Chapter 4

Weighting Functions and

Retrieval Techniques

Goal: Introducing the concept of weighting functions and profile retrieval

techniques for the case of an idealised example of atmospheric temperature
retireval from a microwave satellite nadir sounder.

4.1 Weighting Function for Microwave and

Infrared Satellite Nadir Sounding
As a simple example we will consider in this section the retrieval of atmo-
spheric temperature by a nadir looking microwave satellite sounder. I.e., the
instrument measures brightness temperatures emitted by the Earth surface
and the atmosphere. In order to simplify the example, we will consider a
case (i.e., a spectral region) where the atmospheric absorption is primarily
due to well mixed cases such as O2 or CO2 .
Remember the RTE without scattering:

dIν
= −αν Iν + αν Lν (4.1a)
dz
= −α(Iν − Lν ) (4.1b)

(Also known as ”Schwarzschild” equation.)

39
40 CHAPTER 4. WEIGHTING FUNCTIONS

With Tν (z, z ′ ) as the transmission between z and z ′ is given by:

Z z′ !
Tν (z, z ′ ) = exp − α(z”)dz” (4.2)
z

Now, the RTE can be solved to give the upward directed radiance at the
altitude z ′ . This equation is given by:
Z z
∂Tν (z, z ′ ) ′
Iν (z) = Iν (0) Tν (0, z) + Lν (z ′ ) dz (4.3)
0 ∂z ′
This is what the a satellite nadir sounder in the microwave or infrared spectral
regin will observe (for z → ∞). For a nadir looking satellite sensor this can
be written as:
Z ∞
Iν (∞) = Iν (0) Tν (0, ∞) + Lν (z)Kν (z)dz (4.4)
0

with the Weighting Functions K(z) defined as:

∂Tν (z, ∞)
Kν (z) = . (4.5)
∂z
We can write the measured intensity and the radiative transfer equation
also in terms of brightness temperature:
Z ∞
Tb,ν (∞) = ǫTsurface Tν (0, ∞) + T (z)Kν (z) dz (4.6)
0

where Tb,ν (∞) is the measured brightness temperature at frequency ν, Tsurface

the (actual, physical) temperature of the Earth surface and ǫ the emissivity
of the Earth surface (ǫ = 1 if the surface is a black body). T (z) is the temper-
ature of the atmosphere at hight z, while Tν (0, ∞) is the total transmission
of the atmosphere at frequency ν between the surface (z = 0) and the top of
the atmosphere (z = ∞).
This means that the observed signal will be the sum of a contribution from
the emission of the surface (attenuated by the atmospheric transmission) plus
the contributions from the individual atmospheric heights, weighted by the
corresponding Weighting Function. This is the meaning of the Weighting
Function: It will determine how much one atmospheric layer (the tempera-
ture of this layer in this example here) contributes to the measured signal
(the measured brightness temperature in this example).
4.1. MICROWAVE & INFRARED NADIR 41

For the idealized example discussed in this section we can find an analyt-
ical expression for the weighting functions, as we will see in the following. In
more realistic cases, however, the weighting functions can only be calculated
numerically from the discretized radiative transfer equation.
If the absorption results form a uniformly mixed compound (such as CO2
or O2 ), the absorption coefficient can be assumed to be of the form :

αν (z) = σν n(z) (4.7)

With the number density profile n(z) decreasing exponetially with the height,
given by:
 z
n(z) = n(0) exp − (4.8)
H

(scale height H, typically about 7km)

That means:
 z
αν (z) = αν (0) exp − (4.9)
H

then the transmission will be given by:

Tν (z, ∞) = exp(−αν (z)H) (4.10)

and the Weighting Function are given by:

∂Tν (z, ∞)
Kν (z) = = αν (z) exp(−αν (z)H) (4.11)
∂z

These functions have their maximum at zmax when:

α(zmax )H = 1 ⇒ zmax = H ln(αν (0)H) (4.12)

At the maximum the Weight Function have the value:

1
Kν (zmax ) = ≈ 0, 05/km (4.13)
eH
42 CHAPTER 4. WEIGHTING FUNCTIONS

Weighting Functions for different α

0
100

90

80

70

60
Altitude [km]

3
α = (7 )/km
0
50

α = 49/km
40 0

30
α = 7/km
0

20
α = 1/km
0
10
α0 = (1/49)/km
0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
Weighting Function [1/km]

Figure 4.1: Several Weighting Functions with the

scale height H = 7 km.

4.2 Retrieval of atmospheric temperature pro-

files
In the idealized example obove, we have seen that the measured brightness
temperature at frequency ν is given by the vertical integral over the atmo-
spheric temperatures, weighted by the weighting functions Kν :
Z ∞
Tb,ν (∞) = T (z)Kν (z) dz (4.14)
0
(here we have neglected the surface contribution for simplicity, see the cor-
responding equation above).
To solve this numerically, we will replace the integral by a sum over some
finite layers: X
Tb,j (∞) = Ti Kj,i∆zi . (4.15)
i
4.2. RETRIEVAL OF ATMOSPHERIC TEMPERATURE PROFILES 43

Here Ti is the Temperature at level i, ∆zi is the thickness of layer i, Tb,j (∞) is
the measured brightness temperature at frequency j, and Kj,i is the weighting
function at layer i and frequency j.
This can be written in more compact matrix notation as

y = K · x, (4.16)

where y is a vector containing the measured brightness temperatures (yj =

Tb,j (∞)), x is a vector containg the atmospheric temperature profile (xi = Ti )
and K is the weighting function matrix (with elements Kj,i∆zi ).
This equation thus provides us with a way to calculate the measured
brightness temperature from an atmospheric temperature profile (and is
sometimes called the ‘forward’ equation. In remote sensing we want to go the
oposite way: retrieving the atmospheric temperature profile from the mea-
sured brightness temperature. This can be done by the so-called ‘inversion’
of the forward equation. Naively, this may be done by direct inversion of the
weighting function matrix:
x = K−1 · y. (4.17)
However, not only is the inverse of K only defined if i = j, it is also generally
very sensitive to small errors in y and will not give any meaningfull result.
This is an example of a so called ill-posed problem.
One may try to look for the least squares normal solution

−1
x̂ = KT K · KT y. (4.18)

However, for an ill-posed system this will still not result in a stable solution
in the sense that small errors in y will result in hugh errors in x.
In order to solve this under-determined or ill-posed problem is to provide
some ‘regularisation’. This can be done in a statistical sense by the so called
optimal estimation equation, that selects from the infinite range of possible
solution a particular solution that is statistically optimal:

−1
x̂ = Sx KT KSx KT + Sy (y − Kx0 ) + x0 . (4.19)

Here Sy is the error covariance matrix of the measured brightness temper-

ature (i.e., describing the measurement error), Sx is the error covariance
matrix of the atmospheric temperature profiles, describing the statisticaly
properties of atmospheric temperature variations. x0 is a climatological tem-
perature profile (also known as ‘a priori profile’). If the error covariance
44 CHAPTER 4. WEIGHTING FUNCTIONS

matrices and the climatological a priori profile are known, than the optimal
estimation equation provides an estimate x̂ of the true atmospheric profile
x that is optimal in a statistical sense. In reality, however, the error co-
variance matrices are hardly known, so that they have to be estimated or
approximated.
The Optimal Estimation Equation above can be written as

x̂ = x0 + G (y − Kx0 ) (4.20)

where we have used

−1
G = Sx KT KSx KT + Sy (4.21)

and G is called a pseudo inverse of K or the ‘gain matrix’. If we insert

y = Kx in the equation above we get:

x̂ − x0 = GK (x − x0 ) = A (x − x0 ) . (4.22)

This means that the retrieved or estimated atmospheric profile x̂ (more pre-
cisely x̂ − x0 ) is given by the true but unknown atmospheric profile x (or
again, more precisely x − x0 ), multiplied by some matrix A = GK, called
the averaging kernel matrix. The averaging kernel matrix is of paramount
importance in remote sensing, as it describes the sensitivity and (vertical)
resolution of the instrument or retrieval process. E.g., if the averaging kernal
matrix is close to a unit matrx, than the retrieved profile x̂ is close to the true
(but unknwon) atmospheric profile x. On the other hand, broad averaging
kernels mean that the retrieved atmospheric temperature at a given level is
in reality an average over a broad reagion, i.e., the instrument has only a
poor vertical resolution.