the sidelobe ratio of Terrasar-X is matched, but with slightly
higher resolution.
This spectrum shape defines the actual resolution. There is no fixed
definition of resolution, but a commonly used measure is the -3dB main-
 integrated
Satellite spectrum function window sidelobe
resolution ratio
ISLR dB

ASAR/ERS - range RC α = 0.75 1.02 -15.9

ASAR/ERS - azimuth RC α = 0.75 * antenna pattern 1.11 -21.3
TerraSAR-X* RC α = 0.6 1.19 -25.0
RADARSAT-2* Kaiser β = 2.4 1.03 -17.9
- Kaiser β = 3.55 1.16 -25.0
- Kaiser β = 3 1.11 -21.3
- Rectangular 0.89 -9.7

Table 1: Signal spectrum of default processing of various SAR systems. For

comparison the resolution and integrated sidelobe ratio of some Kaiser
windows are shown. *) Signal is corrected for antenna pattern during
processing and has the same spectra in range and azimuth.

spectrum spectrum
real part of signal

real part of signal

−1 0 1 −1 0 1

←→ ←→
1.20 samples 1.80 samples
(a) rectangular window (b) Hamming window

spectrum spectrum
real part of signal

real part of signal

−1 0 1 −1 0 1

←→ ←→
1.60 samples 1.60 samples
(c) raised cosine window with α = 0.6 (d) Kaiser window with β = 4

Figure 4: Spectrum and impulse response of several windows. It can be clearly

seen the resolution and sidelobe levels of the windows are very
different.

9
10 radar signal theory

real part of signal

real part of signal

samples samples
(a) 1.0 mainlobe distance (b) 1.06 mainlobe distance

Figure 5: Resolution explained by two point scatterers: closer than 1 mainlobe

width they are not separatable, further away they are separatable.

lobe width. This is the same as the distance that two point scatterers
need to be apart to be separatable. This is shown in Figure 5.

2.5 interferometry

When the composition of some scatterers is the same in multiple obser-

vations, the phase difference between them can be directly related to
the difference in signal travel time to the satellite, hence deformation is
measured.
There are some more complications which must be accounted for.
Because the orbit of the satellite is slightly different at each pass the
height of a scatterer influences the signal travel time to the satellite.
This allows the phase to be used to estimate elevation, but also makes
that the elevation must be taken into account when only deformation
is wanted.
When multiple acquisitions are available, and the deformation is mod-
elled using few variables, the elevation can be estimated together with
the deformation. This is further complicated by atmospheric delays,
ambiguous phases, unmodelled deformations, and changing scatterers.
The latter is further discussed in the next section.

2.6 signal, clutter and coherence

In radar interferometry not all pixels of an image are usable. Scatterers

for which the composition does significantly change between observa-
tions are not usable for interferometry, because the composition change
also causes a phase change.
To describe the quality of scatterers on the basis of these changes some
terms are introduced:
• signal to clutter ratio (SCR). SCR is the ratio between the power of
the signal and the power of the clutter, shown in figure 6, where

– Signal is the reflection from the scatterer. In the case of point

scatterers this is a clear definition. In the case of distributed
scattering it is more vague, there a useful definition is that it
is the part of the signal that has a stable composition.
 2.6 signal, clutter and coherence 11

Figure 6: Signal and clutter explained. With s: signal, c: clutter, mA : expected

amplitude, σA : amplitude standard deviation, σφ : phase standard
deviation. From Adam [2003]

– Clutter is the rest of the signal, the unwanted part. This is

the signal that comes from other objects, but ends up in the
same pixel.

• Coherence. Coherence is defined for two images, as the part of the

signal that has the same composition in both images. As formula
for the signals y1 and y2 :

E y1 y∗2
γ= r (2.1)
2 2
E |y1 | E |y2 |

So SCR is defined for one image, while coherence is defined for a pair
of images. In formula form SCR becomes:

s2
SCR = . (2.2)
c2

where s2 is the power of the signal and c2 the power of the clutter.
One inconvenience is that SCR has a range of [0, ∞], while coherence
has a range of [0, 1]. To relieve us from this inconvenience the signal to
signal+clutter ratio (SSCR) is defined:

s2 SCR
SSCR = 2 2
= (2.3)
s +c SCR + 1
SSCR
SCR = − .
SSCR − 1

So coherence is very related to SCR, but for two images. As such it can
be related to SCR or SSCR: [Just and Bamler, 1994, Hanssen, 2001]

p
γ = SSCR1 · SSCR2 (2.4)
1
= r   
1 + SCR−1
1 · 1 + SCR−1
2

where SSCR1 and SCR1 relate to the first image and SSCR2 and SCR2
to the second.
12 radar signal theory

2
experimental
1.8 CEOS93
Hanssen01
1.6

1.4

std of phase [rad]

1.2

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
SSCR

Figure 7: SCR can be related to the standard deviation of the phase error of a
scatterer. Two different formulas for doing this are shown here, both
only valid for SSCR close to 1. Formula 2.6 (CEOS93) and formula 2.7
(hanssen01)

Here the benefit of using SSCR can also be noticed.

Now if the SCR is the same at both observations this reduces to simply:

γ = SSCR. (2.5)

The phase error can be related to the SCR, for high SCR or SSCR close
to 1 the following is valid: [CEOS, 1993]

1 1 1 − SSCR
σψ ≈ √ , σ2ψ ≈ = . (2.6)
2 · SCR 2 · SCR 2 · SSCR

Another estimate, also valid for SSCR close to 1,

1 − SSCR2 2 · SCR + 1
σ2ψ ≈ 2
= , (2.7)
4 · SSCR 4 · SCR2

is directly derived from the following estimation for interferometric

phase in relation to coherence: [Hanssen, 2001, 4.2.31]

1 − |γ|2 1 − SSCR2 2 · SCR + 1

σ2φ ≈ = = . (2.8)
2 |γ|2 2 · SSCR 2 2 · SCR2

The last formula is a bit closer the the experimental values in the
sensible range of SSCR > 0.5, but both are very accurate above SSCR =
0.9, as can be seen in figure 7.
The SCR of a scatterer can be estimated using various methods. For
SCR estimation of corner reflectors on even fields the clutter level is
directly estimated from a neighbourhood that is out of reach of the
mainlobe and sidelobes of the scatterer. The best signal level estimation
is the signal at the location of the corner reflector. See figure 8. This
method is used for calibration of test corner reflectors [CEOS, 1993].
Also the commonly used metric amplitude dispersion can be related
to SCR, assuming the SCR is the same in all observations. Amplitude
 2.6 signal, clutter and coherence 13

Figure 8: SCR estimation window used for calibration of corner reflectors. The
hatched area is used for clutter level estimation, the cross for signal
level.

2
experimental
1.8 theoretical

1.6

1.4
Amplitude dispersion DA

1.2

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
SSCR

Figure 9: Amplitude dispersion related to the SCR. An experimentally derived

relation and a theoretical estimation using formula 2.9 are shown.
The theoretical estimation is accurate for SSCR > 0.9, but degrades
quickly.

dispersion is calculated from a stack of images, using the standard

deviation σA and the mean mA of the amplitudes: [Ferretti et al., 2001]
r
σ c 1 1 − SSCR
DA , A = =√ = (2.9)
mA s 2 · SCR 2 · SSCR

This relation to SCR is also only valid for high SCR, as can be seen in
figure 9.
The coherence between two images can be estimated too. This is done
by comparing the images using a spatial window. This window is often
chosen to cover an area containing a reasonable number of scatterers,
which are then assumed to be stationary as a whole. This window can
also be narrowed down to contain just about one scatterer, by using
a weighted window which puts more weight in the centre. However,
the smaller the window, the more bias will appear in the estimation
[Hanssen, 2001, 4.3.4]. The formula for coherence estimation including
the optional window can be constructed as follows:
Rk ∗
−k s1 (t + u) · s2 (t + u) · w (u) du
γ̂ (t) = qR Rk (2.10)
k 2 2
−k |s1 (t + u)| · w (u) du · −k |s2 (t + u)| · w (u) du

where s1 and s2 are the radar images, k is the window size, and w
the optional window weights. This is the 1D case, the 2D case follows
logically from this.
14 radar signal theory

This gives the complex coherence, and is often converted to real coher-
ence by just taking the absolute value.
Another option is to project the complex coherence on the observation
phase difference. This will cause the coherence estimate to be reduced
when the interferometric phase of the observation is different from the
average of the observation window.
 
|γ| (t) = Re γ (t) · e−φ(t)i (2.11)

where φ (t) is the interferometric phase or phase difference.

This alternative conversion to real coherence in combination with a
high resolution window allows coherence estimation of single scatterers,
without the need for a stationary area surrounding the scatterer.
I N S A R I N V E S T I G AT I O N T O O L S
3
3.1 introduction

InSAR techniques, and especially time series InSAR techniques, pro-

duce large amounts of information. The results contain thousands of
scatterers, each having their own deformation pattern, and also other
properties like height, shape, size, and quality measures. These proper-
ties are estimated using a complicated processing chain. During this
processing decisions are made. For more insight on those processing
decisions the original and intermediate data should be analysed, which
increases the total amount of data involved even more.
An optimal way to effectively make this amount of information in-
terpretable is visually. A good way to get an overview of what is
happening is to plot a single parameter on a map. However, more
advanced visualisation is necessary, since the information is spatially
dense, multivariate, and relative.
It is spatially dense, as the covered area has more spatial detail than can
be shown on a single image. A single house cannot be distinguished on
an image showing the whole province. The information is multivariate,
since for every point the behaviour in time is available, along with
numerous other properties. Finally, the information is relative, a phase
time series has no information on itself, it must be related to a reference
point to contain information. Any point can be chosen as reference, and
interpretability changes with this choice, so the number of visualisation
possibilities is increased dramatically.
An interactive, multi-scale environment in which all this information
can be visualised in multiple ways can help making this information
interpretable, but is currently lacking.The current tools for time series
InSAR analysis used at TU Delft are split in two steps. The radar data
is first processed by DORIS, which creates coregistrated images and
interferograms, after which the DORIS results are loaded into Matlab
for further processing using persistent scatterer techniques.
DORIS, Delft Object-oriented Radar Interferometric Software [DORIS,
1998], is a powerful package for interferometry. However, it is intended
to produce single interferograms, it has no time series analysis. It is
hard to do interactive research or scripted processing, since all settings
need to be passed through input files. Since it is written in C++ the
barrier is high to change parts in the code, and one is limited to the
available options.
Several toolboxes have been built on top of Doris for time series analysis,
including Stamps, and DePSI. These are both written as a combination
of shell-scripts and Matlab code. DORIS plus its successors provide a
processing chain to get from single look complex (SLC) radar data to
unwrapped time series.
None of those toolboxes provide an interactive environment, nor do
they provide the easy interface to the metadata and calculation functions

15
16 insar investigation tools

that is needed for creating such an environment. Since all data come
preprocessed from DORIS the processing functions are not available
in an interactive environment, so to use different settings for filtering,
oversampling, or other DORIS functions, DORIS must be called again.
The functional requirements for this environment are set as follows:
• Give easy access to all data, metadata and functions.
• Quick and interactive loading, calculating and displaying of data
and metadata.
• Ability to handle large datasets faste and efficient.
• Reimplement DORIS functionality to be accessible from Matlab.
• Minimize possible error sources, such as geoid, continental drift,
nonlinearity of topographic phase.
• Processing functions have sufficient precision, least possible error
sources.
• Consistent coordinates, coordinates should not depend on the
used crop of oversampling ratio.
In this study a suite of tools is developed that satisfies these require-
ments, which is described in this chapter.
First the methodology is explained, followed by sections about the
graphical InSAR Inspector tool and the underlying object-oriented InSAR
processing and metadata toolbox. The chapter concludes with a section
about the outlook to the future of the tools, describing where it is used
and experiments with features to be added.

3.2 methodology

To make all data, metadata and functions easily accessible an object

oriented structure is used. The available information can be seen as a
big tree of objects containing the information. A scatterer is an object,
which has properties and functions working with this data, just as a
radar image. There are several programming languages available that
are object oriented and interactive, such as Python and Matlab. For this
toolbox Matlab is used, mainly because of easy reuse of existing code.
In order to assure quick loading, calculating and displaying several
techniques are used, described in the following paragraphs.
Dependent data are computed ‘on the fly’. On various places data is
transparently calculated when it is requested. For example when a part
of an oversampled image is requested the corresponding part of the
original data is loaded and oversampled to the requested oversampling
level. Or when the location on earth of a scatterer is requested it
is calculated using its coordinates and height. The same for phase
corrected for atmosphere and topography, the atmosphere phase is
interpolated, topographic phase is calculated using the height, and both
are subtracted from the original phase. This way a consistent state is
assured, when for example the height of a scatterer changes, the phase
is automatically corrected.
Slow and complex computations are approximated by polynomials.
In DORIS this is already done for some computations, like flat-earth
 3.3 insar inspector 17

phase and h2ph, which are approximated by 2D polynomials. Here

this is taken a step further: 3D polynomials are used for topographic
phase and its derivative h2ph, so their dependency on height is taken
into account. Also for geocoding 3D polynomials are used, both for
[X, Y, Z] and [lat, lon, height] coordinates. These polynomials are used
in dependent data and evaluated transparently.
For handling large datasets all necessary data is stored on disk, and
loaded only when needed.

3.3 insar inspector

The InSAR Inspector is a GUI that provides a way to make the spatially
dense, multivariate, and relative information that is involved in InSAR
processing fully explorable.
An overview map is shown at the main screen, on which all scatterers
are shown. When a scatterer is selected, either on this overview map
or using a unique point ID, all information about a scatterer is shown,
both as numerical data and using plots. All information that is added
during processing is shown, such as quality measures, height, and
deformation. Information that varies per acquisition is plotted, such as
amplitude or deformation. For deeper inspection several options are
available, such as showing the underlying radar image data, calculating
a coherence matrix, and data access from the command line.
The base components for the GUI are:
• Zoomable background.
• Phase time series plot.
• Point information table.
• Additional information plot.
The state and contained information is exposed to the programming
environment as an object. This way it is possible to interact with the
GUI in a programmatic way, and do custom processing on information
shown in the GUI.
The GUI is initiated with a PointSet. This is an object that contains all
information that is necessary for operation, such as access to the radar
images, the mean amplitude, and a set of selected points.
In the following sections the GUI components are described.

3.3.1 Zoomable background

The largest panel contains a zoomable map on which the mean ampli-
tude of the radar images or mean reflectivity map (MRM) is shown, with
the selected scatterers on top, see figure 11. The canvas can be dragged
and zoomed using the mouse. On loading a medium resolution MRM
is shown, a higher resolution version can be loaded when zooming in
to see the full resolution details of the radar image. On the canvas the
following features are shown:
• Extracted scatterers, coloured with a property. Any property can
be used.
18 insar investigation tools

Figure 10: The elements that appear in the InSAR Inspector GUI.

• Selected point and reference point, connected by a line.

• Subpixel position of the point at each acquisition.
Using this map single scatterers can be selected. Depending on the ’use
points’ setting the scatterer is loaded from the PointSet, or will be loaded
from the SLC images on the fly, including resampling to the clicked
subpixel position. This makes it possible to examine even regions which
were not selected using the automatic selection algorithms.
The reference point can be selected the same way with the right mouse
button. A line to the reference point is always shown to remind the
user of the fact that the shown information is relative, and which point
it is relative to.

3.3.2 Phase time series plot.

One plot is dedicated to showing deformation. On this plot the phase of

the selected point with respect to the selected reference point is shown.
The phase is corrected for height and optionally for atmosphere, so
only deformation is left, distorted by noise. The phase is shown in
3 ambiguities, wrapped to [−3π, 3π). This shows how the data is the
unwrapped and allows visual unwrapping, so the correctness of the
unwrapping can be checked, see the example in figure 12.
Depending on the ’estimate height’ setting the height is automatically
estimated upon loading a scatterer. Otherwise the height available in
the PointSet is used.
The height of the selected point can be changed interactively to man-
ually find a correct decomposition of the phase in deformation and
height difference, as shown in figure 13, or to check the validity of an
automatic decomposition.
Figure 11: The zoomable background shows the selected points. The colour
property can be chosen using the dropdown menu, which lists all
point properties available in the pointset. The green line connects
the selected point to the reference point, here a point on a bridge
built in 2010 is shown with as reference a stable point on the ground.
The amplitude peaks at each acquisition are shown, and show that
the position was less stable before the bridge was built, as expected.

Figure 12: Automatic unwrapping clearly makes a mistake in the case shown
here. Points with a suspicious linear deformation velocity can be
inspected and errors can be tracked down.

19
20 insar investigation tools

Figure 13: The phase shown in the plot is corrected for topographic phase. If the
amount of topographic phase is wrong, the height can be changed
until the residual phase is plausible to be deformation. By changing
the height smoothly by sliding the mouse this can be done very
quick.

3.3.3 Point information table

The point information table gives a quick overview of the point proper-
ties extracted during processing. All point properties that are available
in the PointSet are shown.
The table starts with the point ID, a unique identifier based on the
radar coordinates. This doubles as input field to quickly return to a
previously selected point. This way interesting points can be written
down and found back later on.
Following are all numerical point properties, such as the height, defor-
mation rate and ensemble coherence. Only properties with one value
per scatterer are shown, other properties, which have one value per
acquisition, can be shown in the additional information plot.

3.3.4 Additional information plot

This plot is intended to show the relation between multiple properties.

For example the amplitude behaviour versus time or perpendicular
baseline can be seen, or by using colour even both together. Each of the
x, y and colour axes of this plot can be set to a property containing time
series information, such as date, amplitude or perpendicular baseline.
All PointSet properties containing time series information can be chosen
for any of the axes. See figure 14 for examples.

Time selection slider

Using this slider the currently selected radar acquisition can be changed.
This selection is shown in the phase plot, additional plot, and subpixel
position plot on the background. It also determines the displayed image
of the SLC stack. This way changes in any of those properties can be
related.

Settings

Some frequently used settings are shown in the GUI, other settings
need to be accessed through the GUI object via the command line. The
GUI contains toggles for:
 3.4 object oriented insar toolbox 21

(a) (b)

Figure 14: The additional information plot is used for finding a relation between
properties. a) perpendicular baseline shows a clear relation with
amplitude, which is useful for determining the scatterer size. b)
Multiple relations can be shown together by using colour or size in
addition to the x-, and y-axis.

• SLC stack: Toggles loading of the SLC data of the surroundings

of the selected point.
• Use points: Toggles between preselected points and raw data.
• Estimate height: Toggles automatic height estimation.

3.3.5 The GUI object

The GUI object gives access to the selected points and all their prop-
erties, and also to all GUI elements and display functions. Interaction
via the command line is necessary when more flexibility or more user
input is required than can be given through the GUI. For example to
show a custom property on the background, make high quality plots of
selected points, or estimate new parameters or relations. See appendix
A for the implementation details and examples.

3.4 object oriented insar toolbox

The information that is shown in the GUI, and also is used while
processing or post-processing, needs to be stored in an accessible way.
Functions, which are also used by the GUI and processing routines,
must be able to access this information too, so it doesn’t have to be
supplied manually. All this is done in an object oriented structure.
All information is stored in objects, and a lot of processing functions
are added to the object classes, such as algorithms for radar image
processing, scatterer selection, and geocoding.
The toolbox is able to perform the following actions:
• Load metadata from DORIS .res files
• Calculate metadata dependent properties
• Load SLC data with on the fly oversampling and filtering
22 insar investigation tools

PointSet SlcStack
tsinsar_inspector +slc +master_idx
+n_points +Slc
+pointset +n_observations +...
+handles +pnt_line
+slc +make_poly_refpha_h2ph()
+pnt_pixel +xyz2linepixel()
+... +phase +latlonh2linepixel()
+loadbackgrounddetail() +pnt_height +...()
+...()
1

SlcImage 1..*

+data Slc
+raster +precise_orbit
+slc +wavelength
+spectrum_az +date
+spectrum_ra +poly_latlon
+coherence() +poly_refpha
+interferogram() +poly_h2ph
+subtract_refpha() +poly_dopplercentroid
+filter_az_to() +...
Spectrum
+filter_ra_to() +linepixelheight2xyz()
+...() +window +xyz2linepixel()
+bandwidth +make_poly_geocoding()
+shift +...()

Figure 15: The most important classes and their connections are shown in an
UML Class diagram. This shows for each class the properties and
functions, and how they relate to each other.

• Store processing results and intermediate steps in a clear way

• Do radar image processing, such as spectral filtering, reference
phase subtraction, interferogram creation and coherence calcula-
tion
• Do scatterer selection using existing and newly developed algo-
rithms
• Work with sub pixel precision
All functions and data are grouped in several classes of objects, which
use each other in order to break down everything in clear parts. The
structure of the toolbox can be seen in figure 15 as a UML class diagram,
showing the most important classes, properties, and functions. Several
other smaller classes are used, such as a 3D polynomial class for the
storage and evaluation of 3D polynomials, a raster class for the storage
and transformation of coordinates of gridded data, and a disk storage
class to store and read data on disk transparently.
In the following sections the classes and their functionality are ex-
plained.

3.4.1 Slc and SlcStack

In an Slc object all metadata of a radar image is stored. This includes

the observation date, satellite orbit, and sensor and signal parameters,
but also the location and format of the files containing the radar image
data.
Furthermore it provides functions to access the SLC data, and functions
to calculate metadata dependent information. In this category there
is orbit dependent information, such as satellite velocity and looking
 3.4 object oriented insar toolbox 23

angle, but also pixel spacing and resolution, which depend on the
signal parameters. Also functions for geocoding and inverse geocoding
are included, which transform given coordinates using orbit and signal
parameters.
The information is imported from DORIS result files, complemented
with some satellite specific information from a small database, which is
not available in the result files.
The SlcStack is mainly a container for a stack of Slc objects. It provides
convenient access to properties that are either the same for the whole
stack, or most often needed from only the master. These are mostly
properties that are related to the coordinates, for example geocoding or
ground resolution.
To use a stack of radar images the radar data should be coregistrated
and resampled using DORIS, so from that moment the only the master
coordinates have to be used. Any following DORIS processing steps,
such as interferogram creation or topographic phase correction, are
not used, as those steps can be performed more flexible using the
toolbox. Oversampling in DORIS is allowed, the coordinates are auto-
matically converted back to not oversampled coordinates, but it is not
recommended due to the big increase in disk usage.
Topographic phase computation is done using the SlcStack, since the
master SLC is necessary for this.
Geocoding and topographic phase computation is done using 3D poly-
nomials. This way it is very fast even for millions of points, and an
otherwise computationally expensive geoid model can be included
at no extra cost. In chapter 5.2 the theory and implementation of the
computation of these polynomials is discussed.

3.4.2 SlcImage

An SlcImage object holds SLC data in memory, together with under-

lying metadata. It is used to do the actual processing that is done on
image level, before points are selected.
The metadata that is stored is necessary to relate the data with other
data, and for a number of processing functions. The following metadata
is stored:
• The crop position. This gives the coordinates and pixel spacing,
in master coordinates.
• Processing state. Things such as whether or not topographic phase
is subtracted, and the current signal characteristics in the form of
signal spectra.
• The corresponding Slc object.
Many processing functions are implemented in this class. The most
important of these processing functions are:
• Oversampling. Since the radar signal is a continuous defined it
can be perfectly oversampled. This function takes into account
the spectral shift, it adjusts the oversampling filter to match the
spectral shift of the data.
24 insar investigation tools

• Spectral filtering. Spectral filtering is a very powerful tool to get

more out of the radar data. It can be used to reduce sidelobe
levels or increase resolution (chapter 2). Another use is to match
the spectra of two radar images to increase the coherence. This
can be done both in azimuth direction to correct for doppler shift,
or in range direction to correct for incidence angle difference.
• Interferogram computation. Compute the phase difference be-
tween two or more images. Optionally compute coherence and
use that as amplitude.
• Coherence estimation. Estimate the coherence between two im-
ages, either using a given window, or using a window that is
optimised for high resolution coherence estimation.
• IRF correlation. New method for point scatterer selection, details
can be found in chapter 4.
• Peak selection. Find the local maxima and determine their posi-
tions with sub pixel precision. This comes in two flavours, one
that finds all peaks, and one that finds the nearest peaks for a
given list of coordinates. This peak selection is part of the scatterer
selection described in chapter 4.
SlcImages can be used in a stack, in which case the function is per-
formed on all images, or if there is functionality specifically for stacks
that is used. This is the case for interferogram computation or some
forms of spectral filtering.

3.4.3 PointSet

The PointSet is a database with point properties of points that are

extracted from a stack of SLCs. There are several types of properties:
• Base properties that are necessary for correct operation. This is the
case for the point coordinates, phase time series and observation
dates.
• Dependent properties. These are calculated on the fly from the
base properties, and can be used just as normal properties. Some
of these properties are:

– Topographic phase. Calculated using the point coordinates

and the 3D polynomial from the SlcStack.
– Corrected phase. Scatterer phase minus the topographic
phase, and optionally the atmospheric phase.
– Deformation. Corrected phase minus the reference point
phase, converted to meters.
– Geocoded coordinates. Latitude and longtitude calculated
using the point coordinates and a 3D polynomial from the
SlcStack.

• Custom properties. Any custom property can be added during

processing, or afterwards. Examples of such properties are defor-
mation velocity, quality estimations, coordinates per observation
and temperature time series.
 3.4 object oriented insar toolbox 25

shift
window
shape

−1 0 1
bandwidth
Figure 16: A normalised spectrum is defined with three properties: window
shape, bandwidth, and center frequency shift.

All of those properties are either:

• One value per point, such as point height, or point quality mea-
sures.
• One value per observation per point, such as phase or deforma-
tion.
• Only one value per observation, being the same for all points.
Such are observation date or temperature.
Furthermore a place for all metadata is provided, which can hold any
kind of data. Here things such as a mean amplitude image, estimated
atmosphere or reference point are stored.
Since a PointSet might store a large amount of data, any data can
be stored on disk an only be loaded when necessary. This makes it
possible to store millions of points without running out of memory.
This is accomplished using the hdf5prop class, which makes properties
stored on disk usable just as normal variables.

3.4.4 Spectrum

For processing radar images optimally the signal properties are neces-
sary. Most of the radar signal properties are defined by its spectrum,
such as the impulse response, resolution and spectral shift. For process-
ing data only the normalised spectrum is important, which relative to
the sampling rate, and so always goes from -1 to 1. This normalised
spectrum is stored in this object. The absolute frequencies and band-
width can be found in the Slc object, so can also easily be accessed
when necessary.
The spectrum is defined by the following properties:
• Window shape. Defined as a polyline.
• Bandwidth. Relative to the full bandwidth, so a value in the range
(0,1).
• Centre frequency shift. When the spectrum is not zero centred, as
with zero doppler processed SAR images, this property defines
the shift.
The spectrum class is used to represent the actual state of a radar image.
This might be the original spectrum, or a modified spectrum due to
filtering or oversampling.
26 insar investigation tools

It contains several functions to filter data to a new spectrum. Also

functions are included which calculate spectrum dependent properties
such as the impulse response, sidelobe positions and mainlobe width.

3.4.5 Other classes

Furthermore some other useful classes are included, which are more
general and are used on various places. These are described below.
Poly3D is a class that stores a 3D polynomial, and includes a function
to efficiently evaluate it.
A Raster object stores the range and step size of gridded data, and
several functions to make intersections, subdivisions or convert between
coordinates and indices.
RasterData contains gridded data, and the corresponding coordinates
in a Raster object. It is similar to SlcImage, and contain the processing
functions of SlcImage which also can be done on general images, such
as peak selection or gaussian blur.
hdf5prop is a class which handles very large arrays by reading and
writing data on disk transparently. Those variables can be used and
indexed just as normal variables. The data is stored in a HDF5! (HDF5!)
file, which is designed for storing large amounts scientific data in a fast
an structured way.
dependentprop makes dependent properties in a PointSet act as normal
variables. It contains a reference to a function which calculates the
desired property.

3.5 outlook

The toolbox described in this chapter is in a very usable and complete

state, and is now mainly in use at Hansje Brinker.
The DePSI persistent scatterer software has been rewritten around the
toolbox classes, resulting in a modular and much more flexible system.
However, still some things could be improved.
Both on the pro- and postprocessing sides of the InSAR production
chain there are improvements foreseen. On the pre-processing side,
the current implementation of the software is very flexible in its del-
ing with several InSAR implementation chains (focussing, resampling,
cropping, and coregistration can be done through different currently
used recipies, such as used by StaMPS, the DUT implementation of
DePSI, and others, all of which are based on the DORIS software).
An initial implementation in python, to replace these rather complex
implementations has been developed and is currently used at Hansje
Brinker. This preprocessing chain however cannot yet deal with raw
imagery.
On the postprocessing side, several visualisation and reporting tools
were developed, but these have not yet reached maturity. Examples of
these are the pdf-maker which creates standardized reports, several
plotting routines.
 3.5 outlook 27

There are also several future projects that may improve the selection
and characterisation of the point measurements. Below, a few of these
future improvements are outlined.

3.5.1 GIS Integration

Precise geolocation and integration with external data sources is best

served through seamless integration with existing geographical infor-
mation systems in place at the parties using the InSAR results. Cur-
rently, exports towards Google Earth, ESRI Shapefiles, and simple text
documents has been developed.

3.5.2 3D AHN2 viewer

One noteworthy feature that couldn’t be integrated due to MATLAB

limitations is integration with a 3D pointcloud viewer, with the main
purpose of relating the scatterers to AHN2 data. Some tests are de-
scribed here.
To identify the precise location of point scatterers they must be related
to auxiliary real world information. This serves both as a validation
that the point is representing a real object, and for deeper inspection
of the scattering origin. This auxiliary data might be 2D maps, such
as topographic maps or aerial photos, but even better 3D information.
The best 3D information that is available is the Actueel Hoogtebestand
Nederland versie 2 (AHN2), a very detailed and accurate laserscan of
the Netherlands.
Since AHN2 is very detailed and consists of billions of points, it cannot
be simply visualised using of the shelf viewers. The computer graphics
group of the TU Delft has developed a viewer for visualising the AHN2
point clouds, with the possibility of showing other data along.
The Persistent Scatterer InSAR (PSI) results of Delft using TerraSAR-X
have been loaded in the AHN2 viewer, which gives very promising
results. An impression is shown in figure 17.

3.5.3 Bayesian Point Selection

Automatic selection of valid observations, thereby discarding unreal-

istic ’coincidental’ observations from e.g. water or grass lands, can
strongly improve the automation of the InSAR process. The level to
which this discrimination can be made, determines the number of good
observations that are produced. There are several steps foreseen in the
improvement of this process, one of the more promising approaches
being a bayesian point selection algorithm which combines several
available quality measures in an optimal way, in contrast to separate
thresholds on each of those.
Figure 17: TerraSAR-X PSI results of Delft shown along AHN2 in a AHN2
viewer. The coloured dots are point scatterers, the red are from the
descending pass, from a stack of 72 acquisitions; the yellow from the
ascending pass, using 68 acquisitions. The descending pass looks
from the lower right with an incidence angle of 24◦ , the ascending
from the upper left with an incidence angle of 39◦ . The area is
the northern part of the Spoorzone, around the windmill De Roos.
Reflections from roofs, ground-wall corners and train rails can be
identified.

28
P O I N T S C AT T E R E R S E L E C T I O N
4
4.1 introduction

One of the major challenges in PSInSAR is the selection of usable scat-

terers. In this chapter a method is presented for the scatterer selection
in PSInSAR techniques, which can replace amplitude dispersion as
primary selection criterion in existing processing chains.
The selection method that is currently used most is using amplitude
dispersion, followed by a phase based selection criterion and sidelobe
filters for the final selection.
However, this prior quality description is not always accurate enough
to make a decent selection, especially with small stacks. This means the
first selection contains many incoherent scatterers, so strong filtering
on phase is necessary, which might discard interesting scatterers with
strong unmodeled deformation. Also the location of scatters is not
accurately estimated, and multiple copies of the same scatterer might
be included due to sidelobes.
The goal of this study is to develop a method to select coherent scatter-
ers with an accurate quality description and accurate location.
To accomplish this the focus is set on point scatterers. Point scatterers
have a high chance of staying coherent because they are less susceptible
to both temporal and geometrical decorrelation [Hanssen, 2001]. Their
quality can be easily defined by means of signal to clutter ratio. Also
point scatterers have one specific reflection center, and thus can be
located more precisely and interpreted easier. For the detection of point
scatterers the signal characteristics are extensively used to compare the
radar signal with a reconstructed ideal point scatterer, using a spatial
window.
In section 2 the algorithm and implementation is discussed. Section 3
shows the results of IRF correlation by means of examples of processed
real data, and compared to amplitude dispersion. Section 4 relates the
developed measure to the quantifyable measure SCR, supported by
simulation results.

4.2 algorithm

The quality of a point scatterer can be described by its signal to clutter

ratio. For the separation of scatterer signal and clutter signal the corre-
lation between the radar signal and a reconstructed ideal point scatterer
signal is calculated. This ideal signal is equal to the impulse response
of the radar system. A weighted window is used to only include the
local neighbourhood in the calculation and obtain a high resolution.
The weighting function defines the estimation resolution and accuracy,
where a tradeof between those must be found. A wider window can

29
30 point scatterer selection

Figure 18: Estimation window. Only the main lobe is used for SCR estimation.

have higher accuracy, but has a lower resolution. The window is chosen
so that only the mainlobe is included, as shown in figure 18.
The complex impulse response function (IRF) correlation at position x
is defined in 1D as:
R∞ ∗
−∞ s (x + u) · f (u) · w (u) du
ρirfcpx (x) = qR R∞ (4.1)
∞ 2 2
−∞ |s (x + u)| · w (u) du · −∞ |f (u)| · w (u) du

where s (x) is the signal at position x, f is the IRF and f∗ its complex
conjugate, and w is the weighting function.
This equation is constructed by taking the (cross) correlation function
of the signals s and f [Cumming and Wong, 2005, p.23]:

Z∞
ρ (x) = s (u) · f∗ (u − x) du (4.2)
Z−∞
∞
= s (x + u) · f∗ (u) du
−∞

This is then normalised by scaling with the maximum possible value,

just as it is done in coherence calculation:
R∞ ∗
−∞ s (x + u) · f (u) du
ρ (x) = qR R∞ (4.3)
∞ 2 2
−∞ |s (x + u)| du · −∞ |f (u)| du

To this equation weight term w is added that is aligned with the IRF.
For the weighting function the power of the absolute value of the
impulse response w (u) = |f (u)|2 or w (u) = |f (u)| is used. This way
the weight is zero on places where nearby scatterers have no influence,
and high where scatterers have more influence. In other words: it equals
the amount of influence of other scatterers at each distance from the
point scatterer. These windows gives good performance in resolution
and accuracy. Other window function may shift the balance between
resolution or accuracy.
The result is a complex correlation. For a real valued correlation, sig-
nifying the correlation of the signal with an ideal point scatterer with
same phase as the signal, this complex correlation is projected on the
signal phase:

s∗ (x)
 
ρirf (x) = Re ρirfcpx (x) · . (4.4)
|s (x)|
 4.2 algorithm 31

This is an alternative to taking the absolute value, this way the value
of ρirf is reduced when the signal phase differs from the average
phase of the estimation window, which is the case when the scatterer
is influenced by clutter. The value of ρirf can become zero or negative
when the phase difference is very large.

4.2.1 Implementation

The given functions are in 1D, in reality the 2D case is used, taking
both azimuth and range direction into account.
These functions are defined in the continuous domain, because the radar
signal is a continuous signal. For evaluation of the functions a sampled
signal is used. For sufficient accuracy the signal is oversampled at least
2 times. It can be implemented efficiently using two convolutions in
both azimuth and range direction, respectively. The impulse response is
truncated to the main lobe, which has little influence since the weighting
values are very low outside the main lobe. With 2 times oversampling
this results in a window size of about 5 pixels, which is small enough
to be fast in computation.

4.2.2 Stack processing

For estimating the quality of a persistent point scatterer in a stack of

radar images the IRF correlation of all images is computed separately,
and then averaged. This results in an improved estimate, assuming the
point scatterer is persistent. This way scatterers with a high mean IRF
correlation are high quality and persistent all the time, while a lower
mean IRF correlation means that it is either high quality and partially
persistent, or lower quality but persistent throughout. By looking at the
IRF correlation timeseries it is possible to discriminate between those
cases.
The amplitude behaviour is not taken into account. This makes it
independent from amplitude changes between acquisitions. Since that is
the only information that is used for quality estimation using amplitude
dispersion it is a completely independent quality measure, making it
possible to use amplitude dispersion as an additional measure for an
improved quality estimate.

4.2.3 Subpixel position

The center of a point scatterer has the highest IRF correlation, so by

only taking local maxima into account it is possible to select only
one point for a scatterer, and determine it’s position with subpixel
position. The sub pixel position is estimated by fitting a second order
polynomial trough the local maximum and its four neighbour pixels,
and computing the maximum of the polynomial. When the image is
oversampled at least 2 times this leads to a very accurate estimate.
The sub pixel position can be estimated for the whole stack at once by
using the stack average, or for each image separately. In the last case
the obtained information can be used for several purposes, such as
32 point scatterer selection

improved coregistration, detailed speckle tracking, change detection, or

for an additional quality measure.
Usage as an additional quality measure can be accomplished by cal-
culating the stability of the position as a combination of the standard
deviation in azimuth and range direction. This leads to the position
standard deviation:
q
σpos = σ2az + σ2ra (4.5)

where σ2az and σ2ra are the standard deviations of the subpixel position
in all of the images in azimuth and range direction, respectively.

4.3 results

In this section some results are discussed to show the strengths of

the technique. The advantages of treating point scatterers different
from distributed scatterers are shown. Then the accuracy as quality
measurement is compared with amplitude dispersion for different
stack sizes. After that promising results for change detection and quasi
persistent scatterers are shown in more detail.

4.3.1 Focus on point scatterers

The new method is specifically designed to only select point scatterers.

Point scatterers have a high chance of staying coherent because they
are less susceptible to both temporal and geometrical decorrelation.
Also point scatterers have one specific reflection center, and thus can be
located precisely and interpreted easier.
Additionally, distributed scattering can be very valuable too, and can
be added in a densification step. This way no valuable information is
lost, but more difficult scatterers are added later.

4.3.2 Single observation per scatterer

Sidelobes are automatically discarded due to the signal characteristics.

Only one observation is given per scatterer, which is located in the
center of the main lobe. This removes the need for other sidelobe filters.
The separation of the main lobe and sidelobes is effective, as can be seen
in figure 19, and clearly distinguishes between nearby point scatterers
and sidelobes. This separation is best when the radar image is filtered
with a medium resolution filter, such as a Kaiser window with β > 3.
This is because the mainlobe gets wider when the resolution is lower,
while the sidelobes keep the same width, as can be seen in figure 4.
This reduces the similarity between the sidelobes and the main lobe,
and hence the IRF correlation is reduced.

4.3.3 More accurate quality measure for smaller stacks

IRF correlation is an estimate for the coherence of point scatterers.

To test the accuracy of this estimate a comparison is done with other
 4.3 results 33

80

40

height [m]
0

-40

-80

Figure 19: Point scatterers selected using IRF correlation are shown on a MRM.
The sidelobes of the bright scatterer in the center are not selected, but
nearby scatterers which stronger than the sidelobes are selected. In
this case the strong scatterer with visible sidelobes is at the bottom
of a building, while at each story above that scatterer is another
scatterer. The building is the highrise at the Bosboom Tousaintplein
in Delft, a 16 story building, the building has an interior corner, as
can be seen on the photo, which causes strong reflections. The image
is acquired by TerraSAR-X in descending orbit with a resolution of
3x3m.

quality estimates: amplitude dispersion and ensemble coherence. Am-

plitude dispersion is often used for primary point selection, for which
IRF correlation is an alternative. Ensemble coherence is a quality es-
timate based on the phase behaviour, and is a measure for how well
the phase behaviour adheres to a given deformation model. In this
comparison it is used as reference quality, assuming it the best quality
measure available.
For the comparison test a stack of 69 Envisat images covering Rotterdam,
Delft and a part of the Dutch coast is used, as shown in figure 20. This
area is very varied, and includes cities, sea, industry, and rural area.
The stack has been processed using the DePSI software, with a full
pixel processing, which means that all pixels are processed. For every
pixel the local ensemble coherence is calculated, which is the ensemble
coherence to a nearby reference point. A linear deformation model
is used, and height is estimated in a range of 60 m. The ensemble
coherence obtained this way can now be used as a reference quality for
every pixel.
To test the accuracy IRF correlation and amplitude dispersion, points
were selected using both methods for stack subsets of various sizes
between 1 and 69 images. The used images were distributed evenly
over time, to avoid a bias by temporary scatterers.
Of each of those processed sets the best 20 000, 60 000, 120 000, 200 000
and 300 000 pixels were selected, according to either amplitude dis-
persion or IRF correlation. For those selections stacked histograms of
the ensemble coherence were made. This way the distribution between
supposedly good and bad points can be easily seen. The histograms
are shown in figure 21, for stack sizes of 8, 20 and 69 images.
There are several things that stand out in this comparison.
• For small stacks, up to 20 acquisitions, IRF correlation clearly
selects better points, while for bigger stack the difference gets
34 point scatterer selection

Figure 20: Overview of the processed area of 50x30 km, including the
Maasvlakte, Rotterdam and Delft. The stack consists of 69 images
and is acquired by Envisat in the descending pass. The colour shows
IRF correlation, according to eq. (4.4), all points with ρirf > 0.6 are
shown.

smaller. Even for a small stack of 8 images where amplitude

dispersion is unusable, IRF correlation gives quite good results.
• IRF correlation selects much more points with a medium en-
semble coherence. This can be points with a lot of unmodeled
deformation, but more plausible are partially persistent scatterers.
IRF correlation is less sensible to outliers than amplitude disper-
sion, so when scatterers are not present in all images they are still
selected.
• IRF correlation selects a lower percentage of the points with a
very high ensemble coherence, while amplitude dispersion selects
almost all of them. This is because IRF correlation selects only the
point scatterers, while sidelobes and distributed scatterers are left
out.
The scatterers can be divided in several classes based on IRF correlation
and ensemble coherence, to try to clarify the origin of those scatter-
ers: ensemble coherence shows how linear the local deformation is, a
medium value mean either that the phase is noisy, that the deformation
is nonlinear, or that the deformation is linear with some noisy values,
hence partially persistent. The IRF correlation value is the mean of all
acquisitions, so a medium value can mean that it is medium throughout
the stack, or that there are several high and several low values, which
indicates a partially persistent scatterer. The classes created by this
separation are shown in table 2, along with examples.
IRF correlation can already make a reliable scatterer selection for stacks
much smaller than the number of images required for amplitude dis-
persion. Where amplitude dispersion requires at least 20 images, IRF
correlation gives better results already using only 8 images. IRF cor-
relation has higher accuracy as a quality measure for point scatterers.
Combined with the explicit focus on point scatterers it gives a more
discriminating measure for valid measurements. This makes it possible
 (a) IRF correlation (b) Amplitude dispersion
4 4
x 10 x 10
6 6
238 589 152 46 29 21 16 11 9 6 x 104 238 589 152 46 29 21 16 11 9 6 x 104
0.79 0.82
0.73 0.79
5 5
8 aquisitions

0.69 0.76
0.67 0.74
4 0.65 4 0.72
included pixels

included pixels
All All

3 3

2 2

1 1

0 0
0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1
ensemble coherence ensemble coherence
4 4
x 10 x 10
6 6
238 589 152 46 29 21 16 11 9 6 x 104 238 589 152 46 29 21 16 11 9 6 x 104
0.78 0.73
0.71 0.69
20 aquisitions

5 0.66 5 0.66
0.64 0.64
4 0.61 4 0.62
included pixels

included pixels
All All

3 3

2 2

1 1

0 0
0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1
ensemble coherence ensemble coherence
4 4
x 10 x 10
6 6
238 589 152 46 29 21 16 11 9 6 x 104 238 589 152 46 29 21 16 11 9 6 x 104
0.77 0.70
0.70 0.63
69 aquisitions

5 0.65 5 0.59
0.62 0.57
4 0.60 4 0.56
included pixels

included pixels

All All

3 3

2 2

1 1

0 0
0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1
ensemble coherence ensemble coherence

Figure 21: The accuracy of IRF correlation as quality estimate, compared with
amplitude dispersion, for a stack size of 8, 20 and 69 images. For
each stack size overlayed histograms of the ensemble coherence
distribution are shown, with: in gray the total amount of included
pixels, where it is clipped the values are shown in the bars; in blue till
orange the distribution of the best 20 000, 60 000, 120 000, 200 000 and
300 000 points respectively, selected by IRF correlation or amplitude
dispersion. In the legends the thresholds corresponding to those
amounts are shown. (instead of amplitude dispersion amplitude
consistency= 1 − DA is used here). Roughly points with an ensemble
coherence < 0.5 can be regarded as bad.

35
36 point scatterer selection

very nonlinear, nonlinear deformation, linear or no

not unwrappable or partially persistent deformation
1
persistent water,
point scatterers 0.9 thermal deformation stable point scatterers
very high buildings

0.8
partially persistent
or partially persistent
0.7 point scatterers
semi distributed

0.6

IRF correlation
distributed 0.5 Noise
scatterers stable distributed
scatterers
0.4

0.3
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Ensemble coherence (linear deformation)

Table 2: Several scatterer classes based on IRF correllation and ensemble coher-
ence. The classes that most prominent in the test data are shown. One
example of this is water, while water surface scattering just ends up in
as noise, dual bounce reflection on a water surface-wall corner gives
a valid point scatterer, but moving along with the water level, which
even in the Netherlands not stable on mm level.

to get same quality point selection results with smaller stacks, or better
quality results with big stacks. This is especially important for monitor-
ing campaigns, where it is a big burden to have to wait until more than
20 images are acquired, as that can easily take 2 years.

4.3.4 Quality measure per acquisition

The IRF correlation is calculated for every acquisition, before being

averaged to obtain the final estimate. The IRF correlation values per
acquisition can be used too for change detection and to find incoherent
observations in persistent point scatterers. For optimal change detection
the IRF correlation per acquisition should be combined with amplitude
and subpixel position per acquisition.

4.3.5 Sub pixel precision location

For each point scatterer the position is estimated with sub pixel preci-
sion, allowing more precise geocoding.
Besides this also the position can be estimated per acquisition, which
can be used for an additional quality measure, for investigating bigger
movements, or to improve coregistration. Figure 23 shows the estimated
position through time.
The position standard deviation, as defined in equation 4.5, shows a
clear correlation with ensemble coherence, as can be seen in figure 24.
As such it is useful as additional quality estimate.
 phase diff points L64675P73584, L44732P73404 phase diff points L59205P76228, L44732P73404

20 20
phase difference [mm]

phase difference [mm]

10 10

0 0

−10 −10

−20 −20
1 1
Jan10 Jan11 Jan10 Jan11
Time Time
0.8 0.8
IRF correlation

IRF correlation
0.6 0.6

0.4 0.4

0.2 0.2

0 0
Jan10 Jan11 Jan10 Jan11
observation date observation date

(a) (b)

Figure 22: IRF correlation through time can be used for change detection, to
detect appearing and disappearing point scatterers. High IRF corre-
lation values occur also during the incoherent time, this is because
in a single observation random clutter has a fair chance of reflecting
concentrated energy. Here two scatterers from the spoorzone Delft
are shown, where construction of a train tunnel has started early
2011.

−0.4
azimuth [m]

−0.2
0
0.2
0.4
0.3
slant range [m]

0.2

0.1

−0.1
Jul09 Jan10 Jul10 Jan11 Jul11
observation date

(a) (b)

Figure 23: Detected scatterer position through time for a newly appearing point
scatterer in early 2011. The positions before the change are very
varying, while after the change the position stays very much the
same. On a closer look during the coherent period a movement
in range and azimuth can be seen. This movement is about 30
cm in azimuth, to the south, and it is unclear what is causing the
movement, since it seems unlikely that the point has moved so much
in reality. (It is in the new bus station near Delft CS.). In slant range
the movement is about 4 cm, while the deformation measured with
the phase is less than 1 cm, as seen in figure 22a. These positions are
all estimated relative to the coregistration, so care should be taken
when interpreting, however a shift in coregistration correlated with
time seems improbable.

37
38 point scatterer selection

0.8

position deviation
0.6

0.4

0.2

0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ensemble coherence

Figure 24: Position standard deviation, as defined in equation 4.5, can be used
as a quality measure. In this figure the relation with ensemble co-
herence is shown for the stack used in 4.3.3. The problems with
ensemble coherence as quality measure that are shown in table 2 are
applicable here too.

4.3.6 Amplitude calibration is not necessary

Using IRF correlation amplitude differences between acquisitions are

not used, and hence data from satellites wich suffer from unstable
amplitudes, like ERS and RadarSat, can simply be used without prior
amplitude calibration. Also changes in amplitude caused by other
influences won’t influence the selection, such as temperature, moisture,
or observation geometry, which can have a significant effect, as shown
by Perissin [2006]. This in contrast to amplitude dispersion, where the
performance is dependent on correct amplitude calibration, making
selection dependent on complicated calibration algorithms.
When the amplitude from a satellite is stable or well calibrated it is
possible to use amplitude dispersion as an additional quality estimate,
as they are independent measures, with amplitude dispersion only us-
ing the amplitude difference between acquisitions, and IRF correlation
being insensitive to these differences.

4.4 relation to scr

The impulse response correlation can be related to the SCR. By doing

that the accuracy of the phase can be quantified, independent on the
actual phase behaviour. This knowledge can be very useful for various
goals, such as estimation of the deformation using statistic smoothing.
The IRF correlation can be related to SCR through equation 2.4:
p
γ= SSCR1 · SSCR2

where γ can be exchanged with ρirf , SSCR1 is the SSCR of the scatterer,
and SSCR2 = 1. This is because IRF correlation is in fact the coherence
between the signal and a perfect clutter-free signal. So this leads to:

√
ρirf = SSCR (4.6)
SSCR = ρ2irf
 4.5 conclusions 39

0.8

IRF correlation ρ2
0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
SSCR

Figure 25: The relation for ρ2irf as a estimator for SSCR. It is very accurate for
SSCR above 0.7, for lower values it is biased due to the small number
of looks. The thick black line shows the experimentally obtained
relation, the thin black line shows the theoretical relation.

To verify this relation a simulation is done to obtain ρirf for points

with a known SSCR. For the simulation a 2D field of random scatterers
is created, which is then resampled to a much lower sampling rate
and filtered using a Kaiser window with β = 3. This image is used as
clutter image, and is normalised to have a mean amplitude of 1. A point
scatterer image is simulated the same way, but instead of a random
field starting with a single scatterer surrounded by zeros, giving this
point scatterer an amplitude of 1. The signal and clutter are then added
with a given ratio. This is done 10 000 times for 200 different SCR ratios
for the complete simulation. For every SCR ratio the mean ρirf of the
simulations is calculated. The result is shown in figure 25.
As a estimate for SCR the impulse response correlation is very accurate
for SSCR > 0.7, which equals to SCR > 2.3. For lower values it is biased.
At SSCR = 0.7 the phase standard deviation already is 0.6 rad, as can be
read from figure 7. This bias is shown to be dependent on the number
of looks and the actual coherence in the work of Bamler and Hartl [1998,
p.32] or Hanssen [2001, 4.3.4], and agrees with the fact that the effective
number of looks in the weighted window used in the calculation is
very low.

4.5 conclusions

The new method for selection of stable point scatterers based on their
theoretical shape is shown to be very effective. As a quality estimate it
creates new possibilities for processing small stacks, and for large stacks
improvement is shown compared to amplitude dispersion. Besides
being a quality estimate it also neglects sidelobes, solving the problem
of sidelobe filtering inherently. Another benefit is the subpixel precision
position estimation, which is shown to be useful as quality measure
and for increasing geocoding accuracy. Furthermore a quality estimate
per acquisition is available, creating opportunities for change detection.
Since it could be used as a replacement for amplitude dispersion both
methods are compared. The IRF correlation method differs from ampli-
tude dispersion in the following ways:
• Only point scatterers are selected, with sub pixel precision loca-
tion and without sidelobes.
40 point scatterer selection

• As a more accurate quality estimate it makes it possible to process

smaller stacks with good results.
• A quality estimate per acquisition is available. Since the quality
is estimated from the signal characteristics spatially a quality
estimate can be given already for a single image. This makes it
possible to do change detection and find quasi persistent scatter-
ers.
• Amplitude calibration is not necessary since the absolute ampli-
tude is not used for quality estimation. This has major benefits
for ERS and RadarSAT.
It is also possible to use IRF correlation together with amplitude dis-
persion, as they are independent measures, with amplitude dispersion
only using the amplitude difference between acquisitions, and IRF cor-
relation being insensitive to these differences. The effectiveness of this
combination is left for further research.
GEOCODING IMPROVEMENTS
5
5.1 introduction

The most obvious method for identifying scatterers is by looking at

their location on earth. This is a very intuitive method, and can be
easily done by relating the scatterer locations to other data such as
maps, aerial photos or DEMs. The scatterer locations must be estimated
correctly to make this method work. The process of estimating the real
location of a scatterer is geocoding.
Currently there are problems with the geocoding accuracy, resulting in
shifted coordinates and suboptimal relative precision. This is mostly
noticeable in cases where a high accuracy is necessary, such as building
monitoring [Arroyo et al., 2009].
The goal of this chapter is to provide improvements in several pro-
cessing steps that influence geocoding. These steps include the actual
geocoding, which transforms radar coordinates plus height to world
coordinates, and height estimation, which is the separation of the scat-
terer phase into topographic phase and other phase. The selection of
scatterers also influences the location, but this is already covered in
chapter 4.
Geocoding will be done using a 3D polynomial, which is very fast to
evaluate, while the iterative geocoding methods only are used once
for polynomial construction. Also the geoid height will be used as
reference instead of the ellipsoid height. Geocoding is currently done
with respect to the ellipsoid, while heights are mostly expressed with
respect to the geoid. While transformation can be done afterwards that
increases the number of steps and hence the chance of errors and the
amount of necessary computation time.
For accurate height estimation the height difference must be related
to topographic phase. The topographic phase is the phase difference
that is caused by height difference. The way the topographic phase
is calculated will be changed to minimize both structural errors and
computation steps.

5.2 geocoding

Geocoding is done by finding a point on earth such that it corresponds

to a (line, pixel) coordinate and a given height. To accomplish this
efficiently a set of 3D polynomials is set up, with as input the radar
coordinates and height of the scatterer. Upon evaluation this results in
the (latitude, longitude), or (X, Y, Z) coordinate.
For the construction of the polynomial a 3D grid of 1000 points, covering
the whole image with a height range of -500 to 9000 meter, are geocoded
using the traditional iterative method. Then the height of the geoid

41
42 geocoding improvements

above the ellipsoid is calculated for the geocoded coordinates and

subtracted from the original heights. After that the polynomials are
fitted. While it is not obvious to take the geoid height into account
in the traditional method, using a polynomial it can be done almost
without computational cost.
A degree of 5 is used in order to get a high accuracy. For a 3D polyno-
mial that means 56 values are necessary. The maximum error using this
degree in the order of cm, even when the polynomials are set up for a
full frame.
For inverse geocoding another set of 3D polynomials is fitted, giving the
(line, pixel, height) coordinates with as input either (latitude, longitude,
height), or (X, Y, Z) coordinates.

5.3 topographic phase computation

Because the orbit of the satellite during each observation is slightly

different, it is possible to estimate the height difference between scatter-
ers from the phase. The phase change that is caused by the height in
relation with the orbit difference is called the topographic phase.
In DORIS a separation between ellipsoid phase and the topographic
phase is made. While this is useful to work stepwise towards an easier
to unwrap interferogram, it is much easier to combine them when they
both need to be done at once.
Calculating the topographic phase for an (line, pixel, height) coordinate
is done using a 3D polynomial. This polynomial is set up using the
following steps:
1. Geocode a 3D grid of 1000 points, covering the whole image
with a height range of -500 to 9000 meter, using the geocoding
polynomial. Extract the geoid heights from these coordinates and
geocode again using the traditional method. This is necessary
because the polynomial is not accurate enough for phase calcula-
tions.
2. Calculate the distance from those points to the master orbit. This
is easy because the (line, pixel) coordinates are still known.
3. Calculate the distance from the points to the slave orbit, and
subtract them from the distance to the master orbit.
4. This give the parallel baseline as seen from each location, this is
converted to phase and the polynomial is fitted.
5. Repeat step 4 and 5 for each slave orbit.
To estimate the height difference from the phase the derivative of
the topographic phase to height is used, which is sometimes called
H2PH. For ease of use a separate polynomial is set up for this, by
evaluating the topographic phase twice for the gridded points, with
one meter difference in height. To the difference of these phases the
H2PH polynomial is fit.
While creating the topographic phase polynomials optionally the satel-
lite orbits are converted to ETRS, the European Terrestial Reference
System. This corrects for deformation trends due to tectonic plate move-
ment of Europe. As a result for a 100 km wide strip this deformation
 5.4 conclusions 43

trend is corrected with about 2 mm/y. While this reduces the mysteri-
ous trend seen in Envisat results, it doesn’t explain the whole 15 mm/y
that is observed.

5.4 conclusions

Two parts which influence geolocation accuracy have been improved,

one is the geocoding itself and the other deals with height estimation
accuracy. For geocoding the geoid heigth is used as reference, instead
of using the ellipsoid. This makes it compatible with almost all other
places where height is used, like in DEM’s. Height estimation accuracy
is improved for big height differences, where the derivative of the topo-
graphic phase no longer can be assumed constant. The most noticeable
part of the improvements however might the improved speed and ease
of use.
There are a couple of other improvements possible that have been
worked on but are not finished. Those are listed below.
• Orbit averaging. Calculate timing or orbit errors of the master
orbit used for geocoding by using all other orbits, coregistration
polynomials, and a DEM.
• New height estimation methods, decreasing the deformation
model influence and increasing the density of the network.
CASES
6
6.1 introduction

The techniques presented in this paper are beneficial for almost all
InSAR processing, in this chapter several cases which benefit most are
described.
The first case shown is the case that was the main reason for this
research, monitoring the construction of the Delft train tunnel using
InSAR.
Second, a case involving detailed monitoring of unstable storage tanks
in harbours is treated, combining speckle tracking with interferometry
on a very detailed scale.
At last the benefits for the corner reflector experiments undertaken by
the TU Delft are shown.

6.2 delft train tunnel

In Delft the construction of a new tunnel and underground railway

station has started early 2011, which goes very close to the historical
centre and a lot of personal houses. While the direct surroundings
are monitored using traditional methods the monitoring data is not
open to public, and anything further than the first row of buildings is
not monitored at all. A feasibility study for monitoring the area using
InSAR has been done in 2009 [Arroyo et al., 2009], and high resolution
TerraSAR-X images are acquired since April 2009, twice every 11 days,
both ascending and descending. This gives a great opportunity for such
a monitoring project. The main conclusions of the feasibility study were
that the geolocation accuracy of scatterers is a main bottleneck, being
lower than would be expected based on the satellite image resolution,
and, probably even more important, that better filtering is necessary to
remove bad points and keep good points.
The InSAR processing is done again using the new methods for point
selection and geocoding. To make a fair comparison a processing is
done using only the first 16 images that were also available during the
feasibility study. Using all 72 images that are available now also a pro-
cessing is done, giving of course even better results, and show already
the movement that is very likely caused by the tunnel construction.
The difference between the old and the new processing is very large,
as can be seen in figure 26. While using amplitude dispersion with
sidelobe filtering only a handful of scatterers is selected, the results of
the IRF correlation method show almost continuous lines of scatterers.
Also the geocoding looks better, and there are less water scatterers.
Using more acquisitions the quality of course gets better, as can be
seen in figure 27, which show the results of a processed stack of 72

45
 (a)

(b)

Figure 26: Selected scatterers shown on top of AHN2 and GBKN as very ac-
curate reference data. Light points are below 3m NAP, while dark
points are above. A stack of 16 TerraSAR-X images in descending
orbit is used a) Scatterers selected using amplitude dispersion with
sidelobe filtering. This is the data used in Arroyo et al. [2009]. Even
with strong filtering there are points in the water. There is a shift
in geocoding that is variable spacially. b) Scatterers selected using
IRF correlation. The density has been improved manyfold, while no
scatterers in water are selected. Clear straight lines of scatterers are
visible, making it easy to relate them to buildings. There is still a
spacially varying shift due to a small trend in the height estimation
which has to be removed.

46
 6.2 delft train tunnel 47

Figure 27: Scatterers selected using IRF correlation, in green: descending stack
of 72 images, in pink: ascending stack of 68 images. Dark points are
high and light point low. The accuracy is clearly improved compared
to the 16 image stack in figure 26b. The accuracy of geocoding is a lot
better, because the trend in the height estimation is greatly reduced.
A lot of scatterers disappeared because the area on the left has been
demolished for the train tunnel construction. In both ascending and
descending data still a shift is visible, both relating to about a meter
atmospheric delay in range.

images. Since the results using 16 images are already so good, the
improvement is less obvious than the impact the use of a new method
has. The scatterers in the 72 image results are more on straight lines,
which reveals improved local accuracy. Also the absolute accuracy is
higher, the points are accurately positioned in the whole processed
area, while using 16 images a trend in estimated height appeared,
influencing the geocoding. The difference in both relative and absolute
positioning accuracy is mainly caused by height estimation, which is
less accurate with less acquisitions, and depends on the correctness of
the used deformation model.
Besides better point selection and improved geolocation another major
improvement is the ability to easily inspect all scatterers in detail using
the InSAR Inspector. This way it is easy to search for interesting points.
For finding points that go down because of the tunnel construction a
quadratic term is fitted to the deformation in addition to the linear one,
and this quadratic term is selected for the point colour. It is immediately
visible that the area around the spoorzone is accelerating, so one of
the points, near the Binnenwatersloot, is selected to see its timeseries
and other properties, as seen in figure 28. Until January 2011 it is
stable, after that it is going down with about 10 mm/y. To identify the
selected scatterer better a secondary screen can be opened, showing the
scatterers on an aerial image, see figure 29. Also this secondary screen
can be zoomed in further, and points can be selected from here. As can
be seen in the table, the point is on the ground, and it appears to be on
the street.
 Figure 28: Spoorzone data loaded into the InSAR Inspector, with the quadratic
deformation term selected for display.

−3
x 10
4

52.01

1
52.0095
quadratic [m/year2]
Latitude [deg]

0
phase diff points L61635P74247, L66981P67837

52.009 20
−1
phase difference [mm]

10
−2

52.0085 0

−3
−10

−4 −20
4.355 4.3555 4.356 4.3565 4.357 4.3575 4.358 4.3585
Longtitude [deg] Jan10 Jan11
Time

Figure 29: For further identification a secondary screen with an aerial photo is
shown. The white circle denotes the selected point. It is on ground
level on the Binnenwatersloot.

48
 6.3 storage tank monitoring 49

−3
x 10 phase diff points L59876P75970, L66981P67837
4
52.0144
20

phase difference [mm]

52.0142 3
10

52.014
0
2

52.0138
−10
1

quadratic [m/year2]
52.0136
−20
Latitude [deg]

52.0134 0
phase diff points L59819P76009, L66981P67837

52.0132 20
−1

phase difference [mm]

52.013 10

−2
52.0128 0

52.0126 −3 −10

52.0124 −20
−4
4.3505 4.351 4.3515 4.352 4.3525 4.353 4.3535 Jan10 Jan11
Longtitude [deg] Time

Figure 30: Some scatterers on a house in the Spoorsingel, one scatterer on

top of the house, at 7.7 m, and one in front of the house at 1.2 m,
shown with white rings (if you look closely). The high point shows
a seasonal effect, most likely thermal expansion, while the low point
did go down a few mm in January, but appears to be stable again.

Some other points have been selected on a house, for one of the people
who reads this, see figure 30. Neither the top and the bottom front of
the house are moving alarmingly, but there is clear movement visible
in both.

6.3 storage tank monitoring

In harbours very large tanks are used for storage of liquids. These
tanks have to be monitored accurately to guarantee their safety. As an
alternative for traditional measuring techniques a feasibility study for
the use of InSAR for monitoring those tanks has been done by Hansje
Brinker.
A processing using the standard processing chain didn’t give satisfying
results, almost all scatterers on the tanks were discarded because of
unwrapping problems.
A detailed analysis follows, exploring the raw data, alternative unwrap-
ping techniques and other new estimation methods. Also auxiliary
information such as tank diameter and height is incorporated. For
this analysis the toolbox showed to be extremely useful, being flexible
enough to completely custom processing, while it still takes care of all
the metadata, coordinate transformations and easy data access.
The rest of this section will show some results, while showing the ease
of implementation.
Both EnviSat and TerraSAR-X are investigated, but the resolution of
EnviSat turned out to be too low to distinguish separate tanks. Using
TerraSAR-X the tanks are very distinguishable, each tank having at least
a few bright scatterers, and the bigger tanks having the entire top edge
clearly visible, as can be seen in figure 31. The downside of TerraSAR-X
is the short wavelength, which makes that a phase ambiguity already
happens after 16 mm deformation.
 (a)

(b)

Figure 31: Overview of the inspected tanks. a) Reference optical image for
interpretation purposes. b) Selected scatterers on the mean amplitude
radar image. The image was acquired in descending orbit, hence the
satellite is looking from the left. Reference point 1 was chosen for its
recognisable location, but point 2 turned out to be more stable, so is
used instead for further investigation.

50
 6.3 storage tank monitoring 51

The stability of the tanks is monitored using several techniques:

• Interferometric phase between tanks and a carefully selected
reference point.
• Interferometric phase between bottom and top of the same tank.
• Subpixel precision position of a scatterer on the top edge of the
tanks.
• Tilt estimation using the interferometric phase of the whole edge
of the tanks.
Selection of the scatterers has been done using the InSAR Inspector
GUI. This way scatterers could be selected while directly searching
manually for possibilities where the most coherent scatterers are found.
A very simple utility was written to label scatterers with their tank
number and whether on the top or bottom edge of the tank. This could
be easily done because the information about the current scatterer is
exposed via the GUI object.
All selected scatterers are corrected using the known tank height, while
assuming the bottoms of the tanks to be at the same altitude. The
reference scatterer height was interactively estimated in the InSAR
Inspector.
The phase difference between the tanks and the reference point is only
interpretable for part of the tanks, for a lot of tanks it is completely
unstable. This movement is apparently too variable for InSAR to be
useful, at least when X-band radar is used. So other options for using
InSAR were sought.
The phase between the top edge and the bottom edge turned out to
be unwrappable for quite some tanks, but besides that still not all
tanks are covered it is difficult to interpret this deformation. It might be
thermal expansion, but also tilt or floor deformation can be plausible
explanations.
As an attempt to make larger movements visible the subpixel precision
estimated positions on the individual observations are investigated. The
deformations measured that way are with respect to the coregistration,
so while not optimal because the reference is not clear that way, it
still gives an indication of the usability as the reference is sort of an
average over the surroundings. The tanks generally show a movement
of around 7 cm/y over the course of two years, with varying amounts
of noise.
The final and most delicate method attempted is to estimate the tilt
of the top edge through internal phase differences. For this the phase
information from the top edge is extracted from the radar images by
mapping the tank geometry to radar coordinates and requesting the
phase from a stack of oversampled SLC images. The tilt is estimated
from the phases using a 2D periodogram, after the phase from the first
image is subtracted in time from the whole stack. This results in a tilt
estimate in two directions. The estimate in azimuth direction is more
accurate because the strongest scatterers are on the half tank facing
the radar, causing a longer integration length in azimuth, and hence a
higher accuracy.
The results of one of the tanks are shown in figure 32 as an example.
 phase diff laag−ref2, qual: 5.2

6820 20
15

phase difference [mm]

6830
10
Azimuth [px]

6840 5
0
6850
−5

6860 −10
−15
6870 −20

3800 3810 3820 3830 3840 3850 3860 3870 3880 Apr09 Jul09 Oct09 Jan10 Apr10 Jul10
Range [px] Time

(a) (b)

phase diff hoog−laag, qual: 6.5 deformation from subpixel position, hoog
0.6
20

deformation in looking direction [m]

15 0.4
phase difference [mm]

10
0.2
5
0 0
−5
−0.2
−10
−15 −0.4
−20

Apr09 Jul09 Oct09 Jan10 Apr10 Jul10 Apr09 Jul09 Oct09 Jan10 Apr10 Jul10
Time time

(c) (d)

Tiltmap hoog, +: found PS point Tilt of edge

N pi 0.1
Ra (E−W)
0.08 Az (N−Z)
NO
0.06

O 1/2 pi 0.04
tilt [m w.r.t. center]

0.02
ZO
0
Z 0
−0.02

ZW −0.04

−0.06
W −1/2 pi
−0.08

NW −0.1
Apr09 Jul09 Oct09 Jan10 Apr10 Jul10
time
N −pi
Apr Jul Oct Jan 2010 Apr Jul

(e) (f)

Figure 32: The results for one of the tanks, a tank of 18.5 m high and 39 m
wide. a) Point scatterers selected on this tank, red=high, blue=low.
b) Phase difference with reference point, a possible interpretation is
that the deformation is accelerating with about 32 mm/y2 , which
would result in such a pattern. c) The top-bottom difference shows a
seasonal pattern, which can be explained by thermal expansion, the
variation of 8mm on 18.5m relates to a temperature variation of 36◦ .
d) The subpixel movement shows about 7 cm/y, but the accuracy
is probably not high enough to fix ambiguities. e) The tilt of the
top is estimated using the phase of the entire top edge, shown here
for each acquisition, with the first acquisition as reference. f) The
estimated tilt through time, both in azimuth and range direction.
The red dots indicate discarded observations due to low quality of
fit.

52
 6.4 corner reflector experiments 53

Figure 33: The corner reflector experiment in Cabauw. It lasted only 5 observa-
tions before the final EnviSat image was acquired.

6.4 corner reflector experiments

The TU Delft has done a series of experiments with corner reflectors

and compact active transponders (CAT), which both act as perfect
point scatterers. These experiments are done to measure and prove
the accuracy of InSAR deformation measurements by comparing them
with levelling.
For the corner reflector experiments generally not a full processing is
necessary, since the main interest is in the phase differences between
the corner reflectors, so no full network or atmosphere estimation is
necessary.
The IRF correlation method is used to select the point scatters in an
area around the corner reflectors. The resulting pointset is then directly
loaded in the InSAR Inspector to locate the corner reflectors.
An experiment was done in Cabauw in 2010, which only lasted until
the EnviSat mission was finished. The five observations that have been
acquired can be seen in figure 33.
C O N C L U S I O N S A N D R E C O M M E N D AT I O N S
7
The main goal of this research, improving the interpretation and iden-
tification of the measurements carried out using radar interferometry,
has been reached in several aspects.
On one side, in the domain of automated processing, improvements
in valid scatterer selection and geolocation accuracy are accomplished
by using more knowledge about the radar signal, and taking the geoid
height into account. These developments improve the results of auto-
matic processing.
On another side, which is more visible and in the the domain of
development and manual analysis, tools for easy access to data, results
and metadata, and a GUI for visual inspection of processed results, or
even unfinished processing, are developed. This improves the workflow,
making result interpretation and understanding the main task, instead
of all sort of management tasks to access the right data in the right
format, apply the right corrections, and wonder if the right corrections
were made. Since the tools are so flexible, trying something new is often
just a matter of minutes work, improving the speed of development of
new methods. These developments can influence the way people work,
and help them to be more effective in reaching their goals.
Of course not all that can be done is done, and there are a lot of features
waiting to be implemented, new methods to be tried, and implemented
methods to be thoroughly tested. A incomplete list of the most relevant
with respect to this thesis is:
• Improved methods for height estimation, decreasing the deforma-
tion model influence and increasing the density of the network.
• Research on the origin of trends in estimated height. To be thought
of are influence of atmosphere estimation, network construction
and influence of deformation models.
• Easier DEM handling, now DEMs still need to be preprocessed
by DORIS, while this can easily be circumvented.
• Improved filtering, combining available quality measures in an
statistically optimal way, in contrast to separate thresholds on
each of those measures.
• Amplitude variation characterisation with respect to baselines.
Although this was one of the original research ideas it appeared
to be only useful for a small percentage of all scatterers, but this
statement has to be more thoroughly researched.

55
BIBLIOGRAPHY

Nico Adam. Ps processing: Noise budget. 10 2003. (Cited on page 11.)

Ken Arroyo, Prajnaparamita Bhattacharya, Filip Biljecki, Dinesh Kalpoe,
Álvaro Muñoz, Piers Titus van der Torren, Stijn Verlaar, and Hang Yu.
Satellite radar observation feasibility study for large infrastructural
public works, a case study on the delft train tunnel. Geomatics Synthe-
sis Project – Argus Panoptes, 2009. (Cited on pages 41, 45, and 46.)
Richard Bamler and Philipp Hartl. Synthetic aperture radar interferom-
etry. Inverse Problems, 14:R1–R54, 1998. URL http://doris.tudelft.
nl/Literature/bamler98.pdf. (Cited on page 39.)

CEOS. Sar callibration workshop, 1993. (Cited on page 12.)

Ian Cumming and Frank Wong. Digital Processing Of Synthetic Aperture
Radar Data: Algorithms And Implementation. Artech House Publishers,
New York, 2005. ISBN 1580530583. (Cited on pages 8 and 30.)
DORIS. Delft Object-oriented Radar Interferometric Software (DORIS),
User’s manual and technical documentation. Delft Institute of Earth Ob-
servation and Space Systems (DEOS), Delft University of Technology,
Delft, The Netherlands, 1998. (Cited on page 15.)
Alessandro Ferretti, Claudio Prati, and Fabio Rocca. Permanent scatter-
ers in SAR interferometry. IEEE Transactions on Geoscience and Remote
Sensing, 39(1):8–20, January 2001. (Cited on page 13.)
Dirk Geudtner. Die Interferometrische Verarbeitung von SAR-Daten des
ERS-1. PhD thesis, Universität Stuttgart, 1995. Draft. (Cited on
page 8.)
Ramon Hanssen and Freek van Leijen. On the potential of persis-
tent scatterer interferometry for monitoring dikes in the nether-
lands. 9e Nederlands Aardwetenschappelijk Congres 18-19 March,
2008, Veldhoven, 2008. URL http://doris.tudelft.nl/Literature/
hanssen08c.txt. (Cited on page 1.)

Ramon F Hanssen. Radar Interferometry: Data Interpretation and Error

Analysis. Kluwer Academic Publishers, Dordrecht, 2001. (Cited on
pages 1, 5, 11, 12, 13, 29, and 39.)
Andy Hooper, Howard Zebker, Paul Segall, and Bert Kampes. A new
method for measuring deformation on volcanoes and other non-
urban areas using InSAR persistent scatterers. Geophysical Research
Letters, 31:L23611, doi:10.1029/2004GL021737, December 2004. (Cited
on page 1.)
Dieter Just and Richard Bamler. Phase statistics of interferograms
with applications to synthetic aperture radar. Applied Optics, 33(20):
4361–4368, 1994. (Cited on page 11.)
Gini Ketelaar and Ramon Hanssen. Monitoring subsidence due
to gas extraction in Groningen using satellite radar interferome-
try. 8ste Nederlands Aardwetenschappelijk Congres 24-25 April,
2006, Veldhoven, 2006. URL http://doris.tudelft.nl/Literature/
ketelaar06a.pdf. (Cited on page 1.)

57
58 bibliography

V.B.H. (Gini) Ketelaar. Satellite Radar Interferometry: Subsidence Mon-

itoring Techniques. Remote Sensing and Digital Image Processing.
Springer, 2009. (Cited on page 5.)
Daniele Perissin. SAR super-resolution and characterization of urban targets.
PhD thesis, Politecnico di Milano, Italy, 2006. (Cited on page 38.)
L M Th Swart. Spectral filtering and oversampling for radar interfer-
ometry. Master’s thesis, Faculty of Geodetic Engineering, Delft Uni-
versity of Technology, December 2000. URL http://doris.tudelft.
nl/Literature/swart00.pdf. (Cited on page 8.)
 Part I

APPENDIX
TOOLBOX MANUAL
A
On the following pages a full reference manual of the created toolbox
classes is given, including the visual InSAR Inspector tool, and the
SlcStack, Slc, SlcImage, PointSet, Filter, Poly3D, Raster, RasterData,
hdf5prop, and dependentprop classes.

61
Chapter 1

Class: pointset inspector

Visual inspector tool for analysis of PointSet objects. The InSAR Inspector is a GUI that provides
a way to make the spatially dense, multivariate, and relative information that is involved in InSAR
processing fully explorable.
An overview map is shown at the main screen, on which all scatterers are shown. When a scatterer
is selected, either on this overview map or using a unique point ID, all information about a scatterer
is shown, both as numerical data and using plots. All information that is added during processing is
shown, such as quality measures, height, and deformation. Information that varies per acquisition is
plotted, such as amplitude or deformation. For deeper inspection several options are available, such
as showing the underlying radar image data, calculating a coherence matrix, and data access from
the command line.
The GUI object gives access to the selected points and all their properties, and also to all GUI elements
and display functions. Interaction via the command line is necessary when more flexibility or more
user input is required than can be given through the GUI. For example to show a custom property on
the background, make high quality plots of selected points, or estimate new parameters or relations.
psinspector = pointset inspector( pointset ) Start an instance of the PointSet
Inspector on the given PointSet. The returned object can be used to interact with the GUI via
the command line.

1.1 Properties
handles Struct containing all graphics handles. Using this custom data can be added to the plots.
pointset The PointSet with which the GUI was initiated. convenience copies
slc SlcStack object from the pointset
poly Polygon selected using the polygon selection tool image buffers
mrm Stack of RasterData objects shown on the background.
selected1 PointSet containing the selected point.

1
selected2 PointSet containing the reference point.
timesel Selected acquisition index.
pnt custom Custom point attribute for visualisation, can for example be set to interferemetric
phase using pnt custom = wrap(diff(pointset.phase corrected(:,[master slave]),[],2));
obs custom Custom observation attribute for visualisation.
filt Optional selection of the points that are displayed on the background. settings
use points Setting whether to use points from the pointset or raw data.
show slc Setting whether to load and show the slc stack on selecting a point.
estimate height Setting whether to estimate the height difference on the fly or use the given
point height.
subtr atmo Setting whether to correct the phases in the phase plot for atmosphere.
norefpoint Setting whether to subtract the reference point or not.

1.2 loadbackgrounddetail
Load high resolution background around selected point.
gui.loadbackgrounddetail(ovs) Use given oversampling rate for the background. If omitted
it defaults to 1. If ovs == 0 the background will be removed.

1.3 loadaerial
Load an aerial view with selectable points.
gui.loadaerial(zoom) Load an aerial view with Google zoomlevel zoom. Default is 17.

1.4 set orientation

Set the orientation of the map.
tsi.set orientation() Set the orientation to rounded north.
tsi.set orientation('north') Set the orientation to exact north.
tsi.set orientation(dir) Set the orientation to dir degree.

2
Chapter 2

Class: Slc

Radar image class. Stores all metadata of a radar image.

slc = Slc.fromResFiles(basedir, slaveres, ifgres, masterres) Reads radar image
metadata from DORIS .res files.

2.1 Properties
id short identification string: acquisition date as number.
sensor Sensor description string.
date The date of observation (including time) [days].
timeOfDay The time of day of the first pixel, for ms accuracy, also the time used for orbits [s].
timeToFirstPixel two way travel time to first pixel [s].
wavelength Radar wavelength [m].
PRF Pulse Repetition Frequency [Hz].
RSR Range Sampling Rate [Hz].
bandwidthAz Bandwidth in azimuth direction [Hz].
bandwidthRa Bandwidth in range direction [Hz].
poly fDC Doppler centroid frequeny [Hz]. Poly2D object with input in seconds from first pixel,
[line / PRF, pixel / RSR]. For TerraSAR it is given 2D (but not in Doris .res file), for Envisat
only 1D.
preciseOrbit Satellite orbit, n-by-4 array with columns [time of day, x, y, z]
is ascending True if the orbit is ascending, false if descending.
orig Info about original file. .filename .fileformat .window [Raster]
crop Info about crop file. .filename .fileformat .window % Raster

3
rsmp Info about resampled file (to master coordinates). .filename .fileformat .window [Raster]
bperp Perpendicular baseline (from DORIS)
bpar Paralel baseline (from DORIS)
btemp Temporal baseline (from DORIS)
poly coreg Coregistration polynomial, 1x2 Poly2D object for line and pixel.
n coregwin Number of coregistration windows.
poly h2ph h2ph polynomial as calculated by DORIS. Poly2D object. Not used anymore,
superceded by poly3 h2ph.
poly flat flat earth polynomial as calculated by DORIS. Poly2D object. Not used anymore
superceded by poly3 refpha.
poly3 refpha Topographic phase polynomial. Poly3D object. Returns the topographic phase on a
[line, pixel, height] coordinate. Height is relative to the geoid.
poly3 h2ph h2ph polynomial. Poly3D object. Returns the derivative of poly3 refpha to height on
a [line, pixel, height] coordinate.
poly geocoding Geocoding polynomial for geocoding to WGS84 XYZ coordinates. 1x3 Poly3D
object. Returns the XYZ coordinates for a [line, pixel, height] coordinate. Height is relative to
the geoid.
poly latlon Geocoding polynomial for geocoding to WGS84 [lat,lon] coordinates. 1x2 Poly3D
object. Returns the [lat,lon] coordinates for a [line, pixel, height] coordinate. Height is relative
to the geoid.
poly xyz2linepixel Inverse geocoding polynomial from WGS84 XYZ coordinates. 1x2 Poly3D
object. Returns the [line,pixel] coordinates for a [X,Y,Z] coordinate.
poly latlonh2linepixel Inverse geocoding polynomial from WGS84 [lat,lon] coordinates. 1x2
Poly3D object. Returns the [line,pixel] coordinates for a [lat,lon,height] coordinate.
geocoding shift Shift in [line,pixel]
calibrationFactor General calibration factor to apply when reading data.
res slave Content of DORIS slave res file. Only for debugging.
res ifg Content of DORIS interferogram res file. Only for debugging.
res master Content of DORIS master res file. Only for debugging.
spectrumAz Spectrum in azimuth, Filter object. Contains relative bandwidth and spectrum shape.
Not shifted to doppler centroid because that depends on location.
spectrumRa Spectrum in range, Filter object. Contains relative bandwidth and spectrum shape.
velocity Indication of satellite velocity
heading Satellite flight direction, calculated at center of original frame
preciseOrbit etrs Satellite orbit converted to from ITRS to ETRS
look angle Indication of satellite look angle (angle between earth center and look direction)

4
incidence angle Indication of satellite signal incidence angle (angle at which the radar signal
reaches the ground)
spacing az Pixel spacing on ground in azimuth/flight direction, or line spacing [m]. Calculated at
center of original frame.
spacing ra slant Range pixel spacing, in looking direction [m].
spacing ra ground Range pixel spacing projected to geoid [m]. Calculated at center of original
frame.
fDC mean
fDC
poly coregRa
poly coregAz

2.2 clone
Create a clone of self.

2.3 make poly refpha h2ph

Creates polynomials for topographic phase computation. The following polynomials are created by
this function: poly3 refpha and poly3 h2ph. poly3 h2ph is the derivative of refpha to height, and can
be used for iterative estimation of elevation.
accuracy = slc.make poly refpha h2ph( degree, height range, refframe ) Create
topographic phase polynomials and store them in slc. All arguments are optional. degree
overides the polynomial degree of 7. height range the valid height range of [-500 9000],
which includes all places on the earth surface from the Dead Sea to the Mount Everest.
refframe can be ETRS to convert the orbits to ETRS before usage and as such correct for
tectonic plate movement in Europe. The returned structure includes quality descriptions of
the polynomials, with the fields std for standard deviation and max for maximum error.

2.4 readDataRsmp
Read resampled SLC data from disk, and return as an SlcImage.
imc = slc.readDataRsmp( crop, filterwindow ) Read the specified crop from the SLC.
crop must be given as a Raster object. If omitted the whole file is read. If the crop is
oversampled the SLC is automatically oversampled to meet the crop. If filterwindow is
specified the SLC data is spectrally filtered to the given window, which can be used to reduce
sidelobe levels or increase resolution.
Example:

5
imc = slc.readDataRsmp( Raster(5101:6100,2401:2500).oversample(2), kaiser(100, 3) )
Load a crop of 1000x100 pixels, oversampled 2 times, and filtered to a Kaiser window with
beta = 3.

2.5 make poly geocoding.bak

MAKE POLY GEOCODING Creates polynomials for fast geocoding w.r.t. the geoid. The follow-
ing polynomials are created by this function: poly geocoding, poly latlon, poly xyz2linepixel and
poly latlonh2linepixel.
accuracy = slc.make poly geocoding( 'option', value, ... ) Create geocoding
polynomials and store them in slc. Optional arguments are:
shift linepixel Define a shift in the radar coordinates, as [line, pixel].
degree Polynomial degree used, default 5.
height range Valid height range, default [-500 9000], which includes all places on the earth surface
from the Dead Sea to the Mount Everest.
geoid Whether to use the geoid as reference heigth, otherwise the ellipsoid is used. default true.
refframe Can be ETRS to convert the orbits to ETRS before usage.
The returned structure includes quality descriptions of the polynomials, with the fields std for standard
deviation and max for maximum error.

2.6 lph2xyz
Calculate the XYZ coordinate from line, pixel and height above ellipsiod.
[x,y,z] = slc.lph2xyz( line, pixel, height, refframe ) refframe can be ETRS to
convert the orbits to ETRS before usage.
In general the geocoding polynomials can be used instead of this function, either direct or via the
geocoding functions in SlcStack

2.7 make poly geocoding

Creates polynomials for fast geocoding w.r.t. the geoid. The following polynomials are created by
this function: poly geocoding, poly latlon, poly xyz2linepixel and poly latlonh2linepixel.
accuracy = slc.make poly geocoding( 'option', value, ... ) Create geocoding
polynomials and store them in slc. Optional arguments are:
shift linepixel or shift xy Define a shift in the radar coordinates, either as [line, pixel] or as [x,
y] in metre.
degree Polynomial degree used, default 5.

6
height range Valid height range, default [-500 9000], which includes all places on the earth surface
from the Dead Sea to the Mount Everest.
geoid Whether to use the geoid as reference heigth, otherwise the ellipsoid is used. default true.
refframe Can be ETRS to convert the orbits to ETRS before usage.
The returned structure includes quality descriptions of the polynomials, with the fields std for standard
deviation and max for maximum error.

2.8 xyz2azra
Calculate the [line, pixel] coordinates from XYZ coordinates.
[line,pixel,dist] = slc.xyz2azra( xyz, refframe) refframe can be ETRS to convert
the orbits to ETRS before usage.
In general the geocoding polynomials can be used instead of this function, either direct or via the
geocoding functions in SlcStack

2.9 makeInterferograms
make interferograms of an slc stack and save them as png. Input parameters: dirname Directory
where to save images master Master index for single master ifgs, or incremental, positive, all,
or magnitude crop Raster object of crop doFilter do range and azimuth filtering window apply
window, kaiser(N,4) is default flip lr, ud ovs Oversample factor range magnitude range (for mixed),
logarithmic. Adaptive is default ref reference coordinate phaseonly don’t mix phase and magnitude
cohAsMag estimate coherence and use that as magnitude
TODO: Too many input parameters...

7
Chapter 3

Class: SlcStack

The SlcStack class represents a stack of Slc images. Properties of the class apply to the stack as
a whole, such as wavelength and pixel spacing. Properties that are aqcuisition-specific can be found
in the individual Slc objects in the slc array.

3.1 Properties
n images The number of Slc images in the stack.
slc An array of Slc objects with length equal to n images.
date Array of dates (datenum) of all aqcuisitions.
dir The base directory of the raw Slc data.
master idx Index in the Slc array of the master image, to which all the other images are
coregistred.
master The master Slc object
sensor Sensor(s) description string.
wavelength Radar wavelength [m].
PRF Pulse Repetition Frequency [Hz].
RSR Range Sampling Rate [Hz].
is ascending True if the orbits are ascending, false if orbits are descending.
poly geocoding Geocoding polynomial for geocoding to WGS84 XYZ coordinates. 1x3 Poly3D
object. Returns the XYZ coordinates for a [line, pixel, height] coordinate. Height is relative to
the geoid. See Slc/make poly geocoding for technical details.
poly latlon Geocoding polynomial for geocoding to WGS84 [lat,lon] coordinates. 1x2 Poly3D
object. Returns the [lat,lon] coordinates for a [line, pixel, height] coordinate. Height is relative
to the geoid.

8
poly xyz2linepixel Inverse geocoding polynomial from WGS84 XYZ coordinates. 1x2 Poly3D
object. Returns the [line,pixel] coordinates for a [X,Y,Z] coordinate.
poly latlonh2linepixel Inverse geocoding polynomial from WGS84 [lat,lon] coordinates. 1x2
Poly3D object. Returns the [line,pixel] coordinates for a [lat,lon,height] coordinate.
poly h2ph h2ph polynomial as calculated by DORIS. Poly2D object. Not used anymore,
superceded by poly3 h2ph.
poly flat flat earth polynomial as calculated by DORIS. Poly2D object. Not used anymore
superceded by poly3 refpha.
poly3 refpha Topographic phase polynomial. Poly3D object. Returns the topographic phase on a
[line, pixel, height] coordinate. Height is relative to the geoid. See Slc/make poly refpha h2ph
for technical details.
poly3 h2ph h2ph polynomial. Poly3D object. Returns the derivative of poly3 refpha to height on
a [line, pixel, height] coordinate. Difference to the DORIS computed is that multiplying with
m2ph is not necessary.
poly accuracy polynomial accuracy and range information
spacing az Pixel spacing on ground in azimuth/flight direction, or line spacing [m]. Calculated at
center of original frame.
spacing ra slant Range pixel spacing, in looking direction [m].
spacing ra ground Range pixel spacing projected to geoid [m]. Calculated at center of original
frame.
velocity Satellite velocity
heading Satellite flight direction, calculated at center of original frame
look angle Indication of satellite look angle (angle between earth center and look direction)
incidence angle Indication of satellite signal incidence angle (angle at which the radar signal
reaches the ground)

3.2 SlcStack
Create an SlcStack object. The first argument is mandatory, the rest is optional and follows the
key-value convention. Any of the optional arguments can be combined.
slcstack = SlcStack(dir) Creates an SlcStack object from data in directory dir, using
autodetection to determine the data format. Supported formats are DORISstack, DePSI and
Stamps, relying on the specific constructors fromDorisStackDir, fromDepsiDir and
fromStampsDir.
slcstack = SlcStack(dir,'reject',reject) Rejects images based on a boolean or integer
reject array.
slcstack = SlcStack(dir,'fromdate',from,'todate',to) Rejects images taken before
from or after to, both specified as datenum.

9
slcstack = SlcStack(dir,'refframe','ETRS') Use ETRS reference frame for orbits in
computation of refpha and h2ph.

3.3 linepixelheight2xyz
Convert the given coordinates to WGS84 XYZ.
Height is interpreted as height w.r.t. the geoid.
xyz = slcstack.linepixelheight2xyz([line, pixel, height])
[x, y, z] = slcstack.linepixelheight2xyz([line, pixel, height])

3.4 linepixelheight2latlon
Convert the given coordinates to WGS84 [latitude, longtitude].
Height is interpreted as height w.r.t. the geoid.
latlon = slcstack.linepixelheight2latlonh([line, pixel, height])
[lat, lon] = slcstack.linepixelheight2latlonh([line, pixel, height])

3.5 xyz2linepixel
Convert the given coordinates to line, pixel. Coordinates are interpreted as WGS84 XYZ coordinates.
linepixel = slcstack.xyz2linepixel([x, y, z])
[line, pixel] = slcstack.xyz2linepixel([x, y, z])

3.6 latlonh2linepixel
Convert the given coordinates to line, pixel. Coordinates are interpreted as WGS84 coordinates with
height w.r.t. the geoid.
linepixel = slcstack.latlonh2linepixel([lat, lon, height])
[line, pixel] = slcstack.latlonh2linepixel([lat, lon, height]) If the height
column is omitted it is set to 0.

10
Chapter 4

Class: SlcImage

Class for storage and processing of SLC data.

An SlcImage object contains a crop of data of an SLC image, and has functions for processing this
data.
The normal way to obtain an SlcImage is using the readData(Rsmp) function in Slc.

4.1 Properties
raster Coordinates and spacing of the data. Raster object.
data Data array.
slc Slc object of the SLC.
coordinateSystem Coordinate system in which the data is sampled. Can be original or
resampled. resampled means resampled to master coordinates.
subtrflat True if flatearth is removed.
filtAz Azimuth spectrum of data.
filtRa Range spectrum of data.
filtFreq
crop

4.2 oversample
Oversample Slc data. The filter used for oversampling is the MATLAB default with a filter order of
10. The shift of the spectrum in azimuth direction, the doppler shift, is taken into account by shifting
the filter.
imc = imc.oversample(ovs) Oversample data ovs times.

11
4.3 cropTo
Crop the data to the given raster.
imc = imc.cropTo(raster) Crop the data to raster, where raster is a Raster object.

4.4 filterRa
Filter the data in range direction. The filter to be applied is given as a Filter object, which can be
used to construct spectral filters.
imc = imc.filterRa(filter, n) Filter the data in range direction, using the given Filter,
using a n point filter. If n is omitted n=64.

4.5 filterAz
Filter the data in azimuth direction. The filter to be applied is given as a Filter object, which can
be used to construct spectral filters.
imc = imc.filterAz(filter, n) Filter the data in azimuth direction, using the given Filter,
using a n point filter. If n is omitted n=64.

4.6 rewindow
Filter the data to the given window. A very useful window is the Kaiser window, constructed in
MATLAB with window = kaiser(N, beta), where N is sufficiently large and beta defines the
tradeoff between resolution and side lobe levels.
imc = imc.rewindow(window, n) Filter the data so that the spectrum shape is equal to
window, using a n point filter. If n is omitted n=64. Example:
imc = imc.rewindow(kaiser(1025, 3))

4.7 subtractReferencePhase
Subtract the reference phase from the data. The reference phase is calculated using the polynomial
in slc.poly3 refpha, evaluated at height = 0. See Slc.make poly refpha h2ph() for the polynomial
calculation.
imc = imc.subtractReferencePhase() Subtract the reference phase from the data.

4.8 show
Show image or stack of image using Image class.

12
4.9 plotSpectrumRange
Plot the real and theoretical spectrum of the data in range direction.

4.10 plotSpectrumAzimuth
Plot the real and theoretical spectrum of the data in azimuth direction.

4.11 mean
Return mean of a stack.

4.12 std
Return std of a stack.

4.13 mean abs

Return mean of the absolute values of a stack.

4.14 std abs

Return std of the absolute values of a stack.

4.15 value
Return the values at the given coordinates.
val = imc.value(lines, pixels) Return the values at the given coordinates. The values are
interpolated using a cubic kernel if they are not exact on data points.

4.16 clone
Create a clone of self.

13
4.17 filterRaOverlap
Filter all SLC images to have the same spectrum in range (pixel) direction, so to the part of the
spectrum that is overlapping. In this direction the spectrum can be shifted due to incidence angle
differences. For this filtering the inclination and pointing angle of the ground are needed. For this
poly flat is now used, which follows the ellipsoid.
By supplying an additional slope value the incidence angle can be multiplied to get another angle.
This is still experimental.

4.18 nearest peaks

Find the coordinates of the local maxima nearest to the given coordinates.
[lines, pixels] = nearest peaks( self, linepixel ) Find the [line, pixel] coordinates
of the local maxima nearest to the given coordinates. Those coordinates must be supplied as
a n-by-2 array [lines, pixels]. The coordinates are estimated with subpixel precision.

4.19 irf corr

Calculate the impulse response correlation

4.20 coherence
Calculates the coherence between two SLC images.
coh = imc1.coherence(imc2 [, win, type]) Calculate the coherence between imc1 and
imc2, using the window win. If win is omitted a window is constructed based on the impulse
response of the signal, giving a high calculation resolution, and independence from
deformation or other phase signals. type can be projected, complex, or absolute, where
projected is the default.

4.21 peakcoords
Find the coordinates of all local maxima in an image with subpixel precision.
[lines,pixels] = im.peakcoords( mask ) Return the line and pixel coordinates of all local
maxima, based on absolute values. The coordinates are estimated with subpixel precision. An
optional mask can be supplied to mask out unneeded areas.

14
4.22 interferogram
ifg = interferogram( self, master, doFilter, window, doSubtrflat, cohAsMag ) TODO: must be cleaned
up/generalised

4.23 filterAzOverlap
Filter all SLC images to have the same spectrum in azimuth (line) direction, so to the part of the
spectrum that is overlapping. In this direction the spectrum can be shifted due to squint angle or
doppler centroid differences.

4.24 coherencematrix
Calculate the coherence between each two SlcImages in an array. The upper triangle gives the average
coherence over the area, the lower triangle gives the maximum coherence in the area.
cohmat = coherencematrix([imc1, imc2, ...], crop, win, type) Calculate the
coherence between each two SlcImages in the array. crop defines the area that should be
included in the calculation, and must be a boolean matrix of the same size as imc.data. win
and type are described in coherence and are usually omitted.

15
Chapter 5

Class: PointSet

The PointSet class represents a database of point data such as phase, coordinates, etc.
Point data is organized in matlab arrays, or similar reacting object as hdf5prop or dependentprop,
and follows the convention that the first dimension equals the number of points on the pointset
(n points) and the second the number of observations (n observations), i.e. the number of
images in the stack. Data that is the same for all points (such as the aqcuisition date) has its first
dimension equal to one (row vector) and a name that starts with obs , such as in obs date. Data
that is independent of aqcuisition has its second dimension equal to one (column vector) and a name
that starts with pnt , such as in pnt quality. Data that changes with both point and aqcuisition
has no prefix, eg. amplitude and phase. Properties can be added dynamically, and should follow
this convention. The only properties that are exempted are n points, n observations, meta, log,
and slc.
The amplitude is not an element of the PointSet by default as it is not required in any of the main
functions and dependent properties, but it is usually present as a dynamic property. The pnt quality
can mean different things at different processing stanges, but should be the most recent quality
measure of the points in the set. For example, initially it can be the amplitude consistency, and later
after network estimation it will be the ensemble coherence. Separate, specific properties for these
values are likely added in future versions.’

5.1 Properties
n points Number of points.
n observations Number of observations.
meta Structure of meta data that is not attributed to points or observations, such as the mrm or a
polygon that was used to select the current set of points.
log A Logger object that maintains a history of standard output messages.
slc The SlcStack object that the point data is obtained from.
phase The real-valued phase of each point per aqcuisition.

16
pnt line The line coordinate of each point.
pnt pixel The pixel coordinate of each point.
pnt quality The quality of each point.
pnt height The (current estimate of the) height of each point.
pnt lat The estimated lattitude of each point based on the estimated height.
pnt lon The estimated longitude of each point based on the estimated height.
obs date The date of each aqcuisition.
h2ph The height-to-phase of each point per observation.
topo phase The phase corresponding to the current pnt height.
atmo phase The phase corresponding to the estimated atmosphere in meta.
phase notopo The measured phase corrected for the topographic phase topo phase.
phase notopo noatmo The measured phase corrected for the topographic and atmospheric phase
topo phase and atmo phase.
phase corrected The measured phase corrected for the topographic phase and, if available, the
atmospheric phase. This is the best available phase for model estimations.
defo line of sight deformation w.r.t. first observation and w.r.t. meta.refpoint [m]

5.2 PointSet
Create a PointSet object.
pointset = PointSet(slcstack) Creates an empty pointset for an SlcStack object, setting
n observations equal to the images in the stack and n points equal to zero.
pointset = PointSet(slcstack,props) Adds the properties named in the props cell array,
following the naming conventions as outlied in this object’s documentation, and initialized
with nan values.

5.3 append
Add point data to the pointset, increasing the n points number of points. This modifies the pointset
in place.
pointset.append(other) Appends the contents of the other pointset to pointset. The new
value of n points is incremented correspondingly. Data fields of the two pointsets are
merged, and nan values are put where data is missing due to the merge. For example, if
phase exists in other but not in pointset, then phase is added to pointset, initialized
with nan values, and phase data from other is copied into the second part.

17
5.4 add points
Allocates space for extra points in the pointset. This modifies the pointset in place.
index = pointset.add points(n) Increases n points by n and and grows data fields
accordingly with nan values. The index vector points into the modified pointset at the newly
allocated space to aid subsequent filling of data.

5.5 add observations

Allocates space for extra observations in the PointSet. Care must be taken to update the slc property
consistently.

5.6 add property

Dynamically adds data fields to the pointset. This modifies the pointset in place. If a property with
given name already exists it is overwritten.
pointset.add property(name) Adds property with given name, and initializes it with nan
values according to the naming conventions as outlined in the PointSet help text.
pointset.add property(name, prop) Adds property with given name, and value prop.
pointset.add property(name, type) Adds property with given name, and initializes it with
nan values of type according to the naming conventions as outlined in the PointSet help
text.

5.7 has property

Test if pointset contains a certain property.
retval = pointset.has property(name) Returns true if pointset.(name) exists, or false
otherwise.

5.8 del property

Deletes a certain property from the property. This modifies the pointset in place.
pointset.del property(name) Deletes pointset.(name) if it exists.

5.9 clone
Makes a deep copy of the pointset.

18
cloned = pointset.clone() Creates a new pointset cloned that is a replica of the original
pointset. Of the special properties, log is not copied. meta is copied and may share data
with the original pointset if meta contains objects.

5.10 xyz wgs84

Calculates the X,Y,Z coordinates of the points in the coordinate frame of WGS84.
xyz = pointset.xyz wgs84(sel)
[x, y, z] = pointset.xyz wgs84(sel) Return the WGC84 XYZ coordinates of the subset
sel.

5.11 rdnap
Calculate the coordinates in the Dutch RDNAP coordinate system, computed with sub cm accuracy.
rdnap = pointset.rdnap(sel)
[rdx, rdy, nap] = pointset.rdnap(sel) Return the rdnap coordinates of the subset sel.

5.12 subset index

Create a subset of a pointset based on index.
sub = pointset.subset index(I) Direct selection of points based on a boolean or index
vector.

5.13 subset coord

Create a subset of a pointset based on radar coordinates.
sub = pointset.subset coord(C) Select all points that are (within epsilon radius) on the
given (line, pixel) coordinates. If specified as a matrix, C should consist of two columns. If
specified as cell array then each item C{i} should be a two-element vector. If specified as a
string then the format is L...P...\nL...P...\n..., where the dots contain the line and
pixel coordinate times 10, for exampe, 10.23,14.25 becomes L102P143.

5.14 shift geocoding

Change the geocoding to correct for timing errors, orbit errors, atmosphere delays, etc.
out = pointset.shift geocoding('shift linepixel',[lineshift, pixelshift])
Change the geocoding to shift

19
out = pointset.shift geocoding('shift xy',[xshift, yshift]) Change the
geocoding to shift points. xshift is the shift to be applied in west-east direction, yshift in
south-north. Both given in meter.

5.15 plot mrm

TODO: document

5.16 plot highres

make high resolution image of a crop

5.17 subset
Create a subset of a pointset based on a variety of criteria. Arguments are passed by the key-value
convention. Any of the following settings can be combined.
sub = pointset.subset('index',I) Direct selection of points based on a boolean or index
vector.
sub = pointset.subset('coord',C) Select all points that are (within epsilon radius) on the
given (line, pixel) coordinates. If specified as a matrix, C should consist of two columns. If
specified as cell array then each item C{i} should be a two-element vector. If specified as a
string then the format is L...P...\nL...P...\n..., where the dots contain the line and
pixel coordinate times 10, for exampe, 10.23,14.25 becomes L102P143.
sub = pointset.subset(property, limits) Select points based on pnt property. limits
should be a minimum value or a [minimum, maximum] pair. property should be a property
for which a pnt property exists, like quality, height, line, etc.
sub = pointset.subset('crop', raster) Select only points within the extent of raster
sub = pointset.subset('poly', poly) Select only points within the polygon or
multipolygon poly. Also see INMULTIPOLYGON.
sub = pointset.subset('nearlatlon', [lat lon], distance) Select only points within
distance of a given coordinate Distance can be given in meter (eg 100m) or in mainlobes (eg.
3)
sub = pointset.subset('nearlinepixel', [line pixel], distance) Select only
points within distance of a given coordinate Distance can be given in meter (eg 100m) or in
mainlobes (eg. 3)
[sub,sub2,...] = pointset.subset('mindist',d) Select points that are at a certain
minimum distance to each other. Points are added in descending order of quality. After each
addition the radius around the added point is discarded, and the next best point is added,
until the set is exhausted. If specified as a number the implied unit is the mainlobe size: 1 for
one time the mainlobe. If the distance is specified as a string then the unit is specified, and

20
 currently only meters are supported; a distance of 100m is 100 meters in both azimuth and
range. If two or more outputs are requested, the algorithm is repeated for points that are not
selected to create independent subsets. This is possibly useful for testing purposes.
[sub,sub2,...] = pointset.subset('mindist',d,'seedpoints', points) Make sure
the given seedpoints are present in the subset. Points can be given as either PointSet,
LxxxxxPxxxxx or [ lines, pixels].
[sub,sub2,...] = pointset.subset('mindist',d,'seedpoint', points, 'onlytoseedpoints')
Only consider the distance to the given seedpoints, and don’t include the seedpoints themself.

5.18 export csv

Export PointSet data to CSV
pointset.export csv(file, props) Export the properties in the cell array props to file.
Each string in props should be a property, with optionally a printf format string appended.
Example: {pnt line,pnt pixel,pnt lat%0.10g,defo%.3g}
pointset.export csv(..., 'noheader') Don’t include column headers.

5.19 plot interferograms

TODO: improve and document

5.20 plot latlon

TODO: document
gridsize : grid size in [m] range : plot range for colorbar colormap : colors to use function :
e.g. @mean or @abs etc [default @max]

5.21 newfig mrm

TODO: document

5.22 export KML

Exports a transparant plot and a kml file that locates it TODO: document
subset points that should be displayed (indices) raster Raster object

21
Chapter 6

Class: Filter

Spectral metadata and filter class A Filter object contains information about the spectrum of a
signal, and creates and applies filters taking this information into account. The spectrum is defined
as a relative spectrum, which is in the interval [-1 1]. It is defined by a window shape, a bandwith,
and a phase shift.
Filter objects are immutable.
f = Filter(windowAmp, bandwidth, phaseshift, windowFreq) Return a new Filter
object. If any input is omitted the following defaults are used: windowAmp = [1 1]:
rectangular window. bandwidth = 1: full bandwidth. phaseshift = 0: spectrum centered.
windowFreq = linspace(-1, 1, length(windowAmp)): equally spaced window amplitudes.

6.1 Properties
windowAmp Amplitudes of window shape
windowFreq Frequenies of window shape, from -1 to 1.
bandwidth Bandwidth of signal, relative to nyquist. From 0 to 1.
phaseshift Phase shift of center of spectrum.

6.2 shape
Return the shape of the full spectrum.
[a,f] = filter.shape() Return the amplitudes and frequencies of the spectrum in full
resolution.
[a,f] = filter.shape(f in) Return the spectumr interpolated to the given frequencies f in.
If called on filter array, return a and f as cell arrays.

22
6.3 design
Design a FIR filter with the frequency response according to self.

6.4 apply
Apply one spectrum to another, which in fact is multiplication.

6.5 overlap
Return the spectrum that must be applied to get the overlapping spectrum.

6.6 rewindow
Return the spectrum that must be applied to get the given window.

6.7 oversample
Return the spectrum of the oversampled signal.

6.8 shift
Change phaseshift.

6.9 noisebandwidth
Calculate the equivalent noise bandwidth of the filter.

6.10 mainlobewidth
Calculate the width of the mainlobe in fft bins (pixels). The width is defined as up to -3 db (Cumming
p.35)

6.11 sideloberatio
Calculate the integrated sidelobe ratio in dB. The width of the main lobe is defined as null-to-null.

23
6.12 filter
Apply the filter to an image.

6.13 raisedCosine
Raised cosine or generalized Hamming window.

6.14 impulseresponse
Return the impulse response.
irf = filter.impulseresponse(maxdist, oversample) Return the impulse response up to
maxdist samples, oversampled with the given value.

6.15 simulateSidelobes
simulates the sidelobe positions and values of an impulse response with the defined spectrum

24
Chapter 7

Class: Poly3D

3D polynomial, with shift and scale for numerical stability.

7.1 Properties
xshift Shift in 1st dimension
xscale Scale of 1st dimension
yshift Shift in 2nd dimension
yscale Scale of 2nd dimension
zshift Shift in 3rd dimension
zscale Scale of 3rd dimension
poly Array with polynomial values

7.2 Poly3D
Create a Poly3D object.
p = Poly3D( xshift, xscale, yshift, yscale, zshift, zscale, poly ) Create a
Poly3D object with given parameters.

7.3 evaluate
Evaluate the polynomial at the given values.
val = poly3d.evaluate(x, y, z) Evaluate the polynomial at the given values. If poly3d is an
array of Poly3D objects all polynomials are evaluated and the results concatenated. It is
allowed to give a single n-by-3 array as input.

25
Chapter 8

Class: Raster

The Raster class defines a 2D rectangular window with equidistant divisions in both dimensions.
The dimensions are called line and pixel, according to their names in radar sciences.
r = Raster(firstLine:stepLine:lastLine, firstPixel:stepPixel:lastPixel)
Returns an inclusive window (lastLine and lastPixel are included).
r = Raster.ex(firstLine:stepLine:endLine, firstPixel:stepPixel:endPixel)
Returns an exclusive window (endLine and endPixel are excluded).
stepLine and stepPixel default to 1 when they are omitted.

8.1 Properties
firstLine First line included in the raster
firstPixel First pixel included in the raster
stepLine Stepsize in line direction
stepPixel Stepsize in pixel direction
nLines Number of raster elements in line direction
nPixels Number of raster elements in line direction
lastLine Last line included in the raster
lastPixel Last pixel included in the raster

8.2 isInside
sel = raster.isInside(coords) Returns true for coordinates inside the span of the raster.
coords should be a n-by-2 array [line,pixel]

26
8.3 contains
b = raster1.contains(raster2) Returns true if raster2 is fully within the span of raster1.

8.4 index
[I,J] = raster.index(lines,pixels) Returns the indices of the raster cells containing the
coordinates (lines,pixels).
I = raster.index(lines,pixels) Returns linear indices.
see also: locate

8.5 locate
[line, pixel] = raster.locate(I) Returns the location of the linear indices I.
[line, pixel] = raster.locate(I,J) Returns the location of the indices (I,J).
see also: index

8.6 oversample
new = raster.oversample(ovs) Returns a copy of the raster oversampled ovs times. If ovs
has two elements ovs(1) is used in line direction, and ovs(2) in pixel direction.

8.7 resample
new = raster.resample(nLines, nPixels) Returns a raster with the same span, but
containing nLines line elements and nPixels pixel elements.

8.8 bbox
new = raster.bbox(lines, pixels) Returns a raster containing all (lines,pixels) coordinates,
while keeping the stepsize.

8.9 setSteps
new = raster.setSteps(steps) Returns raster with same span, but with stepsize equal to
steps. If steps has two elements steps(1) is used in line direction, and steps(2) in pixel
direction. If steps is a raster the stepsizes of that raster are used.

27
8.10 addBorder
new = raster.addBorder(border) Returns a raster with a border of width border added to
all sides. If border has two elements border(1) is used in line direction, and border(2) in pixel
direction.

8.11 union
new = union([raster1, raster2, ...]) Return a raster spanning all given rasters. Only
works for rasters with same stepsize and compatible start points.

8.12 intersection
Returns intersection of rasters
newraster = intersection([raster1, raster2, ...]) Return Raster containing only
points that are in all given rasters.

8.13 repr
s = raster.repr() Return a string representation of the object.

8.14 makeTiles
tiles = raster.makeTiles( target elements ) Returns an array of Raster objects
spanning the same area as raster. target elements is the approximate number of elements
in each Raster in both directions, and will be rounded to make it fit. If target elements has
two elements target elements(1) is used in line direction, and target elements(2) in pixel
direction.
tiles = raster.makeTiles( target elements, roundfact ) The same, but round the
element count to multiples of roundfact. If roundfact has two elements roundfact(1) is used
in line direction, and roundfact(2) in pixel direction.

28
Chapter 9

Class: RasterData

Class for 2D data with coordinate information.

9.1 Properties
raster Coordinates and spacing of the data. Raster object.
data Data array.
descr Description string.
orientation Orientation, 2 element array [x reverse, y reverse].
scale Scale to convert to [m] or other convenient unit.

9.2 interp
Return the values at the given coordinates.
val = imc.value(lines, pixels) Return the values at the given coordinates. The values are
interpolated using a cubic kernel if they are not exact on data points.
val = imc.value(raster) Return the values at the raster coordinates. The values are
interpolated using a cubic kernel if they are not exact on data points.

9.3 value
Return the values at the given coordinates.
val = im.value(lines, pixels) Return the values at the given coordinates. The values are
interpolated using a cubic kernel if they are not exact on data points.

29
9.4 cropTo
Crop the data to the given raster.
im = im.cropTo(raster) Crop the data to raster, where raster is a Raster object.

9.5 mean
Return mean of a stack.

9.6 std
Return std of a stack.

9.7 mean abs

Return mean of the absolute values of a stack.

9.8 std abs

Return std of the absolute values of a stack.

9.9 show
Show image or stack of image using Image class.

9.10 select polygon

select a polygon with self as background

30
Chapter 10

Class: hdf5prop

Class for transparent file data access. Matlab class to create and access HDF5 datasets transparently
as Matlab variable. Data can be accessed and written with subscript referencing and assignment
methods, just like a matlab variable, only size must be explicitly set or changed.
prop = hdf5prop(file, dataset) Creates a hdf5prop object for the given dataset in the
given HDF5 file.
prop = hdf5prop(file, dataset, mode) Specify the access mode, mode = r : readonly
access (default) mode = rw: readwrite access
prop = hdf5prop(file, dataset, parameter, value, ...) Creates a hdf5prop object
for a new dataset with given parameters.
parameter size - size of the dataset. Necessary. chunk size - size of the chunks of the dataset.
compression - 0-9 compression level. type - type of the dataset to create. This can be a string
such as double, in which case the datatype maps to H5T NATIVE DOUBLE, or it can be a derived
HDF5 datatype. max size - the maximum size of the dataset to create. fill - the fill value.

10.1 Properties
file
dataset
mode
propsize
dataset id

10.2 close
Close file if open.

31
10.3 set extent
Allocate space
hdf5prop.set extent(sz) Extend allocated space to size sz.

32
Chapter 11

Class: dependentprop

Property which is dynamically calculated. Use with empty () to get full evaluated array. Without
() to pass the unevaluated property. Indexing works as with normal arrays, only linear or boolean
indexing for multidimensional arrays is not allowed.

11.1 Properties
parent Object of which the property is dependent.
func Function to call on parent to calculate property data. The function should take care of
indexing.
sizefunc Function to call on parent to calculate property size. Should return array with size in
each dimension.

11.2 dependentprop
Create a dependentprop object.
prop = dependentprop(parent, func, sizefunc) Creates a dependentprop object with
given parameters.

33