Вы находитесь на странице: 1из 173

Polarimetry Tutorial

Earthnet 14-Apr-2007

Polarimetry Tutorial

This section provides access to a wide-ranging tutorial, which aims to provide a grounding in
polarimetry and polarimetric interferometry with a view to stimulating research and
Home development of scientific applications that exploit such techniques.
Background
Airborne Data Sources The tutorial material is made available in PDF format, for which you will require Adobe
Spaceborne Data Sources Acrobat Reader to view. The tutorial is also bundled with the software download. To access
Simulated Data Sources the tutorial, click on a subject heading below.
Polarimetry Tutorial
Course Material
Introduction (70kb)
Documentation
Release notes
Download and Install Part I - Tutorial on Radar Polarimetry
Exploitation Results and
News
1. What Is Polarisation? (400kb)
Contacts and
2. Single vs Multi-Polarisation SAR Data (1.5Mb)
Acknowledgements
3. Speckle Filtering (650kb)
4. Polarimetric Decompositions (4.3Mb)
5. Polarimetric SAR Data Classification (700kb)
6. Envisat/ASAR Dual Polarisation Case (150kb)

Part II - Tutorial on Polarimetric SAR Interferometry

1. Pol-InSAR Training Course (1.1Mb)


2. Single vs Multi-Polarisation Interferometry (2.5Mb)

Part III - Tutorial on Surface Parameter Retrieval

1. Description of Natural Surfaces (300kb)


2. Rough Surface Scattering Models (300kb)
3. Single vs Multi-Polarisation Descriptors (350kb)
4. Estimation of Surface Characteristics (550kb)

Part IV - Glossary (100kb)

Part V - References (350kb)

Part VI - Do It Yourself

Foreword (13kb)
1. Getting Started with POLSARPRO (1.7Mb)
2. Representations of Polarimetric Information (650kb)
3. Speckle Filtering (2.7Mb)
4. Polarimetric Decompositions (1.2Mb)
5. Polarimetric Segmentation (450kb)

http://earth.esa.int/polsarpro/tutorial.html (1 of 2)4/14/2007 2:00:17 PM


Polarimetry Tutorial

6. Envisat/ASAR Dual Polarisation Case (3.6Mb)


7. POLinSAR training course practicals (1.1Mb)

Lecture Notes

This section provides lectures on Basic and Advanced Radar Polarimetry and Polarimetric
SAR Interferometry by Wolfgang-Martin BOERNER, Eric POTTIER, Jong-Sen LEE, Laurent
FERRO-FAMIL and Shane R. CLOUDE.

1. RECENT ADVANCES IN RADAR POLARIMETRY AND POLARIMETRIC SAR INTERFEROMETRY


(3.1Mb)
by Wolfgang-Martin BOERNER

2. BASIC CONCEPTS IN RADAR POLARIMETRY (2.4Mb)


by Wolfgang-Martin BOERNER

3. ADVANCED CONCEPTS (7.7Mb)


by Eric POTTIER, Jong-Sen LEE, Laurent FERRO-FAMIL

4. POL-InSAR Training Course (1.1Mb)


by Shane R. CLOUDE

Higher level pr Print version Last modified: 13-Mar-2007

http://earth.esa.int/polsarpro/tutorial.html (2 of 2)4/14/2007 2:00:17 PM


http://earth.esa.int/polsarpro/Manuals/Introduction.pdf

Embedded Secure Document


The file http://earth.esa.int/polsarpro/Manuals/Introduction.pdf is a secure document that has been
embedded in this document. Double click the pushpin to view.

http://earth.esa.int/polsarpro/Manuals/Introduction.pdf4/14/2007 2:00:42 PM
What is Polarization?

1. WHAT IS POLARIZATION?

1.1 Propagation of a monochromatic plane electromagnetic


wave

1.1.1 Equation of propagation

The time-space behavior of electromagnetic waves is ruled by the Maxwell equations set
defined as
r r r r
r r r ∂B ( r , t ) r r r r r ∂D ( r , t )
∇ ∧ E (r , t ) = − ∇ ∧ H (r , t ) = J T (r , t ) +
r ∂t ∂t
r r r r r r (1)
∇ ⋅ D(r , t ) = ρ (r , t ) ∇ ⋅ B (r , t ) = 0
r r r r r r r r
where E (r , t ), H ( r , t ), D ( r , t ), B (r , t ) are the wave electric field, magnetic field, electric
induction and magnetic induction respectively.
r r r r r r
The total current density, J T ( r , t ) = J a ( r , t ) + J c (r , t ) is composed of two terms. The first one,
r r
J a ( r , t ) , corresponds to a source term, whereas the conduction current density,
r r r r
J c (r , t ) = σ E (r , t ) , depends on the conductivity of the propagation medium, σ . The scalar
r
field ρ (r , t ) represents the volume density of free charges.
The different fields and induction are related by the following relations
( )
r r r r r r r r r r r r
D (r , t ) = ε E (r , t ) + P(r , t ) B(r , t ) = μ H (r , t ) + M (r , t ) (2)
r r r r
The vectors P (r , t ) and M ( r , t ) are called polarization and magnetization, while ε and
μ stand for the medium permittivity and permeability.
In the following, we shall consider the propagation of an electromagnetic wave in a linear
medium (free of saturation and hysteresis), free of sources. These hypothesis imposes that
r r r r r r r r
M ( r , t ) = P ( r , t ) = 0 and J a (r , t ) = 0 .

The equation of propagation is found by inserting (1) and (2) into


r r r r r r r r r r
∇ ∧ (∇ ∧ E ( r , t )) = ∇ (∇ ⋅ E ( r , t )) − ΔE ( r , t ) and is formulated as
r r r r r r
r r ∂ 2 E (r ,t ) ∂ E (r ,t ) ∂ ∇ρ (r ,t )
ΔE (r ,t ) − με − μσ =−1 (3)
∂t 2
∂t ε ∂t

1
What is Polarization?

1.1.2 Monochromatic plane wave solution

Among the infinite number of solutions to the equation of propagation mentioned in (3), we
will study the special case of constant amplitude monochromatic plane waves which is
adapted to the analysis of a wave polarization.
r r
∂ ∇ρ (r ,t ) r
The monochromatic assumption implies that the right hand term of (3) is null = 0,
∂t
i.e. the propagation medium is free of mobile electric charges (e.g. is not a plasma whose
charged particles may interact with the wave).
The propagation equation expression can be significantly simplified by considering the
r r r r
complex expression, E (r ) , of the monochromatic time-space electric field, E (r , t ) , defined
as
( )
r r r r
E ( r , t ) = ℜ E ( r ) e j ωt (4)

The propagation equation mentioned in (3) may then be rewritten as


r r ⎛ σ ⎞ r r r r 2 r r
Δ E (r ) + ω 2 με ⎜1− j ⎟ E (r ) = Δ E (r ) + k E (r )
⎝ εω ⎠
(5)
ω σ
with k = 1− j
v εω
Here appears the concept of complex dielectric constant
σ ω ε ′′
ε = ε ′ − jε ′′ = ε − j then k = 1− j = β − jα (6)
ω v ε
In a general way, a monochromatic plane wave , with constant complex amplitude,
r r
E 0 = E à e jδ , propagating in the direction of the wave vector, kˆ , has the complex following
form
r r r rr r r
E ( r ) = E 0 e − j k ⋅r with E ( r ) ⋅ kˆ = 0 (7)

One may verify that such a wave satisfies the propagation equation given in (5). Without any
loss of generality, the electric field may be represented in an orthonormal basis
( xˆ , yˆ , zˆ ) defined so that the direction of propagation kˆ = zˆ . The expression of the electric field
becomes
r r
E ( z ) = E 0 e −αz e − jβz with E 0 z = 0 (8)
It may be observed from (8) that β acts as the wave number in time domain, while α
corresponds to an attenuation factor. Back to time domain, this expression becomes in
vectorial form
⎡ E 0 x e − az cos(ωt − kz + δ x ) ⎤
r ⎢ ⎥
E ( z , t ) = ⎢ E 0 y e − az cos(ωt − kz + δ y )⎥ (9)
⎢ 0 ⎥
⎣ ⎦

2
What is Polarization?

The attenuation term is common to all the elements of the electric field vector and is then
unrelated to the wave polarization. For this reason, the medium is assumed to be loss free,
α = 0 , in the following
⎡ E 0 x cos(ωt − kz + δ x ) ⎤
r
E (r , t ) = ⎢⎢ E 0 y cos(ωt − kz + δ y )⎥⎥ (10)
⎢⎣ 0 ⎥⎦

1.1.3 Spatial evolution of a plane wave vector: helicoidal trajectory

At a fixed time, t = t 0 , the electric field is composed of two orthogonal sinusoidal waves
with, in general, different amplitudes and phases at the origin.

y$
x$

Ex ( z, t )
0
z$

E y ( z, t )
Figure 1 Spatial evolution of a monochromatic plane wave components.

Three particular cases are generally discriminated:


• Linear polarization: δ = δ y − δ x = 0 + mπ

The electric field is then a sine wave inscribed within a plane oriented with an angle φ with
respect to x̂
⎡cos φ ⎤
r
E ( z 0 , t ) = E 02x + E 02y ⎢⎢ sin φ ⎥⎥ cos(ωt 0 − kz + δ x ) (11)
⎢⎣ 0 ⎥⎦

3
What is Polarization?

y$
x$ r
E( z , t )

0 Ex ( z, t )
z$

Figure 2 Spatial evolution of a linearly (horizontal) polarized plane wave.

• Circular polarization: δ = δ y − δ x = 0 + mπ / 2 and E 0 x = E 0 y

In this case, the wave has a constant modulus and is oriented with an angle φ (z ) with respect
to the x̂ axis
r
E ( z, t 0 ) = E02x + E02y and φ ( z ) = ± (ωt à − kz + δ x ) (12)

y$ r
x$ E( z , t )

Ex ( z, t )
0
z$

E y ( z, t )
Figure 3 Spatial evolution of a circularly polarized plane wave.

The wave rotates circularly around the ẑ axis.


• Elliptic polarization: Otherwise
The wave describes a helicoidal trajectory around the ẑ axis.

4
What is Polarization?

y$ r
x$ E( z , t )

Ex ( z, t )
0
z$

E y ( z, t )
Figure 4 Spatial evolution of a elliptically polarized plane wave.

1.2 Polarization ellipse

1.2.1 Geometrical description

The former paragraph introduced the spatial evolution of a plane monochromatic wave and
showed that it follows a helicoidal trajectory along the ẑ axis. From a practical point of view,
three-dimensional helicoidal curves are difficult to represent and to analyze. This is why a
characterization of the wave in the time domain, at a fixed position, z = z 0 is generally
preferred.
y$ r
x$ E( z , t )
r
y$
E( z0 , t )

0 x$

z0 z$

Figure 5 Temporal trajectory of a monochromatic plane wave at a fixed abscissa z = z 0 .

The temporal behavior is then studied within an equiphase plane, orthogonal to the direction
of propagation and at a fixed location along the ẑ axis. As time evolves, the wave propagates
"through" equi-phase planes nd describe a characteristic elliptical locus as shown in Figure 5.
The nature of the wave temporal trajectory may be determined from the following parametric
r
relation between the components of E ( z0 , t )

5
What is Polarization?

2
⎛ E y ( z0 , t ) ⎞
2
⎛ E x ( z0 , t) ⎞ E ( z , t)E y ( z0 , t)
⎜⎜ ⎟⎟ − 2 x 0 cos(δ y − δ x ) + ⎜ ⎟ = sin(δ y − δ x ) (13)
E E E ⎜ E ⎟
⎝ 0x ⎠ 0x 0 y ⎝ 0y ⎠
The expression in (13) is the equation of an ellipse, called the polarization ellipse, that
describes the wave polarization.
The polarization ellipse shape may be characterized using 3 parameters as shown in Figure 6.

E0 x

τ
ẑ E0 y
φ

A

Figure 6 Polarization ellipse.

- A is called the ellipse amplitude and is determined from the ellipse axis as

A = E02x + E02y (14)

⎡ π π⎤
- φ ∈ ⎢− , ⎥ is the ellipse orientation and is defined as the angle between the ellipse major
⎣ 2 2⎦
axis and x̂
E0 x E0 y
tan 2φ = 2 cos δ with δ = δ y − δ x (15)
E02x − E02y

⎡ π⎤
- τ ∈ ⎢0, ⎥ is the ellipse aperture, also called ellipticity, defined as
⎣ 4⎦
E0 x E0 y
sin 2τ = 2 sin δ (16)
E02x + E02y

6
What is Polarization?

1.2.2 Sense of rotation


r
As time elapses, the wave vector E ( z0 , t ) rotates in the ( xˆ , yˆ ) to describe the polarization
r
ellipse. The time-dependent orientation of E ( z0 , t ) with respect to x̂ , named ξ (t ) is shown in
Figure 7.

r
E (z0 , t )

ẑ ξ (t)

r
Figure 7 Time-dependent rotation of E ( z 0 , t ) .

The time-dependent angle may be defined from the components of the wave vector in order to
determine its sense of rotation.
E y ( z0 , t ) E 0 y cos(ωt − kz 0 + δ y )
tan ξ (t ) = = (17)
E x ( z0 , t) E0 x cos(ωt − kz 0 + δ x )
The sense of rotation may then be related to the sign of the variable τ
∂ξ (t ) ⎛ ∂ξ (t ) ⎞ E0 x E0 y
∝ − sin δ ⇒ sign⎜ ⎟ = − sign (τ ) with sin 2τ = 2 2 sin δ (18)
∂t ⎝ ∂t ⎠ E 0 x + E02y

By convention, the sense of rotation is determined while looking in the direction of


∂ξ (t )
propagation. A right hand rotation corresponds then to > 0 ⇒ (τ , δ ) < 0 whereas a left
∂t
∂ξ (t )
hand rotation is characterized by < 0 ⇒ (τ , δ ) > 0 .
∂t
Figure 8 provides a graphical description of the rotation sense convention.

7
What is Polarization?

ŷ ŷ

ẑ ẑ
x̂ x̂

ŷ ŷ

x̂ x̂

ẑ ẑ

(a) (b)
Figure 8 (a) Left hand elliptical polarizations. (b) Right hand elliptical polarizations.

1.2.3 Quick estimation of a wave polarization state

A wave polarization is completely defined by two parameters derived from the polarization
ellipse
⎡ π π⎤
- its orientation, φ ∈ ⎢− , ⎥
⎣ 2 2⎦
⎡ π π⎤
- its ellipticity τ ∈ ⎢− , ⎥ , with sign (τ ) indicating the sense of rotation
⎣ 4 4⎦
The ellipse amplitude A can be used to estimate the wave power density.
The following procedure provides a quick (calculation free) way to roughly estimate a wave
polarization.
Three cases may be discriminated from the knowledge of δ = δ y − δ x , EOx , EOy

• δ = 0, π

8
What is Polarization?

⎛ E0 y ⎞
The polarization is linear since τ = 0 and the orientation angle is given by φ = tan −1 ⎜⎜ ⎟⎟ if
⎝ E0 x ⎠
⎛ E0 y ⎞
δ = 0 and φ = −a tan⎜⎜ ⎟⎟ if δ = π
⎝ E 0 x ⎠
π
• δ =± and E 0 x = E 0 y
2
π
The polarization is circular, since τ = ± and the sense of rotation is given by sign (δ ) .
4
If δ < 0 , the polarization is right circular, whereas for δ > 0 the polarization is left circular.
• Otherwise
If δ < 0 , the polarization is right elliptic, whereas for δ > 0 the polarization is left elliptic.

1.2.4 Canonical polarization states

In practice the axes x̂ and ŷ are generally referred to as the horizontal, ĥ and vertical vˆ
directions.

(a) (b)
Figure 9 (a) Horizontal polarization (b) Vertical polarization.

(a) (b)
Figure 10 (a) Linear + 45° polarization. (b) Linear - 45° polarization.

9
What is Polarization?

(a) (b)
Figure 11 (a) Right circular polarization. (b) Left circular polarization.

(a) (b)
Figure 12 (a) Right elliptical –45 °polarization. (b) Left elliptical +45 °polarization.

1.3 Jones vector

1.3.1 Definition

The representation of a plane monochromatic electric field under the form of a Jones vector
aims to describe the wave polarization using the minimum amount of information.
r
A Jones vector, E , is defined from the time-space vector E ( z , t ) as

( )
r r r
E = E (0) with E ( z , t ) = ℜ E ( z ) e jωt (19)
r
From the formulation of E ( z , t ) given in (10), E can be written as

⎡ E e jδ x ⎤
E = ⎢ 0 x jδ y ⎥ (20)
⎢⎣ E 0 y e ⎥⎦
The definitions of a polarization state from the polarization ellipse descriptors or from a Jones
vector are equivalent.
A Jones vector can be formulated as a two-dimensional complex vector function of the
polarization ellipse characteristics as follows :

10
What is Polarization?

⎡cos φ cosτ − j sin φ sin τ ⎤


E = Ae jα ⎢ ⎥ (21)
⎣sin φ cosτ + j cos φ sin τ ⎦
Where α is an absolute phase term.
The Jones vector may be written under a more effective matrix form
⎡cos φ − sin φ ⎤ ⎡ cosτ ⎤
E = Ae jα ⎢
cos φ ⎥⎦ ⎢⎣ j sin τ ⎥⎦
(22)
⎣ sin φ

1.3.2 Orthogonal polarization states and polarization basis

1.3.2.1 Orthogonal Jones vectors


Two Jones vectors, E 1 and E 2 are orthogonal if their hermitian scalar product is equal to 0,
i.e.
E1 E 2 = 0

(23)
with † the transpose conjugate operator.
From the definition of a Jones vector given in (22), it is straightforward to remark that the
orthogonality condition implies that ellipse parameters of E 1 and E 2 satisfy
π
φ 2 = φ1 + and τ 2 = −τ 1 (24)
2
One may remark that the orthogonality condition does not depend on the absolute phase term
of each Jones vector, α 1 and α 2 , i.e. if E 1 and E 2 are orthogonal, then E 1 and E 2 e jψ are
orthogonal too, for any value of ψ .

1.3.2.2 Polarization basis


According to the definition of a Jones vector from the time-space electric field given in (19),
any Jones vector expressed in the orthonormal basis ( xˆ , yˆ ) as
E = E x xˆ + E y yˆ (25)
A Jones vector defined in the basis ( xˆ , yˆ ) , E ( xˆ , yˆ ) in the may defined from the unitary vector
associated to the horizontal direction, x̂
⎡cos φ − sin φ ⎤ ⎡ cosτ j sin τ ⎤
E ( xˆ , yˆ ) = Ae jα ⎢ xˆ
cos φ ⎥⎦ ⎢⎣ j sin τ cosτ ⎥⎦
(26)
⎣ sin φ
This expression may be further developed
⎡cos φ − sin φ ⎤ ⎡ cosτ j sin τ ⎤ ⎡e − jα 0 ⎤
E ( xˆ , yˆ ) = A⎢ ⎢ ⎥ xˆ
cos φ ⎥⎦ ⎢⎣ j sin τ cosτ ⎥⎦ ⎣ 0
(27)
⎣ sin φ e jα ⎦

The orthogonal Jones may be expressed from (24) as

11
What is Polarization?

⎡ ⎛ π⎞ ⎛ π ⎞⎤
⎢cos⎜ φ + 2 ⎟ − sin⎜ φ + 2 ⎟⎥ ⎡ cosτ − j sin τ ⎤ ⎡e − jα 0 ⎤
E ⊥( xˆ , yˆ ) = A⎢ ⎝ ⎠ ⎝ ⎠⎥
⎢ ⎢ ⎥ xˆ
cosτ ⎥⎦ ⎣ 0
(28)
⎢ sin⎛⎜ φ + π ⎞⎟ cos⎛⎜ φ + π ⎞⎟ ⎥ ⎣− j sin τ e jα ⎦
⎢⎣ ⎝ 2⎠ ⎝ 2 ⎠ ⎥⎦

or
⎡cos φ − sin φ ⎤ ⎡ cosτ j sin τ ⎤ ⎡e − jα 0 ⎤
E ⊥( xˆ , yˆ ) = A⎢ ⎢ ⎥ yˆ
cos φ ⎥⎦ ⎢⎣ j sin τ cosτ ⎥⎦ ⎣ 0
(29)
⎣ sin φ e jα ⎦
The matrices associated to the φ ,τ , α angular variables

cos φ − sin φ ⎤ cosτ j sin τ ⎤ ⎡ − jα


0 ⎤
[U (φ )] = ⎡⎢ [U (τ )] = ⎡⎢ [U (α )] = ⎢e ⎥ xˆ
cos φ ⎥⎦ cosτ ⎥⎦
(30)
⎣ sin φ ⎣ j sin τ ⎣ 0 e jα ⎦

belong to the group of (2×2) Special Unitary complex matrices SU(2) and have the following
important properties :
- [U ] = +1

- [U ]−1 = [U ]†
- [U ( x)]−1 = [U (− x)]
Two Jones vectors, u and v with unitary norms, form a polarization basis if they result from
the transformation of the ( xˆ, yˆ ) basis
u = [U (φ )][U (τ )][U (α )]xˆ and v = [U (φ )][U (τ )][U (α )] yˆ (31)
Or equivalently if
u = [U (φ )][U (τ )][U (α )]xˆ and v = [U (φ + π )][U (−τ )][U (α )]xˆ (32)
It can be remarked that a polarization basis can uniquely defined by a single vector
u = [U (φ ,τ , α )]xˆ , provided that the second element of the basis v verifies v = u ⊥ .
One has to point out that the definition of a polarization basis provided in (31) and (32)
requires that both elements of the basis are constructed using the same absolute phase value
α . This condition is not necessary for u and v to be orthogonal but may involve important
problems for the analysis of polarimetric response if it is not respected.
Example :
Let R be the Jones vector associated to a right circular polarization

π 1 ⎡1 ⎤
R = [U (φ = 0)][U (τ = − )][U (α = 0)]xˆ = ⎢ ⎥ (33)
4 2 ⎣− j ⎦
Then the other element of the orthonormal basis is
π π 1 ⎡− j ⎤
R ⊥ = [U (φ = + )][U (τ = + )][U (α = 0)]xˆ = ⎢ ⎥ (34)
2 4 2⎣1 ⎦

12
What is Polarization?

It can observed that R ⊥ is slightly different from the usual definition of a left circular
polarization Jones vector L
π
1 ⎡1 ⎤ j
L= ⎢ ⎥ = R ⊥ e 2
(35)
2 ⎣ j⎦
Both Jones vector depict a left circular polarization state but R ⊥ only may be coupled to R to
form a polarization in the sense it was defined in (31) and (32).

1.3.3 Polarization ratio

1.3.3.1 Definition
An efficient way to characterize a Jones vector polarization state is to build its polarization
ratio defined as
Ey E0 y j (δ y −δ y )
ρ= = e (36)
Ex E0 x
The polarization ratio may be written as a function of the polarization ellipse parameters as
sin φ cos τ + j cos φ sin τ
ρ= (37)
cos φ cos τ − j sin φ sin τ
Canonical polarization states can be easily discriminated from the knowledge of ρ :
• Arg ( ρ ) = 0 + mπ
The polarization is linear and the orientation angle is given by φ = tan −1 ( ρ )
π
j
• ρ =e 4

The polarization is circular, sign ( Arg ( ρ )) .


If Arg ( ρ ) < 0 , the polarization is right circular, whereas for Arg ( ρ ) > 0 the polarization is
left circular.
• Otherwise
If Arg ( ρ ) < 0 , the polarization is right elliptic, whereas for Arg ( ρ ) > 0 the polarization is left
elliptic.

13
What is Polarization?

1.3.3.2 Representation of polarization states


Canonical polarization states are given in the following table

Polarization Unitary Jones Orientation Ellipticity Polarizatio ratio


vector uˆ( xˆ , yˆ ) ρ( xˆ , yˆ )
States (φ) ( τ)

⎡1 ⎤
⎢0⎥
Horizontal (H) ⎣ ⎦ 0 0 0

⎡0⎤
⎢1 ⎥ π
Vertical (V) ⎣ ⎦ 2 0 ∝

1 ⎡1⎤
⎢⎥ π
Linear +45° 2 ⎣1⎦ 0 1
4

1 ⎡ −1⎤
⎢ ⎥ 3π
Linear -45° 2 ⎣ 1⎦ 0 -1
4

1 ⎡1 ⎤
⎢ ⎥ π
Left circular 2 ⎣ j⎦ ? j
4

1 ⎡ 1⎤
⎢ ⎥
Right circular 2 ⎣− j ⎦ ? −π -j
4

A polarization map may also be built from the representation of polarization states in a
complex plane.

14
What is Polarization?

ℑ(ρ ( x̂ , ŷ ) )

ℜ(ρ ( x̂ , ŷ ) )
0

Figure 13 Polarization map in the real and imaginary polarization ratio plane.

1.3.3.3 Orthogonal polarization states and polarization basis


A Jones vector components may be expressed as a function of its polarization as follows
⎡cos φ cosτ − j sin φ sin τ ⎤ 1 ⎡1⎤
E = Ae jα ⎢ ⎥ = Ae jα ' ⎢ρ ⎥
⎣sin φ cosτ + j cos φ sin τ ⎦
(38)
1+ ρ
2
⎣ ⎦

where the absolute phase term is modified in order to account for the polarization ratio
argument
α ' = α + arg(cos φ cos τ − j sin φ sin τ ) (39)
The orthogonal Jones vector is given by
⎡− sin φ cos τ + j cos φ sin τ ⎤ 1 ⎡1⎤
E ⊥ = Ae jα ⎢ ⎥ = Ae jα '' ⎢ρ ⎥
⎣ cos φ cos τ + j sin φ sin τ ⎦
(40)
1+ ρ⊥
2
⎣ ⊥⎦

with α ' ' = α + arg(− sin φ cos τ + j cos φ sin τ )


With ρ ⊥ the orthogonal polarization ratio defined as
1
ρ⊥ = − (41)
ρ*
A polarization basis (uˆ , vˆ) may then be defined from a vector polarization ratio as follows
u = [U ( ρ )][U (α )]xˆ and v = [U ( ρ )][U (α )] yˆ (42)
with

15
What is Polarization?

1 ⎡1 − ρ*⎤
[U ( ρ )] = [U (φ )][U (τ )] = ⎢ ⎥ (43)
1+ ρ ⎣ρ 1 ⎦
2

A polarization basis can uniquely defined by the polarization ration of a single vector u ,
provided that the second element of the basis v verifies v = u ⊥ .
Note that the use of the following transformation matrix
ρ ⎡1 1⎤
⎢ρ ρ ⊥ ⎥⎦ (44)
1+ ρ
2

would lead to the same polarization state for v , but to different α phase terms for u and v .

1.3.4 Change of polarimetric basis

One of the main advantages of radar polarimetry resides in the fact that once a target response
is acquired in a polarization basis, the response in any basis can be obtained from a simple
mathematical transformation and does not require any additional measurements.
A Jones vector, E ( xˆ , yˆ ) = E x xˆ + E y yˆ expressed in the ( xˆ, yˆ ) orthonormal polarimetric basis,
transforms to E ( uˆ ,vˆ ) = E u uˆ + E v vˆ in the (uˆ , vˆ) orthonormal basis, with uˆ given by
uˆ = [U (φ )][U (τ )][U (α )]xˆ , by the way of a Special Unitary transformation.
The coordinates E u and E v can be determined according to the following expression
E ( uˆ ,vˆ ) = E u uˆ + E v vˆ ⇒ E ( xˆ , yˆ ) = E u [U (φ ,τ , α )] xˆ + E u [U (φ ,τ , α )] yˆ = E x xˆ + E y yˆ
(45)
⇒ E u = [U (φ ,τ , α )] −1 E x and E v = [U (φ ,τ , α )] −1 E y

Finally
E ( uˆ ,vˆ ) = [U ( xˆ , yˆ ) →( uˆ ,vˆ ) ] E ( xˆ , yˆ ) with
(46)
[U ( xˆ , yˆ )→( uˆ ,vˆ ) ] = [U (φ ,τ , α )] −1 = [U ( −α )][U ( −τ )][U ( −φ )]

Similarly a change of polarimetric basis from ( aˆ , bˆ) to (uˆ , vˆ) can be operated using a
transformation matrix as follows
aˆ = [U (φ a ,τ a , α a )]xˆ = [U a ]xˆ
⇒ uˆ = [U u ][U a ] −1 aˆ ⇒ E (uˆ ,vˆ ) = [U a ][U u ] −1 E (uˆ ,vˆ ) (47)
ˆu = [U (φ u ,τ u ,α u )]xˆ = [U u ]xˆ

Note that transformation matrices can also be built from the polarization ratio as shown in the
former paragraph.

16
What is Polarization?

1.4 Stokes vector

1.4.1 Real representation of a plane wave vector

In the previous section, we presented the representation of the polarization state of a plane
monochromatic electric field by means of the complex Jones vector. As it can be observed in
(20), the Jones vector is determined by two complex quantities. Consequently, if the goal of a
given system is to measure the Jones vector of the received wave, this system must record the
amplitude and the phase of the incoming wave.
The availability of coherent systems able to measure the amplitude and phase of the incoming
waves is relatively recent. In the past, only non-coherent systems were available. These
systems are only able to measure the power of an incoming wave. Consequently, it was
necessary to characterize the polarization of a wave only by power measurements. This
characterization is carried out by the so-called Stokes vector.
Given the Jones vector E of a given wave, we can create the hermitian product as follows

T* ⎡ E E* Ex E *y ⎤
E⋅E = ⎢ x x* ⎥ (48)
⎢⎣ E y Ex E y E *y ⎥⎦

giving as a result a 2×2 hermitian matrix. At this point, if we consider the Pauli group of
matrices
⎡1 0 ⎤
σ0 = ⎢ ⎥ (49)
⎣0 1 ⎦
⎡1 0 ⎤
σ1 = ⎢ ⎥ (50)
⎣0 −1⎦
⎡0 1⎤
σ2 = ⎢
0⎥⎦
(51)
⎣1
⎡0 − j⎤
σ3 = ⎢
0 ⎥⎦
(52)
⎣j
It is possible to decompose (48) as follows
1 1⎡ g +g g 2 − jg3 ⎤
E⋅E
T*
= { g0σ 0 + g1σ1 + g 2σ 2 + g3σ 3} = ⎢ 0 1
g 0 − g1 ⎥⎦
(53)
2 2 ⎣ g 2 + jg3

where the parameters {g0, g1, g2, g3} receive the name of Stokes parameters. From (53), the
Stokes vector, denoted by gE
⎡ E 2+ E 2⎤
⎡ g0 ⎤ ⎢ ⎥
x y

⎢g ⎥ ⎢ E 2 − E 2 ⎥
gE = ⎢ 1 ⎥ = ⎢
x y ⎥
⎢ ⎥ (54)
⎢ g2 ⎥
⎢ ⎥ ⎢ {
2ℜ Ex E *y ⎥ }
⎣⎢ g3 ⎦⎥ ⎢ * ⎥
{
⎣⎢ −2ℑ Ex E y ⎦⎥ }

17
What is Polarization?

where the following relation can be established


g 02 = g12 + g 22 + g 32 (55)
The relation given at (55) establishes that in the set {g0, g1, g2, g3} there are only three
independent parameters. The Stokes parameter g0 is always equal to the total power (density)
of the wave; g1 is equal to the total power in the linear horizontal or vertical polarized
components; g2 is equal to the power in the linearly polarized components at tilt angles ψ=45
degrees or 135 degrees and g3 is equal to the power in the left-handed and right-handed
circular polarized component in the plane wave. If any of the parameters {g0, g1, g2, g3} has a
non-zero value, it indicates the presence of a polarized component in the plane wave.
The Stokes parameters are sufficient to characterize the magnitude and the relative phase, and
hence, the polarization of a wave. As it can be observed in (54), the Stokes parameters can be
obtained from only power measurements. Consequently, the Stokes vector is capable to
characterize the polarization state of a wave by 4 real parameters. The next section presents
the relations existing between the Stokes parameters {g0, g1, g2, g3} and the polarization
ellipse parameters.

1.4.2 Relation between the Stokes vector and the polarization ellipse

The Stokes vector given at (54) can be written as follows


⎡ g 0 ⎤ ⎡ E0 x + E0 y ⎤
2 2

⎢ g ⎥ ⎢ E2 − E2 ⎥
gE = ⎢ 1⎥
= ⎢ 0x 0y ⎥
(56)
⎢ g 2 ⎥ ⎢ 2 E0 x E0 y cos (δ ) ⎥
⎢ ⎥ ⎢ ⎥
⎣⎢ g3 ⎦⎥ ⎢⎣ 2 E0 x E0 y sin (δ ) ⎥⎦
If now, we consider the expression presented at (15) and (16), the Stokes vector can be
written as a function of: the polarization ellipse orientation angle φ and ellipse aperture angle τ
and the polarization ellipse aperture A
⎡ A ⎤
⎢ A cos 2φ cos 2τ ⎥
( ) ( )⎥
gE = ⎢
⎢ A sin ( 2φ ) cos ( 2τ ) ⎥
(57)
⎢ ⎥
⎣⎢ A sin ( 2τ ) ⎦⎥
At Section 1.3.2.2, we represented a given Stokes vector as the product of three unitary
matrices belonging to the special unitary SU(2), see (27) and (30). By using the existing
homorphism between the group SU(2) and the group O(3) of real orthogonal matrices, given
by
1
(
⎡⎣O3 ( 2θ ) ⎤⎦ p , q = Tr ⎡⎣U 2 (θ ) ⎤⎦ σ p ⎡⎣U 2 (θ ) ⎤⎦ σ q
2
T*
) (58)

we can write the Stokes vector of a particular polarization state as follows

18
What is Polarization?

⎡1 0 0 0 ⎤ ⎡1 0 0 0 ⎤ ⎡1 0 0 0 ⎤
cos ( 2τ ) 0 − sin ( 2τ )
g E = A ⎢ 00 2 cos( 2φ ) − sin ( 2φ ) 0 ⎥ ⎢0 ⎥ ⎢0 1 0 0 ⎥g
sin ( 2φ ) cos ( 2φ ) 0 cos ( 2α ) − sin ( 2α )
⎢ 0
⎥ ⎢ 00 0 1 0
⎥ ⎢0 ⎥ uˆx
⎣0 0 0 0 ⎦⎣ sin ( 2τ ) 0 cos( 2τ ) ⎦ ⎣ 0 0 sin ( 2α ) cos ( 2α ) ⎦

(59)
where g uˆ represents the Stokes vector associated with the horizontal polarization (59) can be
x

rewritten compactly as
g E = A2 ⎡⎣O4 ( 2φ ) ⎤⎦ ⎡⎣O4 ( 2τ ) ⎤⎦ ⎡⎣O4 ( 2α ) ⎤⎦ g uˆ (60)
x

1.4.2.1 Orthogonal Stokes vectors


At Section 1.3.2.1 we defined the orthogonal Jones vectors. As observed in (24), the
orthogonality can be established in terms if the angles defining the polarization ellipse.
Consequently, given the Stokes vector of a given polarization state gE, see (57), the
orthogonal Stokes vector is
⎡ A ⎤
⎢ − A cos 2φ cos 2τ ⎥
( ) ( )⎥
gE ⊥ =⎢
⎢ − A sin ( 2φ ) cos ( 2τ ) ⎥
(61)
⎢ ⎥
⎣⎢ − A sin ( 2τ ) ⎦⎥

1.4.2.2 Canonical polarization states


The Stokes vector for the canonical polarization states are presented in the following formula

Polarization Unitary Jones Stokes vector


vector uˆ( xˆ , yˆ ) gE
States

⎡1 ⎤
⎢1 ⎥
⎡1 ⎤ ⎢ ⎥
⎢0⎥ ⎢0⎥
Horizontal (H) ⎣ ⎦
⎢ ⎥
⎣0⎦

⎡1⎤
⎢ −1⎥
⎡0⎤ ⎢ ⎥
⎢1 ⎥ ⎢0⎥
Vertical (V) ⎣ ⎦
⎢ ⎥
⎣0⎦

19
What is Polarization?

⎡1 ⎤
⎢0⎥
1 ⎡1⎤ ⎢ ⎥
⎢⎥ ⎢1 ⎥
Linear +45° 2 ⎣1⎦
⎢ ⎥
⎣0⎦

⎡1⎤
⎢0⎥
1 ⎡ −1⎤ ⎢ ⎥
⎢ ⎥ ⎢ −1⎥
Linear -45° 2 ⎣ 1⎦
⎢ ⎥
⎣0⎦

⎡1 ⎤
⎢0⎥
1 ⎡1 ⎤ ⎢ ⎥
⎢ ⎥ ⎢0⎥
Left circular 2 ⎣ j⎦
⎢ ⎥
⎣1 ⎦

⎡1⎤
⎢0⎥
1 ⎡ 1⎤ ⎢ ⎥
⎢ ⎥ ⎢0⎥
Right circular 2 ⎣− j ⎦
⎢ ⎥
⎣ −1⎦

1.4.3 Representation of Stokes vectors: The Poincaré sphere

As it has been mentioned, the Stokes vector is completely determined by three independent
parameters. Consequently, a three dimensional representation of the Stokes vector is possible.
This representation receives the name of Poincaré sphere.
If we consider the expression for the Stokes vector at (57), it can be observed that the three
parameters {g1, g2, g3} can be considered as the spherical coordinates of a point in a sphere of
radius g0. Figure 14 presents an scheme of this representation. From this figure, it can be
clearly observe which is the effect of the polarization ellipse angles φ and τ, where the
longitude and latitude of the point defining the polarization state are related with 2φ and 2τ.
An interesting aspect to highlight about the Poincaré sphere is the representation of
orthogonal polarization states. Taking into account the expressions presented at (57) and (61)
it can be observed that two orthogonal polarization states are represented by antipodal points
in the Poincaré sphere.

20
What is Polarization?


z$
gE
g3

0 2τ g2 y$

g1 2φ

x$

Figure 14 Poincaré sphere.

Finally, Figure 15 gives the representation of some canonical polarization states within the
Poincaré sphere.
NORTH POLE
LEFT CIRCULAR
POLARISATION z$ LC
NORTHERN HEMISPHERE
LEFT ELLIPTICAL
POLARISATIONS

-45° 0 +45°
EQUATOR y$
LINEAR H
POLARISATIONS
x$

SOUTHERN HEMISPHERE
RIGHT ELLIPTICAL
POLARISATIONS
RC
SOUTH POLE
RIGHT CIRCULAR
POLARISATION

Figure 15 Canonical polarization states represented at the Poincaré sphere.

21
Single vs. Multi polarization sar data

2. SINGLE VS. MULTI POLARIZATION SAR DATA

2.1 Scattering Coefficient vs. Scattering Matrix

In the previous chapter of this document, we dealt with the description and the
characterization of electromagnetic waves. As it was shown, one of the main properties of a
transverse electromagnetic wave is the vectorial nature of the electromagnetic field, which is
called polarimetry.
An electromagnetic wave travels in time and space. In this voyaging through the space, it may
happen that the wave can reach a particular target, then interacting with it, see Figure 1. As a
consequence of this interaction, part of the energy carried by the incident wave is absorbed by
the target itself, whereas the rest is reradiated as a new electromagnetic wave. Due to the
interaction with the target, the properties of the reradiated wave can be different from those of
the incident one. Then, the question which rises at this point is if these changes could be
employed to characterize or identify the target. In particular, we are interested in the changes
concerning the polarization of the wave. In the following, we present the way in which the
interaction of an electromagnetic wave and a given target can be represented.
Incident Wave
r r r jk r
E i (r ) = E0i e( i ) Scattered Wave
r r r jk r
E s (r ) = E0s e( s )

Target

Figure 1 Interaction of an electromagnetic wave and a target.

2.1.1 Single Polarization Image: Scattering Coefficient

Before to define the interaction of electromagnetic waves with the nature, it is necessary to
introduce two important concepts concerning the idea of target, since they will determine the
way in which they shall be characterized. Given a radar configuration as depicted by Figure
1, it may happen the target of interest to be smaller than the coverage of the radar system. In
this situation, we consider the target as an isolated scatterer and from a point of view of power
exchange, this target is characterized by the so-called radar cross section. Nevertheless, we
can find situations in which the target of interest is significantly larger that the coverage
provided by the radar system. In these occasions, it is more convenient to characterize the

1
Single vs. Multi polarization sar data

target independently of his extend. Hence, in these situations, the target is described by the so-
called scattering coefficient.
The most fundamental form to describe the interaction of an electromagnetic wave with a
given target is the so-called radar equation. This equation establishes the relation between the
r
power which the target intercepts from the incident electromagnetic wave E i and the power
r
reradiated by the same target in the form of the scattered wave E s . The radar equation
presents the following form
PG A
Pr = t t
σ r2 (1)
4π Rt 4π Rr
2

where Pr represents the power detected at the receiving system. The term
PG
t t
(2)
4π Rt2
r
is determined by the incident field E i and it consists of its power density expressed in terms
of the properties of the transmitting system. The different terms in (2) are: the transmitted
power Pt, the antenna gain Gt and the distance between the system and the target Rt. On the
contrary, the term
Ar
(3)
4π Rr2
contains the parameters concerning the receiving system: the effective aperture of the
receiving antenna Ar and the distance between the target and the receiving system Rr.
The last term in (1), i.e, σ, determines the effects of the target of interest on the balance of
powers established by the radar equation. Since (2) is a power density, i.e., power par unit
area and (3) is dimensionless, the parameter σ has units of area. Consequently, σ consists of an
effective area which characterizes the target. This parameter determines which amount of
power is intercepted from the density (2) by the target and reradiated. This reradiated power is
finally intercepted by the receiving system (3), according to the distance Rt. An important fact
which arises at this point is the way the target reradiates the intercepted power in a given
direction of the space. In order to be independent of this property, the radar cross section shall
be referenced to and idealized isotropic scatterer. Thus, the radar cross section of an object is
the cross section of an equivalent isotropic scatterer that generates the same scattered power
density as the object in the observed direction
rs 2
E
σ = 4π R 2 r 2 = 4π S
2

i
(4)
E
r2
where E represents the intensity of the electromagnetic field and S is the complex scattering
amplitude of the object. The final value of σ is a function of a large number of parameters
which are difficult to consider individually. A first set of these parameters are concerned with
the imaging system:
• Wave frequency f.

2
Single vs. Multi polarization sar data

• Wave polarization. This dependence is specially considered later.


• Imaging configuration, that is, incident (θi,φi) and scattering (θs,φs) directions.
A second set of parameters are related with the target itself
• Object geometrical structure.
• Object dielectrical properties.
Then, the radar cross section σ is able to characterize the target being imaged for a particular
frequency, and imaging system configuration.
The radar equation, as given by (1), is valid for those cases in which the target of interest is
smaller than the radar coverage, that is, a point target. For those targets presenting an extend
larger than the radar coverage, we need a different model to represent the target. In these
situations, a target is represented as an infinite collection of statistically identical point targets.
Figure 2 presents an scheme of this type of targets.

Figure 2 Interaction of an electromagnetic wave and an extended target.

r
As it can be observed in Figure 2, the resulting scattered field E s results from the coherent
addition of the scattered waves from every one of the independent targets which model the
extended scatterer. In order to express the scattering properties of the extended target
independently of its area extend, we considerer every elementary target as being described by
a differential radar cross section dσ. In order to separate the effects of the target extend, we
consider dσ as the product of the averaged radar cross section per unit area σ0 and the
differential area occupied by the target ds. Then, the differential power received by the
systems due to an elementary scatterer can be written as
PG A
dPr = t t
σ 0 ds r 2 (5)
4π Rt 2
4π Rr
Hence, to find the total power received from the extended target we need to integrate over the
illuminated area A0

3
Single vs. Multi polarization sar data

PG A
Pr = ∫∫ t t
σ 0 r 2 ds (6)
A0
4π Rt 2
4π Rr

It must be noted that the radar equation at (1) represents a deterministic problem, whereas (6)
considers a statistical problem. Eq. (6) represents the average power returned from the
extended target. Hence, the radar cross section per unit area σ0, or simply scattering
coefficient, is the ratio of the statistically averaged scattered power density to the average
incident power density over the surface of the sphere of radius Rr
rs 2
σ E
4π Rr2
σ0 = = r 2 (7)
A0 A0 Ei

The scattering coefficient σ0 is a dimensionless parameter. As in the case of the radar cross
section, the scattering coefficient is employed to characterize the scattered being imaged by
the radar. This characterization is for a particular frequency f, polarization of the incident and
scattered waves and incident (θi,φi) and scattering (θs,φs) directions.

2.1.2 Different Emission-Reception Polarization States

As it has been shown in the previous section, the characterization of a given scatterer by
means of the radar cross section σ or the scattering coefficient σ0 depends also on the
r
polarization of the incident field E i . As one can observe in (4) and (7), these two coefficients
are expressed as a function of the intensity of the incident and scattered fields. Consequently,
σ and σ0 shall be only sensible to the polarization of the incident fields through the effects the
polarization has over the power of the related electromagnetic waves. Hence, if we denote by
p the polarization of the incident field and by q the polarization of the scattered field, we can
define the following polarization dependent radar cross section and scattering coefficient
respectively
rs 2
E qp 2
σ qp = 4π R 2 r 2 = 4π S qp (8)
i
Eqp
rs 2
σ qp E
4π Rr 2 qp
σ qp0 = = ri 2 (9)
A0 A0 E qp

2.1.3 General Case: Scattering Matrix

As it has been shown in the previous two sections, a given target of interest can be
characterized by means of the radar cross section or the scattering coefficient depending on
the nature of the scatterer itself, see (4) and (7). Additionally, in (8) and (9) it has been shown
that these two coefficients depend also on the polarization of the incident and the scattered
electromagnetic fields. A closer look to these expressions reveals that these two coefficients

4
Single vs. Multi polarization sar data

depend on the polarization of the electromagnetic fields only through the power associated
with them. Thus, they do not exploit, explicitly, the vectorial nature of polarized
electromagnetic waves. Consequently, in order to take advantage of the polarization of the
electromagnetic fields, that is, their vectorial nature, the scattering process at the target of
interest must be considered as a function of the electromagnetic fields themselves.
In the previous chapter, it was shown that the polarization of a plane, monochromatic, electric
field could be represented by the so-called Jones vector. Additionally, a set of two orthogonal
Jones vectors form a polarization basis, in which, any polarization state of a given
electromagnetic wave can be expressed. Therefore, given the Jones vectors of the incident and
the scattered waves, E i and E s respectively, the scattering process occurring at the target of
interest is expressed as follows
e − jkr ⎡ ⎤ e − jkr ⎡ S ⊥⊥ S ⊥ // ⎤ i
ES = S E i
= E
r ⎢⎣ S // ⊥ S //// ⎥⎦
(10)
r ⎣ ( ⊥ ,// ) ⎦

where the matrix [S(⊥,//)] receives the name of scattering matrix and the entries of this matrix
Spq, for p, q∈(⊥,//), are the so-called complex scattering coefficients or complex scattering
amplitudes. The diagonals elements of the scattering matrix receive the name of co-polar
terms, since they relate the same polarization for the incident and the scattered fields.
Nevertheless, the off-diagonal elements are known as cross-polar terms as they relate
orthogonal polarization states. Finally, the term e − jkr r takes into account the propagation
effects both, in amplitude and phase. It must be taken into account that the relation expressed
by (10) is only valid for the far field zone, where the planar wave assumption is conisdered
for the incident and the scattered fields. Considering (8), the elements of the scattering matrix
can be related with the radar cross section of a given target as follows
σ qp
p, q ∈ ( ⊥, // )
2
Sqp = (11)

As it can be deduced from the previous equation, the characterization of a given target by
means of the scattering matrix allows the possibility to explore the phase information
provided by the phase of complex scattering coefficients, and no only the intensity or
amplitude. As one can observe, the polarimetric scattering equation presented at (10) involves
the Jones vectors of the incident and the scattered fields, which characterize their polarization
properties in a given coordinates systems. In order to be correct, these two Jones vectors must
be expressed within the same coordinates reference. As a result, the scattering matrix shall be
associated to a particular coordinates system. In (10), we consider the coordinates system
centered at the target. Hence, the basis (⊥,//) refers to a plane respect the coordinates systems
centered in the target to which the fields and the scattering matrix are referred.
The following example demonstrates the importance of the phase parameters. Let’s consider a
trihedral and a dihedral. These two targets present the radar cross section coefficients and
scattering matrices given in Figure 3 in the polarization basis formed with the horizontal and
vertical polarization states, which are parallel to the x̂ and ŷ axis, respectively. As it can be
concluded from Figure 3, the trihedral and the dihedral cannot be differentiated in terms of
the radar cross section coefficients, whereas they are seen as different objects if they are
analyzed by means of the corresponding scattering matrices. The conclusion which can be

5
Single vs. Multi polarization sar data

extracted at this point is that polarimetry opens the door to consider phase measurements to
characterize the targets.
ŷ ŷ

x̂ x̂

Trihedral Dihedral
⎡σ xx σ xy ⎤ ⎡1 0 ⎤ σ
⎡ xx σ xy ⎤ ⎡1 0 ⎤
⎢σ ⎥ = 4π ⎢ ⎥ ⎢σ ⎥ = 4π ⎢ ⎥
⎣ yx σ yy ⎦ ⎣0 1 ⎦ ⎣ yx σ yy ⎦ ⎣0 1 ⎦
⎡1 0 ⎤ ⎡1 0 ⎤
⎡ S( x , y ) ⎤ = ⎢ ⎡ S( x , y ) ⎤ = ⎢
⎣ ⎦ 0 1⎥ ⎣ ⎦ 0 −1⎥
⎣ ⎦ ⎣ ⎦
Figure 3 Trihedral and dihedral polarimetric characterization.

Since the scattering matrix [S(⊥,//)] is employed to characterize a given target, it can be
parameterized as follows
⎡ S⊥ // e ( ⊥// ⊥⊥ ) ⎤
j ϕ −ϕ
− jkr ⎡ S⊥⊥ e jϕ⊥⊥ S ⊥ // e jϕ⊥// ⎤ e− jkr e jϕ⊥⊥ S⊥⊥
⎡ S( ⊥ ,// ) ⎤ = e ⎢ ⎥ = ⎢ ⎥
⎣ ⎦ r jϕ jϕ
r 3 ⎢ S // ⊥ e j (ϕ// ⊥ −ϕ⊥⊥ ) S //// e j (ϕ //// −ϕ⊥⊥ ) ⎥
⎢⎣ S // ⊥ e // ⊥ S //// e //// ⎥⎦ 1424 ⎣
(12)
Absolute Phase Term 144444 42444444 3⎦
Relative Scattering Matrix

The absolute phase term in (12) is not considered as an independent parameter since it
presents an arbitrary value due to its dependence on the distance between the radar and the
target. Consequently, it is assumed that the scattering matrix can be parameterized by 7
independent parameters: the amplitudes {|S⊥⊥|, |S⊥//|, |S//⊥|, |S////|} and the relative phases {(ϕ⊥//-
ϕ⊥⊥), (ϕ//⊥-ϕ⊥⊥),(ϕ////-ϕ⊥⊥)}. As a conclusion, a given target of interest is determined by 7
independent parameters in the most general case and an absolute value.
It is important, at this point, to analyze some particular aspects about the definition of the
matrix [S(⊥,//)] and the relation about the different coordinates systems which can be defined to
describe the scattering process characterized in (10).
As it was already highlighted in the previous two sections of this chapter, the radar cross
section and the scattering coefficients depend on the direction of the incident and the scattered
waves. When considering the matrix [S(⊥,//)], the analysis of this dependence is of extreme
importance since it also involves the definition of the polarization of the incident and the
scattered fields. Since (10) considers the polarized electromagnetic waves themselves, it is
mandatory to assume a frame in which the polarization is defined. There exist two principal
conventions concerning the framework where the polarimetric scattering process is
considered: Forward Scatter Alignment (FSA) and Backscatter Alignment (BSA), see Figure
4. In both cases, the electric fields of the incident and the scattered waves are expressed in
local coordinates systems centered on the transmitting and receiving antennas, respectively.

6
Single vs. Multi polarization sar data

All coordinate systems are defined in terms if a global coordinate system centered inside the
target of interest.

(a)

(b)
(c)
Figure 4 Reference frameworks: (a) FSA, (b) Bistatic BSA and (c) Monostatic BSA.

The FSA convention, see Figure 4a, also called wave-oriented since it is defined relative to
the propagating wave, is normally considered in bistatic problems, that it, in those
configurations in which the transmitter and the receiver are not located at the same spatial
position.
The bistatic BSA convention framework, see Figure 4b, is defined, on the contrary, respect to
the radar antennas in accordance with the IEEE standard. The advantage of the BSA
convention is that for a monostatic configuration, also called backscattering configuration,
that is, when the transmitting and receiving antennas are collocated, the coordinated systems
of the two antennas coincide, see Figure 4c. This configuration is preferred in the radar
polarimetry community. In the monostatic case, the scattering matrix in the FSA convention,
[S(⊥,//)]FSA, can be related with the same matrix referenced to the monostatic BSA convention
[S(⊥,//)]BSA as follows
⎡ −1 0⎤
⎡ S( ⊥,// ) ⎤ = ⎢ ⎥ ⎡⎣ S( ⊥,// ) ⎤⎦ FSA (13)
⎣ ⎦ BSA
⎣ 0 1 ⎦

As it has been mentioned previously, in the radar polarimetry community, the monostatic
BSA convention (backscattering) is considered as the framework to characterize the scattering

7
Single vs. Multi polarization sar data

process. The reason to select this configuration is due to fact that the majority of the existing
polarimetric radar systems operate with the same antenna for transmission and reception. One
important property of this configuration, for reciprocal targets, is reciprocity, which states that
[ S⊥ // ]BSA = [ S// ⊥ ]BSA (14)
[ S⊥ // ]FSA = − [ S// ⊥ ]FSA (15)

Then, the formalization of the scattering process given by (10), in the monostatic case under
the BSA convention, reduces to
e − jkr ⎡ ⎤ e − jkr ⎡ S ⊥⊥ S ⊥ // ⎤ i
ES = S E i
= E
r ⎢⎣ S ⊥ // S //// ⎥⎦
(16)
r ⎣ ( ⊥ ,// ) ⎦

In the same sense, equation (12) takes the form


⎡ S⊥ // e ( ⊥// ⊥⊥ ) ⎤
j ϕ −ϕ
− jkr ⎡ S⊥⊥ e jϕ⊥⊥ S ⊥ // e jϕ⊥// ⎤ e− jkr e jϕ⊥⊥ S⊥⊥
⎡ S( ⊥ ,// ) ⎤ = e ⎢ ⎥= ⎢ ⎥
⎣ ⎦ r jϕ ⊥ // jϕ
r 3 ⎢ S ⊥ // e j (ϕ⊥// −ϕ⊥⊥ ) S //// e j (ϕ //// −ϕ⊥⊥ ) ⎥
⎣⎢ S⊥ // e S //// e //// ⎦⎥ 1424 (17)

Absolute Phase Term 144444 42444444 3⎦
Relative Scattering Matrix

The main consequence of (17) is that in the backscattering direction, a given target is no
longer characterized by 7 independent parameters, but by 5. These are: the amplitudes {|S⊥⊥|,
|S⊥//|, |S//⊥|} and the relative phases {(ϕ⊥//-ϕ⊥⊥),(ϕ////-ϕ⊥⊥)} and one additional absolute phase.

2.2 Additional Information

2.2.1 Span: Joint Intensity Information

As it was indicated at the beginning of Section 2.1.1, a central parameter when considering
the scattering process occurring at a given target consists of the scattered power. For single
polarization systems, the scattered power is determined by means of the radar cross section or
the scattering coefficient. Nevertheless, a polarimetric radar has to be considered as a multi
channel system. Consequently, in order to determine the scattered power, it is necessary to
consider all the data channels, that is, all the elements of the scattering matrix. The total
scattered power, in the case of a polarimetric radar system is know as span, being defined in
the most general case as

( )
SPAN ⎡⎣ S( ⊥,// ) ⎤⎦ = trace ⎛⎜ ⎡⎣ S( ⊥,// ) ⎤⎦ ⎡⎣ S( ⊥ ,// ) ⎤⎦ ⎞⎟ = S⊥⊥ + S⊥ // + S // ⊥ + S ////
T* 2 2 2 2
(18)
⎝ ⎠
where trace(.) represents the trace of a matrix, consisting of the addition of the elements of the
principal diagonal. In the backscattering case, due to the reciprocity theorem, the span reduces
to

( )
SPAN ⎡⎣ S( ⊥ ,// ) ⎤⎦ = S ⊥⊥ + 2 S ⊥ // + S ////
2 2 2
(19)

8
Single vs. Multi polarization sar data

The main property of the span is that it is polarimetrically invariable, that is, it does not
depend on the polarization basis employed to describe the polarization of the electromagnetic
waves.
Now, if we consider the definition (11) into (19) we get

( )
SPAN ⎡⎣ S( ⊥ ,// ) ⎤⎦ = 4π (σ ⊥⊥ + 2σ ⊥ // + σ //// ) (20)

|Shh| |Shv| |Svv|


(dB) (dB) (dB)
-30dB -15dB 0dB

Span
Figure 5 Intensities of the elements of the scattering matrix measured in the basis (h,v) and the resulting span.

Therefore, the span presents the same limitations as the radar cross section in order to
represent the polarimetric information contained in the scattering matrix, that is, the important

9
Single vs. Multi polarization sar data

information provided by the measurement of the relative phases is completely lost.


Additionally, since it is polarimetrically invariable, it does not contain polarimetric
information at all.
One of the main applications of the span is to present the multidimensional information
provided by the scattering matrix in a single image. But, as shown in the previous paragraph,
this image does not contain any polarimetric information at all. Figure 5 presents an example
of the span image.

2.2.2 Lexicographic Color-Coded Representation

In the previous section, we showed that it was possible to present part of the information
contained in the scattering matrix [S(⊥,//)] in a single image by means of the span of the
scattering matrix, but at the expense to loose all the polarimetric information. In what it
follows, we introduce two alternatives to represent part of the multidimensional information
provided by the scattering matrix in a single image. These two alternatives are based on the
decomposition RBG of color images.

2.2.2.1 RBG color coding


The RGB color model is an additive color model in which red, green, and blue light are
combined in various ways to create other colors. The very idea for the model itself and the
abbreviation "RGB" come from the three primary colors in additive light models. The concept
to retain is that we can create any color of the spectra by combining the three primary colors:
red, green and blue light, that is, color can be considered as a three dimensional space. Figure
6 presents the color space, coded in the RGB base.
Blue

White

Green
Red

Figure 6 RBG color cube.

Consequently, three different parameters can be represented in a single color image by coding
each initial parameters as one color of the RGB space. Therefore, the color in the resulting
image can be interpreted in terms of the three selectedparameters.

2.2.2.2 Polarimetric information coded as a color


The color representation presented above is now considered to represent the polarimetric
information provided by the scattering matrix [S(⊥,//)]. The lexicographic color-coded

10
Single vs. Multi polarization sar data

representation considers the intensity of the elements of [S(⊥,//)], in the backscattering


direction, as the dimensions of the color space. A possible codification could be
S⊥⊥ → Red
2

2 S⊥ // → Green
2
(21)
S //// → Blue
2

despite any other codification is also valuable. Again, it must be mention that, since we are
considering the intensity of the different elements of the scattering matrix, this codification is
not able to reflect the information that can be provided by the relative phases of elements.

2.2.3 Importance of Relative Phase Terms: Pauli Color-Coded


Representation

In the following we present a color-coded representation of the information provided by the


scattering matrix which takes into account the information provided by the relatives phases of
the different entries of the matrix. Any 2×2 symmetric matrix can be parameterized as a
function of three parameters {a, b, c} as follows
⎡ S11 S12 ⎤ ⎡ a + b c ⎤
⎢ ⎥ =⎢
a − b ⎥⎦
(22)
⎣ S12 S 22 ⎦ ⎣ c

If we now consider the representation of the scattering matrix in the backscattering case
presented in (16), we can obtain the following equivalences
S ⊥⊥ + S ////
a=
2
S − S //// (23)
b = ⊥⊥
2
c = 2 S ⊥ //
The Pauli color-coded representation considers the codification in color of the intensity of the
parameters in (23) as follows
b → Red
2

c → Green
2
(24)
a → Blue
2

Finally, Figure 8a presents an example or the Pauli Color coded representation of the image
channels presented in Figure 5.

2.3 Polarization Basis and Polarimetric Signature

11
Single vs. Multi polarization sar data

2.3.1 Acquisition According to Orthogonal Jones Vectors: Polarization


Basis

As it was demonstrated in the preceding chapter, the Jones vector corresponding to a given
electromagnetic field can be expressed in an orthonormal basis (⊥,//) as
E = E⊥ ⊥ˆ + E// //ˆ (25)

where the unitary vectors ⊥ ˆ and //ˆ form the so-called polarization basis. If now, we consider
the polarimetric scattering equation, see (10), the Jones vectors corresponding to the incident
and the scattered waves, E i and E s respectively, are expressed in the orthonormal basis
(⊥,//). Consequently, the scattering matrix [S(⊥,//)] specifies how the target transforms the
polarization of the incident field to the scattered field when both are expressed in the basis
(⊥,//). As a result, it is said that the scattering matrix [S(⊥,//)] has been acquired with respect to
the polarization basis (⊥,//).

2.3.2 Acquisition in ONE Polarization Basis: Representation in ANY


Polarization Basis

The scattering properties of a given target, as demonstrated, are contained within the
scattering matrix [S(⊥,//)], which, as shown previously, is measured in the particular
polarization basis (⊥,//). Since, there exist an infinite number of orthonormal polarization
bases, the question raising at this point is that whether it is possible or not to infer the
polarimetric properties of the given target in any polarization basis from the response
measured at a particular basis, for instance (⊥,//). This question presents an affirmative
answer. The possibility to synthesize any polarimetric response of a given target from its
measurement in a particular orthonormal basis represents the most important property of
polarimetric systems in comparison with single-polarization systems. The most important
consequence of this process is that the amount of information about a given scatterer can be
increased, allowing a better characterization and study. This polarization synthesis process is
based on the concept of change of polarization basis presented in the previous chapter. In
what it follows, we shall consider the polarization synthesis process in the backscattering
direction or monostatic configuration of a radar system.

12
Single vs. Multi polarization sar data

Target
ri
Incident Wave
r r i ( jki r ) [S ]
E (r ) = E0e
kˆi
kˆs = − kˆi Scattered Wave
rs r r s ( jk r ) r − jk r
E (r ) = E0 e s
= E0s e( i )

TRANSMITTER
RECEIVER
Figure 7 Polarization synthesis process in the backscattering direction.

Before to describe the polarization synthesis process in the backscattering direction, it is


necessary to analyze the scattering process given by (10) with respect to the direction of
propagation of the incident and the scattered fields. As observed in Figure 7, the incident
field propagates in the direction given by the unitary vector kˆi , whereas the scattered one
propagates in the opposite direction, given by kˆ . Consequently, this difference in the
s

propagation direction must be taken into account when defining the polarization state of the
wave. Given a Jones vector propagating in the direction k̂ , the Jones vector of the same wave,
but which propagates in the direction −k̂ is obtained as

()
kˆ → − kˆ E kˆ = E * − kˆ ( ) (26)

From now, we shall not consider the scattering matrix referred to the coordinates system
centered at the target, but in the axis coordinate centered in the transmitting/receiving system.
Consider a polarimetric radar system as depicted by Figure 7, which transmits the
electromagnetic waves in the following orthonormal basis (A, A⊥). In this particular basis, the
incident and scattered fields are related by the scattering matrix as follows
⎡ S AA S AA⊥ ⎤ i
E(sA, A⊥ ) = ⎢ E = ⎡S ⎤ Ei
S A⊥ A⊥ ⎥⎦ ( A, A⊥ ) ⎣ ( A, A⊥ ) ⎦ ( A, A⊥ )
(27)
⎣ S AA⊥
As shown in the previous chapter, given the Jones vector measured in a particular basis, for
instance (A, A⊥), it is possible to derive it in any other polarization basis (B, B⊥) by means of
the following simple mathematical transformation
E( B , B⊥ ) = ⎡⎣U ( A, A⊥ ) a ( B , B⊥ ) ⎤⎦ E( A, A⊥ ) (28)

Then, if we consider (28) for the incident Jones vector in (27), we get
E(iB , B⊥ ) = ⎡⎣U ( A, A⊥ ) a ( B , B⊥ ) ⎤⎦ E(i A, A⊥ ) (29)

13
Single vs. Multi polarization sar data

E(i A, A⊥ ) = ⎡⎣U ( B , B⊥ ) a ( A, A⊥ ) ⎤⎦ E(iB , B⊥ ) (30)

In order to apply the transformation basis procedure to the scattered field E(sA, A⊥ ) we need to
consider that it propagates in the opposite direction as the incident field E(i A, A⊥ ) . Hence, in
order to consider both fields in the same frame of reference, that is, their polarizations are
referred to the same coordinate system we need to consider (26). As a result, the
transformation basis procedure applies to the scattered field as follows
*
E(sB , B⊥ ) = ⎡⎣U ( A, A⊥ ) a ( B , B⊥ ) ⎤⎦ E(sA, A⊥ ) (31)
*
E(sA, A⊥ ) = ⎡⎣U ( B , B⊥ ) a ( A, A⊥ ) ⎤⎦ E(sB , B⊥ ) (32)

Now, introducing the results of (30) and (32) into (27) we get
*
⎡U ( B , B ) a ( A, A ) ⎤ E(sB , B ) = ⎡ S( A, A ) ⎤ ⎡U ( B , B ) a ( A, A ) ⎤ E(iB , B ) (33)
⎣ ⊥ ⊥ ⎦ ⊥ ⎣ ⊥ ⎦⎣ ⊥ ⊥ ⎦ ⊥

−1
E(sB , B⊥ ) = ⎜⎛ ⎣⎡U ( B , B⊥ ) a ( A, A⊥ ) ⎦⎤ ⎟⎞ ⎣⎡ S( A, A⊥ ) ⎦⎤ ⎣⎡U ( B , B⊥ ) a ( A, A⊥ ) ⎦⎤ E(iB , B⊥ )
*
(34)
⎝ ⎠

Since the transformation matrix ⎡⎣U ( A, A⊥ ) a ( B , B⊥ ) ⎤⎦ is unitary, i.e., [U ] = [U ] , we get


−1 *T

(a) |Shh+ Svv| |Shh- Svv| |Shv|

14
Single vs. Multi polarization sar data

(b) |Saa+ Sbb| |Saa- Sbb| |Sab| (c) |Sll+ Srr| |Sll- Srr| |Slr|
Figure 8 Polarization synthesis. (a) Measured response at the linear polarization basis (h,v). (b)
Synthesis of the response in the linear basis rotated 45 degree (a,b). (c) Synthesis in the circular
polarization basis (l,r).

T
E(sB , B⊥ ) = ⎡⎣U ( B , B⊥ ) a ( A, A⊥ ) ⎤⎦ ⎡⎣ S( A, A⊥ ) ⎤⎦ ⎡⎣U ( B , B⊥ ) a ( A, A⊥ ) ⎤⎦ E(iB , B⊥ ) (35)

In the previous equation, we can clearly identify, then,


T
⎡ S( B , B ) ⎤ = ⎡U ( B , B ) a ( A, A ) ⎤ ⎡ S( A, A ) ⎤ ⎡U ( B , B ) a ( A, A ) ⎤ (36)
⎣ ⊥ ⎦ ⎣ ⊥ ⊥ ⎦ ⎣ ⊥ ⎦⎣ ⊥ ⊥ ⎦

The transformation expressed in (36) receives the name of con-similarity transformation. This
transformation allows to synthesize the scattering matrix in an arbitrary basis (B, B⊥), from its
measure in the basis (A, A⊥).
If instead of an arbitrary basis, we consider the linear orthogonal polarization basis ( xˆ, yˆ ) , the
transformation matrix from ( xˆ, yˆ ) to an arbitrary elliptical basis can be parameterized as

⎡U ( xˆ , yˆ ) a ( A , A ) ⎤ = ⎡U (φ ) ⎤ ⎡U (τ ) ⎤ ⎡U (α )⎤
⎣ ⊥ // ⎦ ⎣ ⎦⎣ ⎦⎣ ⎦
⎡cos φ − sin φ ⎤ ⎡ cosτ j sin τ ⎤ ⎡e− jα 0 ⎤ (37)
=⎢ ⎥⎢ ⎥⎢ ⎥
⎣ sin φ cos φ ⎦ ⎣ j sin τ cosτ ⎦ ⎣ 0 e− jα ⎦

Figure 8 presents an example of the application of the con-similarity transformation given at


(36) to synthesize the polarimetric response at different polarization basis. The polarimetric
information is represented by means of the Pauli color-coded representation. The original
polarimetric set, presented in Figure 8a is obtained in the linear polarization basis (h,v),
where h stands for the horizontal polarization and v for the vertical polarization. Using (37)
the response to two different polarization basis is synthesized. Figure 8b presents the
response to the orthonormal basis (a,b) where a indicates the linear polarization at 45 degrees

15
Single vs. Multi polarization sar data

and b the linear polarization at -45 degrees. Finally, Figure 8c shows the response to the
circular polarization basis (l,r), where l refers to the left circular polarization and r to the right
circular polarization.

2.3.3 Complete Polarimetric Characterization of Each Pixel: Polarization


Signature

In the previous section, we have employed the powerful con-similarity transformation to


synthesize the polarimetric response just in two alternative polarization bases: the 45 degrees
rotated linear basis (a, b) and the circular polarization basis (l, r). Nevertheless, this section
aims to exploit all the information which can be extracted from this transformation through
the exploration of the space of all possible polarization bases. The objective of this process is
to present a new way to characterize a given scatterer.

Figure 9 presents the scheme of a general polarimetric radar employed to measure a given
target, characterized by a particular scattering matrix. In this scheme, the Jones vector A
refers to the polarization emitted by the transmitting system and B indicates the Jones vector
containing the polarization with respect to which the receiving antenna receives the scattered
field E . In this context, the power received at the receiving system is obtained by means of
2
P = α BT E (38)

where α contains all those terms which do not depend on the polarization of the antennas A
and B . Using the characterization of the target by means of the scattering matrix, it is
possible to rewrite (38) as follows
Target
[S ]
E
A
B

TRANSMITTER RECEIVER
Figure 9 General polarimetric system configuration.

PAB = α BT [ S ] A
2
(39)

In the frame of (39), two different power measurements are defined:

16
Single vs. Multi polarization sar data

• The co-polarized power. In this configuration, we assume that the transmitting and
receiving antennas are characterized by the same polarization state. Hence, (39) turns
out to be

PCO = α AT [ S ] A
2
(40)

• The cross-polarized power. In this case, it is assumed that the receiver antenna
receives with the orthogonal polarization of the transmitting system. Thus, we write
(39) as

PX = α A⊥T [ S ] A
2
(41)

The co- and cross-polarized powers can be generates synthetically, since, given the linear
polarization basis ( xˆ, yˆ ) , we can write an arbitrary polarization state

⎡cos φ − sin φ ⎤ ⎡ cosτ j sin τ ⎤


A=⎢ xˆ
cos φ ⎥⎦ ⎢⎣ j sin τ cosτ ⎥⎦
(42)
⎣ sin φ
and
⎡ ⎛ π⎞ ⎛ π ⎞⎤
⎢cos ⎜ φ + 2 ⎟ − sin ⎜ φ + 2 ⎟ ⎥ ⎡ cos ( −τ ) j sin ( −τ ) ⎤
⎝ ⎠ ⎝ ⎠⎥
A=⎢ ⎢ ⎥ xˆ (43)
⎢ ⎛ π⎞ ⎛ π ⎞ ⎥ ⎣ j sin ( −τ ) cos ( −τ ) ⎦
⎢ sin ⎜ φ + ⎟ cos ⎜ φ + ⎟ ⎥
⎣ ⎝ 2⎠ ⎝ 2⎠⎦

Consequently, we can obtain the co- and cross-polarized powers, PCO and PX, as a function of
the angles φ and τ, that is, the ellipse orientation and the ellipse aperture which define the
polarization ellipse of an electromagnetic wave. We have not considered here the phase term
α.
Section 2.5 presents the polarization signatures for the canonical scattering mechanisms.

2.4 Optimal Polarization States and The Poincaré Sphere

2.4.1 Optimal Polarization States: Deduction Oriented Parameterization

As it has been shown in the previous section of this chapter, the received power at the receiver
system varies according to the polarization of the transmitted and incident waves, and on the
characteristics of the target under study through its scattering matrix [S], (39). In the case of
the polarization signature, we particularized the study of the received power to two special
cases. On the one hand, we have considered the situation in which the transmitting and
receiving systems employ the same polarization. On the other hand, the case in which both
systems use orthogonal polarizations has been analyzed. The polarization signatures for the
co-polar and cross-polar configurations, given by (40) and (41) respectively, explore all the

17
Single vs. Multi polarization sar data

space of possible polarizations. Nevertheless, the characterization of a particular scatterer can


be also performed by studying the so culled characteristic polarization states. These states are
defined as such wave polarization states giving as a result maximum or minimum received
powers.
Given a particular polarization state for an electromagnetic wave, the corresponding Jones
vector can be represented in terms of the polarization ratio ρ as presented in the previous
chapter. Then for the Jones vector A
e − jα
⎡1⎤
A= ⎢ ⎥
1 + ρ ⎣ρ ⎦
2 (44)

The Jones vector of the orthogonal state A⊥ is


⎡− ρ * ⎤
e + jα
A⊥ =
2 ⎢ ⎥ (45)
1+ ρ ⎣ 1 ⎦

If now, we consider the scattering matrix as


⎡S S XY ⎤
[ S ] = ⎢ S XX SYY ⎥⎦
(46)
⎣ XY

the co-polar PCO and cross-polar PX can be respectively written as follows


2
PCO = α S XX + 2 ρ S XY + ρ 2 SYY (47)

( )S
2
PX = α ρ * S XX − 1 − ρ + ρ SYY
2
XY (48)

Then, in order to derive the characteristic polarization states, we need to calculate the
following derivatives
∂PCO
=0 (49)
∂ρ
∂PX
=0 (50)
∂ρ
As one can observe, (47) and (48) consist of bilinear forms. Consequently, the processes
presented at (49) and (50) to derive the characteristic polarization states shall present two
solutions. The next two sections analyze the characteristics states resulting from (49) and (50).

2.4.1.1 Characteristics Polarization States in the Co-polar Configuration


The co-polar power PCO presents two polarizations states resulting in maximum received
power. These two states are called COPOL MAX and are represented by the pair of
polarization states (K,L). Then,
PKK ⇒ Global Maximum (51)
PLL ⇒ Local Maximum (52)

The COPOL MAX polarizations states are orthogonal.

18
Single vs. Multi polarization sar data

Additionally, the co-polar power PCO has two characteristic polarization states for which the
received power is zero. This pair of polarization states are named COPOL NULLS and are
represented by (O1,O2). These polarization states give as a result
POO11 = 0 (53)
P O2
O2 =0 (54)

The main characteristic of these states is that they are not mutually orthogonal.

2.4.1.2 Characteristics Polarization States in the Cross-polar Configuration


In the case of the cross-polar power PX, it is possible to find three pairs of orthogonal
polarization states which result into characteristic polarization states.
The first pair of polarization states results into maximum received power at the receiver
system. This pair of polarization states receives de name of XPOL MAX and are represented
by the pair (C1,C2). These states result into
PCC11⊥ ⇒ Global Maximum (55)
PCC22⊥ ⇒ Local Maximum (56)

The states (C1,C2) are mutually orthogonal.


The second pair of polarization states gives null revived power. This set of polarization states
is known as XPOL NULL and are represented by (X1,X2), resulting into
PXX11⊥ = 0 (57)
PXX22⊥ = 0 (58)

As first established by Kennaugh, the XPOL NULL and the COPOL MAX represent the same
pair of polarization states. Consequently, (X1,X2) consist also in orthogonal polarization states.
Finally, in the case of the cross-polar power, it is possible to define a third pair of
polarizations states which can be taken as characteristics. These polarization states result into
a minimum received power and are know as XPOL SADDLE. They are represented by
(D1,D2) and
PDD11⊥ ⇒ Global Minimum (59)
P D2⊥
D2 ⇒ Local Minimum (60)

Again, (D1,D2) consist of mutually orthogonal polarization states.

2.4.2 Representation on The Poincare Sphere

In the former chapter, we presented the representation of a given polarization state into the so-
called Poincare Sphere. In the following, we are going to make use of this representation to
show that the characteristic polarization states pairs given previously present a particular
shape into the Poincare sphere. Figure 10 presents the representation of the characteristic
polarization states within the Poincare sphere.

19
Single vs. Multi polarization sar data

V C1
O1

LX 2

D1 D2
U
O2
X1
K

C2
Figure 10 Representation in the Poincare Sphere of the five pairs of characteristic polarization
states.

In the previous figure, it can be clearly observed that all the pairs except (O1,O2) represent
pairs of orthogonal polarization states since they consist of antipodal points of the Poincare
Sphere. Additionally, it is possible to recognize that the orthogonal pairs of polarization states
XPOL NULL (equal to the COPOL MAX) and XPOL MAX define a plane within the Poincare
sphere which also contains the non-orthogonal pair of COPOL NULLS.

2.4.3 Polarization Fork

In the previous section, we have considered the representation of the five pairs of
characteristic polarization states within the Poincare Sphere. Huynen introduced the concept
of Polarization Fork by considering only the representation of the pairs (K,L) or (X1,X2) and
the non-orthogonal set (O1,O2). Figure 11 gives the representation of the polarization fork. In
order to better understand this concept it is helpful to consider the parameterization of the
scattering matrix proposed by Hyunen.
Given an arbitrary scattering matrix, we want to find the rotation matrix [U], in the sense of
(37), which gives as a result a diagonal scattering matrix. In general, the diagonalization of a
matrix is performed by the standard eigen-decomposition of matrices. Nevertheless, in the
backscattering case under the BSA convention, we have to be aware that we are dealing with
electromagnetic waves traveling in opposite directions. Consequently, the diagonalization of
the scattering matrix must be done according to the con-similarity transformation, as
presented in Section 2.3.2. In this particular situation, the diagonalization of the scattering
matrix is done with the pseudo eigen-decomposition, that is,

20
Single vs. Multi polarization sar data

[S ] X = λ X * (61)

where λ refers to the pseudo-eigenvalues of [S] and X to the corresponding pseudo-


eigenvectors. Since the scattering matrix consists of a 2×2 complex symmetric matrix in the
backscattering case under the BSA convention, (61) presents two solutions. As demonstrated
by Huynen, the pseudo-eigenvectors of the scattering matrix correspond to the XPOL NULL,
i.e., (X1,X2). Since these two polarization states are orthogonal, it is only necessary to specify
one of them. Therefore, the con-similarity transformation to diagonalize the scattering matrix
is
T
⎡ S( X , X ) ⎤ = ⎡U ( X , X ) a ( X ,Y ) ⎤ ⎡ S( X ,Y ) ⎤ ⎡U ( X , X ) a ( X ,Y ) ⎤ (62)
⎣ 1 2 ⎦ ⎣ 1 2 ⎦ ⎣ ⎦⎣ 1 2 ⎦
where, the unitary matrix ⎡⎣U ( X1 , X 2 ) a ( X ,Y ) ⎤⎦ takes the form

⎡U ( X , X ) a ( X ,Y ) ⎤ = [ X 1 X2] (63)
⎣ 1 2 ⎦
and it can be parameterized as given by (37), i.e., it can be expressed in terms of the three
angles {φ, τ, α}. Huynen parameterized the resulting diagonal matrix as follows
⎡1 0 ⎤ jξ
⎡ S( X , X ) ⎤ = m ⎢ 2 ⎥
e (64)
⎣ 1 2 ⎦
⎣0 tan γ ⎦
where m represents the maximum polarization and the angle γ is the target ship angle. Finally
ξ consists of an absolute phase. As a result, Huynen parameterized the scattering matrix in
terms of the 5 Euler parameters {m, γ, φ, τ, α} and the absolute phase ξ.
As is can be observed in Figure 11, the angle γ determines the angle between the line formed
by the states (X1,X2) and (O1,O2), both considered to be over the same section or plane of the
Poincare sphere. Additionally, the three Euler parameters {φ, τ, α} consist of the rotations to
be done over every one of the three axis defining the space in order to bring the polarization
state given by x, corresponding to a linear polarized wave, to the polarization state given by
X1.

21
Single vs. Multi polarization sar data


z$

X1
O2
0 y$

x$ X2

O1

Figure 11 Polarization fork.

2.5 Canonical Scattering Mechanisms

A real target presents always a complex scattering response as a consequence of its complex
geometrical structure and its reflectivity properties. Consequently, the interpretation of this
response is obscure. As it shall be presented in a future chapter, a possible solution to
interpretate this response is to decompose it into the response of canonical mechanisms. These
scattering mechanism are characterized by presenting a simple scattering response. This
section presents the list of this canonical scattering mechanisms and its characterization by the
scattering matrix and the polarimetric signatures. The scattering matrix shall be presented in
three different orthogonal polarimetric bases:
• Linear polarization basis (h,v) where h and v stand for the horizontal and vertical
polarizations respectively.
• Linear rotated basis (a,b) where a and b stand for the 45 degrees linear and the -45
degrees linear polarizations respectively.
• Circular polarization basis (l,r) where l and r stand for the left circular and the right
circular polarizations respectively.

22
Single vs. Multi polarization sar data

2.5.1 Sphere, Plane, Trihedral


vˆ t

hˆ t

Figure 12 Trihedral orientated horizontally.

Scattering matrices of a sphere, a plane or a trihedral in the three polarization basis:


Linear rotated polarization
Linear polarization basis (h,v) Circular polarization basis (l,r)
basis (a,b)
⎡1 0 ⎤ ⎡1 0 ⎤ ⎡0 j⎤
[ S ] = ⎢0 ⎥ [ S ] = ⎢0 ⎥ [S ] = ⎢ j 0 ⎥⎦
⎣ 1⎦ ⎣ 1⎦ ⎣
Co-polar and cross-polar signatures of a sphere, a plane or a trihedral:

PCO PX

Figure 13 Co-polar and Cross-polar signatures corresponding to a sphere, plane or trihedral.

23
Single vs. Multi polarization sar data

2.5.2 Dipole

2.5.2.1 Dipole orientated in the vertical axis


vˆ t

hˆ t

Figure 14 Dipole orientated in the vertical axis.

Scattering matrices of a dipole orientated in the vertical axis in the three polarization basis:
Linear rotated polarization
Linear polarization basis (h,v) Circular polarization basis (l,r)
basis (a,b)
⎡1 0 ⎤ 1⎡1 −1⎤ 1⎡ 1 − j⎤
[ S ] = ⎢0 ⎥ [ S ] = 2 ⎢−1 1 ⎥⎦
[ S ] = 2 ⎢− j 1 ⎥⎦
⎣ 0⎦ ⎣ ⎣
Co-polar and cross-polar signatures of a dipole orientated in the vertical axis:

PCO PX

Figure 15 Co-polar and Cross-polar signatures corresponding to a dipole oriented vertically.

24
Single vs. Multi polarization sar data

2.5.2.2 Oriented dipole


In this case, we consider the dipole oriented with an angle φ:
vˆ t

l
φ hˆ t

Figure 16 Dipole oriented with a φ angle.

Scattering matrices of a dipole oriented with an angle φ in the three polarization basis:
Linear rotated polarization
Linear polarization basis (h,v) Circular polarization basis (l,r)
basis (a,b)
⎡ 1 ⎤ ⎡1 1 ⎤
⎢ cos φ sin 2φ ⎥ ⎢ + cos φ sin φ − cos 2 φ ⎥
2
1 ⎡e j 2φ −j ⎤
[S ] = ⎢ 1 2
⎥ [S ] = ⎢ 2 1 2
⎥ [S ] = 2 ⎢ ⎥
⎢ sin 2φ ⎣−j e − j 2φ ⎦
sin 2 φ ⎥ ⎢ − cos φ sin φ ⎥
1
− cos 2 φ
⎣⎢ 2 ⎦⎥ ⎣⎢ 2 2 ⎦⎥

Co-polar and cross-polar signatures of a dipole oriented with an angle φ:

PCO PX

Figure 17 Co-polar and Cross-polar signatures corresponding to a dipole oriented with a φ angle.

25
Single vs. Multi polarization sar data

2.5.3 Dihedral
vˆ t

hˆ t

Figure 18 Dihedral oriented in the horizontal axis.

Scattering matrices of a dihedral oriented in the horizontal axis in the three polarization basis:
Linear rotated polarization
Linear polarization basis (h,v) Circular polarization basis (l,r)
basis (a,b)
⎡1 0⎤ ⎡0 −1⎤ ⎡1 0 ⎤
[ S ] = ⎢0 −1⎥⎦
[ S ] = ⎢−1 0 ⎥⎦
[ S ] = ⎢0 ⎥
⎣ ⎣ ⎣ 1⎦
Scattering matrices of a dihedral oriented with an angle φ in the three polarization basis:
Linear rotated polarization
Linear polarization basis (h,v) Circular polarization basis (l,r)
basis (a,b)
⎡ cos 2φ sin 2φ ⎤ ⎡ sin 2φ − cos 2φ ⎤ ⎡ e j 2φ 0 ⎤
[ S ] = ⎢ sin 2φ − cos 2φ ⎥⎦
[ S ] = ⎢ − cos 2φ − sin 2φ ⎥⎦
[ ] ⎢
S = ⎥
− j 2φ
⎣ ⎣ ⎣ 0 e ⎦
Co-polar and cross-polar signatures of a dihedral oriented in the horizontal axis:

PCO PX

Figure 19 Co-polar and Cross-polar signatures corresponding to a dihedral oriented in the


horizontal axis.

26
Single vs. Multi polarization sar data

2.5.4 Right Helix


vˆ t

hˆ t

Figure 20 Right helix.

Scattering matrices of a right helix oriented with an angle φ in the three polarization basis:
Linear rotated polarization
Linear polarization basis (h,v) Circular polarization basis (l,r)
basis (a,b)
1 ⎡1 − j⎤ 1 ⎡ − j −1⎤ ⎡0 0 ⎤
[ S ] = 2 e − j 2φ ⎢ − j −1⎥⎦
[ S ] = 2 e− j 2φ ⎢ −1 j ⎥⎦
[ S ] = ⎢0 −e− j 2φ ⎥⎦
⎣ ⎣ ⎣
Co-polar and cross-polar signatures of a right helix oriented at 0 degrees:

PCO PX

Figure 21 Co-polar and Cross-polar signatures corresponding to a right helix.

27
Single vs. Multi polarization sar data

2.5.5 Left Helix


vˆ t

hˆ t

Figure 22 Co-polar and Cross-polar signatures corresponding to a left helix.

Scattering matrices of a left helix oriented with an angle φ in the three polarization basis:
Linear rotated polarization
Linear polarization basis (h,v) Circular polarization basis (l,r)
basis (a,b)
1 ⎡1 j⎤ 1 ⎡ j −1 ⎤ ⎡ e − j 2φ 0⎤
[ S ] = 2 e− j 2φ ⎢ j −1⎥⎦
[ S ] = 2 e− j 2φ ⎢ −1 − j ⎥⎦
[S ] = ⎢ ⎥
⎣ ⎣ ⎣ 0 0⎦

Co-polar and cross-polar signatures of a left helix oriented at 0 degrees:

PCO PX

Figure 23 Co-polar and Cross-polar signatures corresponding to a left helix.

28
Speckle Filtering

3. SPECKLE FILTERING

3.1 Need for speckle filtering

Unlike optical remote sensing images, characterized by very neat and uniform features,
SAR images are affected by speckle. Speckle confers to SAR images a granular aspect
with random spatial variations. Figure 1 shows an example of single polarization speckled
SAR images.

2 2
Figure 1 Single-look S11 (top) and S12 (bottom)images.

The intensity images displayed in Figure 1 show a poor contrast, as well as a random
aspect, that reduce the possibilities of visual interpretation and analysis of the scene under
consideration.
The discrimination of different natural media by comparing intensity to a fixed threshold
leads, in general to numerous errors due to the high variability of SAR speckled response.
Speckle phenomenon also affects the phase of scattering coefficients and corrupts
polarimetric information.
The image of Arg ( S11 ) shown in Figure 2 indicates that the absolute phase of a scattering
coefficient is highly random and does not contain evident information. Speckle does not
affect similarly different polarimetric channels, as shown in the between channel relative
*
phase image, Arg ( S11 S 22 ) and in the color coded image, built from the three polarimetric
2 2
and R = S 11 , G = S 12 , B = S 22
2 2 2
channel intensities, S 11 , S 12

1
Speckle Filtering

*
Figure 2 Arg ( S11 ) (top) and Arg ( S11 S 22 ) (middle) images. Color coded image
R = S11 , G = S12 , B = S 22
2 2 2
(bottom).

Speckle corrupts polarimetric observables (phase and intensity) in an important way. Specific
procedures have to be used to retrieve relevant polarimetric information and to reduce the
randomness of the acquired signals.

3.2 Simple speckle model

3.2.1 Single-polarization multiplicative speckle model

Speckle confers a random aspect to SAR images, but may not be considered as a simple noise.
It is, in fact, tightly related to the SAR measurement principle.
Synthesized SAR data may be considered as the result of the integration of a scene coherent
response within each resolution cell, resulting from the convolution of the SAR impulse
response with the coherent contribution of each elementary scatterer, as illustrated in Figure
3. As the number of contributing scatterers, within a resolution cell, tends to be large (it is the
case for common resolution SAR measurements), the resulting integrated response is random
in phase and amplitude and is shown to follow, over homogeneous areas, a Normal
distribution.

2
Speckle Filtering

Im

Ei

E tot = ∑ E i
i
E1
Re

Figure 3 Principle of coherent integration.

A speckled response is usually represented under the form of a simple product model
y = x.n
(1)
Where y represents a complex speckled scattering coefficient, x the original unspeckled
scattering coefficient and n the multiplicative speckle contribution.
The speckle term, n is composed of independent real and imaginary parts, following both real
centered Normal distribution N C (0, 1 / 2) .

The corresponding speckled intensity, Y , is


Y = yy ∗ = xx ∗ nn∗
(2)
⇒ Y = X nn ∗
Over homogeneous areas, X is considered to be constant and the speckled intensity follows
an exponential probability density function
Y

−1
p (Y ) = X e X (3)

Its first two moments are given by


E (Y ) = X E (nn ∗ ) = X
(4)
Var (Y ) = X 2 Var (nn ∗ ) = X 2

3.2.2 Polarimetric multiplicative speckle model

This speckle model may be extended to the polarimetric case by considering that polarimetric
channels are affected by independent multiplicative speckle components :

3
Speckle Filtering

⎡ S11 ⎤ ⎡ x11 ⎤ ⎡ n11 ⎤ ⎡ S11 2 ⎤ ⎡ X (n n ∗ ) ⎤


⎢S ⎥ = ⎢ ⎥⎢n ⎥ ⇒ ⎢ 2⎥ ⎢
11 11 11
∗ ⎥
⎢ 12 ⎥ ⎢ x12 ⎥ ⎢ 12 ⎥ ⎢ S12 ⎥ = ⎢ X 12 (n12 n12 ) ⎥ (5)
⎢⎣ S 22 ⎥⎦ ⎢⎣ x22 ⎥⎦ ⎢⎣n22 ⎥⎦ ⎢ S 2 ⎥ ⎢ X (n n ∗ )⎥
⎣⎢ 22 ⎦⎥ ⎣ 22 22 22 ⎦
One may note that the multiplicative assumption is in general not valid to model speckled
correlation terms.

3.3 Principle of scalar speckle filtering

3.3.1 Incoherent averaging

As presented in the former paragraph, a speckled intensity, Y , may be considered as a


random variable whose mean value equals the unspeckled intensity, X , but affected by a
large variance due to speckle.
The principle of speckle filtering consists of reducing the variance of Y in order to improve
the estimate of its mean.
The sample mean, Y , is defined as the empirical average of L independent realizations of a
speckled intensity as follows
1 L
Y = ∑ Yi
L i =1
(6)

It can be shown that, over homogeneous areas, this estimate of X follows a Gamma density
function
L −1 L Y
LL X Y −
p( Y ) = e X
(7)
Γ( L ) X L

and has the following two first moments


E( Y ) = X
(8)
Var ( Y ) = X 2 / L

It is possible to observe from (8) that as the number of independent samples, L, reduces to 1
the variance of the estimate intensity increases, whereas incoherent averaging over L
independent realizations permits to reduce the variance of a speckled intensity in a significant
way.
The quantity L = var( I ) / E ( I ) 2 is called the Equivalent Number of Looks (ENL) and is a
measure of speckle importance.

4
Speckle Filtering

stdev ( Y ) 10n
10 4

9
3.5
8
3
7
2.5
6

5 2

4
1.5
3
1
2
0.5
1

0 0

E( Y )
0 2 4 6 8 10

Figure 4 Occurrence (scattering diagram) of single-look of intensity moments over a homogeneous


areas.

The occurrence plot displayed in Figure 4 clearly shows that the standard deviation and the
mean of sampled intensities are linearly related over homogeneous areas. The slope of this
linear relation is L and equals 1 in the case of single-look data sets.

3.3.2 Boxcar and J. S. Lee filters

3.3.2.1 The boxcar filter


The boxcar filter is a direct application of the incoherent averaging described by (6) to the
case of an image.
~
Filtered intensity estimates, X i , j , are constructed by computing the sample mean over each
pixel neighborhood, defined by a sliding window of ( N w × N w ) pixels
Nw / 2 Nw / 2
~ 1
X i , j = Yi , j
Nw
=
N w2
∑ ∑Y
p=− N w / 2 q=− N w / 2
i+ p , j +q (9)

where the subscripts i and j correspond to the considered pixel row and column index
respectively.

2
Figure 5 S11 filtered image using a boxcar filter.

5
Speckle Filtering

Figure 5 shows an intensity image obtained using a (7×7) boxcar filter. This image shows
enhanced contrast and lower random aspect.
As it can be seen in Figure 5, the boxcar filter is characterized by two main limitations :
- sharp edges are generally blurred
- point scatterers are over filtered and transformed to spread targets
Solutions to these limitations are offered by the refined Lee filter.

3.3.2.2 J.S. Lee adaptive filter


J. S. Lee's filter determines the unspeckled intensity estimate that minimizes the mean squared
error
~ 2
X−X (10)

This MMSE filter is based on a linearized speckle model leading to the following estimate
expression
~
X = Y N +k Y − Y N
w
( w
) (11)

where k is an adaptive filtering coefficient, based on local statistics, given by


var( X ) var Y − E ( Y )σ n
2 2

k= =
var Y var Y [1 + σ n2 ]
(12)
1
with σ n = the a priori speckle variance
2

L
~
Over homogeneous areas, var( X ) = 0 ⇒ k = 0 and X = Y N , whereas over point targets
w
~
and highly heterogeneous areas, k = 1⇒ X = Y and the pixel intensity remains unaffected by
the filtering procedure.

0
1

Figure 6 Examples of directional masks.

In order to reduce the sensitivity of the adaptive filtering coefficient, k , to isolated


heterogeneities, this filter uses directional masks to determine the most homogeneous part of
the sliding window where local statistics have to be estimated. This modification permits to
preserve relatively sharp edges.

6
Speckle Filtering

The Lee filter results displayed in Figure 7 demonstrate the effectiveness of this adaptive
filtering approach

2
Figure 7 S11 filtered image using a (7×7) Lee filter.

3.4 Extension to the polarimetric case

3.4.1 Second order polarimetric representations

Speckle filtering is based on incoherent averaging and requires handling statistical second
order representations. The intensity information used in the scalar case has to be extended to
the vector case when dealing with two or more polarization channels in order to take into
account the different intensities as well as the cross-correlation related information.
A simple way to build an incoherent polarimetric representation consists in vectorizing the
scattering matrix to create a target vector and computing the corresponding covariance matrix
(details on matrix incoherent representations may be found in the appendix).
[S ] ⇒ k 3L , k 3P
[
with k 3 L = S11 2 S12 S 22 ]
T

(13)
and k 3 P =
1
[S11 + S 22 S11 − S 22 2S12 ]T
2
The (3×3) covariance matrix, [C3 ] , is defined as :

[C3 ] = k 3 L k 3L

(14)
where † represents the transpose conjugate operator. One may note that [C3 ] is not a usual
covariance matrix that would require the use of an expectation operator, but this
representation may be understood as an incoherent polarimetric form obtained from a single
realization of k 3L .
The (3×3) coherency matrix, [T3 ] , is defined as :

[T3 ] = k 3 P k 3 P

(15)

7
Speckle Filtering

One may note that the coherency matrix is in fact the single realization covariance matrix of
k 3P .
Covariance and coherency matrices are similar (related by a similarity transformation) and
may be used in an equivalent way
⎡ 1 1 ⎤
⎢ 0⎥
2 2
−1 ⎢ ⎥
k 3 P = [ A]k 3 L ⇒ [T3 ] = [ A][C3 ][ A] = [ A][C3 ][ A] with [ A] = ⎢

0 0 1⎥ (16)
1 1
⎢ − 0⎥
⎢⎣ 2 2 ⎥⎦

3.4.2 Polarimetric boxcar and Lee's filters

3.4.2.1 Polarimetric boxcar filter


The extension of the boxcar filter to the polarimetric case is straightforward. The estimated
covariance matrix is given by
Nw / 2 Nw / 2
~ 1
[C ]i , j = [C ]i , j
Nw
=
N w2
∑ ∑ [C ]
p=− N w / 2 q=− N w / 2
i+ p, j+q (17)

where the subscripts i and j correspond to the considered pixel row and column index
respectively.

3.4.2.2 J. S. Lee polarimetric filter


J. S. Lee proposed to estimate the unspeckled covariance matrix according to the following
expression
[C ] = [C ] + k ([C ] − [C ] )
~
(18)
where k remains a scalar coefficient computed from the span statistics, span = C11 + C 22 + C33 .

This approximation allows to filter polarimetric data in a fast and simple way and avoids
additional coupling (or cross-talk) between the polarimetric channels.
Figure 8 shows improved color coded images processed through the boxcar and J. S. Lee
filters.

8
Speckle Filtering

Figure 8 Color-coded, R = S11 , G = S12 , B = S 22


2 2 2
, filtered image using a (7×7) boxcar (top)
Lee's filter (bottom).

3.5 Consequences of polarimetric speckle filtering

It was seen in former paragraphs that it is necessary to reduce polarimetric variables random
aspect by speckle filtering prior to any interpretation of polarimetric information. The
incoherent averaging of [T3 ] or [C3 ] matrices has an important impact on their polarimetric
properties.

3.5.1 Interpretation loss vs. information gain

Speckle filtering may cause a loss of polarimetric information by destroying the relation
between and [T3 ] or [C3 ] matrices.

A coherency matrix is fully defined by 9 real coefficients : its three diagonal terms and three
complex correlation coefficients. In the case of a single look coherency matrix, all three
correlation coefficients have unitary modulus and one of their phase may be obtained by a
linear combination of the remaining two, leaving 5 degrees of freedom. A relative scattering
matrix [ S rel ] and single-look [T3 ] or [C3 ] matrices may be related in a unique way as shown
in the following example :
⎡C11 C12 C13 ⎤ ⎡ C11 m12 e jφ12 m13 e jφ13 ⎤
⎢ ⎥
[C ] = ⎢⎢C12* C 22 C 23 ⎥⎥ = ⎢m12 e − jφ12 C 22 m23 e j (φ13 −φ12 ) ⎥
⎣⎢C13
* *
C 23 C33 ⎦⎥ ⎢⎣ m13e 13 m23 e − j (φ13 −φ12 )
− jφ
C33 ⎥

(19)

with mij = Cii C jj

9
Speckle Filtering

And
⎡ C 22 / 2 e − jφ12 ⎤
[S rel ] = e − jφ [S ] = ⎢
1
C11
⎥- (20)
− jφ
⎢⎣ C 22 / 2 e 12 C33 e − jφ13 ⎥⎦

The scattering mechanism may then be interpreted by comparing [S rel ] to canonical


examples.

After speckle filtering, this may not be true anymore. In a general case, the modulus of
correlation coefficients is inferior to one and the phase terms are linearly independent

S ij S kl*
2
≤ S ij
2
S kl
2
(
and Arg S ij S kl* )≠ (
Arg S ij S kl* ) (21)

In such a case, the coherency matrix is said to be distributed and cannot be related to a
coherent scattering matrix.
The correlation coefficient displayed in Figure 9 shows a varying modulus over the selected
scene, indicating that the degree of correlation might be related to the nature of the scattering
medium. The additional information contained in the cross-correlation terms will be exploited
by incoherent decomposition theorems to extract even more characteristics from polarimetric
data sets.

10
Speckle Filtering

* 2 2
Figure 9 Argument (top) and modulus (bottom) of S11S 22 / S11 S 22 after application of a

Lee filter.

3.5.2 Partially polarized waves and polarimetric signature

As it is demonstrated in the appendix, polarization signatures may be built indifferently from


coherent scattering matrix or incoherent Kennaugh matrix representations.
The Co-polarization polarimetric signature is given by
2
PC = E (φ ,τ ) [ S ] E (φ ,τ ) or PC = g (φ ,τ ) [ K ] g (φ ,τ )
t t
(22)

Let us assume that the Kenaugh matrix [K ] results from the incoherent averaging of N
single-look matrices [ K i ] . The Co-polarization polarimetric signature of [K ] may then be
expressed as
PC = g (φ ,τ ) [ K ] g (φ ,τ ) = g (φ ,τ ) [ K ] g (φ ,τ ) = g (φ ,τ ) g r (φ ,τ )
t t t
(23)

where g r (φ ,τ ) stands for the average reflected Stokes vector and may be formulated as

1 N 1 N
g r (φ ,τ ) = ∑ i
N i =1
[ K ] g (φ ,τ ) = ∑ g (φ ,τ )
N i =1 ri
(24)

Knowing that each reflected Stokes vector may be written as a function of the corresponding
Jones vector components, that is

[
g ri = E ri1 + E ri 2
2 2
E ri1 − E ri 2
2 2
(
2ℜ E ri1 E ri* 2 ) (
− 2ℑ E ri1 E ri* 2 )]
T
(25)

The average reflected Stokes vector may be expressed as

g r = E r1 [ 2
+ Er 2
2
E r1
2
− Er 2
2
(
2ℜ Er1 Er*2 ) (
− 2ℑ Er1 Er*2 )]
T

= [g r 0 gr3 ]
T (26)
g r1 gr2

From the Cauchy-Schwarz inequality, we know that


g r20 ≥ g r21 + g r22 + g r23 (27)
Relation (27) indicates that the average polarization-dependent intensity of a wave,
g r21 + g r22 + g r23 may be inferior to its total average intensity g r20 (when the reflected Stokes
vector are not parallel).
Such waves are said to be partially polarized and may be decomposed as follows
g r = [g r 0 g r 3 ] = [(α g r 0 g r 3 ] + [(1 − α ) g r 0 0 0 0]
T T T
g r1 gr2 g r1 gr2 (28)

11
Speckle Filtering

Where [(α g r 0 g r1 gr2 g r 3 ] represents


t
a fully polarized wave, with
α 2 g r20 = g r21 + g r22 + g r23 and [(1 − α ) g r 0 0 0 0] is associated to a completely unpolarized
t

(insensitive to polarization) term.


The decomposition factor
g r21 + g r22 + g r23
α= , 0 ≤α ≤1 (29)
g r20

is called the wave degree of polarization.


The co-polarization polarimetric signature mentioned at the beginning of this paragraph can
then be rewritten as a function of the degree of polarization
⎡α g r 0 ⎤
⎢ g ⎥
PC = g (φ ,τ ) g r (φ ,τ ) = g (φ ,τ ) ⎢ r1 ⎥ + (1 − α ) g 0 g r 0
t t
(30)
⎢ gr2 ⎥
⎢ ⎥
⎣ gr3 ⎦
When the reflected Stokes vectors are not parallel, i.e. the averaged Kenaugh matrices do not
correspond by a scalar factor, α ≠ 1 and it the copular intensity is strictly superior to. In such
cases, one cannot determine co-pol nor cross-pol nulls and the polarization signature is said to
show a pedestal.

12
Polarimetric Decompositions

4. POLARIMETRIC DECOMPOSITIONS

4.1 Coherent Decompositions

4.1.1 Purpose of the Coherent Decompositions

The objective of the coherent decompositions is to express the measured scattering matrix by
the radar, i.e. [ S ] , as a the combination of the scattering responses of simpler objects
k
[ S ] = ∑ ci [ S ]i (1)
i =1

In (1), the symbol [ S ]i stands for the response of every one the simpler objects, also known as
canonical objects, whereas ci indicates the weight of [ S ]i in the combination leading to the
measured [ S ] . As observed in (1), the term combination refers here to the weighted addition
of the k scattering matrices. In order to simplify the understanding of (1), it is desirable that
the matrices [ S ]i present the property of independence among them to avoid that a particular
scattering behavior to be present in more than one matrix [ S ]i . Often the most restrictive
property of orthogonality is imposed.
As it has been already highlighted in this document, the scattering matrix [S ] can
characterize the scattering process produced by a given target, and therefore the target itself.
This is possible only in those cases in which both, the incident and the scattered waves are
completely polarized waves. Consequently, coherent target decompositions can be only
employed to study the so-called coherent targets. These scatterers are also known as point or
pure targets.
In a real situation, the measured scattering matrix by the radar [ S ] corresponds to a complex
coherent target. Only in a few occasions, this matrix will correspond to a simpler or canonical
object, which a good example is, for instance, the trihedrals employed to calibrate SAR
imagery. Nevertheless, in a general situation, a direct analysis of the matrix [ S ] , with the
objective to infer the physical properties of the scatterer under study, is shown very difficult.
Thus, the physical properties of the target under study are extracted and interpreted through
the analysis of the simpler responses [ S ]i and the corresponding coefficients ci in (1).

The decomposition exposed in (1) is not unique in the sense that it is possible to find a
{ }
number of infinite sets [ S ]i ; i = 1,K, k in which the matrix [ S ] can be decomposed.

Nevertheless, only some of the sets {[ S ] ; i = 1,K, k}


i
are convenient to interpret the
information contained in [ S ] . In the next, we shall detail three of these sets which lead to: the

1
Polarimetric Decompositions

Pauli, the Krogager and the Cameron decompositions. As mentioned above, these
decompositions of the scattering matrix can be only employed to characterize coherent
scatterers. Therefore, we will finish this section presenting an algorithm to detect such a
targets where these decomposition theorems are relevant.

4.1.2 The Pauli Decomposition

4.1.2.1 Description of the Pauli Decomposition


The Pauli decomposition expresses the measured scattering matrix [ S ] in the so-called Pauli
basis. If we considered the conventional orthogonal linear (h,v) basis, in a general case, the
{ }
Pauli basis [ S ]a , [ S ]b , [ S ]c , [ S ]d is given by the following four 2×2 matrices

1 ⎡1 0⎤
[ S ]a = ⎢ 1 ⎥⎦
(2)
2 ⎣0
1 ⎡1 0⎤
[ S ]b = ⎢ −1⎥⎦
(3)
2 ⎣0
1 ⎡0 1⎤
[ S ]c = ⎢ 0 ⎥⎦
(4)
2 ⎣1
1 ⎡0 −1⎤
[ S ]d = ⎢ 0 ⎥⎦
(5)
2 ⎣1
In this document, it has been always considered that S hv = Svh , since reciprocity applies in a
monostatic system configuration. In this situation, the Pauli basis can be reduced to a basis
composed by the matrices (2), (3) and (4), that is,

{[ S ] ,[ S ] ,[ S ] }
a b c (6)

Consequently, given a measured scattering matrix [ S ] , it can be expressed as follows

⎡S Shv ⎤
[ S ] = ⎢ Shh Svv ⎥⎦
= α [ S ]a + β [ S ]b + γ [ S ]c (7)
⎣ hv

where
Shh + Svv
α= (8)
2
S hh − Svv
β= (9)
2
γ = 2 Shv (10)

From (8), (9) and (10) it can be easy show that the span of [ S ] can be obtained as
2 2 2 2 2 2
SPAN = S hh + Svv + 2 S hv = α + β + γ (11)

2
Polarimetric Decompositions

4.1.2.2 Interpretation of the Pauli Decomposition


The interpretation of the Pauli decomposition must be done according to the matrices in the
basis given at (6) and the corresponding coefficients (8), (9) and (10).
The matrix [ S ]a corresponds to the scattering matrix of a sphere, a plate or a trihedral. In
general, [ S ]a is referred to single- or odd-bounce scattering. Thus, the complex coefficient α,
given at (8), represents the contribution of [ S ]a to the final measured scattering matrix. In
2
particular, the intensity of this coefficient, i.e., α , determines the power scattered by targets
characterized by single- or odd-bounce.
The second matrix, [ S ]b represents the scattering mechanism of a dihedral oriented at 0
degrees. In general, this component indicates a scattering mechanism characterized by double-
or even-bounce, since the polarization of the returned wave is mirrored respect to the one of
the incident wave. Consequently, β stands for the complex coefficient of this scattering
2
mechanism and β represents the scattered power by this type of targets.

Finally, the third matrix [ S ]c corresponds to the scattering mechanism of a diplane oriented at
45 degrees. As it can be observed in (4), and considering that this matrix is expressed in the
linear orthogonal basis (h,v), the target returns a wave with a polarization orthogonal to the
one of the incident wave. From a qualitative point of view, the scattering mechanism
represented by [ S ]c is referred to those scatterers which are able to return the orthogonal
polarization, from which, one of the best examples is the volume scattering produced by the
forest canopy. The coefficient γ represents the contribution of [ S ]c to [ S ] , whereas γ stands
2

for the scattered power by this type of scatterers.

4.1.2.3 Representation of the Polarimetric information by means of the Pauli


decomposition
The Pauli decomposition of the scattering matrix is often employed to represent all the
polarimetric information in a single SAR image.
The polarimetric information of [S ] could be represented by the combination of the
2 2 2
intensities S hh , Svv and 2 S hv in a single RGB image, i.e., every of the previous
intensities coded as a color channel. The main drawback of this approach is the physical

3
Polarimetric Decompositions

|Shh| (dB) |Shv| (dB) |Svv| (dB)


-30dB -15dB 0dB

|Shh+ Svv| |Shh- Svv| |Shv|


2 2
Figure 1 Intensities of the polarimetric channels S hh , S vv and the combination of them in an
RGB image.

2 2 2
interpretation of the resulting image in term of S hh , Svv and 2 S hv . Consequently, an
2 2 2
RGB image can be formed with the intensities α , β and γ , which, as explained before,
correspond to clear physical scattering mechanisms. Thus, the resulting color image can be
employed to interpret the physical information from a qualitative point of view. The most
employed codification corresponds to
2
α → Red (12)
2
β → Blue (13)

4
Polarimetric Decompositions

2
γ → Green (14)

Then, the resulting color of the RGB image is interpreted in terms of scattering mechanism as
given by (12), (13) and (14). Figure 1 presents an example of this codification

4.1.3 The Krogager Decomposition

4.1.3.1 Description of the Krogager Decomposition


Krogager has proposed an alternative to factorize the scattering matrix as the combination of
the responses of a sphere, a diplane and a helix. The last two components present an
orientation angle θ. If we consider the scattering matrix expressed in the linear orthogonal
basis (h,v), the Krogager decomposition presents the following formulation

⎦ {
⎡ S( h , v ) ⎤ = e jϕ e jϕs ks [ S ] + kd [ S ] + kh [ S ]
⎣ s d h }
⎧ ⎡1 0 ⎤ ⎡ cos 2θ sin 2θ ⎤ ⎡ 1 ± j⎤⎫ (15)
= e jϕ ⎨e jϕs ks ⎢ ⎥ + kd ⎢ ⎥ + kh e m j 2θ ⎢ ⎥⎬
⎩ ⎣0 1 ⎦ ⎣ sin 2θ − cos 2θ ⎦ ⎣± j 1 ⎦ ⎭
If we compare (15) with (7), it can be observed that both decompositions present the same
number of independent parameters, i.e., six. In the case of the Pauli decomposition, the
complex coefficients α, β and γ, whereas in the case of the Krogager decomposition, the three
angles φ, φs and θ and the three real coefficients ks, kd and kh. The phase φ is referred as the
absolute phase, whose value depends on the distance between the radar and the target under
study. Due to the arbitrary value that this phase can present, it is often considered that the
Krogager decomposition presents 5 independent parameters given by {φs, θ, ks, kd, kh} plus
the absolute phase given by φ.
In order to calculate the value of the parameters {φs, θ, ks, kd, kh} plus the absolute phase φ, a
reformulation of (15) with the objective to simplify the process is next presented. Now, if we
consider the measured scattering matrix expressed in the circular polarization basis (r,l), the
Krogager decomposition is then

⎡S S rl ⎤ ⎡ S rr e rr

S rl e jϕrl

⎡ S( r ,l ) ⎤ = ⎢ rr =⎢ ⎥
⎣ ⎦
⎣ S rl Sll ⎥⎦ ⎢⎣ S rl e jϕrl − Sll e ( rr ) ⎥⎦
j ϕ +π

(16)
⎪⎧ jϕs ⎡ 0 j ⎤
jϕ ⎡e j 2θ 0 ⎤ ⎡e j 2θ 0 ⎤ ⎫⎪
= e ⎨e k s ⎢ ⎥ + kd ⎢ ⎥ + kh ⎢ ⎥⎬
⎪⎩ ⎣ j 0⎦ ⎣ 0 −e − j 2θ ⎦ ⎣ 0 0 ⎦ ⎪⎭

From (16), it can be easily observed that the response of the sphere can be obtained from S rl

ks = Srl (17)

The terms Srr and Sll represent, directly, the diplane component of the decomposition (16), but
two cases of analysis must be considered according to the difference in absolute value of Srr
and Sll. This is necessary in order to accommodate the difference in the scattered power in the
right and the left circular polarizations. When Sll represents the diplane component, it occurs
that Srr > Sll . Hence

5
Polarimetric Decompositions

kd+ = Sll (18)


+
k = Srr − Sll
h (19)

And the helix component presents a left sense. On the contrary, when it is Srr the term which
represents the diplane component, it occurs that Sll > Srr

kd− = Srr (20)


kh− = Sll − Srr (21)

and the helix has a right sense. Finally, from (16), the phase components are
1
ϕ= (ϕrr + ϕll + π ) (22)
2
1
θ = (ϕrr − ϕll − π ) (23)
4
1
ϕ = ϕrl − (ϕrr + ϕll + π ) (24)
2
In order to relate the formulations of the Krogager decomposition presented in (15) and (16),
the following relations are useful
1
S rr = jS hv +( Shh − Svv ) (25)
2
1
Sll = jShv − ( Shh − Svv ) (26)
2
j
Srl = ( Shh + Svv ) (27)
2

4.1.3.2 Interpretation of the Krogager Decomposition


The interpretation of the Krogager decomposition must be done according to the coefficients
{φs, θ, ks, kd, kh}, plus the absolute phase φ.
The absolute phase φ can contain information about the scatterer under study. But, since its
value depends also on the distance between the radar and the target, it is considered as a
irrelevant parameter.
The parameters φs and ks characterize the sphere component of the Krogager decomposition.
On the one hand, the phase φs represents a displacement of the sphere respect to the diplane
and the helix components. On the other hand, the real parameter ks represents the contribution

6
Polarimetric Decompositions

ks (dB) kd (dB) kh (dB)

-30dB -15dB 0dB

ks(dB) kd(dB) kh(dB)


2 2 2
Figure 2 Intensities corresponding to the Krogager decomposition k s , kd and kh , and the
combination of them in an RGB image. Images are shown in a dB scale.

of the sphere component to the final scattering matrix [ S ] . Consequenty, |ks|2 is interpreted as
the power scattered by the sphere-like component of the matrix [ S ] .

The phase parameter θ stands for the orientation angle of the deplane and the helix
components of the Krogager decomposition.
Finally, the coefficients kd and kh correspond to the weights of the diplane and the helix
components. Thus, |kd|2 and |kh|2 are interpreted as the power scatterered by the diplane- and
the helix-like components of the Krogager decomposition.

7
Polarimetric Decompositions

4.1.3.3 Representation of the Polarimetric information by means of the Krogager


decomposition
The information provided by the Krogager decomposition can be also employed to form a
RGB color-coded image representing the polarimetric information. In this case, the phase
terms are discarded and only the coefficients {ks, kd, kh} are considered in the following way
2
k s → Red (28)
2
kd → Blue (29)
2
kh → Green (30)

Figure 2 gives an example of this codification.

4.1.4 The Cameron Decomposition

4.1.4.1 Description of the Cameron Decomposition


The Cameron decomposition performs a factorization of the measured scattering matrix [ S ]
based on two basic properties of radar targets: reciprocity and symmetry.
A radar target is considered reciprocal when the diagonal terms of the measured scattering
matrix are equal, i.e., the reciprocity theorem applies. For a scattering matrix measured in the
orthogonal linear (h,v) basis
S hv = Svh (31)
whereas for the orthogonal circular (r,l) basis
Srl = Slr (32)
The reciprocity assumption applies in the case of monostatic SAR systems, where the
transmitting and receiving antennas are located in the same position. Consequently, all the
scatterers can be considered as reciprocal when imaged by a monostatic SAR system.
A scattering is considered symmetric when the target has an axis of symmetry in the plane
orthogonal to the direction between the radar and the target. The symmetry of a scatterer can
be also considered in the frame of the Pauli decomposition (7). Hence, a scatterer is
considered symmetric if it exists a rotation which cancels the projection of [ S ] in the
component [ S ]c of the Pauli decomposition.

Since monostatic SAR imagery considerers only reciprocal scatterers, we present, in the
following, the Cameron decomposition applied to reciprocal targets, i.e., [ S ] is symmetric.

Given a scattering matrix measured in the orthogonal linear (h,v) basis


⎡S Shv ⎤
[ S ] = ⎢ Shh Svv ⎥⎦
(33)
⎣ hv

we consider the vector form of [ S ] as follows

8
Polarimetric Decompositions

⎡ Shh ⎤
r ⎢S ⎥
S = ⎢ hv ⎥ (34)
⎢ S hv ⎥
⎢ ⎥
⎢⎣ Svv ⎥⎦
Hence, the Pauli decomposition given at (7) can be formulated in a vector form as follows
r r r r
S = α S a + β Sb + γ S c (35)
The Cameron decomposition states that a reciprocal target can be decomposed as the sum of
two components as follows
r r max r min
S = A ⎡⎣cosτ S sym + sin τ S sym ⎤
⎦ (36)

Considering the inner vector product as (.,.), and the vector norm as ||.||, the different
parameters of (36) are obtained as follows
r
A= S (37)
r max r r
S sym = α Sa + ε Sb (38)
r max
The matrix S sym is called the largest or maximum symmetric component of [ S ] and it is
obtained by means of
ε = β cos θ + γ sin θ (39)
and
βγ * + β *γ
tan ( 2θ ) = 2 2 (40)
β −γ
r min r
The matrix S sym is called the minimum symmetric component of S . Finally, the factor cos τ
r r r min
is the degree of symmetry of S and it measures the degree to which S deviates from S sym and
is obtained as
r r min
cosτ =
( rS , Sr )
sym

min
(41)
S S sym

4.1.4.2 Representation of Symmetric Scatterers


r
An arbitrary symmetric scatterer Ssym can be decomposed according to
r
ˆ ( z ) a ∈ R + ρ ,ψ ∈ ( −π , π ]
S sym = ae j ρ ⎡⎣ R (ψ ) ⎤⎦ Λ (42)

where a indicates the amplitude of the scattering matrix, ρ is the nuisance phase and ψ is the
scatterer orientation angle. The matrix ⎡⎣ R (ψ ) ⎤⎦ denotes the rotation operator. Finally, the
ˆ ( z ) , expressed in the linear polarization basis, is
normalized vector Λ

9
Polarimetric Decompositions

⎡1 ⎤
⎢ ⎥
1 ⎢0 ⎥
Λ( z) =
ˆ z ∈ C, z ≤ 1 (43)
1 + z ⎢0 ⎥
⎢ ⎥
⎣z⎦
Consequently, the complex quantity z in (43) can be employed to characterize the symmetric
scatterer under consideration. The following list presents the values of z some some canonical
targets
• Triedral
ˆ (1)
Sˆa = Λ (44)

• Diplane
ˆ ( −1)
Sˆb = Λ (45)

• Dipole
ˆ ( 0)
Sˆl = Λ (46)

• Cylinder

ˆ ⎛1⎞
Sˆcy = Λ (47)
⎜ ⎟
⎝2⎠
• Narrow diplane

ˆ ⎛− 1 ⎞
Sˆnd = Λ (48)
⎜ ⎟
⎝ 2⎠
• Quarter wave device
ˆ ( j)
Sˆ1 4 = Λ (49)

4.1.4.3 Classification based on the Cameron Decomposition


On the basis of the factorization of the measured scattering matrix [S ] , (36), and the
representation of the maximum symmetrical component as a complex quantity z, Cameron
r max
proposed a classification scheme for the maximum symmetrical component S sym . This
classification scheme is based on comparing the quantity z of the matrix under study with
those corresponding to the targets given from (44) to (49). In order to compare the measured z
and the scattering responses of the targets of reference, the following metric must be
considered
1 + z * zref
d ( z, zref ) = (50)
2 2
1+ z 1 + zref

Finally, the measured scatterer z is classified according to the shortest distance d(z,zref).
Figure 3 presents the Cameron’s classification scheme based on the metric (50).

10
Polarimetric Decompositions

⎡S S hv ⎤
[ S ] = ⎢ Shh Svv ⎥⎦
⎣ hv

T Right Helix
F
Symmetrical Match Helix Left Helix

Asymmetrical
T F

Maximum
Symmetrical
Component
Triedral

Dihedral

Narrow dihedral
Shortest distance
d(z,zref)
Dipole

Cylinder

¼ Wave

Figure 3 Cameron’s classification scheme.

4.1.5 Relevance of Coherent Decompositions, Touzi Criterion

As it has been noticed, the previous coherent decompositions can be only employed to
analyze pure targets whose scattering response is completely determined by the measured
scattering matrix [ S ] . Consequently, when working with SAR imagery, it is necessary to
determine whether a particular pixel is a pure target or, on the contrary, it belongs to a
distributed scatterer. For the first case, coherent decompositions can be employed to study the
physics of the scatterer. Nevertheless, the analysis of distributed scatterers must be performed
by means of the so-called incoherent decompositions.
A qualitative way to differentiate pure from distributed scatterers is to consider their physical
nature. A rough division would be to consider the man-made targets as pure targets, whereas
natural targets can be considered as distributed. For man-made targets we understand all type
of artificial man-structures in a SAR image such buildings, power lines, train tracks or cars.
On the contrary, for instance, forests, agricultural areas, bare soils or water have to be
considered as distributed targets.
Touzi has proposed a technique to identify pure targets in a SAR images, based on the
Cameron decomposition. This technique determines the nature of a target on the maximum
r max
symmetrical component of the Cameron decomposition S sym . In the analysis of a given target,
Touzi differentiates two cases to perform such a study. On the one hand, pure targets whose

11
Polarimetric Decompositions

response occupy only one pixel and, on the other hand, targets whose response extends among
several pixels.

4.1.5.1 Coherent Test for Point Targets


A resolution cell, or pixel, in a SAR images is formed by the coherent addition of the
responses of the elementary scatterers within the resolution cell. In those cases in which there
is no a dominant scatterer, the statistics of the response is given by the complex Gaussian
scattering model, giving rise to the so-called speckle.
Nevertheless, the resolution cell can present a point target, which dominates the response of
the resolution cell. In this case, the scattering response is due to the coherent combination of
two components: the one due to the dominant scatterer and the coherent combination due to
the clutter, which is given by the complex Gaussian scattering statistics model. The statistics
of the resulting combination receives the name of Rician model. Figure 4 compares the
response with and without the presence of a point scatterer within the resolution cell.
Imaginary part Imaginary part
Speckle

Point target response


Speckle
Final response
Final response
Real part Real part

(a) (b)

Figure 4 Coherent response of a given resolution cell (a) without a dominant scatterer, (b) with a
dominant scatterer.

As observed in Figure 4b, despite the point target (blue arrow) can dominate the response of
a resolution cell (red arrow), the effect of the clutter (cloud of black arrows) must be taken
into consideration. It is necessary, hence, to compare the response of the point target with the
one of the clutter. Hence, a signal to clutter power ratio, denoted by S/C, must be defined.
Touzi established that this ratio must be higher than 15dB. The reason to choose this value is
that the response of the resolution cell will present a phase variation smaller than π/4 with
respect to the response of the point target.
In order to be, as much as possible, independent from the conditions in which the SAR image
has been taken, Touzi proposed to apply the previous S/C criteria on the maximum
polarization return |m|2, which represents the polarization state giving maximum returned
power from the scatterer under consideration. This parameter, representing the pure target
response, can be obtained from the diagonalization of the measured scattering matrix [ S ] , via
its eigendecomposition, as established by Hyunen. According to Huynen, the eigenvalues of
the covariance matrix [ S ] can be parameterized as follows

λ1 = me j 2υ (51)
− j 2υ
λ2 = m tan γ e
2
(52)

12
Polarimetric Decompositions

where λ1 and λ2 are the maximum and the minimum eigenvalues, respectively. In (51) and
(52), m is the maximum polarization, υ is the skip angle and γ is the polarisability angle.
Consequently, |m|2 can be derived from the magnitude of the maximum eigenvalue of [ S ] .
Finally, the power corresponding to the clutter can be derived from a neighboring area not
affected by the point scatterer.

4.1.5.2 Coherent Test for Distributed Targets


For distributed scatterers, Touzi proposed to measure its coherency also in terms of the
r max
maximum symmetric component S sym derived from the Cameron decomposition. On the
basis of its definition, (38), Touzi defined the so-called degree of coherence of a distributed
scattered as follows

( )
2 2
2 2
α −β + 4 α ⋅ β*
psym = (53)
2 2
α +β

The degree of coherence is generated by a moving window. In the resulting map, those pixels
presenting psym close to 1 represent areas in which the scatterer is locally coherent. Hence,
the coherent decomposition theorems can be applied in this case.
⎡S Shv ⎤
[ S ] = ⎢ Shh Svv ⎥⎦
⎣ hv

r max r r
S sym = α S a + ε Sb

Symmetric scattering

T
F Dist. target T
coherence

Point target T
coherence

Non-coherent Scatterer Coherent Scatterer

Figure 5 Cameron’s classification scheme.

4.1.5.3 Coherent Test Algorithm


On the basis of the coherent test presented in the previous two points, Touzi proposed an
algorithm to determine which pixels in a SAR image are coherent, and, therefore, the coherent
decomposition theorems can be employed to analyze them. Figure 5 presents the flow chart
of the algorithm.

13
Polarimetric Decompositions

4.2 Incoherent Decompositions

4.2.1 Purpouse of the incoherent decompositions

As explained previously, the scattering matrix [ S ] is only able to characterize the so-called
coherent or pure scatterers. On the contrary, this matrix can not be employed to characterize,
from a polarimetric point of view, the so-called distributed scatterers. This type of scatterers
can be only characterized, statistically, due to the presence of speckle noise. Since speckle
noise must be reduced, only second order polarimetric representations can be employed to
analyze distributed scatterers. These second order descriptors are the 3×3, Hermitian average
covariance [C3 ] [T3 ] and the coherency [T3 ] matrices. These two representations of the
polarimetric information are equivalent.
The complexity of the scattering process makes extremely difficult the physical study of a
given scatterer through the direct analysis of [C3 ] or [T3 ] . Hence, the objective of the
incoherent decompositions is to separate the [ C3 ] or [T3 ] matrices as the combination of
second order descriptors corresponding to simpler or canonical objects, presenting an easier
physical interpretation. These decomposition theorems can be expressed as
k
[C3 ] = ∑ pi [C3 ]i (54)
i =1
k
[T3 ] = ∑ qi [T3 ]i (55)
i =1

where the canonical responses are represented by [C3 ]i and [T3 ]i , and pi and qi denote the
coefficients of these components in [ C3 ] or [T3 ] , respectively. As in the case of the
coherent decomposition, it is desirable that these components present some properties. First of
all, it is desirable that the components [C3 ]i and [T3 ]i correspond to pure targets in order to
simplify the physical study. Nevertheless, this is not absolutely necessary. In addition the
components [C3 ]i and [T3 ]i are independent, or in a more restrictive way, orthogonal.

The bases in which [C3 ] or [T3 ] are decomposed, i.e., {[C ] ; i = 1,K, k}
3 i and

{[T ] ; i = 1,K, k} are not unique. Consequently, different decompositions can be presented. In
3 i

the following we detail: the Freeman, the Huynen and the Eigenvector-eigenvalue
decompositions.

4.2.2 The Model-based Freeman Decomposition

4.2.2.1 Description of the Freeman Decomposition


The Freeman decomposition models the covariance matrix as the contribution of three
scattering mechanisms

14
Polarimetric Decompositions

• Volume scattering where a canopy scatterer is modeled as a set of randomly oriented


dipoles.
• Double-bounce scattering modeled by a dihedral corner reflector.
• Surface or single-bounce scattering modeled by a first-order Bragg surface scatterer.
The volume scattering from a forest canopy is modeled as the contribution from an ensemble
of randomly oriented thin dipoles. The scattering matrix of an elementary dipole, expressed in
the orthogonal linear (h,v) basis, when horizontally oriented, has the expression
⎡R 0⎤
[ S ] = ⎢ 0h Rv ⎥⎦
(56)

For a thin dipole, (58) reduces to
⎡1 0 ⎤
[ S ] = ⎢0 ⎥ (57)
⎣ 0⎦
Now, if we considerer a set of randomly oriented dipoles, characterized by the previous
scattering matrix and oriented according to a uniform phase distribution, the covariance
matrix of the ensemble of thin dipoles can be modeled by
⎡1 0 1 3⎤
[C3 ] v = f v ⎢ 0 2 3 0 ⎥⎥

(58)
⎢⎣1 3 0 1 ⎥⎦

where fv corresponds to the contribution of the volume scattering to the |Svv|2 component. The
covariance matrix [C3 ] v presents rank 3. Thus, the volume scattering can not be
characterized by a single scattering matrix of a pure target.
The second component of the Freeman decomposition corresponds to the double-bounce
scattering. In this case, a generalized corner reflector is employed to model the scattering
process. The diplane itself is not considered metallic. Hence, we consider that the vertical
surface has reflection coefficients Rth and Rtv for the horizontal and the vertical polarizations,
whereas the horizontal one presents the coefficients Rgh and Rgv for the same polarizations.
Additionally, two phase components for the horizontal and the vertical polarizations are
considered, i.e., e j 2γ h and e j 2γ v , respectively. The complex phase terms γ h and γ v account
for any attenuation or phase change effect. Hence, the scattering matrix of the generalized
dihedral is
⎡ e j 2γ h Rgh Rth 0 ⎤
[S ] = ⎢ 0 e j 2γ v ⎥
Rgv Rtv ⎦⎥
(59)
⎣⎢
which gives rise to the covariance matrix of the double-bounce scattering component. After
normalization respect to the Svv component, this covariance matrix can be written as follows
⎡α 2 0 α⎤
⎢ ⎥
[C3 ]d = fd ⎢ 0 0 0⎥ (60)
⎢α* 0 1 ⎥⎥
⎣⎢ ⎦

15
Polarimetric Decompositions

where
Rgh Rth
α = e j 2(γ h −γ v )
(61)
Rgv Rtv

and fd corresponds to the contribution of the double-bounce scattering to the |Svv|2 component
2
f d = Rgv Rtv (62)

As it can be observed, the covariance matrix [C3 ]d has rank 1, as it can be represented by the
scattering matrix given at (59).
The third component of the Freeman decomposition consist of a first-order Brag surface
scatterer modeling surface scattering. The scattering mechanism is represented by the
scattering matrix
⎡R 0⎤
[ S ] = ⎢ 0h Rv ⎥⎦
(63)

Consequently, the covariance matrix corresponding to this scattering component is
⎡β 2 0 α⎤
⎢ ⎥
[C3 ]s = fs ⎢ 0 0 0⎥ (64)
⎢ α* 0 1 ⎥⎥
⎢⎣ ⎦
where fs corresponds to the contribution of the double-bounce scattering to the |Svv|2
component
2
f s = Rv (65)

and
Rh
β= (66)
Rv

As in the case for the double-bounce scattering mechanism, since the matrix [C3 ]s presents
rank 1, therefore, it is completely represented by the scattering mechanism presented at (63).
Hence, the Freeman decomposition expresses the measured covariance matrix [C3 ] as follows

[C3 ] = [C3 ] v + [C3 ]d + [C3 ]s (67)

4.2.2.2 Interpretation of the Freeman Decomposition


The term fv in (58) corresponds to the contribution of the volume scattering of the final
covariance matrix [C3 ] . Hence, the scattered power by this component can be written as
follows
8 fv
Pv = (68)
3

16
Polarimetric Decompositions

From (60), it can be concluded that the power scattered by the double-bounce component of
[C3 ] has the expression
(
Pd = f d 1 + α
2
) (69)

Finally, the power scattered by the surface-like component is

(
Ps = f s 1 + β
2
) (70)

Consequently, the scattered power Pv, Pd and Ps can be employed to generate a RGB image to
present all the color-coded polarimetric information in a sole image. Figure 6 presents an
example of the Freeman decomposition.

Pv (dB) Pd (dB) Ps (dB)


-30dB -15dB 0dB

Pv(dB) Pd(dB) Ps(dB)


Figure 6 Intensities corresponding to the Freeman decomposition Pv, Pd and Ps and the
combination of them in an RGB image. Images are shown in a dB scale.

From (68), (69) and (70) it can be observed, that the Freeman decomposition maintains the
span or total scattered power

17
Polarimetric Decompositions

2 2 2
SPAN = S hh + Svv + 2 S hv = Pv + Pd + Ps (71)

The Freeman decomposition presented in (67) presents 5 independent parameters {fv, fd, fs, α,
β} and only 4 equations. Consequently, some hypothesis must be considered in order to find
the values of {fv, fd, fs, α, β}. Figure 7 presents the scheme employed to invert the Freeman
decomposition.

{S hh
2
, Svv
2
, Shv
2
, S hh Svv* }
5 parameters
4 equations

{ f v , f d , f s ,α , β }
2
f v = 3 S hv

4 parameters
3 equations

if ℜ {S hh Svv* } ≥ 0 ⇒ α = −1 if ℜ {S hh Svv* } < 0 ⇒ β = 1

3 parameters
3 equations

α = −1 { fd , f s , β } { f d , f s ,α } β =1
Single-bounce Double-bounce
scattering dominates scattering dominates

Figure 7 Inversion of the Freeman decomposition parameters.

4.2.3 Phenomenological Huynen Decomposition

4.2.3.1 Description of the Huynen Decomposition


The Phenomenological Huynen decomposition represents the first attempt to consider
decomposition theorems for the analysis of distributed scatterers. The Huynen decomposition
considers the concept of “wave dichotomy”, exporting it to the study of distributed scatterers.
On a first stage, the phenomenological Huynen decomposition considers a particular
parameterization of a distributed scatterer. In the case of the covariance matrix, this
parametrization is
⎡ 2 A0 C −j D H + j G ⎤
⎢ ⎥
[T3 ] =⎢ C + j D B0 + B E + j F ⎥ (72)
⎢⎣ H − j G E −j F B0 − B ⎥⎦

The set of nine independent parameters of this particular parameterization allows a physical
interpretation of the target under consideration. The following list presents the information
provided by each one of the parameters:
• A0: Represents the total scattered power from the regular, smooth, convex parts of the
scatterer.

18
Polarimetric Decompositions

• B0: Denotes the total scattered power for the target’s irregular, rough, non-conex
B

depolarizing components.
• A0+ B0: Gives roughly the total scattered power.
• B0+B: Total symmetric depolarized power.
B

• B0-B: Total non-symmetric depolarized power.


B

• C, D: Depolarization components of symmetric targets


o C: Generator of target global shape (Linear).
o D: Generator of target local shape (Curvature).
• E, F: Depolarization components due to non-symmetries
o E: Generator of target local twist (Torsion).
o F: Generator of target global twist (Helicity).
• G, H: Coupling terms between target’s symmetric and non-symmetric terms
o G: Generator of target local coupling (Glue).
o H: Generator of target global coupling (Orientation).
The phenomenological Huynen decomposition expresses the measured covariance matrix
[T3 ] as follows
[T3 ] = [T0 ] + [TN ] (73)

The matrix [T0 ] refers to a pure target, that is, a target which can be also completely
characterized by a corresponding scattering matrix. Its parameterization is
⎡ 2 A0 C −j D H + j G ⎤
[T0 ] = ⎢⎢ C + j D B0T + BT ET + jFT

⎥ (74)
⎢⎣ H − j G ET + jFT B0T − BT ⎥⎦

Consequently, [T0 ] presents rank 1. The matrix [TN ] is called N-target (for symmetric
targets) and it corresponds to a distributed scatterer. Since [TN ] does not have rank equal to
1, it does not present an equivalent scattering matrix. The parameterization of [TN ] , given
(73) and (74), is
⎡0 0 0 ⎤
⎢ EN + jFN ⎥⎥
[TN ] = ⎢0 B0 N + BN (75)
⎢⎣0 EN + jFN B0 N − BN ⎥⎦

One on the main properties of the N-target [TN ] is that it is invariant under rotations of the
antenna coordinate system about the line of sight, i.e., it is roll-invariant. Mathematically, this
property can be expressed as

19
Polarimetric Decompositions

−1
⎡⎣TN (θ ) ⎤⎦ = ⎡⎣U 3R ⎤⎦ [TN ] ⎡⎣U 3R ⎤⎦
⎡1 0 0 ⎤ ⎡0 0 0 ⎤ ⎡1 0 0 ⎤
(76)
⎢ ⎥ ⎢
= ⎢0 cos 2θ sin 2θ ⎥ ⎢0 B0 N + BN ⎥ ⎢
EN + jFN ⎥ ⎢ 0 cos 2θ − sin 2θ ⎥⎥
⎢⎣0 − sin 2θ cos 2θ ⎥⎦ ⎢⎣0 EN + jFN B0 N − BN ⎥⎦ ⎢⎣ 0 sin 2θ cos 2θ ⎥⎦

where ⎡⎣TN (θ ) ⎤⎦ has the form

⎡0 0 0 ⎤
⎢ ⎥
⎡⎣TN (θ ) ⎤⎦ = ⎢ 0 B0 N (θ ) + BN (θ ) EN (θ ) + jFN (θ ) ⎥ (77)
⎢ 0 EN (θ ) + jFN (θ ) B0 N (θ ) − BN (θ ) ⎥
⎣ ⎦
As it can be observed, the rotated N-target (77) presents the same structure as the original N-
target (75).

2 A0 B0 + B

(C 2
+ D
2
) 2 A0 (H 2
+ G
2
) 2 A0

20
Polarimetric Decompositions

-30dB -15dB 0dB

2 A0 (C 2
+ D
2
) 2 A0 (H 2
+ G
2
) 2 A0

Figure 8 Elements of the parameterization of the coherency matrix provided by the Huynen decomposition.

4.2.3.2 Barnes Decomposition and Uniqueness of the Hyunen Decomposition


As given by (73), the Huynen decomposition factorizes the measured scattering matrix into a
rank 1 pure target [T0 ] and into a distributed N-target [TN ] . [TN ] is characterized by
presenting a rank larger than 1 and being roll-invariant.
In terms of spaces of vectors, the fact that [TN ] is roll-invariant can be interpreted as the
fact that the vector space generated by [TN ] and the vector space generated by the pure
target [T0 ] are mutually orthogonal. Additionally this orthogonality is maintained under
rotations about the line of sight. Therefore, the question which arises at this point is that
whether the structure proposed by Huynen, (73), is unique or not, in the sense that whether a
different decomposition with the same structure is possible or not.
Given an arbitrary vector q, it belongs to the orthogonal space of the rotated N-target, i.e., the
space generated by the pure target [T0 ] if
−1
⎣⎡TN (θ ) ⎦⎤ q = 0 ⇒ ⎣⎡U 3 ⎦⎤ [TN ] ⎣⎡U 3 ⎦⎤ q = 0
R R
(78)

The condition imposed by (78) is accomplished for any vector q such that
−1
⎡⎣U 3R ⎤⎦ q = λ q (79)
−1
Eq. (79) indicates that q is an eigenvector of the matrix ⎡⎣U 3R ⎤⎦ . This matrix presents the
following three eigenvectors

21
Polarimetric Decompositions

⎡1 ⎤ ⎡0⎤ ⎡0⎤
1 1
q1 = ⎢⎢0⎥⎥ q2 = ⎢1 ⎥ q =
⎢ ⎥
⎢ j⎥
(80)
2⎢ ⎥
3
2
⎢⎣0⎥⎦ ⎢⎣ j ⎥⎦ ⎢⎣ 1 ⎥⎦

Consequently, (78), (79) and (80) show that there exist three ways in which the measured
covariance matrix [T3 ] can be factorized into a pure target [T0 ] and a distributed N-target
[TN ] , as proposed by Huynen in (73).

The Huynen type decomposition, fruit of choosing the vector q1, to generate the orthogonal
space of the distributed N-target, corresponds to the original decomposition proposed by
Huynen, in which the pure target [T0 ] presents the structure given by (74) and the N-target
has the structure given at (75). The normalized target vector corresponding to [T0 ] for q1 has
the following structure
⎡ 2 A0 ⎤
[T ] q1 1 ⎢ ⎥
k01 = = ⎢ C + j D ⎥ (81)
q1T * [T ] q1 2 A0 ⎢
⎣ H − j G ⎥⎦
In this case, the target vector (81) correspond to such targets where Shh+Svv↑0. The
normalized targets vectors corresponding to q2 and q3 are respectively
⎡ C − G + j H −j D ⎤
[T ] q2 ⎢ 1 ⎥
k02 = = ⎢ B0 + B − F + j E ⎥ (82)
q2T * [T ] q22 ( B0 − F ) ⎢ E − j B0 − j B − j F ⎥⎦

⎡ H + D + j C + j G ⎤
[T ] q3 1 ⎢ ⎥
k03 = = ⎢ E + j B0 + j B + j F ⎥ (83)
q3 [T ] q3
T*
2 ( B0 + F ) ⎢ B0 − B + F + j E ⎥⎦

The target vectors in (82) and (83) are associated with helical type scattering behaviors.

22
Polarimetric Decompositions

(C ) +( ) ( B0 ) ( B0 )
2 2 2 2 2 2
− G H − D + B − F + E − B − F + E
2 ( B0 − F ) 2 ( B0 − F ) 2 ( B0 − F )

-30dB -15dB 0dB

(C ) +( ) ( B0 ) ( B0 )
2 2 2 2 2 2
− G H − D + B − F + E − B − F + E
2 ( B0 − F ) 2 ( B0 − F ) 2 ( B0 − F )

Figure 9 Images concerning the Barnes decomposition.

4.2.3.3 Interpretation of the Huynen and Barnes Decompositions


As given by (73), the Huynen decomposition has to be interpreted in terms of the pure target
[T0 ] and the distributed N-target [TN ] . The idea behind the Huynen decomposition is to
extract, from the measured covariance matrix, a scattering mechanism which can be
characterized by a single scattering matrix. The remainder, which is also a true covariance
matrix is considered as a noise component.
Due to the nature of the decomposition itself, it is reasonable to use the Huynen
decomposition to analyze human-made areas. This type or areas are characterized by
presenting a high density of pure targets, which can be studied, then, by the Huynen
decomposition. On the contrary, natural scenes are dominated by distributed scatterers. The

23
Polarimetric Decompositions

structure of the Huynen decomposition tends to consider this type of targets as a noise
component.

4.2.4 Eigenvector-Eigenvalue based Decomposition

4.2.4.1 Description of the Eigenvector-Eigenvalue Decomposition


The eigenvector-eigenvalue based decomposition is based on the eigen decomposition of the
coherency matrix [T3 ] . According to the eigen decomposition theorem, the 3×3 Hermitian
matrix [T3 ] can be decomposed as follows

[T3 ] = [U 3 ][Σ3 ][U 3 ]


−1
(84)

The 3×3, real, diagonal matrix [ Σ3 ] contains the eigenvalues of [T3 ]


⎡λ1 0 0⎤
[Σ3 ] = ⎢⎢ 0 λ2 0 ⎥⎥ (85)
⎢⎣ 0 0 λ3 ⎥⎦
where ∞ > λ1 > λ2 > λ3 > 0 .

The 3×3 unitary matrix [U 3 ] contains the eigenvectors u i for i=1,2,3 of [T3 ]
[U 3 ] = [ u 1 u2 u 3] (86)

The eigenvectors u i for i=1,2,3 of [T3 ] can be formulated as follows


T
u i = ⎡⎣cos α i sin α i cos β i e jδi sin α i cos β i e jγ i ⎤⎦ (87)

Considering the expressions (85) and (86), the eigen decomposition of [T3 ] , i.e., (84), can
be written as follows
3
[T3 ] = ∑ λi u i u *i T (88)
j =1

where the symbol *T


stands for complex conjugate. As (88) shows, the rank 3 matrix [T3 ]
can be decomposed as the combination of three rank 1 coherency matrices formed as
[T3 ]i = u i u *i T (89)
which can be related to the pure scattering mechanisms given at (87).
The eigenvalues (85) and the eigenvectors (86) are considered as the primary parameters of
the eigen decomposition of [T3 ] . In order to simplify the analysis of the physical
information provided by this eigen decomposition, three secondary parameters are defined as
a function of the eigenvalues and the eigenvectors of [T3 ] :

24
Polarimetric Decompositions

• Entropy
3
λi
H = −∑ pi log 3 ( pi ) pi = 3
(90)
i =1
∑λ
k =1
k

where pi, also called the probability of the eigenvalue λi, represent the relative importance
of this eigenvalue respect to the total scattered power, since
3
SPAN = S hh + Svv + 2 S hv = ∑ λk
2 2 2
(91)
k =1

• Anisotropy
λ2 − λ3
A= (92)
λ2 + λ3
• Mean alpha angle
3
α = ∑ piα i (93)
i =1

The eigen decomposition of the coherency matrix is also referred as the H/A/α decomposition.

4.2.4.2 Interpretation of the Eigenvector-Eigenvalue Decomposition


The interpretation of the information provided by the eigen decomposition of the coherency
matrix must be performed in terms of the eigenvalues and eigenvectors of the decomposition
or in terms of H/A/α. Nevertheless, both interpretations have to be considered as
complementary.
The interpretation of the scattering mechanisms given by the eigenvectors of the
decomposition, u i for i=1,2,3, i.e., (87), is performed by means of a mean dominant
mechanism which can be defined as follows
jγ T
u 0 = λ ⎡⎣cos α sin α cos β e jδ sin α cos β e ⎤⎦ (94)

where the remaining average angles are defined in the same way as α
3 3 3
β = ∑ pi β i δ = ∑ piδ i γ = ∑ piγ i (95)
i =1 i =1 i =1

The mean magnitude of the mechanism is obtained as


3
λ = ∑ pi λi (96)
i =1

An example of the mean scattering mechanism can be observed in Figure 10.

25
Polarimetric Decompositions

λ cos (α ) λ sin (α ) cos ( β ) λ sin (α ) sin ( β )

Figure 10 Main scattering mechanism provided by the eigenvector-eigenvalue based decomposition..

The study of the mechanism given in (94) is mainly performed through the interpretation of
the mean alpha angle, since its values can be easily related with the physics behind the
scattering process. The next list reports the interpretation of α:
• α→0: The scattering corresponds to single-bounce scattering produced by a rough
surface.
• α→π/4: The scattering mechanism corresponds to volume scattering.
• α→π/2: The scattering mechanism is due to double-bounce scattering.
The second part in the interpretation of the eigen decomposition is performed by studying the
value of the eigenvalues of the decomposition. A given eigenvalue corresponds to the
associated scattered power to the corresponding eigenvector. Consequently, the value of the
eigenvalue gives the importance of the corresponding eigenvector or scattering mechanism.
The ensemble of scattering mechanisms is studied by means of the entropy H and the
anisotropy A.

26
Polarimetric Decompositions

probability probability
1 3/7
H=0 H=0.91
A=0 A=0.5
1/7

p1 p2 p3 p1 p2 p3

probability probability
1/2 1
H=0.95 H=1
A=0 A=0
1/4
1/3

p1 p2 p3 p1 p2 p3

Figure 11 Entropy (H) and Anisotropy (A) values for four different configurations of the eigenvalues.

The entropy H determines the degree of randomness of the scattering process, which can be
also interpreted as the degree of statistical disorder. In this way:
• H→0:
λ1 = SPAN λ2 = 0 λ3 = 0 (97)
Consequently, the scattering matrix [T3 ] presents rank 1 and the scattering process
corresponds to a pure target.
• H→1:
λ1 = SPAN 3 λ2 = SPAN 3 λ3 = SPAN 3 (98)
In this situation, the scattering matrix [T3 ] presents rank 3, that is, the scattering process
is due to the combination of three pure targets. Consequently, [T3 ] corresponds to the
response of a distributed target.
• 0<H<1: In this case, the final scattering mechanism given by [T3 ] results from the
combination of the three pure targets given by u i for i=1,2,3, but weighted by the
corresponding eigenvalue.
Figure 11 presents four different configurations of the eigenvalues and the corresponding
entropy values.
The anisotropy A, (92), is a parameter complementary to the entropy. The anisotropy
measures the relative importance of the second and the third eigenvalues of the eigen
decomposition. From a practical point of view, the anisotropy can be employed as a source of
discrimination only when H>0.7. The reason is that for lower entropies, the second and third
eigenvalues are highly affected by noise. Consequently, the anisotropy is also very noisy.
Again Figure 11 presents the anisotropy value for four different configurations of the
eigenvalues. In the figure, it can be clearly observe the way anisotropy discriminate two
different configurations presenting the same value of entropy.

27
Polarimetric Decompositions

Entropy H Anisotropy A

0 0.5 1.0 0 0.5 1.0

Average alpha angle α

0 45° 90°

Figure 12 Entropy (H), Anisotropy (A) and alpha (α) images.

28
Polarimetric SAR data Classification

5. POLARIMETRIC SAR DATA CLASSIFICATION

5.1 Classification of polarimetric scattering mechanisms

- Direct interpretation of decomposition results


- Cameron classification
- Lee classification based on Freeman decomposition

5.1.1 H-A-α classification of scattering mechanisms

Cloude and Pottier proposed an algorithm to identify in an unsupervised way polarimetric


scattering mechanisms in the H- α plane. The key idea is that entropy arises as a natural
measure of the inherent reversibility of the scattering data and that α can be used to identify
the underlying average scattering mechanism.

Figure 1 Optical image (left), polarimetric color coded image (right)of the Oberpfaffenhofen test site.

The H- α classification plane is sub-divided into 8 basic zones characteristic of different


scattering behaviors. The basic scattering mechanism of each pixel of a polarimetric SAR
image can then be identified by comparing its entropy and α parameters to fixed thresholds.
The different class boundaries, in the H- α plane, have been determined so as to discriminate
surface reflection (SR), volume diffusion (VD) and double bounce reflection (DB) along the
α axis and low, medium and high degree of randomness along the entropy axis. Detailed
explanations, examples and comments concerning the different classes can be found in the
publication from Cloude and Pottier.
Figure 2 shows the H- α plane and the occurrence of the studied polarimetric data into this
plane.

1
Polarimetric SAR data Classification

α (° )
90

45

H A
0 0.5 1

Figure 2 Polarimetric decomposition main parameters: H, A and α .

Grey zones in the H- α plane correspond to unfeasible areas.


It can be seen, in Figure 3, that the largest densities in the occurrence plane correspond to
volume diffusion and double bounce scattering with moderate and high randomness. Medium
and low entropy surface scattering mechanisms are also frequently encountered in the scene
under examination.

2
Polarimetric SAR data Classification

α(°)
90

80
DR 70
6 3 1
60

50
VD 7 4 2
40

30
8 5
SR 20

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 3 H- α scattering mechanism identification plane (top). Polarimetric data occurrence in the H-
α plane (bottom).

Data distribution in the H- α plane does not show, for the considered scene, distinct natural
clusters belonging to a single scattering mechanism class. Therefore, identification results
may highly depend on segmentation thresholds.
Results of the unsupervised identification procedure are presented in Figure 4.
It can be observed in Figure 4 that the proposed segmentation in the H- α plane permits to
identify in a macroscopic way the type of scattering mechanism. Agricultural fields and bare
soils are characterized by surface scattering. Scattering over forested areas is dominated by
volume diffusion while urban areas are mainly characterized by double bounce scattering. It
may be noted that the identification process slightly overestimates volume diffusion and
double bounce scattering over surfaces.

3
Polarimetric SAR data Classification

Figure 4 Unsupervised scattering mechanism identification in the H- α plane.

5.2 Polarimetric data statistical segmentation

5.2.1 Maximum Likelihood Supervised segmentation

5.2.1.1 Bayes optimal decision rule


The problem of optimal segmentation over a SAR image over a fixed number of clusters may
be formulated as follows.
How to assign, in an optimal way, a SAR image pixel, p , to one of the M possible clusters
{Θ1 ,L, Θ M }, according to a SAR observable, x ?
The answer is given by Bayes optimal decision rule using the conditional probability of the
different clusters
(
Decide p ∈ Θ i if P (Θ i x ) > P Θ j x ) ∀j ≠ i (1)
A pixel is assigned to the most probable cluster conditionally to the observation of x over the
pixel under consideration.
The probability of error related to this decision is obtain from the a posteriori probabilities of
the unselected clusters by
(
P(error x) = ∑ P Θ j x P( x) ) (2)
j ≠i

It is obvious, that assigning a pixel to the cluster with the highest a posteriori probability
minimizes the conditional error probability.
The total probability is then computed using

4
Polarimetric SAR data Classification

P(error x) = ∫ P (error x)P ( x ) dx (3)

The set of a posteriori probabilities P(Θ i x ) is generally uneasy to derive. One prefers then to
express such quantities using Bayes rule
P(Θ i x ) P( x) = P(x Θ i ) P(Θ i ) (4)
Inserting this expression in (1), the expression of the optimal decision rule becomes
Decide p ∈ Θ i if P(x Θ i ) P(Θ i ) > P(x Θ j ) P(Θ j ) ∀j ≠ i (5)

Where P(x Θ i ) is the likelihood of the observation x , given Θ i and P(Θ i ) is the prior
probability of Θ i .

Additionally, if the prior probabilities are supposed to be equal, the optimal decision rule may
written using only cluster likelihood as
Decide p ∈ Θ i if Θ i = Arg max P(x Θ i ) (6)

5.2.1.2 Supervised segmentation scheme


A supervised segmentation scheme may be represented as a three-step process described in
the following.
First training data, separated into M distinct groups or classes, are supplied to the
segmentation algorithm. Training inputs may be :
- user supplied under the correct format
- sampled from the image under study
- sampled from SAR images acquired over other scenes
During the second phase, the segmentation algorithm learns from the training data sets how to
discriminate the M classes and establishes decision rules. In the case of supervised
segmentation based on Bayes optimal decision rule, the segmentation algorithm learns the
different statistical quantities useful to compute the expression mentioned in (6).
Once the learning phase has converged, the algorithm assigns each pixel of the SAR image
under study to one of the M clusters provided in the first phase, according to decision
procedure derived during the second step.

5.2.1.3 Supervised ML segmentation of [S] matrix images


In the case of [S] matrix images, the SAR observation x is replaced by a 3 element complex
target vector k = [ S11 , 2 S12 , S 22 ]T .
It has been verified that when the radar illuminates an area of random surface of many
elementary scatterers, k can be modeled as having a multivariate complex circular gaussian
probability density function N C (0, [Σ]) of the form :

5
Polarimetric SAR data Classification

exp(−k [Σ]−1 k )

P(k ) = (7)
π q [Σ ]
where q stands for the number of elements of k , equal to three in the monostatic case,
represents the determinant, and [Σ] = E ( k k ) is the global (3×3) covariance matrix of k .

The likelihood of a target vector k given a cluster Θ i is then given by

exp(−k [Σ i ] −1 k )

P(k Θ i ) = (8)
π q [Σ i ]
With [Σ i ] the covariance matrix of cluster Θ i computed during the learning phase. In
practice the actual value of [Σ i ] remains unknown and the covariance matrix is replaced by
its maximum likelihood estimate [Σˆ ] defined as i

1
[Σˆ i ] = ∑ kk

(9)
ni p∈Θ i

With ni the number of pixels belonging to the training cluster Θ i .

From a computational point of view, it is generally preferable to deal with log-likelihood


function, L(k Θ i ) instead of the expression given in (8)

(
L(k Θ i ) = − ln [Σˆ i ] − tr [Σˆ i ] −1 k k − q ln π

) (10)

(
Where tr [Σˆ i ] −1 k k

) stands for the trace of [Σˆ ] i
−1
k k and equals k [Σˆ i ]−1 k .
† †

The logarithm function being strictly increasing with its argument, the optimal decision rule is
transformed, in the case of [S] matrix segmentation, to
Decide p ∈ Θ i if Θ i = Arg min d k [Σˆ i ] with ( )
( )
d k [Σˆ i ] = + ln [Σˆ i ] + tr [Σˆ i ] −1 k k( †
) (11)

( )
Where the variable d k [Σˆ i ] may be assimilated to a statistical distance, derived from the
opposite of log-likelihood function without constant terms.
During the segmentation process a pixel is then assigned to the minimum distance cluster,
having the closest statistics.
The supervised segmentation algorithm may be summarized as follows

6
Polarimetric SAR data Classification

Initialize pixel distribution


over M custers
fro m training data sets

For each cluster


1
[Σˆ i ] =
Ni
∑ k k † ∈ Θi

For each pixel, k ∈ Θm if


d (k , Θi ) < d ( k / Θ j ) j = 1,L, M j ≠i

Figure 5 Supervised ML segmentation scheme.

5.2.1.4 Supervised ML segmentation of [T] or [C] matrix images


It has been shown that assuming that target vectors have a N C (0, [Σ]) distribution, a sample
n-look covariance matrix [C ] = 1 / n ∑ k k , follows a complex Wishart distribution with n

degrees of freedom, WC (n, [Σ]) , given by


n−q
n qn [T ] exp(−tr (n[Σ] −1 [T ]))
p ([T ]) =
K (n, q) [Σ]
n

q
with K (n, q) = π q ( q −1) / 2 ∏ Γ(n − i + 1)
i =1

Where Γ ( .) represents the Gamma function.


A development similar to the one presented in the case of scattering matrices leads to the
following decision rule
Decide p ∈ Θ i if Θ i = Arg max L [T ] [Σˆ i ] with ( )
( ) ( )
L [T ] [Σˆ i ] = − n ln [Σˆ i ] − ntr [Σˆ i ] −1 [T ] + qn ln n + (n − q) ln [T ] − ln K (n, q )
(13)

With [Σˆ i ] , the maximum likelihood estimate of the coherency matrix defined as
1
[Σˆ i ] =
ni
∑ [T ]
p∈Θi
(14)

Taking the opposite of the lower expression of relation (13) and removing terms that do not
depend on the cluster under test, the optimal decision rule becomes
Decide p ∈ Θ i if Θ i = Arg min d [T ] [Σˆ i ] with ( )
( ) (
d [T ] [Σˆ i ] = + ln [Σˆ i ] + tr [Σˆ i ] −1 [T ] ) (15)

7
Polarimetric SAR data Classification

It may be noted that, since covariance and coherence matrices are related by a similarity
transformation, the form of the statistical distance expressed in (15) is invariant under a
change from coherency matrix form to covariance matrix form and (15) is adapted to both [T]
and [C] representations.
The supervised segmentation algorithm may be summarized as follows
Initialize pixel distribution
over M custers
fro m training data sets

For each cluster


1
[Σi ] =
ˆ
Ni
∑[T ] ∈ Θi

For each pixel, [T ] ∈ Θm if


d ([T ], Θi ) < d ([T ] / Θ j ) j = 1, L, M j≠i

Figure 6 Supervised ML segmentation scheme.

5.2.1.5 Supervised segmentation performance


The performance of a supervised segmentation process is generally estimated via a confusion
matrix computed from the training data.
Each element of the training data set is processed through the supervised classifier. The
confusion matrix summarizes the percentage of pixels belonging to an original cluster, Θ i ,
effectively affected to one of the M possible clusters, Θ j , with j = 1L M . The confusion
matrix of an ideal, error-free, segmentation is diagonal.
User-defined training data sets are generally oriented towards the discrimination of different
kinds of medium and may not represent all scattering behaviors encountered over a natural
scene. Therefore, some parts of a SAR image may have statistics that do not fit with any of
the proposed cluster ones. In this case, the assignment of the corresponding pixels may lead to
irrelevant interpretation of the segmentation results. One possible solution to this problem
consists in creating a reject class. A pixel is assigned to this reject class if it is too much
distant, from a statistical point of view, from all the learning clusters.
A possible metric may be defined from the statistical distances defined in (11) and (15). A
pixel belongs to the reject class if its distance to the closest cluster exceeds a given factor of
the within-class distance standard deviation.
This decision process may be summarized as follows
(
Decide p ∈ Θ rej if d [T ] [Σˆ i ] > σ di .r )
with Θ i = Arg min d [T ] [Σˆ i ] ( ) (16)

8
Polarimetric SAR data Classification

2
⎛ 1 ⎞
and σ 2
di =
1
Ni
∑ ( )
d [T ] [Σˆ i ] − ⎜⎜
2
∑ ( )
d [T ] [Σˆ i ] ⎟⎟
p∈Θ i ⎝ Ni p∈Θ i ⎠
Where the variable r represents a user-defined rejection factor.

5.2.2 Maximum Likelihood unsupervised segmentation

It has been showed, in former paragraphs that a pixel can be assigned to a cluster in an
optimal way, according to the following decision rule.
Decide p ∈ Θ i if Θ i = Arg max P(x Θ i ) (17)

In the case of an unsupervised segmentation scheme, the likelihood of a cluster, P(x Θ i ) ,


cannot be estimated since its calculation requires to compute the ML estimate of the cluster
coherency matrix from the samples contained in the cluster.
An optimal solution to the unsupervised segmentation problem consists in distributing the
pixels of an image over the set of M clusters so as to maximize the global likelihood defined
as the product of all individual likelihood functions.
A rigorous resolution requires to test all the possible combinations and to select the one
corresponding to the maximum joint likelihood value. This optimal solution cannot be applied
due to the unrealistic computational load it involves.
Alternative solutions based on sub-optimal iterative optimization procedures are generally
preferred.

5.2.3 Principle of statistical K-means clustering

The K-means procedure is an iterative optimization algorithm described by the following


synopsis :

9
Polarimetric SAR data Classification

Initialize pixel distribution


over M clusters

Co mpute each
cluster center

Assign each pixel


to the closests cluster

Convergence ?
No
Yes

Figure 7 Synopsis of the K-means clustering algorithm.

The algorithms begins with the initialization of the image pixel distribution over the M
clusters. This distribution may be done in a random way or according to user specifications.
Once all pixels are affected, the different cluster centers are computed according to the
processed data type. Each pixel is then affected to the closest cluster according to a distance
measure. The convergence of the algorithm is then tested using stability metrics. If a
termination criterion is met, the segmentation stops, otherwise a new iteration starts over from
the class center computation step.
The K-means algorithm aims to optimize a global function by iteratively optimizing local
expressions. It is known this type of techniques may get stuck into locally stable states and
fail to determine the optimal pixel distribution. The initialization of the pixel distribution into
N classes is then a critical stage of the K-mean clustering algorithm. An adequate
initialization permits a fast convergence and provides correctly segmented clusters.
The convergence of the algorithm is evaluated by testing a condition of termination. Such a
criterion may be defined from the estimation of the classification quality, or consist in a
maximum number of iterations or in a sufficiently low number of pixels that are differently
classified from one iteration to the other.

5.2.4 Unsupervised ML segmentation of [T] or [C] matrix images

The K-means clustering algorithm may be adapted to the segmentation of [T] or [C]
polarimetric representations as shown on the following synopsis

10
Polarimetric SAR data Classification

Initialize pixel distribution


over M custers

For each cluster


1
[Σˆ i ] =
Ni
∑[T ] ∈ Θi

For each pixel, [T ] ∈ Θm if


d ([T ], Θi ) < d ([T ] / Θ j ) j = 1, L, M j≠i

Convergence ?
No
Yes

Figure 8 Unsupervised Maximum Likelihood segmentation scheme.

The class center initialization step consists in computing each cluster coherency matrix ML
estimate. Pixels are then affected to a cluster according to the ML distance measure derived in
(15).

5.2.4.1 Unsupervised Wishart H-α segmentation


The particularity of the identification procedure based on H and α , introduced in a former
paragraph, resides in the estimation of the type of observed media from a physical
interpretation of canonical scattering mechanisms using robust indicators. Nevertheless, the
analysis of natural scenes using this unsupervised approach may reach some limitations:
- The arbitrarily fixed linear decision boundaries in the H- α plane may not fit data
distribution. A natural cluster corresponding to similar targets may lie across a frontier in the
decision plane. In this case, pixels with very similar characteristics may be assigned, in an
almost random way, to different classes due slightly different locations in the H- α plane.
This effect can be observed in Figure 4 where the variability in natural media polarimetric
features lead to noisy classification results.
- Even if the computation of H and α requires fully polarimetric data, these two parameters
do not represent the whole polarimetric information. The use of other indicators such as the
span or specific correlations coefficients may improve the classification results in a significant
way.
Segmentation procedures based on the whole coherency matrix statistics permit to overcome
the limitations mentioned above. Nevertheless, it is shown in the following, that the physical
interpretation of the scattering phenomenon permits to enhance in a significant way the
performance of statistical segmentation schemes.

11
Polarimetric SAR data Classification

The unsupervised segmentation scheme mentioned above is initialized in an efficient way


with the results of the unsupervised identification of a scattering mechanism, using H, A and
α . This initialization provides 8 stable clusters relating to the underlying physical scattering
mechanism.

Initialize pixel distribution


over 8 clusters

For each cluster


compute [Σˆ i ] Segmented H-α plane

Assign each pixel


to a cluster Θi

Convergence ?
No
Yes

Figure 9 Unsupervised Wishart H- α segmentation scheme.

C1 C2 C3 C4 C5 C6 C7 C8

Figure 10 Wishart H- α segmentation results.

An important improvement in the segmentation accuracy can be observed in the image


presented in Figure 10. The main kinds of natural media are clearly discriminated by the
Wishart H- α segmentation scheme. This unsupervised classification algorithm modifies the

12
Polarimetric SAR data Classification

decision boundaries in an adaptive way to better fit the natural distribution of the scattering
mechanisms and takes into account information related to the back-scattered power.

5.2.4.2 Unsupervised Wishart H-A-α segmentation


The ML Wishart segmentation may be further improved by explicitly including the anisotropy
information during the segmentation procedure. As mentioned previously, the anisotropy
indicates the relative importance of secondary mechanisms obtained from the expansion of a
coherency matrix. This polarimetric indicator is particularly useful to discriminate scattering
mechanisms with different eigenvalue distributions but with similar intermediate entropy
values. In such cases, a high anisotropy value indicates two dominant scattering mechanisms
with equal probability and a less significant third mechanism, while a low anisotropy value
corresponds to a dominant first scattering mechanism and two non-negligible secondary
mechanisms with equal importance.
Among the different approaches tested, the best way to introduce the anisotropy information
in the classification algorithm consists in implementing two successive segmentation
procedures as shown in Figure 11.

Initialize pixel distribution


over 8 clusters

Segmented H-α plane

A > 0.5 ⇒ Χm → Χm +8
⇒ 16 clusters

Anisotropy

Figure 11 Unsupervised Wishart H-A- α segmentation scheme.

Polarimetric data are first segmented according to the algorithm presented in the former
paragraph. Once this procedure has converged, the 8 resulting clusters are split into 16 ones

13
Polarimetric SAR data Classification

by comparing the anisotropy of each pixel to a threshold fixed to 0.5. The 16 segments are
then used to initialize a second Wishart ML segmentation procedure.

C1 C2 C3 C4 C5 C6 C7 C8

C9 C10 C11 C12 C13 C14 C15 C16

Figure 12 Wishart H-A- α segmentation results.

The segmentation results presented in Figure 12 show an enhanced description of the


Oberpfaffenhofen scene. The introduction of the anisotropy in the clustering process permits
to split large segments into smaller clusters discriminating small disparities in a refined way.
Several kinds of agricultural fields are separated. The runway and other low intensity targets
are distinguished from other surfaces. Buildings are discriminated from other types of
scatterers present in urban areas. The Wishart H-A- α classification scheme gathers into
segments pixels with similar statistical properties, but does not provide any information
concerning the nature of the scattering mechanism associated to each cluster.

14
ENVISAT/ASAR Dual Polarisation Case

6. ENVISAT/ASAR DUAL POLARISATION CASE

6.1 ASAR dual polarization data formats

6.1.1 Supported ASAR data formats

6.1.1.1 Alternating Polarization modes


The ASAR instrument may be configured to acquire data according to one of the Alternating
Polarization (AP) modes.
Classical fully polarimetric SAR devices perform data acquisition with two emission and
reception polarization channels. ASAR, instead, emits signals on a single polarization channel
(H or V) while reception is performed on one or two orthogonal polarizations (H and/or V).
The different possible configurations are summarized in the following figure.

PP Received Emitted
configuration polarizations polarization

PP 1 H

PP2 V

1
ENVISAT/ASAR Dual Polarisation Case

H H

PP3

V V

where PP stands for Partial Polarization or Partially Polarimetric.

6.1.1.2 Supported AP data formats


Partial polarization data sets are delivered under different formats adapted to specific types of
application.
Among the different AP data formats, three are supported by PolSARpro and can be exploited
to provide an interpretation of the polarimetric properties of scattering.
• Alternating Polarization mode Single-look complex (APS) data format

Data type Sets of complex scattering coefficients corresponding to the selected PP


configuration

Geometry Slant Range

Coverage 100 km along-track, 56 - 100 km across-track

Pixel spacing Equal to Resolution : e.g 7.8m * 4m

• Alternating Polarization mode Precision (APP) image format

Data type Images of real reflectivity corresponding to the selected PP configuration

Geometry Ground Range

Coverage 100 km along-track, 56 - 100 km across-track

Pixel spacing 12.5 m * 12.5 m

2
ENVISAT/ASAR Dual Polarisation Case

• ASAR Alternating Polarization mode ellipsoid Geocoded (APG) image

Data type Images of real reflectivity corresponding to the selected PP configuration

Geometry Geocoded Ground Range

Coverage 100 km * 100 km

Pixel spacing 12.5 m * 12.5 m

6.1.2 Relation with fully polarimetric representations

Fully polarimetric quantities are measured from two acquisitions performed with orthogonal
polarization states (generally H and V) in order to form each pixel scattering matrix [S]
The coherent scattering matrix modifies the polarization of an incident wave according to the
following expression
E s = [S ] E i (1)

The scattering matrix is built from two acquisitions as follows


⎡S S hv ⎤
[ S ] = ⎢ hh ⎥ = [E sh E sv ]
⎣ S vh S vv ⎦
(2)
⎡S S hv ⎤ ⎡1⎤ ⎡ S hh ⎤ ⎡ S hh S hv ⎤ ⎡0⎤ ⎡ S hv ⎤
= ⎢ hh ⎥ = ⎢ ⎥ = ⎢S =
S vv ⎦ ⎢⎣0⎥⎦ ⎣ S vh ⎦ S vv ⎥⎦ ⎢⎣1⎥⎦ ⎢⎣ S vv ⎥⎦
with E sh and E sv
⎣ S vh ⎣ vh
where E sh and E sv correspond to the scene response to normalized incident horizontally and
vertically polarized waves respectively.

6.1.2.1 APS format


APS format images are composed of coherent scattering coefficients and my then be
expressed directly from the expression given in (2) according to the selected PP configuration
• PP1
A pixel is represented by a Jones vectors formed with two of the scattering matrix coefficients
⎡S S hv ⎤ ⎡1⎤ ⎡ S hh ⎤
E PP1 = ⎢ hh =
S vv ⎥⎦ ⎢⎣0⎥⎦ ⎢⎣ S vh ⎥⎦
(3)
⎣ S vh

• PP2
A pixel is represented by a Jones vectors formed with two of the scattering matrix coefficients

3
ENVISAT/ASAR Dual Polarisation Case

⎡S S hv ⎤ ⎡0⎤ ⎡ S hv ⎤
E PP 2 = ⎢ hh =
S vv ⎥⎦ ⎢⎣1⎥⎦ ⎢⎣ S vv ⎥⎦
(4)
⎣ S vh

• PP3
A pixel is represented by a two element complex vector formed with two of the scattering
matrix coefficients
⎡1⎤
[S hh S hv ] ⎢ ⎥
E PP 3 = ⎣0⎦ = ⎡ S hh ⎤
⎡0 ⎤ ⎢ S ⎥
(5)
[S vh S vv ] ⎢ ⎥ ⎣ vv ⎦
⎣1⎦
One may note that, oppositely to the vector E s , the Jones vectors E sh and E sv are received in
a polarization basis.
It is then possible to represent E sh and E sv into any desired polarization basis using classical
Special Unitary transformation matrices

E = [U (φ ,τ , α )] E (6)
Such a transformation may also be applied to E s in PP3 configuration, but without the
physical interpretation associated to the formalism of changes of polarization basis. This
might be particularly useful to perform a transformation similar to the one linking [C] and [T]
matrices

′ 1 ⎡ 1 1⎤ 1 ⎡ S hh + S vv ⎤
Es = ⎢− 1 1⎥ E s = ⎢ ⎥ (7)
2⎣ ⎦ 2 ⎣ S vv − S hh ⎦

6.1.2.2 APP and APG formats


APP and APG format images are composed of intensities and my then be expressed from the
square moduli of the Sinclair matrix elements, according to the selected PP configuration
• PP1
⎡S 2

I PP1 = ⎢ hh 2⎥ (8)
⎢⎣ S vh ⎥⎦

• PP2
⎡ S vv 2 ⎤
I PP 2 =⎢ 2⎥ (9)
⎣⎢ S hv ⎥⎦

• PP3

4
ENVISAT/ASAR Dual Polarisation Case

⎡S 2

I PP 3 = ⎢ hh 2 ⎥ (10)
⎢⎣ S vv ⎥⎦
One may note that APP and APG mode data do not exactly correspond to intensities derived
from equivalent APS data, due to the slant range to ground range mapping transformation.

6.2 Speckle filtering

6.2.1 APP and APG format data filtering

As it was seen in the former paragraph, APP and APG format data consist of incoherent
intensity values that are fit to incoherent speckle filtering.
The formulation of the boxcar and Lee filters can be adapted to intensity variables.

6.2.1.1 APP and APG boxcar filter


~
Filtered intensity vector estimates, I PPi , are constructed by computing the sample mean over
each pixel neighborhood, defined by a sliding window of ( N w × N w ) pixels
~
I PPi = I PPi Nw (11)

6.2.1.2 APP and APG Lee filter


The adaptation of Lee MMSE filter to APP and APG data formats leads to the following
intensity vector estimate expression
~
(
I PPi = I PPi N + k I PPi − I PPi N
w w
) (12)

where k is an adaptive filtering coefficient based on local statistics and given by


var I − E 2 ( I )σ n2
k=
var I [1 + σ n2 ]
(13)
1
with σ = the a priori speckle variance
2
n
L
where I represents the sum of the components of the vector I PPi .
The two components of the intensity vector I PPi may be filtered independently by applying
separate scalar filters.

6.2.2 APS format data filtering

APS format data consist of coherent scattering coefficients that cannot be directly used as
inputs of a speckle filter.

5
ENVISAT/ASAR Dual Polarisation Case

Similarly to the case of target vectors filtering, a incoherent covariance matrix may be built
from the APP coherent vector, E PPi , as

[C 2 ] PPi = E PPi E PPi



(14)
The express of the boxcar and Lee filters are similar to those introduced for the (3 × 3) [C]
and [T] matrices.

6.2.2.1 APP and APG boxcar filter


~
Filtered estimates, [C 2 ] PPi , are constructed by computing the sample mean over each pixel
neighborhood, defined by a sliding window of ( N w × N w ) pixels
~
[C 2 ] PPi = [C 2 ] PPi Nw (15)

6.2.2.2 APP and APG Lee filter


The adaptation of Lee MMSE filter to APP and APG data formats leads to the following
intensity vector estimate expression
~
( )
[C 2 ] PPi = [C 2 ] PPi N + k [C 2 ] PPi − [C 2 ] PPi N
w w
(16)

where k is an adaptive filtering coefficient, based on local statistics, given by


var span − E 2 ( span )σ n2
k=
var span [1 + σ n2 ]
(17)
1
with σ = the a priori speckle variance
2
n
L
where span represents the trace of the covariance matrix.

6.3 Incoherent decomposition

Averaged [C 2 ] matrices do not correspond to an PP scattering mechanism and may the be


decomposed in order to extract a pure PP representation.
A [C 2 ] may be decomposed on its eigenvector basis as follows

[C 2 ] = [V2 ][Λ 2 ][V2 ]† = λ1v1v1† + λ2 v2 v2† (18)


where [ Λ 2 ] and [V2 ] represent ( 2 × 2) real eigenvalue and special unitary eigenvector
matrices respectively.
Each unitary eigenvector, vi , may be parameterized using 3 real parameters

vi = e jξ [cosα i sin α i e jδ i ]T (19)


The eigenvalue set may also be parameterized using two real variables
λ1 + λ 2 = Tr[C 2 ] = span and H = − P1 log 2 ( P1) − P 2 log 2 ( P2 ) (20)

6
ENVISAT/ASAR Dual Polarisation Case

λi
with Pi =
λ1 + λ2
where H represents the scattering mechanism entropy while Pi is the pseudo-probability of
one of the decomposed orthogonal scattering mechanisms, by convention P1 ≥ P2 .
The anisotropy, A, may also be used to characterize the scattering phenomenon
P1 − P2
A= (21)
P1 + P2
One may note that the joint use of H and A is redundant since the characterization of a two
probability set requires only a single real parameter.
The physical signification of H and A is identical to the one derived for the decomposition of
(3 × 3) [C] and [T] matrices.
Similarly to the (3 × 3) case, an average scattering mechanism may be rebuilt from the pseudo-
probability set as
(α , δ ) = P1 (α 1 , δ 1 ) + P2 (α 2 , δ 2 ) (22)
On the opposite, the interpretation of (α , δ ) is totally different from one PP configuration to
the other.

6.4 Supervised segmentation

It has been demonstrated, in the chapter dedicated to polarimetric SAR data classification, that
statistical supervised classification of data requires the derivation of data ML statistics.

6.4.1 APS format data segmentation

6.4.1.1 Segmentation of coherent data


APS format consist of complex two-element vector composed of coherent scattering
coefficients and may be represented, in a general way, as follows
⎡S ⎤
E = ⎢ pq ⎥ (23)
⎣ S rs ⎦
where the subscripts p, q, r, s represent polarization channels.
It has been verified that when the radar illuminates an area of random surface of many
elementary scatterers, E can be modeled as having a multivariate complex circular gaussian
probability density function N C (0, [Σ]) of the form

exp(− E [Σ] −1 E )

P (k ) = (24)
π q [Σ ]

7
ENVISAT/ASAR Dual Polarisation Case

where q stands for the number of elements of E , equal to three in the monostatic case,
represents the determinant, and [Σ] = E ( E E ) is the global (3×3) covariance matrix of E .

The likelihood of a target vector E given a cluster Θ i is then given by

exp(− E [Σ i ] −1 E )

P (E Θ i ) = (25)
π q [Σ i ]
With [Σ i ] the covariance matrix of cluster Θ i computed during the learning phase. In
practice the actual value of [Σ i ] remains unknown and the covariance matrix is replaced by
its maximum likelihood estimate [Σˆ ] defined as
i

1
[Σˆ i ] = ∑ EE

(26)
ni p∈Θi

With ni the number of pixels belonging to the training cluster Θ i .

From a computational point of view, it is generally preferable to deal with log-likelihood


function, L(E Θ i ) instead of the expression given in Erreur ! Source du renvoi introuvable.

(
L(E Θ i ) = − ln [Σˆ i ] − tr [Σˆ i ]−1 E E − q ln π

) (27)

The logarithm function being strictly increasing with its argument, the optimal decision rule is
transformed, in the case of E vector segmentation, to :

Decide p ∈ Θ i if Θ i = Arg min d E [Σˆ i ] with ( )


( ) (
d E [Σˆ i ] = + ln [Σˆ i ] + tr [Σˆ i ]−1 E E

) (28)

( )
Where the variable d E [Σˆ i ] may be assimilated to a statistical distance, derived from the
opposite of log-likelihood function without constant terms.
The supervised segmentation algorithm may be summarized as follows

8
ENVISAT/ASAR Dual Polarisation Case

Initialize pixel distribution


over M custers
fro m training data sets

For each cluster


1
[Σˆ i ] = ∑ E E ∈ Θi

Ni

For each pixel, E ∈ Θm if


d ( E, Θ i ) < d ( E / Θ j ) j = 1,L, M j≠i

Figure 1 Supervised ML segmentation scheme

6.4.1.2 Segmentation of incoherent data

It has been shown that assuming that E vectors have a N C (0, [Σ]) distribution, a sample n-
look covariance matrix [C 2 ] = 1 / n ∑ E E † , follows a complex Wishart distribution with n
n

degrees of freedom, WC (n, [Σ]) , given by


n−q
n qn [C 2 ] exp(−tr (n[Σ] −1 [C 2 ]))
p ([C 2 ]) = n
K ( n, q ) [ Σ ]
(29)
q
with K (n, q) = π q ( q −1) / 2 ∏ Γ(n − i + 1)
i =1

Where Γ( .) represents the Gamma function.


A development similar to the one presented in the case of coherent vectors leads to the
following decision rule
Decide p ∈ Θ i if Θ i = Arg max L [C 2 ] [Σˆ i ] with ( )
( ) ( )
L [C 2 ] [Σˆ i ] = −n ln [Σˆ i ] − ntr [Σˆ i ] −1 [C 2 ] + qn ln n + (n − q) ln [C 2 ] − ln K (n, q)
(30)

With [Σˆ i ] , the maximum likelihood estimate of the coherency matrix defined as
1
[Σˆ i ] =
ni
∑ [C
p∈Θ i
2 ] (31)

Taking the opposite of the lower expression of relation (30) and removing terms that do not
depend on the cluster under test, the optimal decision rule becomes :
Decide p ∈ Θ i if Θ i = Arg min d [C 2 ] [Σˆ i ] with ( )
d ([C 2 ] [Σˆ i ] = + ln [Σˆ i ] + tr [Σˆ i ]−1 ( [C ])2
(32)

9
ENVISAT/ASAR Dual Polarisation Case

The supervised segmentation algorithm may be summarized as follows


Initialize pixel distribution
over M custers
fro m training data sets

For each cluster


1
[Σˆ i ] =
Ni
∑ [C2 ] ∈ Θi

For each pixel, [C2 ] ∈ Θ m if


d ([C2 ], Θi ) < d ([ C2 ] / Θ j ) j = 1,L , M j≠i

Figure 2 Supervised ML segmentation scheme

6.4.2 APP and APG data segmentation

The segmentation presented is simplified by assuming that APP and APG intensities are
independent. In this case, their joint probability is given by
n−q
n qn [C I 2 ] exp(−tr (n[Σ I ] −1 [C I 2 ]))
p ([C I 2 ]) =
K ( n, q ) [ Σ I ]
n

(33)
q
with K (n, q) = π q ( q −1) / 2 ∏ Γ(n − i + 1)
i =1

Where [C I 2 ] and [Σ I ] are ( 2 × 2) real matrices with off-diagonal elements equal to zero and
built as follows
⎡I 0⎤ ⎡ E (I ) 0 ⎤
[C I 2 ] = ⎢ 1 ⎥ and [Σ I ] = ⎢ 1
E (I 2 )⎥⎦
(34)
⎣0 I2 ⎦ ⎣ 0
The optimal decision rule for APP and APG format data is then given by
(
Decide p ∈ Θ i if Θ i = Arg min d [C I 2 ] [Σˆ I i ] with )
] + tr ([Σˆ ) (35)
d ([C I 2 ] [Σˆ I i ] = + ln [Σˆ I i Ii ]−1 [C I 2 ]

10
1

POL-InSAR TRAINING COURSE


Shane R. CLOUDE
AEL Consultants
26 Westfield Avenue, Cupar, Fife, KY15 5AA
Scotland, UK
Tel/Fax : +44 1334 650761
e-mail : aelc@mac.com, web : http://web.mac.com/aelc

1 Introduction..................................................................................................................................2
2 Background Theory .....................................................................................................................3
3 Algorithms for Optimum Interferogram Generation ...................................................................6
4 POLInSAR for Bare Surface scattering .......................................................................................9
5 POLInSAR for Random Volume scattering ..............................................................................10
6 POLInSAR : 2-layer combined surface and random volume scattering....................................11
7 Algorithms for Vegetation Parameter Retrieval ........................................................................15
8 Forest Height Inversion Algorithm ............................................................................................16
8.1 DEM differencing ..............................................................................................................16
8.2 Height compensated for extinction ....................................................................................17
8.3 Height Compensated for Vertical Structure.......................................................................18
8.4 Height from Coherence Amplitude only............................................................................18
8.5 Robust Inversion Accounting for Extinction/Vertical Structure........................................19
9 POLInSAR Data Processing ......................................................................................................20
9.1 Step 1 : Reading POLInSAR Data Files ............................................................................20
9.2 Step 2 : Generating an Interferogram.................................................................................22
9.3 Step 3 : Flat Earth Removal ...............................................................................................23
9.4 Step 4: Vertical Wavenumber Estimation..........................................................................23
9.5 Step 5: Complex Coherence Estimation ............................................................................26
9.6 Step 6: Coherence Bias and Convergence .........................................................................27
9.7 Step 7: Algorithm 1 : DEM differencing ...........................................................................28
9.8 Step 8: Algorithm 2 : Coherence Amplitude Inversion .....................................................31
9.9 Step 9 : Ground Phase estimation using dual polarisations ...............................................32
9.10 Step 10: Algorithm 3: Phase and Coherence Inversion .....................................................33
9.11 Step 11: Polarisation Selection ..........................................................................................36
9.12 Step 12: Generalisation to the Coherence Loci..................................................................37
10 Conclusions............................................................................................................................39
11 POLInSAR Bibliography : August 2005 ...............................................................................39
2

1 INTRODUCTION
In this course you will learn about a radar remote sensing technique called polarimetric
interferometry [Cloude 1997,1998,2004, Papathanassiou 1998]. When used with an imaging
synthetic aperture radar (SAR) system, it is usually termed Polarimetric Interferometric SAR or
POLInSAR for short [Papathanassiou 2001]. POLInSAR has important applications in the remote
measurement of vegetation properties such as forest height [Papathanassiou 2005] and biomass
[Mette 2004] and developing future applications in agriculture [Williams 2005, Preiss 2005],
snow/ice thickness monitoring [Dall 2003, Papathanssiou 2005] and urban height and structure
applications [Schneider 2005]. As its name suggests, the technique combines two separate radar
technologies, polarimetry and interferometry. The former involves switching the polarisation state
of transmit and receive channels to measure differences in backscatter due to orientation, shape and
material composition [Cloude 1996, Pottier 2005]. This leads ultimately to measurement of the 2x2
complex scattering matrix [S], from which we can synthesise the response of the image pixel to
arbitrary polarisation combinations. Such radars are termed S-matrix or Quadpol systems (since
four complex channels are measured by the radar, usually all possible combinations of horizontal
and vertical linear polarisations HH, HV, VH and VV respectively)
On the other hand, radar interferometry [Bamler 1998] involves coherently combining
signals from two separated spatial positions (defining the so called baseline of the interferometer) to
extract a phase difference or interferogram. In radar this can be achieved in two main
configurations, so called along-track interferometry, which involves time displacements between
separated antennas along the flight direction of the platform leading to velocity estimation.
Alternatively we can perform across-track interferometry, involving lateral separation of antennas
and leading to spatial information relating to the elevation of the scatterer above a reference ground
position. In POLInSAR interest centre mainly on across-track geometries but in principle it can be
applied to along track configurations as well.
POLInSAR differs from conventional interferometry in that it allows generation of
interferograms for arbitrary transmit/receive polarisation pairs. It turns out that the phase of an
interferogram changes with the choice of polarisation and consequently we can extract important
bio and geophysical parameters by interpreting this change in the right way. We shall see that
consequently the combination of interferometry with polarimetry is greater than the sum of its parts
and that POLInSAR allows us to overcome severe limitations of both techniques when taken alone.
This is especially true in the important area of remote sensing of vegetated land surface, where
polarimetry suffers from the inherent high entropy problem [Cloude 1996] while standard
interferometry remains underdetermined i.e. the interferogram depends on many possible physical
effects, no one of which can be identified from the data itself [Treuhaft 1996, 2000]. POLInSAR
offers a window on a new way to overcome these limitations. In this course we develop the theory
step by step, illustrating each stage with processed data and allowing the user to repeat the
processing stages
The course begins with a review of the basis theory and notation before running through a
12 step practical training course covering most of the major steps involved in processing and
analysis of POLInSAR data.
3

2 BACKGROUND THEORY
POLInSAR algorithms make use of signal coherence (or equivalently phase and local phase
variance) rather than backscattered power [Zebker 1992, Hagberg 1995,Askne 1997, Touzi 1999].
For this reason we begin with a review of the techniques and problems associated with the
estimation of coherence from radar data. We follow closely the approach developed in [Touzi
1999]. Starting with any two co-registered single look complex (SLC) data channels s1 and s2 the
coherence is formally defined as shown in equation 1
E(s1s*2 )
γ˜ = γe iφ = - 1)
E(s1s1* ). E(s2 s*2 )

where E(..) is the expected value and 0 ≤ γ ≤ 1. In practice the sample coherence is frequently used
as a coherence estimate of 1, as shown in equation 2
L

∑s s *
1i 2i

δ˜ = δe iχ = L
i=1
L
- 2)
∑s s . ∑s*
1i 1i s *
2i 2i
i=1 i=1

where i is the sample number and we have only a finite number L independent signal measurements
available. Equation 2 represents the maximum likelihood (ML) estimate of coherence and under
some general statistical assumptions provides an estimate that is asymptotically unbiased. For
jointly complex Gaussian processes (s1,s2) the probability density function (pdf) of δ can then be
derived as a function of the true coherence value γ and the number of samples L as shown in 3
[Touzi 1999]

p(δ γ ) = 2(L −1)(1− γ 2 ) L δ(1− δ 2 ) L−2 F(L,L :1: δ 2γ 2 ) - 3)

where F is a special mathematical function called the hypergeometric function [Touzi 1999]. More
significant for POLInSAR is estimation of the bias in the coherence magnitude, derived from 3 as
the first moment of δ and shown in equation 4

Γ(L)Γ(1+ 12 )
E(δ ) = 1
3 F2 (3/2,L,L;L + 2 ;1;δ )(1− δ )
2 2 L
- 4)
Γ(L + 12 )

where pFq is the generalised hypergeometric function. The behaviour of this function can also be
obtained through straightforward numerical simulation of Gaussian distributed signals (so avoiding
the need to calculate pFq), shown for example in figure 1. Here we see that the estimate is
consistently biased towards higher values (in the extreme case of 1–look estimation the coherence
estimate is equal to unity and so always overestimated). However, importantly for us, the bias
decreases with increasing number of independent samples L and with increasing underlying
coherence γ.
A second important consequence of equation 3 is estimation of the variance of the sample
coherence magnitude. This is required to assess the precision of parameters estimated from the
coherence and impacts on the accuracy of derived products such as vegetation height. The most
4

accurate way to estimate this variance is as shown in equation 5. Note that the second term can be
estimated from equation 3 while the first is shown in equation 6

var(δ ) = E(δ 2 ) − E (δ ) var(δ ) = E (δ 2 )− E (δ )


2 2
- 5)
Γ(L)Γ(1+ 1)
E(δ 2 ) = 3 F2 (2,L,L;L + 1;1;δ )(1− δ )
2 2 L
- 6)
Γ(L + 1)

Similar considerations lead us to the following expression for the pdf of the corresponding
interferometric phase, shown in 7

Γ(L + 12 )(1− γ 2 ) L β (1− γ 2 ) L 1 ⎧ −π < χ ≤ π


p( χ | φ, γ ) = 1 + F(L,1; ;β 2 ) where ⎨ -7)
L+ 2π 2 ⎩β = γ cos( χ − φ )
2 π Γ(L)(1− β )2 2

which can be used to formally estimate the variance of the phase estimate as shown in 8
π
varφ = ∫χ 2
p( χ | φ, γ )dχ - 8)
−π
For a small number of looks, the hypergeometric functions can be replaced by simpler
trigonometric functions [Lee 1994], but for L > 4, as generally required for POLInSAR
applications, the full calculation is required. While these equations provide the most accurate
method of assessing bias and complex coherence variance, often we assume zero bias (by using
sufficient averaging) and estimate the variance by making use of simpler equations for speedier
computation.

L= 4

L= 8

L= 50

Figure 1: ML Coherence Bias as a function of coherence and number of looks

In particular, the Cramer-Rao bounds provide lower limits on the variance for coherence and phase
and have been derived in [Seymour 1994] to provide the simpler formulae shown in equation 9

(1− γ 2 ) 2 1− γ 2
varγ = ≤ var(δ ) varφ = ≤ var( χ ) - 9)
2L 2Lγ 2
5

In conclusion we note that for phase based processing, it is always better to operate at high
coherence and avoid low coherences, the latter involving not only increased variance but also
severe bias issues that can distort the phase information. It is a key limitation of polarimetry that
scattering by vegetation leads to low coherences for all polarization channels (because of so called
depolarization). This severely limits our ability to use polarimetric phase information over
vegetated land surfaces. Interferometry on the other hand allows us to partially control coherence
via baseline selection. POLInSAR exploits this advantage to obtain high coherence in multiple
polarization channels.
The above considerations for coherence estimation are important in POLInSAR, the major
distinguishing feature of which is that we add an extra stage in the construction of the two SLC
channels s1 and s2. In general, for a QUADPOL data set, we take as input the three calibrated SLC
images shh, shv and svv and generate projections of these onto user defined complex weight vectors
w1 and w2 before calculating the coherence defined as shown in equation 10

(s1hh + s1vv ) (s1 − s1 ) ⎫


s1 = w11 + w12 hh vv + w13 2s1hv = w1T .k1 ⎪⎪
2 2 ⎬
(s 2
+ 2
) (s 2
− svv2 )
+ w 2 2shv = w 2 .k 2 ⎪⎪
s
s2 = w 2
1
+ w22 3 2
hh vv hh T

2 2 ⎭ - 10)
E(s1s*2 )
⇒ γ˜ (w1,w 2 ) =
E(s1s1* ). E(s2 s*2 )

The weight vectors w1 and w2 define user selected scattering mechanisms at ends 1 and 2 of the
across-track baseline. In general w1 and w2 can be different and both parameterised as complex
unitary vectors of the form shown in 11 [Cloude 1996, 1998]

w = [w1 w 3 ] = [cos α sin α sin βe iμ ]


T
w2 sin α cos βe iε - 11)

Table I shows important examples of the weight vectors for coherence estimation in the commonly
used linear, Pauli and circular bases. This table can be used together with equation 10 to generate
interferograms in different polarisation channels. However, it is a feature of POLInSAR algorithm
development that use is often made of more general w vectors than those shown, derived for
example as eigenvectors for coherence optimisation [Cloude 1997, Tabb 2001,2002, Colin 2003,
Gomez Dans 2005] or through a prior model studies of scattering from vegetated terrain [Williams
1999,2000]. For this reason we need to keep the more general notation of equation 11 so as to be
able to consider arbitrary vectors in the formation of an interferogram. We now turn to consider
such optimisation algorithms in more detail and to briefly assess their implications for coherence
estimation and validation.

Polarisation α β ε μ 
Selection
HH 45o 0o 0 0o 0.707 0.707 0
HV 90o 90 o 0 0o 0 0 1
VV 45o 180o 0 0o 0.707 -0.707 0
HH+VV 0o 0o 0 0o 1 0 0
HH-VV 90o 0o 0 0o 0 1 0
LL 90o 45 o 0 90o 0 0.707 0.707i
LR 0o 0o 0 0o 1 0 0
RR 90o 45 o 0 -90o 0 0.707 -0.707i
Table I: Example scattering mechanisms used for POLInSAR
6

3 ALGORITHMS FOR OPTIMUM INTERFEROGRAM GENERATION


Polarimetric Interferometry is a special case of multi-channel coherent radar processing [Reigber
2000]. Such problems are characterised by multi-dimensional covariance matrices [Lee 1994, 1999,
2003]. In polarimetric SAR (POLSAR) for example, interest centres on the 3 x 3 hermitian
coherency matrix [T], unitarily equivalent to the covariance matrix [C] as employed in multivariate
statistical analyses [Lee 1999]. This is the basic building block in polarimetric interferometry and so
we designate this matrix as Λ 1 to indicate how it relates to fully polarimetric measurements but
made at only 1 spatial position. In single baseline POLInSAR we then add a second measurement at
a displaced position 2. This is now characterised by a 6 x 6 coherency matrix Λ 2 as shown in
equation 12. We see that this 6 x 6 matrix can be naturally partitioned into 3 sub matrices each of
size 3 x 3. This formulation then scales in a natural way for multi-baseline POLInSAR by
expansion of the governing coherency matrix Λ N to a 3N x 3N complex system (4N x 4N for
bistatic multi-baseline POLInSAR) as shown in 12
⎡ T1 Ω12 K Ω1N ⎤
⎡ T1 Ω12 ⎤ ⎢ ⎥
⎢ Ω12
*
T2 L Ω 2N ⎥
. Λ1 = [T] → Λ 2 = ⎢ * ⎥ → ΛN = - 12)
⎣Ω12 T2 ⎦ ⎢ M M O M ⎥
⎢ * ⎥
⎣Ω1N Ω2N L TN ⎦
*

Returning now to the important case of Λ2, two of the sub-matrices, T1 and T2 are Hermitian and
relate to the polarimetry from positions 1 and 2 while the third Ω12 is a complex 3 x 3 matrix that
contains information about the variation of interferometric coherence and phase for all possible
weight vectors w1 and w2 as shown in equation 13.

⎡ T11 Ω12⎤ *T
w1 Ω12 w 2
Λ 2 = ⎢ *T ⎥ ⇒ γ˜(w1,w 2 ) = - 13)
⎣Ω12 T22 ⎦ w1*T T11 w1 . w *T
2 T22 w 2

This relation leads to an important choice of approach to algorithm development in POLInSAR. In


the first case if we know the vectors w1 and w2 in advance, then we can directly estimate the
coherence using equation 10 with the same InSAR fluctuation statistics and bias outlined in 4 and 8.
However, often we wish to determine ‘optimum’ weight vectors from the data itself and it follows
from 13 that to do this we require estimates of the three 3 x 3 matrices T11, T22 and Ω12. This opens
up a much wider discussion about the fluctuation statistics and bias arising from the fact that only
estimates and not true matrix values can be used in 13. For example, to estimate the submatrices we
must first estimate the full 6 x 6 coherency matrix Λ 2. This estimate, Z, is obtained as a
straightforward extension of equation 2 and is shown in 14, where we have L independent S-matrix
(and hence scattering vectors u) available [Lee 1999].

1 L
[Z ] = ∑ u j u*Tj - 14)
L j =1

For Gaussian statistics, this estimate [Z] of Λ follows a complex Wishart distribution [Lee 1999], so
that its pdf is given for the general q dimensional case by equation 15
7

LL det([Z]) L − q exp(−L.Trace([Λ]−1 [Z ]))


pN ([Z] | Λ) =
K(L,q)det([Λ]) L - 15)
K(L,q) = π 0.5q(q −1)
Γ(L )..Γ(L - q +1)

For interferometry we have seen q =2 leads to the statistics shown in 4 and 8. For S matrix
polarimetry q = 3, while for single baseline POLInSAR q = 6. One key point to note is that the
minimum number of data samples (Lmin) required for adequate estimation of covariance matrices in
multi-variate problems increases with the dimensionality q. Hence there are ever increasing
demands on the number of looks required to obtain good estimates of derived products from multi-
dimensional coherency matrix analysis.
One important application of this approach is the calculation of the optimum coherences in
POLInSAR. The most general formulation of this was first presented in [Cloude 1997, 1998] and is
summarised in equation 16. Here we first state the problem mathematically, which is to choose w1
and w2 so as to maximise the coherence magnitude, defined from the complex coherence as a
function of the three sub-matrices T11 T22 and Ω12 as shown. This can be mathematically solved by
using a Lagrange multiplier technique as shown and leads to the calculation of the required w
vectors as eigenvectors of a pair of matrices, themselves defined as products of the composite
matrices. Hence in order to calculate these we require first estimation of the 6x6 matrix Λ2.
*T
w1 Ω12 w 2
max
*T *T
w1 w 2
w1 T11 w1 .w 2 T22 w 2

(
L = w1 Ω12 w1 + λ1 w1 T11 w1 −1 + λ2 w 2 T22 w 2 −1
*T *T
) ( *T
)
- 16)
⎧ ∂L
⎪⎪ ∂ w *T = Ω12 w 2 + λ1T11 w1 = 0
⇒⎨ 1
⎪ ∂L*T = Ω12*T
w1 + λ2T22 w 2 = 0
⎪⎩ ∂ w 2
⎧T −1Ω*T T −1Ω w 2 = λ1λ*2 w 2
⇒ ⎨ 22−1 12 11−1 12
⎩ T11 Ω12T22 Ω12 w1 = λ1λ2 w1
*T *

To illustrate the increased sample requirements we show a simple example based on numerical
estimation of the eigenvalues and eigenvectors in equation 16 for an underlying coherency matrix
Λ2 of the form shown in 17

⎡ 1 0 0 0.9e
i π4
0 0 ⎤
⎢ iπ

⎢ 0 1 0 0 0.6e 3
0 ⎥
⎢ 0 0 1 0 0 0.4e 2 ⎥

[Λ 2 ] = ⎢ −i π4 ⎥ - 17)
⎢0.9e 0 0 1 0 0 ⎥
⎢ 0 0.6e
−i π3
0 0 1 0 ⎥
⎢ −i π ⎥
⎣ 0 0 0.4e 2 0 0 1 ⎦

which we see has a trivial solution with eigenvectors w1 = w2 equal to the basis vectors (1,0,0),
(0,1,0) and (0,0,1) respectively and the optimum triplet of coherences are 0.9 0.6 and 0.4. If we now
8

use this matrix to generate complex Gaussian noise with the same underlying coherency matrix and
use 16 to estimate the eigenvalues for varying number of looks then we obtain the results shown in
figure 2. Here we plot the coherence estimates as a function of increasing number of looks. We
show two sets of curves. In black we show the estimates based on a prior knowledge of the w
vectors, using the projection and coherence estimation of equation 10. We see good convergence
with the absence of any significant bias (each value on the curve is obtained as the mean of 256
realisations of the process so as to reduce the variance and hence expose any underlying bias
issues). In the second case we show estimates based on the assumption of no a prior knowledge of
the vectors i.e. when the vectors themselves must also be estimated from the raw data. Here we see
a much slower convergence onto the true values with significant bias for a low number of looks. We
see that the first 2 coherences are consistently overestimated while the third is underestimated. Only
after a large number of looks does the bias reduce. For example for 10% coherence accuracy we
require in excess of 40 looks, even for high coherence values. However, with care and sufficient
averaging it has been shown that such optimisation can lead to better phase estimation for digital
elevation model (DEM) generation [Nico 2000 ]

Figure 2 : Simulated Example of Optimum Coherence Estimation based on a prior knowledge of w


vectors (black) and vectors derived from eigenvalue estimation (colour)

In order to obtain an optimization approach that has less bias for a given number of samples, it is
necessary to reduce the effective dimensionality of the problem. Several authors have proposed
adopting the a prior assumption that w1 = w2. i.e. that the optimum coherence vector remains
unknown but we assume that it doesn’t change with baseline [Sagues 2000, 2001, Flynn 2002,
Pascual 2002, Colin 2003, Gomez Dans 2005 ]. This idea is supported on physical grounds for short
baselines in the absence of temporal decorrelation i.e. for single pass or low frequency sensors
where the scattering does not change significantly over the effective angular width of the baseline.
This approach calls for a new mathematical formulation of the optimization process. One approach
is based on a straightforward extension of the Lagrange multiplier technique to constrain w1 = w2
9

[Gomez Dans 2005]. This leads by manipulation of 16 to a set of w vectors given as eigenvectors of
the composite matrix shown in 19

(T11 + T22 )
−1

12 + Ω12
*T
)w = −λ w - 19)

One problem with 19 is that the eigenvalue is not the coherence, but its real part and so the
optimization is phase sensitive. For this reason a second related approach based on maximization of
the phase difference as a function of polarization vector w. has been developed. In this case the
optimum vector is found by solving a phase parameterised eigenvalue problem of the form shown
in 20. [Flyn 2002, Colin 2003 ]

⎧ 1
⎪[Ω H ] = 2 (Ω12e + Ω12 e )
iφ1 *T −iφ1

[Ω H ]w = λ[T]w ⎨ -20)
1
⎪ [T] = (T11 + T22 )
⎩ 2

This has been shown to be equivalent to calculating the numerical radius of the complex matrix
1 1
− −
A = T 2 Ω12T 2 . A proposed algorithm for finding this optimum state has been presented in [Colin
2003, 2005]. One drawback in this approach is that φ1 is a free parameter and so either search or
iterative methods must be used to secure the global optimum. This adds to the computational
complexity for each pixel.
A third related approach has been proposed based on a sub-space Monte Carlo searching
algorithm [Sagues 2000,2001]. This limits the search for the optimum (again assuming w1 = w2) to
the diagonal elements of Ω12 i.e. to copolarised or crosspolarised combinations across the whole
Poincaré sphere. This again acts to effectively limit the dimensionality of the problem and
demonstrates less bias than the full Lagrange multiplier method. Finally, phase centre super-
resolution techniques based on the ESPRIT algorithm have also been proposed to find the optimum
w vectors. [Yamada 2001].
In all these cases a sub-optimum solution is obtained compared to the unconstrained
Lagrange multiplier method but often with better numerical stability. Given the general increased
processing overhead of employing optimization, it is of interest to investigate the potential benefits
of employing an optimization approach over simple linear, Pauli and circular options (table I). To
do this we look at a set of analytical solutions for the full optimizer.
In the previous section we formulated an important optimisation problem in POLInSAR,
namely to investigate the maximum variation of coherence with polarisation by solving an
eigenvalue problem. In this section we look at some canonical problems of interest in the remote
sensing of land surfaces and try and use the mathematical solutions obtained to conclude as to the
potential of optimisation versus standard coherence estimation in POLInSAR. We consider three
important problems, scattering from non-vegetated surfaces, random volume scattering and finally a
2-layer surface+volume mixture which more closely matches the behaviour of natural vegetated
land surfaces.

4 POLINSAR FOR BARE SURFACE SCATTERING


We start by considering the simplest case of non-vegetated terrain. Under the assumption of surface
scattering only, the polarimetry can then be characterised as a reflection symmetric random media
with a coherency matrix [T] of the form shown in equation 21 [Cloude 1996, 2004 ]. The
10

interferometry (following range spectral filtering [Gatelli 1994] and assuming no temporal or SNR
decorrelation ) is characterised by a single parameter, the ground phase φ as shown in 21.

K = T11−1Ω12T11−1Ω12
*T

- 21)
−1 −1
⎡t11 t12 0⎤ ⎡t11 t12 0 ⎤ ⎡t11 t12 0⎤ ⎡t11 t12 0 ⎤ ⎡1 0 0⎤
⎢ ⎥ iφ ⎢ * ⎥⎢ ⎥ −iφ ⎢ * ⎥ ⎢ ⎥
= ⎢t12* t 22 0 ⎥ e ⎢t12 t 22 0 ⎥.⎢t12* t 22 0 ⎥ e ⎢t12 t 22 0 ⎥ = ⎢0 1 0⎥
⎢⎣ 0 0 t 33 ⎥⎦ ⎢⎣ 0 0 t 33 ⎥⎦ ⎢⎣ 0 0 t 33⎥⎦ ⎢⎣ 0 0 t 33⎥⎦ ⎢⎣0 0 1⎥⎦

It follows from equation 16 that the optimum coherences are obtained as eigenvectors of the matrix
[K] as shown. By multiplying terms we see that the matrix [K] is just the 3 x 3 identity matrix. This
implies that all polarisations have the same interferometric coherence and POLInSAR plays no role
in surface scattering problems. This is not quite true in practice for two important reasons: in
practice there will be polarisation dependent SNR decorrelation. In fact, recently [Hajnsek 2005] it
has been suggested that such SNR coherence variations with polarimetry be used for quantitative
InSAR surface parameter estimation. This formulation assumes that the scattering from the surface
occurs within a thin layer. If there is significant penetration into the surface then volume scattering
effects can occur and this will lead to volume decorrelation effects (see below). These effects have
been observed for land ice [Dall 2003] and snow studies [Papathanassiou 2005] where the surface is
non-vegetated but covered by a low loss scattering layer. Nonetheless, equation 21 demonstrates
how for bare surface scattering POLInSAR plays only a secondary role. More interesting for
application of natural land surfaces is to consider the presence of volume scattering due to
vegetation cover.

5 POLINSAR FOR RANDOM VOLUME SCATTERING


In case we consider scattering from a volume, interest centres on the special case of a random
volume i.e. one with macroscopic azimuthal symmetry [Cloude 1996]. In this case the polarimetric
coherency matrix [T] is diagonal. However more care is required over consideration of the
interferometric phase in Ω12. We now must include the effects of volume decorrelation due to the
random vertical distribution of scatterers [Treuhaft 1996,2000, Cloude 2003]. In this case the
interferometry must include a complex integral I2 normalised by a real integral I1 as shown in
equation 22.

K = T11−1Ω12T11−1Ω12
*T

⎡1 0 0 ⎤ ⎡t11 0 0 ⎤ ⎡ t111 0 0 ⎤ ⎡t11 0 0⎤


1 ⎢ 11 ⎥ ⎢ ⎥1⎢ ⎥ ⎢
t

= ⎢0 1
0 ⎥I2 ⎢ 0 t 22 0 ⎥ ⎢0 1
0 ⎥I2*⎢ 0 t 22 0⎥
I1 ⎢ t 22
1 ⎥ ⎢
I t 22
- 22)
⎣0 0 t 33 ⎦ ⎣
0 0 t 33 ⎥⎦ 1 ⎢⎣ 0 0 1 ⎥ ⎢
t 33 ⎦ ⎣
0 0 t 33 ⎥⎦

2 ⎡1 0 0⎤
I2 ⎢ ⎥
= ⎢ 0 1 0⎥
I1
⎢⎣0 0 1⎥⎦

Here we again see that [K] is proportional to the identity matrix but this time the eigenvalues (all
equal) are given by a ratio of integrals over the vertical distribution. This ratio is just the volume
11

decorrelation displaying an increase in phase variance and a vegetation bias to the ground phase
determined by two parameters, namely the height of the vegetation and its mean extinction
coefficient σ as shown in equation 23. Here the vertical interferometric wavenumber kz [Bamler
1998] appears as a function of the baseline to wavelength ratio B/λ as well as the sensor height H
and angle of incidence θ.
2σh v h v 2σz'

e cosθ o
∫e cosθ o
e ikz z' dz'
2σ e iφ (zo )
hv 2σ z'

∫e
I ikz z' cosθ o
γ˜ (w) = 2 = 0
2σh v h v 2σz'
= e dz'
I1 − cosθ o (e 2σ hv / cosθ o −1)
e cosθ o
∫e cosθ o
dz'
0

0
- 23)
⎧ 2σ
⎪ p=
p e p1 hv −1 ⎪ cosθ
= where ⎨ p1 = p + ikz = γ˜ v
p1 e phv −1 ⎪ 4 πΔθ 4 πBn
⎪⎩ kz = ≈
λ sin θ λH tan θ

where Δθ is the angular separation of the baseline end points from the surface pixel. Note that the
vegetation is characterized by a height hv and mean extinction rate σ as shown, both parameters of
interest in remote sensing. Again however we note that this coherence is independent of
polarisation, [K] has 3 degenerate eigenvalues and POLInSAR plays no role in the analysis of
random volume scattering. This statement has to be modified in the presence of oriented volumes
[Treuhaft 1999, Cloude 2000] i.e. ones with a preferred orientation of scattering elements such as
occur in some agricultural crops and even in forestry applications at low frequencies [Cloude 2000].
In such cases POLInSAR does indeed play a role for volume scattering, with [K] developing 3
distinct eigenvalues. However for the treatment of forestry applications at L band and above such
orientation effects are small and the random volume assumption is justified [Papathanssiou 1999,
2000].
In conclusion, both bare surfaces and random volumes lead to a degenerate eigenvalue
spectrum for the matrix [K]. It is only when we combine these two effects together that we start to
see the potential benefits of employing POLInSAR processing.

6 POLINSAR : 2-LAYER COMBINED SURFACE AND RANDOM VOLUME


SCATTERING

In the general case when combined surface and volume scattering occurs then POLInSAR
coherence optimisation becomes useful as we now demonstrate. In this 2-layer case or random-
volume-over-ground (rvog) model approach [Treuhaft 2000, Cloude 2003], the observed coherence
is given by a mixture formula as shown in 24. Here the ground phase φ and complex volume
coherence γ˜ v are combined with a new real parameter μ, the ratio of effective surface (i.e. all
scattering contributions with a phase centre located at φ) to volume scattering. In effect when μ = 0
we resort to the case of random volume scattering while when μ tends to infinity then we resort to
the surface scattering case. Interest centres on the intermediate case because here we have an
unknown but constant complex number contribution from the volume scattering combined with a
polarisation dependent surface term. By isolating the polarisation dependent terms the resulting
coherence then lies along a straight line in the complex coherence plane as shown in 24.
12

γ˜ v + μ(w ) μ(w )
γ˜(w ) = e

= e iφ [γ˜ v + (1− γ˜ v )] - 24)
1+ μ(w ) 1+ μ(w )

This straight line model has been successfully tested on varied forest data sets [Isola 2001,
Papathanassiou 2001, 2005] and seems to be a good fit for L and P band POLInSAR forestry
applications. It is interesting to note how the coherence varies as we adjust the single parameter μ
along this line. Figure 3 illustrates three important cases. In all three we first note how the
coherence starts for small μ at some value depending on the volume scattering contribution (0.8 in
the example). It then initially decreases with increasing surface contribution until reaching a turning
point after which it increases with μ, always approaching unity as μ tends to infinity.
In figure 3 we superimpose three important special cases of the eigenvalue spectrum of [K]
for this scenario. In the top we show the case when μ is always small (i.e. when there is strong
volume scattering with high extinction masking the surface contributions). Here we see that as we
adjust polarisation (w) then μ will also change and the optimiser has an incentive to select the
minimum μ channel to maximise coherence.At the other extreme, when μ is large and surface
scattering dominates, we see that the optimiser has an incentive instead to maximise μ in order to
maximise coherence. A more interesting case (and one that occurs often in practice for L band
forestry applications) is the intermediate zone when the variation of μ (the μ spectrum) includes the
turning point. In this case the coherence can be maximised by either increasing or decreasing μ
depending on circumstances.
We make two important conclusions from this, firstly that in the mixed surface + volume
scattering case the coherence varies with polarisation and so optimisation plays a role in POLInSAR
analysis. Secondly we see that we cannot simply associate the maximum coherence with for
example the maximum value of μ. Both maxima and minima of μ can lead to the optimum
coherence, depending on the circumstances. However it follows that if we can estimate the μ
spectrum for any problem then we can compare the max/min with the values for the standard
channel (linear, Pauli etc) to quantify the potential benefits of employing optimisation techniques.
Determination of the extreme points of the μ spectrum is related to a classical problem in
radar polarimetry, namely contrast optimisation [Novak 1990]. The solution to this is obtained as
the eigenvalues of the product of the inverse volume times the surface polarimetric coherency
matrices as shown in 25.

⎡ ⎤⎫
⎡1 0 0⎤ ⎢1 0 0 ⎥⎪
⎢ ⎥ 1 ⎢ 1 ⎥⎪ ⎡ ⎤
TV = mI1⎢0 κ 0⎥ ⇒ Tv = −1
⎢0 0 ⎥⎪ ⎢ t11 t12 0⎥
mI1 ⎢ κ ⎥⎪⎪
⎢⎣0 0 κ⎥⎦ 1 ⎥⎬ ⇒ T −1T = 1 ⎢ t12 ⎥
*
t 22
⎢0 0 ⎢ 0⎥ - 25)
⎣ κ ⎦⎪ I1m ⎢ κ κ
v s

⎡t11 t12 0⎤ ⎪ ⎢0 0
t 33 ⎥
⎢ ⎥ ⎪ ⎣ κ⎦
Ts = ⎢t12* t 22 0 ⎥

⎢⎣ 0 0 t 33 ⎥⎦ ⎪⎭

Under the assumption of a random volume and reflection symmetric surface scattering component,
the eigenvalues of this matrix can be determined analytically as shown in equation 26.
13

⎧ ⎛ 2 2 ⎞
⎪ μ = 1 ⎜ t + t 22 + ⎛⎜ t − t 22 ⎞⎟ + 4 t12 ⎟ ⎧ iδ e iφ o (γ v + μ1 )
⎪ 1 2I1m ⎜ 11 κ ⎝ 11 κ ⎠ κ ⎟⎠ ⎪ γ1e 1 =
⎪ ⎝ 1+ μ1
⎛ 2 ⎞ ⎪
⎪ 1 ⎜ t 22 ⎛ 2
t 22 ⎞ 4 t12 ⎟ ⎪ iδ 2 e (γ v + μ 2 )
iφ o
⎨μ2 = t11 + − ⎜ t11 − ⎟ + ⇒ ⎨γ 2e = - 26)
⎪ 2I1m ⎜ κ ⎝ κ ⎠ κ ⎟ ⎪ 1+ μ2
⎝ ⎠ iφ o
⎪ 1 ⎛ t 33 ⎞ ⎪ iδ 3 e (γ v + μ 3 )
⎪ μ3 = ⎜ ⎟ ⎪γ 3e = 1+ μ3
⎪ I1m ⎝ κ ⎠ ⎩

Equally importantly, the eigenvectors of this matrix indicate the w vectors that should be employed
in POLInSAR to secure these extreme coherence values. We note from 26 that the optimum
contrast solutions are not generally the simple HH, HV and VV channels and this supports
investigation of optimisation techniques based on full quadpol data acquisition for POLInSAR
Processing.
14

μ3 μ2 μ1

μ3 μ2 μ1

μ3 μ2 μ1

Figure 3 : Variation of Coherence with small (top), large(centre) and intermediate (lower) μ values
15

7 ALGORITHMS FOR VEGETATION PARAMETER RETRIEVAL


We have seen in equation 26 that the coherence of vegetated land surfaces depends on several
important vegetation and surface parameters. The two most important of these are the mean
vegetation height and the true ground topographic phase. These two are important products for
scientific and commercial applications in their own right but also offer the possibility of estimating
important secondary products such as vegetation biomass [Mette 2004]. Here we summarise the
main algorithms used for generation of vegetation height and ground topography products from
single baseline POLInSAR.
In general, we obtain POLInSAR products by employing model based parameter estimation,
whereby using a scattering model M with parameters p we make observations of a set o and then
obtain estimates of the parameters by inverting the model so that formally we can write

p = M −1 o - 27)

where the inverse is often approximated by using a least squares approach so that the parameters are
chosen so as to minimise the difference between the observations and the model predictions. In our
case M is of the form given by the ‘rvog’ 2-layer coherence model so that we can formulate 27 as
shown in equation 28

iδ 1
e iφ o (γ (hv ,σ ) + μ1 ) ⎫ ⎛φo ⎞
γ1e = = f1 (φ o ,hv ,σ ) ⎪ ⎜ ⎟
1+ μ1 ⎪ ⎜ hv ⎟ ⎛ γ˜1 ⎞
γ 2e iδ 2 =
e iφ o
(γ (h v , σ ) + μ 2 ) ⎪ ⎜ σ ⎟ ⎜ ⎟
= f 2 (φ o ,hv ,σ ) ⎬ ⇒ p = ⎜ ⎟,o = ⎜ γ˜ 2 ⎟ - 28)
1+ μ2 ⎪ ⎜ μ1 ⎟ ⎜ ⎟
⎝ γ˜ 3 ⎠
iδ 3
e (γ (hv ,σ ) + μ3 )
iφ o
⎪ ⎜ μ2 ⎟
γ 3e = = f 3 (φ o ,hv ,σ )⎪ ⎜ ⎟
1+ μ3 ⎭ ⎝ μ3 ⎠

Note that we have 6 parameters and 6 observations (3 complex coherences). It is clear also that we
should choose the three coherence values to be as different as possible so as to maximise stability of
the inversion. This can be achieved either through physical knowledge of the problem (e.g. using
HV for the volume channel and HH for a surface dominated channel) or via use of an appropriate
optimizer as discussed in equation 26. Of particular importance as products are the first two
elements of p, namely ground topography and mean vegetation height, although the m estimates
have recently been proposed for foliage penetration [Cloude 2004] and surface parameter estimation
[Cloude 2005]. We can obtain estimates of these (i.e. invert M) via various inversion strategies as
we now show.
To start we need to estimate the phase φο. There are two basic ways to do this. The first is to
select a polarisation channel via choice of a weight vector ws where it is assumed μ is very large. At
P-band for example HH is often employed [Cloude 2000]. At L band HH can again be used or HH-
VV if Quadpol data is available. In this case φ̂ the estimate of φο is then simply obtained as

φˆ = arg(γ˜ w )
S
- 29)

The problem here is to find the best polarisation wS. Extensive analysis of L and P band data sets
have shown that in general there is no single wS that yields an unbiased estimate of φο. For this
16

reason a second class of algorithms have been developed where we attempt residual bias removal
using the ‘rvog’ model by employing a pair of complex coherence values as follows.
As equation 28 represents a straight line in the complex plane it follows that one simple
algorithm is to employ a straight line fit to a pair of coherence values at polarisations wS and wV,
where the latter is volume dominated and the former surface dominated. The intersection of this line
with the unit circle of the complex coherence plane yields 2 candidate points for the ground phase.
To resolve the ambiguity between these two points, it must be assumed in advance that the wS
combination has a phase centre closer to the ground than wV In this case we can solve for φo by
using the first two equations (for γ1 and γ2) in 28 and setting μ1 = 0. This leads to a quadratic
equation as shown in 30. In a more general approach, coherences for 3 or more polarisation
combinations (w vectors) can be combined to yield an over-determined least squares estimate of the
line fit. Either a simple standard least squares fit between the real and imaginary or a total least
squares fit to the 2-D problem can be employed. This LS line is then used to find the 2 intersection
points with the unit circle and again the ambiguity is resolved by allocating a rank ordering to the
coherence values. Usually it is assumed that HV is higher in the canopy than HH and this can be
used to resolve the ambiguity. Note however that for a small number of looks and low coherences
such a strategy may be noisy (see figure 21).

φˆ = arg(γ˜ w − γ˜ w (1− Lw )) 0 ≤ Lw ≤ 1
V S S S

−B − B 2 − 4 AC
AL2w S + BLw S + C = 0 ⇒ Lw S = - 30)
2A
2 2
A = γ˜ w S −1 B = 2Re((γ˜ wV − γ˜ w S ).γ˜ *w S ) C = γ˜ wV − γ˜ w S

Having obtained an estimate of the ground phase φο. the second stage is then to estimate the height

8 FOREST HEIGHT INVERSION ALGORITHM


There are 3 main approaches used to estimate vegetation height from POLInSAR data:

8.1 DEM DIFFERENCING

In this approach we employ the same idea as used in the estimation of the ground phase to isolate a
polarisation channel that scatters from the top of the canopy and hence generate a height estimate
directly as shown in equation 31

arg(γ w V ) − φˆ 4 πΔθ 4 πBn


hv = , kz = ≈ - 31)
kz λ sinθ λRsin θ

where wV is a user selected polarisation, assumed to be located at the top of the vegetation. Often
this is taken to be HV, as this channel is dominated by volume scattering. Alternatively, phase
optimisation based on the ESPRIT algorithm [Yamada 2001] or numerical radius estimation [Colin
2005] can be used. Note however that HV phase centre can lie anywhere between half the tree
height and the top of the canopy itself. The exact location depends on two properties of the
vegetation, namely the mean wave extinction and vertical canopy structure variation. In case the
trees have a high thin canopy then extinction is small but the phase centre is high due to the
structure. On the other hand, when the canopy extends over the full tree height then the phase centre
can be at half the true height for low density (small extinction) through to the top of the canopy for
17

dense vegetation (high extinction). This ambiguity is inherent to single baseline methods and to
combat this we need to employ model based correction methods as we now consider.

8.2 HEIGHT COMPENSATED FOR EXTINCTION

In this approach we use the extinction variation to compensate both density and structure variations
but obtain a reliable estimate of height [Cloude 2002, Papathanassiou 2001]. In this way we
sacrifice accuracy of retrieval of extinction for robustness in the height parameter. To do this we use
the algorithm shown in 32. Here we make use of the previously estimated ground phase and the full
‘rvog’ model to match the model against observation of a coherence in a channel γ(w), which we
expect to be volume dominated. However, in doing so we have no idea of the surface-to-volume
scattering ratio in that channel. Consequently there are an infinity of possible candidate ‘volume-
only’ points along the line. These are parameterised by the parameter 0 ≤ λ ≤ 1 in equation 32. If λ
= 0 then we assume that γ(w) is the volume only coherence itself (μ = 0). At the other extreme the
volume-only point can lie on the unit circle at the far end of the coherence line with phase φ2 as
shown in 32.

p e p1 hv −1
h v ,σ
( ˆ
)
min L1 (λ,w ) = γ˜ (w) + λ e iφ 2 − γ˜ (w) − e iφ
ˆ

p1 e phv −1
⎧ 2σ
⎪ p= - 32)
⎪ cosθ 4 πΔθ 4 πBn
where ⎨ p1 = p + ik z ,k z = ≈
⎪φˆ = arg(γ˜ − γ˜ 1− L ) λ sinθ λRsin θ
⎪⎩ 2 w V w S w V
( )
In order to resolve these multiple solutions the user must select a value of λ. The usual choice is to
set λ = 0 by assuming that w = wV corresponds to a state with μ = 0 (HV for example), in which
case inversion simplifies as shown in 33.


p e p1 hv −1 ⎪ p = 2σ
min L1 (λ = 0) = γ˜ w V − e iφ
ˆ
where ⎨ cos θ - 33)
h v ,σ p1 e phv −1 ⎪⎩ p1 = p + ik z

In this expression kz and θ are known from the radar geometry and there are 2 unknowns, hv and σ.
As stated above, the parameter σ absorbs variations in density and structure of the vegetation and
hence tends to be noisy. Equation 33 can be applied to invert real data in one of 2 ways, either by
iterative search methods (such as the simplex method), starting with an initial guess and then
converging on the minimum norm solution, or by use of look up tables (LUT). In the latter case the
range of extinction values used in the LUT should be extended to accommodate variations of
structure as well as expected extinction rates for the radar centre frequency. Failure to do this can
lead to LUT boundary effects causing apparent saturation of the estimate. A way to overcome this
LUT limitation will be presented in equation 37. Note that since the coherence and phase are both
required to obtain a solution, then this approach is sensitive to calibration errors, especially in
coherence [Cloude 2002]. However this approach has demonstrated, for airborne and chamber
based data, the best height retrieval accuracy of all POLInSAR algorithms to date [Mette 2004,
Papathanassiou 2005, Sagues 2000].
18

8.3 HEIGHT COMPENSATED FOR VERTICAL STRUCTURE

As an alternative strategy we can fix the extinction at some mean value σ appropriate for the radar
frequency and then ascribe all variations in volume coherence to vertical structure. One way to do
this is to use empirical extinction relationships published in the literature and derived from a mean
of airborne measurements on different forest types [Bessette 2001]. This leads to a relation of the
form shown in equation 34, where A is the 2-way total extinction (in dB), θ the angle of incidence, f
the frequency in MHz and α and β regression coefficients. These are shown in table II for H and V
polarisations. Note that these relationships are derived for low frequency operation (valid for L
band and below). We see that in fact there is a slight differential extinction with VV having a higher
extinction than HH. However the effect is small and in practice we can ignore this and take the
mean extinction for input to the ‘rvog’ model.

A cos θ = α . f β - 34)

Attenuation factors
HH 0.18 0.53
VV 0.3 0.47
Table II :Mean Extinction Regression Parameters

Relations such as this, or more direct EM modelling of propagation extinction in random media
[Williams 2000], can be used to estimate the mean extinction rate in dB/m for an estimated canopy
thickness. The simplest way to then model variations in vertical structure is to allow for an offset
canopy from the reference ground phase [Cloude 2001,2002]. In this way height estimation is
augmented by a second parameter, canopy thickness d (0 ≤ d ≤ hv) in a modified ‘rvog’ model
inversion as shown in 35.

iφˆ ikz ( h v − d ) p e
p1 d
−1 ⎪ p = 2σ
min L1(λ = 0) = γ˜ w v − e e where ⎨ cosθ - 35)
h v ,d p1 e pd −1 ⎪⎩ p1 = p + ikz

This model has been applied to height retrieval for tree species with high thin canopies such as
mature Scots Pine in the Glen Affric region of Scotland [Cloude 2001 ].

8.4 HEIGHT FROM COHERENCE AMPLITUDE ONLY

In the discussion around equation 30, we highlighted one potential problem with ground phase
estimation in low coherence regions, namely that in come circumstances it can be difficult to
resolve the ambiguity between the two intersection points on the unit circle. In these cases, two
height solutions can be found for the same point. To avoid errors of this type we can isolate such
points by checking solutions for both intersection points and flagging those that suggest ambiguous
height solutions. However we then need to employ an alternative height estimation algorithm on
these points. One possible technique is to ignore the phase of the coherence completely and to select
a polarisation channel with expected low surface to volume scattering ratio (HV for example). The
coherence amplitude in this channel is then compared with the random volume prediction to obtain
a height estimate. One limitation of this method is that the estimate remains sensitive to density and
vertical structure variations. Two main options are available, in the first the extinction is set to zero
and we obtain the simple ‘sinc’ coherence model. In the second we can employ a mean extinction as
in table II. In any case, the height estimate is then obtained as a solution of the following equation:
19

⎧ 2σ
p e −1
p1 h v ⎪ p=
min L1 = γ˜ w v − where ⎨ cosθ
hv p1 e phv −1 ⎪ p = p + ik ,k = 4 πΔθ ≈ 4 πBn - 36)
⎩ 1 z z
λ sinθ λRsin θ

As it ignores phase and is sensitive to extinction and structure variations, this method is the least
robust of the algorithms, but as stated above its main role has been as a back up solution when other
approaches fail.

8.5 ROBUST INVERSION ACCOUNTING FOR EXTINCTION/VERTICAL STRUCTURE

We can use the above observations to develop a hybrid approach based on fusion of the coherence
amplitude and DEM differencing algorithm. This algorithm is much faster and easier to implement
than the full ‘rvog’ inversion (equation 32) and yet is also robust to variations in extinction or
vertical structure as we now demonstrate. Although based on an approximation, this algorithm
nonetheless gives height estimates within 10% accuracy, matching the bounds achievable with
current air and space borne sensors [Papathanassiou 2005].
The algorithm requires selection of two interferograms, one for a surface dominated channel
ws and the second for a volume dominated channel wv. The forest height can then be estimated as
shown in equation 37

arg(γ˜ w v ) − φˆ 2sinc −1 ( γ˜ wv )
hv = +ε
kz kz

where
- 37)
φˆ = arg(γ˜ wV ( )
− γ˜ w S 1− Lw S ) 0 ≤ Lw S ≤ 1

−B − B 2 − 4 AC
AL2w S + BLw S + C = 0 ⇒ Lw S =
2A
2 2
A = γ˜ w S −1 B = 2Re((γ˜ w V − γ˜ w S ).˜γ *w S ) C = γ˜ w V − γ˜ w S

We see that the height is obtained as a sum of 2 components. The first is just the height from phase
or DEM difference between the estimated ground topography point and estimated volume only
complex coherence. Importantly, unlike the case in equation 31, this polarisation need not be at the
top of the vegetation and to compensate this, the height estimate is augmented by the second
coherence amplitude term. This is obtained by matching the observed coherence amplitude to the
simple zero extinction ‘sinc’ model. This stage requires comparison to a 1-D LUT, but unlike the
extinction LUT in equation 33, its range is set by the first zero of the sinc function. In this way all
observed coherences can be matched to an effective height and no LUT boundary effects occur.
Choice of the factor ‘ε’ weighting the two components is very important and should be chosen to
provide robustness to extinction variations. In the zero extinction case it is easy to show that it
should be chosen as 0.5, in which case equation 37 gives the exact result for height, regardless of
canopy offset structure. In the more general nonzero extinction case ε should be reduced. In the
limit of infinite extinction ε tends to zero and the phase centre term tends to the true height.
However, practical extinction levels at L band and below are less than 1dB/m (see equation 34). In
this case we have found that choice of ε = 0.4 keeps the height error variations with extinction
20

below 10%. By adopting a constant ε value we then avoid problems of matching extinction to local
variations and hence save considerable computation time.
Given its balance between accuracy and ease of computation, we adopt equation 37 as our standard
test algorithm for tree height estimation. We now demonstrate its effectiveness by application to
POLInSAR test data.

9 POLINSAR DATA PROCESSING


In the previous section we covered in some detail the theoretical background to POLInSAR. based
vegetation height and ground topography estimation, culminating in a robust and efficient algorithm
in equation 37. Here we illustrate and reinforce the ideas introduced by giving an example of a full
data processing chain. The data we use is from a 3-D coherent SAR simulator, details of which can
be found in [Williams 1999, 2000, 2005,Cloude 2004,2005]. We choose an L band vegetated scene
at 45 degrees angle of incidence containing 3 separate layers. The first is a rough dielectric ground,
modelled as a tilted Bragg scattering surface. Above this is a short 0.5m vegetation layer modelling
understorey. Both these layers cover the whole scene of 100m x 100m. Finally over the central part
of the image (56m x 56m) we place a random canopy of branches, similar in structure to a hedge i.e.
without major trunk elements. This canopy forms the main source of volume decorrelation in the
scene and its regular height of 10m provides a convenient check of the height retrieval accuracy of
the various algorithms introduced earlier. The test scenario has the geometry shown in figure 1.
Here we see the flat underlying ground with the 10m hedge in the centre of the scene. This layer is
comprised of a random Gaussian distribution of branches with mean length of 1.5m and standard
deviation of 0.2m with a density or mean volume fraction of 0.2. The L band signal (λ = 0.23061m)
illuminates the scene at θ = 45 degrees incidence from an altitude of 3km, similar to the geometry
used by the airborne E-SAR system from DLR in Germany [Papathanassiou 1997]. A 10m
horizontal offset baseline is used for the interferometry, again reflecting the typical flight geometry
used in repeat pass L band POLInSAR by the ESAR system. The SAR simulator allows
convolution of the scattered field with an instrument point spread function, chosen in this case with
a resolution of 0.6905m in azimuth and 1.3811m in ground range. The image pixel size is then
sampled at 0.5m x 0.5m in ground range and azimuth. These values are typical of those used for
airborne sensors.

With these preliminaries we now turn to the processing chain itself:

9.1 STEP 1 : READING POLINSAR DATA FILES

The first step in POLInSAR is to prepare the appropriate data files. As we need to employ phase
information, use is made of single look complex (SLC) data files as opposed to multilook amplitude
products. Note also that the well-known Stokes matrix compression scheme as used by JPL-
AIRSAR, although useful for polarimetry, is not suitable for POLInSAR studies. As a consequence
POLInSAR requires three calibrated (for both radiometry and polarimetry) SLC data files for each
spatial position (usually HH, HV and VV). In common with conventional interferometry, the
second spatial position (slave track) defining the offset baseline needs to be co-registered with the
first (master track). This processing step is common to all polarisation channels and will lead to a
small residual loss of coherence due to misalignment errors of the tracks (usually around 1/10 of a
pixel accuracy). Such errors (and those associated with signal to noise and quantisation) need in
practice to be assimilated into the design of POLInSAR systems. A suggested framework for this
has been presented in [Krieger 2005, Cloude 2005]. In practice, the largest error source in
POLInSAR has been shown to be temporal decorrelation [Papathanassiou 2003]. In this tutorial
however we assume perfect co-registration and ignore such effects. In the simulator perfect co-
21

registration is possible due to the exact geometry employed and temporal effects can be avoided by
maintaining exact scatterer locations between baseline end points.

Hence for a full POLInSAR analysis we require six SLC data files, three from the master and three
from the slave. Figure 5 shows a SAR image of the test scene for the 4 channels HH, HV,VH and
VV provided by the simulator. Note that for calibrated data HV = VH as required by the reciprocity
theorem for backscatter and so only one of the crosspolar channels need be used in the analysis.
However providing separate HV and VH data channels can be useful in practice. In the modelling
context it can be used to check the co-ordinate system used in the simulator. For forward scattering
alignment (FSA) HV and VH will be equal in magnitude but 180 degrees out of phase whereas for
the backscatter alignment (BSA) HV = VH in both amplitude and phase. For real SAR data, the
HV/VH coherence can also be used as a check of signal to noise ratio in the crosspolar channel. If
the HV/VH coherence is high then the SNR is high and vice versa.

In figure 5 we can clearly see the increased backscatter from the hedge layer and note the shadow
region at the rear of the hedge due to its 10m elevation. We note too a bright band at the front of the
hedge in HH and VV. This is due to a second order ground volume interaction. The simulator
accounts for three levels of scattering [Williams 1999, 2000], direct scattering from volume and
surface, second order surface-volume interactions and third order surface-volume-surface
interactions. While the third order interactions are generally small it is important to model correctly
the first and second order interactions, which requires careful calculation of the effective reflection
coefficient from the rough surface and also correct modeling of the polarimetric phase for the
second order or dihedral component.
22

θ = 45o

hv=10m

Bragg Surface Scattering

B = 10m

h=3km r1
r2

θ
y

Figure 4 : POLInSAR Simulation Geometry

9.2 STEP 2 : GENERATING AN INTERFEROGRAM

The next step is to generate a complex interferogram s1s*2 using equation 10. For illustration
purposes we generate an interferogram for the HV polarization channel. The raw phase of the
product is then shown in figure 6. Here we see two features of importance. Firstly we see a
background phase variation across the whole scene, which is a function of range only and comprises
one complete fringe or 2π phase variation. The second feature we note is the phase noise associated
23

with the canopy layer. This phase noise is due to volume decorrelation and later we shall use this
decorrelation to extract information about the height of the vegetation using POLInSAR.

9.3 STEP 3 : FLAT EARTH REMOVAL

It is possible to remove the background phase variations by employing the geometry of figure 4 to
calculate the expected phase variation for a flat surface and then removing this phase by multiplying
the interferogram by the complex conjugate of the so called ‘flat earth’ phase. From the geometry of
figure 4, this phase variation can be calculated as


e
iφ fe
= exp(i (r2 − r1))
λ
r1 = h 2 + y 2 - 38)
r2 = h 2 + (y + B) 2

where y is the ground range co-ordinate, which we vary across the scene. When we form the
−iφ
modified interferogram s1.s*2e fe we obtain the phase image shown in figure 7. Here we see that
the phase of the flat surface is now constant at zero degrees. This becomes our ground phase
reference across the whole scene. We can again see the phase noise due to the vegetation layer but
note in addition that there is a bias or offset to the mean phase of around 1 radian in this region.
This is called vegetation bias and reflects that fact that the mean phase centre in the vegetation lies
above the ground (positive phase). Again, in POLInSAR we make use of the variations of this phase
centre with polarisation to estimate the vegetation height. Before we can estimate height from phase
however we need to calculate the scale factor or vertical wavenumber kz, which relates phase to
height via the relation φ = kz.h

9.4 STEP 4: VERTICAL WAVENUMBER ESTIMATION

We can calculate the sensitivity of the interferometer to height variations from the geometry of the
baseline as follows. [Bamler 1998]
B
Δθ = tan−1 (tan θ + ) − θ
[1]
h - 39)
4πΔθ
kz =
λ sin θ

We see that this is a function of angle of incidence, for a fixed baseline B, kz is higher in the near
range where θ is small and decreases in the far range when θ increases. However for this small
scene of only 100m range swath from 3km altitude the variation of angle of incidence is small and
hence kz is approximately constant with a value of 0.1282 for a 10m baseline. Hence a vegetation
bias of 1 radian corresponds to a height of 7.8m. With kz calculated we can then turn the phase
image of figure 7 into an equivalent height map. Figure 8 shows a histogram of the height of the
pixels in the canopy zone. Here we see a mean of 5m, half the canopy height, with a significant
spread due to the phase noise caused by volume decorrelation.
24

Figure 5 : SAR Images for the 4 polarisation channels and two baseline positions1 and 2 of the
simulated scene
25

Figure 6 : Raw Phase of the Interferogram for HH Polarisation

Figure 7 : Interferometric Phase Following Flat Earth Removal


26

Figure 8 : Histogram of Phase Centre heights in Canopy Region


We see from this example that the phase of the vegetation bias does not correspond simply to the
height itself, but generally underestimates the true height by an amount that depends on extinction,
structure and the contribution from surface scattering. Further we don’t always know the ground
reference phase (in this case zero due to knowledge of the exact geometry) and so in general to
estimate height in a robust way we need to take the phase relative to some reference point. To do
this and further improve the height estimate we first need to include coherence amplitude as well as
phase in the inversion.

9.5 STEP 5: COMPLEX COHERENCE ESTIMATION

In order to estimate coherence we employ equation 2, which generates a complex ratio, the phase of
which is just the mean phase over the selected pixels and the amplitude of which lies in the range 0
to 1 and relates to the quality of the phase through its local variance. When the summation in the
numerator is zero, we have complete decorrelation and a coherence of zero. In this case there is no
significant phase information in the selected pixel. At the other extreme when the neighbouring
pixels all have the same phase then the coherence is unity and we have a deterministic phase. We
can visualise this coherence information in two ways. Firstly we can make an image of the
coherence amplitude. This is traditionally viewed as a grey scale image with white = 1 and black =
0 and gives a direct visual interpretation of areas of high and low coherence. However such an
image misses completely the important contribution made by phase. To overcome this we employ a
mapping of the complex coherence inside a unit circle in the complex plane, with radius equal to the
coherence amplitude and phase the polar angle of the coherence point. This circle diagram is shown
in figure 9. This has the advantage of displaying both phase and coherence amplitude on the same
diagram but is limited to a pixel-by-pixel view rather than an image. Nonetheless it an important
27

tool in POLInSAR as it allows us to view the variation of both phase and coherence with
polarisation and hence check the robustness of inversion algorithms based on the shape of the
coherence region, that is the region formed inside the unit circle by changes over the complete set of
polarisation combinations.

Complex Coherence

Figure 9: Polar Representation of Complex Coherence inside the unit circle

9.6 STEP 6: COHERENCE BIAS AND CONVERGENCE

To calculate coherence we need to specify a window size for the local averaging process around the
pixel under consideration. Coherence is related to phase variance and hence is akin to estimation of
second order statistics of a stochastic variable. The choice of window size is therefore crucial to the
quality of the estimate of coherence obtained. For example, in the extreme case we might choose a
window size of 1x1 and just take the pixel phase itself. In this case all estimated coherences will be
unity, but we do not obtain a good estimate of the true value of coherence and have an extreme case
of bias in the estimate. As the window size increases we will therefore still obtain a slight
overestimation (see figure 1) until for some window size, which depends on the underlying
coherence value itself, we obtain convergence to an unbiased estimate with variance limited by the
Cramer Rao bounds in equation 9. To illustrate this process, figure 10 shows histograms of the
estimated coherence in the canopy region for various window sizes. The window used is square of
integer dimension N, where we have used N = 3,7,11, and 15.
Note that the N = 3 window is too small. We see a residual bias in the peak of the histogram
and a very wide distribution of coherence points. N = 7 shows some improvement but still a bias
offset. N= 11 starts to show some convergence. This we can see because for N = 15 we see little
change in the peak of the histogram which occurs around 0.7, although of course there is some
narrowing of the width of the distribution for the larger window. For this reason we select N = 11
for further investigations. Figure 11 shows a coherence image of the whole scene for this window
28

size. Note that the coherence over the surface regions is unity while the coherence over the canopy
area reduces due to volume decorrelation. In the phase image we see a smoothed estimate of the
vegetation bias already noted in figure 7.

Figure 10 : Histograms of Coherence Estimate versus window size for HV interferogram


Having now established estimates of both phase and coherence we use the data to illustrate the
accuracy of three important POLInSAR inversion algorithms. In the first case we take a simple
DEM difference between polarisations to try and estimate canopy height. In the second we employ
just the coherence amplitude and finally we combine the phase and coherence amplitude together in
the algorithm described in equation 37.

9.7 STEP 7: ALGORITHM 1 : DEM DIFFERENCING

In this approach we need to choose 2 polarisation channels, one with a phase centre close to the
ground, obtained for polarisation ws, and the other with a phase centre close to the canopy top for
polarisation wv. Differencing the interferograms and normalising by kz then provides a direct
estimate of relative height. There are many ways to choose these two polarisation but here we give
a typical example of the logic involved. The cross-polarised or HV channel can often be expected to
have a high phase centre, as the ratio of surface to volume scattering is generally small in this case.
Hence the ‘upper’ interferogram can be formed from HV polarisation. To select a lower channel we
use the fact that dihedral or double bounce scattering has a phase centre lying on the surface and so
as long as the double bounce is stronger than the volume in the selected channel then the phase
centre will lie close to the ground. The HH-VV channel is closely matched to such a dihedral
component and hence we can choose the second interferogram in this polarisation combination.
Figure 12 shows the resulting height estimate.
29

Figure 11: Coherence Image for HV Interferogram using an 11x11 window


(amplitude (upper) and phase (lower))
30

Figure 12 : Height Estimate from Difference of DEMs in HV ands HH-VV channels

Figure 13 : Height Estimate from Coherence Amplitude


31

We note two important points:

1. The height is severely underestimated using this algorithm. The phase centres are only
separated by a few metres inside the canopy. Physically this is due to the presence of a
strong volume scattering component in all polarisation channels and also because the HV
phase centre lies not at the top but approximately half way up the volume (see figure 8).
2. The second key point is that sometimes the height estimate is negative. This arises because
the HV phase centre lies below the HH-VV and our assumption of the ‘upper’ and ‘lower’
phase centres is reversed. This will also have implications for the estimation of ground
topography in the full ‘rvog’ inversion scheme (see figure 15).

9.8 STEP 8: ALGORITHM 2 : COHERENCE AMPLITUDE INVERSION

As an alternative to the DEM differencing approach we can use the coherence amplitude in a
channel we believe to have only volume scattering present (μ = 0). Again we can adopt the HV
channel as a good approximation to this. We then use the relation between height and coherence for
a known extinction in the layer. If we assume zero extinction then we obtain the ‘sinc’ relation
shown in figure 13. When we apply this relation to the HV coherence we obtain the height estimates
shown in figure 14.

Figure 14 : Height estimate from HV Coherence

Here we see a much better estimate, closer to the true height of 10m. However several points are
overestimated and also such an approach is sensitive to density (extinction) and structure (crown
depth) variations in the vegetation. For this reason we turn finally to consider the combination of
phase and coherence in equation 37 for robust height estimation.
32

9.9 STEP 9 : GROUND PHASE ESTIMATION USING DUAL POLARISATIONS

The first stage in better using phase information is to try and locate the true ground position. This
we can do using just two interferograms in two polarisations, similar to the DEM differencing
approach. However this time we employ both the phase and coherence of the two interferograms to
compensate the volume offset in the ground channel using equation 30. If we choose HV as the
‘volume’ dominated channel and HH-VV the ‘surface’ dominated channel we obtain the ground
phase estimate shown in figure15 (again using an 11 x 11 window)

Figure 15 : Ground Phase Estimate Based on HV and HH-VV Interferograms

We note that the true ground phase in this example is zero and can see that most of the vegetation
bias has in fact been removed. However, the phase estimate is noisy, due to inversions of the HV
and HH-VV phase centres and also to density variations in the canopy. In general therefore ground
phase estimation requires a more sophisticated approach than just the simple dual polarisation line
fit. In a more sophisticated approach we can use multiple polarisation channels to fit the line
parameters in an overdetermined system of equations. In our case however we adopt a simpler
solution, namely to filter the phase jumps in figure 15 by using a median filter (note that a
conventional smoothing or mean filter would be inappropriate for such phase filtering and the
median approach provide a better way to suppress phase jumps of the type observed in figure 15).
Figure 16 shows the result of applying a strong 21 x 21 median filter. Here we see some residual
phase errors but note that the phase jumps in figure 15 have been removed. Having obtained a better
estimate of the true ground position, we can now employ equation 37 to combine phase and
coherence information in a way that is robust to structure and density fluctuations.
33

Figure 16: Median Filtered ground Phase Estimate

9.10 STEP 10: ALGORITHM 3: PHASE AND COHERENCE INVERSION

Using figure 16 as an estimate of the ground phase φο in equation 37, we can then estimate the two
components required for height. Firstly we obtain the height of the HV phase centre, now from the
estimated ground position, as shown in figure 17. This again confirms a height around half way up
the canopy at 5m. The difference between this result and figure 8 is that we are now compensating
for arbitrary ground topography variations and hence are providing a more robust algorithm
approach. As outlined in the discussion around equation 37, we must now compensate this
underestimation of height by adding a fraction of the coherence based height estimate (ε = 0.4).
When combined, we obtain the final height image shown in figure 18. Here we see a good estimate
of the true 10m height of the canopy. This is confirmed by plotting histograms of the height
estimates over the canopy region for the various algorithms as shown in figure 19. Here we confirm
that the DEM difference is the poorest method, the coherence amplitude and phase/coherence
methods provide similar mean values but the latter has smaller dispersion and, as mentioned earlier,
is more robust to vertical structural variations. In figure 20 we show sample azimuth transects
through the scene. Here again we see the comparative performance of the various algorithms. Note
that the difference between the coherence only and coherence/phase methods may in fact be used to
estimate crown occupancy of the volume. In this case the two estimates are similar and so the
estimated crown occupancy is 100%, which is correct in this simulation where the canopy
components were distributed from ground to top.
34

Figure 17: Height Estimate from Ground Phase

Figure 18: Height Estimate from Combined Phase and Coherence


35

Figure 19: Histograms of Height Estimation over Canopy Region for 3 POLInSAR Algorithms

Figure 20: Azmuth Transects through the height profiles for R = 55m (true height = 10m)
36

9.11 STEP 11: POLARISATION SELECTION

In the previous analysis we have employed just two polarisation channels, HV selected as a volume
dominated channel and HH-VV as a surface dominated component. In practice there may be other
options available. For example it is common to operate imaging radars in a dual polarisation mode
where the transmitter emits a single polarisation but the receiver has 2 orthogonal channels so
enabling a co and cross-polar measurement. The JAXA ALOS-PALSAR for example is able to
transmit H and receive H and V so obtaining only 2 polarisation combinations HH and HV. It is
interesting to see the change in height retrieval performance with this restricted combination. The
test data set can be used to investigate this. HV is still used as the volume dominated channel but
now we use only HH in place of HH-VV for the surface channel. Figures 21 to 23 show the results
obtained.

Figure 21 : Ground Phase Estimation using HH and HV channels

Figure 22 : Height Estimation using Phase+Coherence for HH and HV channels


37

Figure 23 : Histograms of Canopy Height estimation using HH and HV channels


We see some degradation of performance, related primarily to the smaller phase centre separation
between the channels. This is particularly noticeable in the ground phase estimation, which is now
much noisier. Nonetheless with the same median filtering the height histograms (figure 23) show
acceptable performance for mean height estimation.
This was for restricted polarisation switching. At the other extreme, we face the option for
Quadpol systems of being able to locally adapt the polarisation vectors w to optimise the phase
centre separation and hence improve the height estimation further. There are several algorithms
being developed to do this and details can be found in the literature [Cloude 1997, Papathanassiou
1999, 2001, Stebler 2002, Gomez Dans 2005] Here we simply illustrate the potential for such
adaptive techniques by looking at coherence loci diagrams.

9.12 STEP 12: GENERALISATION TO THE COHERENCE LOCI

In order to investigate the potential for using polarisations other than those listed in table I, we need
to calculate the boundary of the coherence region inside the unit circle for all polarisation vectors w.
In general we can allow different w vectors at either end of the baseline (equation 16) but to
simplify the analysis we follow the suggestion in equations 19 and 20 to use the same polarisation
at either end of the baseline. The boundary of the region can then be found using equation 20 [Flynn
2002, Tabb 2001, 2002, Colin 2003]. For each value of phase (equivalent to rotating the unit circle),
we obtain the maximum and minimum eigenvalues of the matrix product shown below in equation
40. Having found the eigenvectors corresponding to these eigenvalues we can then use these as w
vectors to estimate coherence and plot their complex coherences inside the unit circles. By
repeating this calculation for all φ values we obtain a shape, the boundary of the coherence region.
We can then see where our ‘standard’ polarisations lie within this boundary to assess the potential
for using adaptive techniques to improve the inversion.
38

⎧ 1
⎪[Ω H ] = 2 (Ω12e + Ω12 e )
iφ 1 *T −iφ1

[T]−1[Ω H ]w = λ w ⎨ - 40)
1
⎪ [T] = (T11 + T22 )
⎩ 2

We use the test data set to illustrate this scheme. Since the coherence regions are plotted inside the
unit circle of figure 9 it is a pixel based analysis and so we must select a pixel for analysis. Figure
24 shows a typical example for a pixel location in the canopy region.

Figure 24 : Coherence region for Pixel in Canopy region (green = HV, red = HH)

Here we show the region boundary in black, calculated using equation 40. In green we show the HV
coherence and in blue the HH coherence as used in the height analysis. In blue we show a line
between the HV coherence and median filtered topographic phase point. We see that this lies close
to the true phase point at zero. We see that in this case optimisation may indeed help the retrieval
accuracy. It would be better to use w vectors that correspond to the ends of the major axis of the
elliptical coherence region. This would maximise the separation of the phase centres and provide a
better ground phase estimate. However as the region shape changes from pixel to pixel, this requires
an adaptive process that adds considerably to the processing overhead.
39

10 CONCLUSIONS
In this tutorial we have introduced ina systematic way most of the major concepts required for an
understanding of the application of POLInSAR to vegetation parameter estimation. We began by
considering the role of coherence in interferometry and polarimetry before exploring in some detail
models explaining the variation of coherence with polarisation for surface, volume and the
important case of mixed surface and volume scattering. It is in this last case that POLInSAR plays
an important role.
We then summarised the main approaches to estimation of vegetation height and ground
topography using POLInSAR and developed a fast robust algorithm, based on the random volume
over ground or ‘rvog’ 2-layer coherence model (equation 37). To illustrate these concepts and
demonstrate a typical 12 step processing chain we then employed simulated L-band POLInSAR
data for a mixed surface and volume scattering scene. We used this to look at issues of coherence
and phase estimation before employing the data for a test of various height and ground phase
inversion techniques. We concentrated mainly on using dual polarised data but completed the
tutorial by looking at the idea of a coherence region and how the polarisation diversity offered by
Quadpol systems can be used to further improve performance via the use of adaptive processing
techniques.

11 POLINSAR BIBLIOGRAPHY : AUGUST 2005


Askne J, P B Dammert, L M Ulander, G Smith”C-Band Repeat Pass Interferometric SAR
Observations of the forest”, IEEE Transactions on Geoscience and Remote Sensing, Vol. 35., Jan.,
pp. 25-35, 1997

Bamler R, P. Hartl, “Synthetic Aperture Radar Interferometry”, Inverse Problems, 14, R1-R54,
1998

Bessette L A, S Ayasli “Ultra Wide band P-3 and Carabas II Foliage Attenuation and backscatter
Analysis”, Proceedings of IEEE Radar Conference, 2001, pp 357-362

Cloude S R, E. Pottier, "A Review of Target Decomposition Theorems in Radar Polarimetry",


IEEE Transactions on Geoscience and Remote Sensing, Vol. 34 No. 2, pp 498-518, March
1996

Cloude S R, K P Papathanassiou, "Polarimetric Optimisation in Radar Interferometry", Electronics


Letters, Vol. 33, N0. 13, June 1997, pp 1176-1178

Cloude S R, K.P. Papathanassiou, "Polarimetric Effects in Repeat-Pass SAR Interferometry",


IEEE International Symposium on Geoscience and Remote Sensing (IGARSS '97), Singapore,
August 1997

Cloude S R , K P Papathanassiou, “Polarimetric SAR Interferometry”, IEEE Transactions on


Geoscience and Remote Sensing, Vol 36. No. 5, pp 1551-1565, September 1998

Cloude S R, K P Papathanssiou, A Reigber, “Polarimetric SAR Interferometry at P Band for


Vegetation Structure Extraction”, Proceedings of 3rd European SAR Conference EUSAR 2000,
Munich, Germany, May 2000, pp 249-252
40

Cloude S R, K P Papathanssiou, W M Boerner “A Fast Method for Vegetation Correction in


Topographic Mapping Using Polarimetric Radar Interferometry”, Proceedings of 3rd European SAR
Conference EUSAR 2000, Munich, Germany, May 2000, pp 261-264

Cloude S R, K P Papathanassiou, W M Boerner, “The Remote Sensing of Oriented Volume


Scattering Using Polarimetric Radar Interferometry”, Proceedings of International Symposium on
Antennas and Propagation, ISAP 2000, Fukuoka, Japan, pp 549-552, August 2000

Cloude S R, I.H. Woodhouse, J. Hope, J.C. Suarez Minguez, P. Osborne, G. Wright,


“The Glen Affric Radar Project: Forest Mapping using dual baseline polarimetric radar
interferometry”, ESA Symposium on “Retrieval of Bio and Geophysical Parameters from SAR for
Land Applications”, University of Sheffield, England, September 11-14, 2001 pp 333-338

Cloude S R, “Calibration Requirements for Forest Parameter Estimation using POLinSAR”,


Proceedings of CEOS CalVal Workshop, BNSC Conference Centre, London, September 2002

Cloude, K.P. Papathanassiou, A 3-Stage Inversion Process for Polarimetric SAR


Interferometry, Proceedings of European Conference on Synthetic Aperture Radar, EUSAR'02,
pp. 279-282, Cologne, Germany, 4-6 June 2002.

Cloude S R, K.P. Papathanassiou, “ A 3-Stage Inversion Process for Polarimetric SAR


Interferometry”, IEE Proceedings, Radar, Sonar and Navigation, Volume 150, Issue 03, June
2003, pp 125-134

Cloude S R, “Radar Polarimetry and Interferometry : A Tutorial Introduction”, IEEE


Geoscience and Remote Sensing Newsletter, June 2004

Cloude S R, D.G. Corr, M.L. Williams, “Target Detection Beneath Foliage Using Polarimetric
SAR Interferometry”, Waves in Random Media, volume 14, issue 2, pages S393 - S414., 2004

Cloude S R, M.L.Williams, “The Negative Alpha Filter: A New Processing Technique for
Polarimetric SAR Interferometry”, IEEE Geoscience and Remote Sensing Letters, Vol. 2,
April 2005, pp 187-191

Cloude S R, G Krieger, K Papathanassiou, “A Frammework for Investigating Sp[ace-Borne


POLInSAR using the ALOS PALSAR sensor”, Proceedings of IEEE Geoscience and Remote
Sensing Symposium (IGARSS 2005), Seoul, South Korea, 25-29 July 2005

Colin E., C Titin-Schneider, W Tabbara, “Investigation of Different Interferometric Coherence


Optimisation Methods”, Proceedings of 1st ESA Workshop on Applications of SAR Polarimetry
and Polarimetric Interferometry (POLInSAR 03), January 2003, SP-529,
http://earth.esa.int/workshops/polinsar2003/

Colin E., C Titin-Schneider, W Tabbara, “ Coherence optimization Methods for Scattering Centre
Separation in Polarimetric Interferometry”, Journal Of Electromagnetic Waves and Applications
(JEWA), in Press, 2005
41

Dall J, K P Papathanassiou, H Skriver, “Polarimetric SAR Interferometry Applied to Land Ice: First
Results”, proceedings of IEEE Geoscience and Remote Sensing Symposium (IGARSS ’03),
Toulouse, France, 2003, Vol III, pp 1432-1434

Flynn T., Tabb M., Carande R., “Coherence region Shape Estimation for Vegetation Parameter
Estimation in POLINSAR”, Proceedings of IGARSS 2002, Toronto, Canada, pp V 2596-2598

Gatelli F A Monti Guarnieri, F parizzi, P pasquali, C Prati, F Rocca, “The Waveneumber shift in
SAR Interferometry”, IEEE Trans. GRS-32, pp 855-865, July 1994

Gomez-Dans J.L., S Quegan, “Constraining Coherence Optimisation in Polarimetric Interferometry


of Layered Targets”, Proceedings of 2nd ESA Workshop on Applications of SAR Polarimetry and
Polarimetric Interferometry, POLInSAR 05, January 2005,
http://earth.esa.int/workshops/polinsar2005/

Hagberg J O, L Ulander, J Askne“Repeat-Pass SAR Interferometry over Forested Terrain”, IEEE


Transactions on Geoscience and Remote Sensing, Vol. 33. No. 2, pp. 331-340, 1995

Hajnsek I, S R Cloude, “The Potential of InSAR for Quantitative Surface Parameter Estimation”,
Canadian Journal of Remote Sensing, Vol. 31, No. 1, pp 85-102, February 2005

Isola M., S.R Cloude, “Forest Height Mapping using Space Borne Polarimetric SAR
Interferometry”, Proceedings of IEEE International Geoscience and Remote Sensing Symposium
(IGARSS 2001), Sydney, Australia, Vol., July 2001

Krieger G., Kostas Papathanassiou, Shane Cloude, Alberto Moreira, Hauke Fiedler, Michael
Völker, Spaceborne Polarimetric SAR Interferometry: Performance Analysis and Mission
Concepts, Proceedings of 2nd ESA POLInSAR Workshop, Frascati, January 2005,
http://earth.esa.int/workshops/polinsar2005/

Lee J S, K W Hoppel, S A Mango, A Miller , “Intensity and Phase Statistics of Multi-Look


Polarimetric and Interferometric SAR Imagery”, IEEE Trans GE-32, pp. 1017-1028, 1994

Lee J S, M R Grunes, T L Ainsworth, L J Du, D L Schuler, S R Cloude, “Unsupervised


Classification using Polarimetric Decomposition and the Complex Wishart Distribution”, IEEE
Transactions Geoscience and Remote Sensing, Vol 37/1, No. 5, p 2249-2259, September 1999

Lee J S, S.R. Cloude, K.P. Papathanassiou, M.R. Grunes, I. H. Woodhouse, “Speckle Filtering and
Coherence Estimation of POLInSAR Data for Forest Applications”, IEEE Transactions on
Geoscience and Remote Sensing, Vol. 41, No. 10, pp 2254-2263, October 2003

Mette T, K Papathanassiou, I Hajnsek, “Biomass Estimation from POlInSAR over Heterogeneous


Terrain”, Proceedings of IEEE Geoscience and Remote Sensing Symposium (IGARSS 2004),
Anchorage, Alaska, 20-24 September, 2004

Mette T., K Papathanassiou, I Hajnsek, H Pretzsch, P Biber, “Applying a Common Allometric


Equation to Convert height from POLInSAR data to Forest Biomass”, Proceedings of IEEE
Geoscience and Remot Sensing Symposium (IGARSS 2004), Anchorage, Alaska, 2004
42

Nico G., J M Lopez-Sanchez, J Fortuny,D Tarchi,D leva,A J Sieber, “Assessment of the Impact of
Polsarimetric Coherence Optimisation on Phase Unwrapping”, Proceedings of 3rd European
Conference on Synthetic Aperture Radar (EUSAR 2000), pp 523-526, May 2000

Novak L, M.C. Burl, "Optimal Speckle Reduction in Polarimetric SAR Imagery", IEEE
Transactions AES Vol. 26, pp. 293-305, March 1990

Pascual C., E Gimeno-Nieves, J M Lopez-Sanchez, “The Equivalence Between the Polarisation


Subspace Method (PSM) and Coherence Optimisation in Polarimetric Radar Interferometry”,
Proceedings of 4th European Synthetic Aperture Radar Conference, EUSAR 2002, pp 589-592

Papathanassiou K P, A Reigber, R Scheiber, R Horn, A Moreira, S R Cloude“Airborne


Polarimetric SAR Interferometry”, , Proceedings of IEEE Symposium on Geoscience and Remote
Sensing (IGARSS), Seattle, USA, July 6-10, 1998

Papathanassiou K P, A. Reigber, S.R. Cloude “Vegetation and Ground Parameter Estimation using
Polarimetric Interferometry Part 1: The Role of Polarisation”, Proceedings of ESA CEOS SAR
Workshop, Toulouse, France, October 1999.

Papathanassiou K P, A. Reigber, S.R. Cloude “Vegetation and Ground Parameter Estimation using
Polarimetric Interferometry Part 2 : Parameter Inversion and Optimal Polarisations”, Proceedings
of ESA CEOS SAR Workshop, Toulouse, France, October 1999.

Papathanassiou K P, S R Cloude, A Reigber, “Single and Multi-Baseline Polarimetric SAR


Interferometry over Forested Terrain”, Proceedings of 3rd European SAR Conference EUSAR
2000, Munich, Germany, May 2000, pp 123-126

Papathanassiou K P, S.R. Cloude, “Single Baseline Polarimetric SAR Interferometry”, IEEE


Transactions Geoscience and Remote Sensing, Vol 39/11, pp 2352-2363, November 2001

Papathanassiou K P, S R Cloude, “The Effect of Temporal Decorrelation on the Inversion of Forest


Parameters from POlInSAR Data””, Proceedings of IEEE International Geoscience and Remote
Sensing Symposium (IGARSS 2003), Toulouse, France, July 21-25, 2003

Papathanassiou K P, S.R. Cloude , A Liseno, T. Mette , and H. Pretzsch, “Forest Height Estimation
by means of Polarimetric SAR Interferometry: Actual Status and Perspectives”, Proceedings of 2nd
ESA POLInSAR Workshop, Frascati, Italy, January 2005,
http://earth.esa.int/workshops/polinsar2005/

Papathanassiou K P, I. Hajnsek , Thomas Nagler , and Helmut Rott “Polarimetric SAR


Interferometry for Snow Cover Parameter Estimation”, Proceedings of 2nd ESA Workshop on
Applications of SAR Polarimetry and Polarimetric Interferometry, POLInSAR 05, January 2005,
http://earth.esa.int/workshops/polinsar2005/

Pottier E., L Ferro-Famil, S Cloude, I Hajnsek, K Papathanassiou, A Moreira, T Pearson, Y


Desnos, “PolSARpro v2.0: The Polarimetric SAR Data Processing and Educational Toolbox”,
Proceedings of IEEE Geoscience and Remote Sensing Symposium (IGARSS 2005), Seoul, South
Korea, 25-29 July 2005
43

Preiss M, N J Stacy, “Scene Coherence at X-Band from Repeat Pass Polarimetric Interferometry”,
Proceedings of IEEE Geoscience and Remote Sensing Symposium (IGARSS 2005), Seoul, South
Korea, 25-29 July 2005

Reigber A, K P Papathanassiou, S R Cloude, A Moreira, ”SAR Tomography and Interferometry for


the Remote Sensing of Forested Terrain”, Proceedings of 3rd European SAR Conference EUSAR
2000, Munich, Germany, May 2000, pp 137-140

Sagues L, J M Lopez-Sanchez, J Fortuny, X Fabregas, A Broquetas, A J Sieber, “Indoor


experiments on Polarimetric SAR Interferometry”, IEEE GRS-38, pp 671-684, March 2000

Sagues L, J M lopez-Sanchez, J Fortuny, X Fabregas, A Broquestas, A J Sieber, “Polarimetric


Radar Interferometry for improved Mine Detection and Surface Clutter Rejection”, IEEE GRS-39,
pp 1271-1278, June 2001

Schneider R Z, K P Papathanassiou, I hajnsek, A Moreira, “Polarimetric Interferometry over Urban


Areas: Information Extraction using Coherence Scatterers”, Proceedings of IEEE International
Geoscience and Remote Sensing Symposium (IGARSS 2005),Seoul, Korea, 25-29 July 2005

Seymour S., Cumming I.G., “Maximum Likelihood Estimation for SAR Interferometry”,
Proceedings of IEEE-IGARSS’94, Pasadena, USA

Stebler, O., Meier, E., Nueesch, D., “Multi-baseline polarimetric SAR interferometry - first
experimental spaceborne and airborne results”. ISPRS Journal of Photogrammetry and Remote
Sensing, 56(3), 2002

Tabb M, R Carande, “Robust Inversion of Vegetation Structure Parameters from Low Frequency
Polarimetric Interferometric SAR”Proceedings of IEEE International Geoscience and Remote
Sensing Symposium (IGARSS 2001), Sydney, Australia, Vol., pp , July 2001

Tabb M., Flynn T., Carande R., ”Direct Estimation of Vegetation Parameters from Covariance Data
in POLINSAR”, Proceedings of IGARSS 2002, Toronto, Canada, pp III 1908-1910

Tabb M., J Orrey, T Flynn, R Carande, “Phase Diversity: A Decomposition for Vegetation
Parameter Estimation using Polarimetric SAR Interferometry”, Proceedings of 4th European
Synthetic Aperture Radar Conference, EUSAR 2002, pp 721-724

Touzi R, A Lopes, J Bruniquel, P W Vachon, “Coherence Estimation for SAR Imagery”, IEEE
Transactions Geoscience and Remote Sensing, Vol. 37/1, pp 135-149, January 1999

Treuhaft R N, S. Madsen, M. Moghaddam, J.J. van Zyl, “Vegetation Characteristics and


Underlying Topography from Interferometric Data”, Radio Science, Vol. 31, Dec, pp. 1449-1495,
1996

Treuhaft R N, S R Cloude, “The Structure of Oriented Vegetation from Polarimetric


Interferometry”, IEEE Transactions Geoscience and Remote Sensing, Vol 37/2, No. 5, p 2620,
September 1999

Treuhaft R N, P. Siqueria, “Vertical Structure of Vegetated Land Surfaces from Interferometric and
Polarimetric Radar”, Radio Science, Vol. 35(1), pp 141-177, January 2000
44

Williams M.L., “Prediction and Observation of SAR Clutter from Vegetation Canopies”,
Proceedings of IGARSS ’99, Hamburg, Germany, pp 1983-1985

Williams M. L.,“Simulating Low Frequency SAR Clutter from a Pine Forest”, Proeedings of 3rd
European SAR Conference (EUSAR), 23-25 May, 2000, Munich, Germany, pp 149-152

Williams, M L S R Cloude, ‘Predictions of SAR Polarimetry and InSAR Coherence for a Model
Wheat Canopy”, Proceedings of IEEE Geoscience and Remote Sensing Symposium (IGARSS
2005), Seoul, South Korea, 25-29 July 2005

Yamada H, Y Yamaguchi, E Rodriguez, Y Kim, W M Boerner,“Polarimetric SAR Interferometry


for Forest Canopy Analysis by Using the Super-resolution Method”
IEICE Transactions on Electronics, VOL.E84-C, No.12, 2001, pp1917-1924, December 2001

Zebker H A, J Villasenor, “Decorrelation in Interferometric Radar Echoes”, IEEE Transactions on


Geoscience and Remote Sensing, Vol. 30. No. 5, pp. 950-959, September 1992
http://earth.esa.int/polsarpro/Manuals/2_Single_vs_Multi_Polarization_Interferometry.pdf

Embedded Secure Document


The file http://earth.esa.int/polsarpro/Manuals/2_Single_vs_Multi_Polarization_Interferometry.pdf is a
secure document that has been embedded in this document. Double click the pushpin to view.

http://earth.esa.int/polsarpro/Manuals/2_Single_vs_Multi_Polarization_Interferometry.pdf4/14/2007 2:16:43 PM
http://earth.esa.int/polsarpro/Manuals/1_Description_Of_Natural_Surfaces.pdf

Embedded Secure Document


The file http://earth.esa.int/polsarpro/Manuals/1_Description_Of_Natural_Surfaces.pdf is a secure
document that has been embedded in this document. Double click the pushpin to view.

http://earth.esa.int/polsarpro/Manuals/1_Description_Of_Natural_Surfaces.pdf4/14/2007 2:17:10 PM
http://earth.esa.int/polsarpro/Manuals/2_Rough_Surface_Scattering_Models.pdf

Embedded Secure Document


The file http://earth.esa.int/polsarpro/Manuals/2_Rough_Surface_Scattering_Models.pdf is a secure
document that has been embedded in this document. Double click the pushpin to view.

http://earth.esa.int/polsarpro/Manuals/2_Rough_Surface_Scattering_Models.pdf4/14/2007 2:17:26 PM
3. SINGLE VS MULTI-POLARIZATION
DESCRIPTORS

3.1 Polarisation and Surface Scattering

The main problem for the quantitative estimation of soil moisture and/or surface roughness
from SAR data lies in the separation of their individual effects on the backscattered signal.
Polarimetry plays here an important role as it allows either a direct separation or a
parameterisation of roughness and moisture effects within the scattering problem. The
scattering problem of electromagnetic waves from randomly rough surfaces has been an
actual research topic over decades and is still not satisfactorily solved due to the lack of an
exact closed-form solution. However, for many practical application, approximate solutions
are sufficient. In the absence of any direct relationship between surface parameters and
backscattering signal models empirical relations have been developed. In the field of radar
remote sensing the most common approximation methods are based on the evaluation of
backscattering amplitudes considering single or dual-channel SAR data. The choice of an
appropriate scattering model is essential for the quantitative estimation of the surface
parameters. On the one hand it must contain enough physical structure while on the other
hand it should have a right balance between the amount of parameters needed for their
description and available observables. As natural surfaces are complex stochastic objects a
priori information and assumptions can be used to simplify the inversion problem. Hence, for
some of the scattering model in order to obtain an accurate estimate of soil moisture,
information about the surface roughness was required or roughness has been considered as a
disturbing effect and several conditions have been developed in order to minimise its
influence (Hajnsek, 2001).
However, an independent estimate of roughness conditions is not possible by using only a
single polarisation and single- frequency SAR system. Increasing the amount of obervables by
using a fully polarimetric data set, the amount of surface parameters and its estimation
quantity increases. However, the main limitation of using models based on polarimetric
backscattering amplitudes is their insufficiency to deal - at leas in a practical way - with
diffuse or secondary scattering processes, depolarisation effects caused by the surface
roughness, and the presence of multiplicative and / or additive noise components.
A large class of natural surface scatterers, is characterised by secondary and/or multiple
scattering effects. With increasing surface roughness, relative to the wavelength, the effect of
multiple scattering becomes stronger, generating an adequate |HV| scattering component.
Dihedral scattering due to small correlation lengths characterised by |HH| > |VV|, and/or
diffuse scattering (|HV| contribution) affects the backscattered signal. Also, effects induced by
the presence of vegetation cover can not be accounted with surface scattering models. Both
effects lead to a violation of the required conditions of most models and/or to biased
estimation of the roughness and moisture parameters.
3.1.1 Polarimetric Coherence

The first way followed in order to extend the observation space was by incorporating
polarimetric phase information in form of (second-order) correlation (coherence) coefficients
between different polarisations. Indeed, correlation coefficients at different polarisations have
been reported to be sensitive to surface parameters. The HH-VV correlation coefficient

|< S HH SVV >|
γ HHVV := ∗ ∗ (1)
< S HH S HH > < SVV SVV >

has been evaluated with respect to its sensitivity to dielectric constant and/or rms height
condition of rough surfaces in (Borgeaud, 1994). An increased correlation to soil moisture
conditions within a certain roughness range has been reported. In contrary, (HH+VV)-(HH-
VV) correlation coefficient
(
|< S HH + S VV ) (S HH − S VV )

>|
γ ( HH +VV )( HH −VV ) := (2)
2 2
< S HH + S VV > < S HH − S VV >

has been found to correlate with surface roughness, independent on the surface moisture
content and the local incidence angle (Hajnsek, 2001).
Further investigations in (Mattia, 1997) demonstrated that the circular polarisation correlation
coefficient

|< S LL S RR >|
γ LLRR := ∗ ∗ (3)
< S LL S LL > < S RR S RR >

is widely independent on the dielectric properties of rough surfaces and depends only on the
surface roughness. Its independence from the azimuth topographic tilt has been reported in
(Schuler 2002). Thus, γLLRR can be used for robust quantitative surface roughness estimates.

More recently the polarimetric scattering anisotropy obtained from the eigen-values of the
polarimetric coherence matrix has been evaluated to be a sensitive estimate for the surface
roughness (Cloude 1999). This is in accordance with the observations about the circular
polarisation correlation coefficient as the anisotropy A can be interpreted as a generalised
rotation invariant expression of γLLRR
in case where the coherency matrix is diagonal.
λ2 − λ3
A := λ , λ3 Eigenvalues of [T] (4)
λ2 + λ3 2
Both parameters, γLLRR and A, are by definition normalised between 0 and 1 and have – in first
order - a direct linear relationship to the surface roughness
1 − γLLRR = ks = 1 − A for 0 ≤ ks ≤ 1 (5)

The price to be payed for using the rotation invariant anisotropy is its very ‘noisy’ nature
especially at the low range of its range i.e. for smooth surfaces (Hajnsek 2002). In this case
the amplitudes of the secondary scattering mechanisms, expressed by the second and third
eigenvalues are very small, down to -25 [dB] or even less. As this is close to the system noise
floor, A is strongly affected by additive noise effects. However, by additive filtering the
influence can be reduced.
Concerning the valuable surface roughness range, radar sensors operating at lower
frequencies (L-band) allow the coverage of a sufficiently wide range of natural surfaces, as it
can be seen in Tab 1.

ks ks
Band ID Wavelength [cm] for rms height 0.5
[cm] for rms height 4 [cm]
X 3 1.04 8.37
C 5 0.52 4.19
S 10 0.31 2.51
L 23 0.14 1
P 68 0.05 0.37
Tab.1: The surface roughness range for natural surface calculated for different wavelength
Based on these experimental observations a polarimetric scattering model using as a starting
point the SPM (Bragg-Model) an extended Bragg model (X-Bragg), has been developed,
which allows quantitative estimation of roughness and dielectric constant over a wide range of
natural bare surfaces. X-Bragg assumes reflection symmetric surfaces, where the axis of
symmetry is defined by the mean normal to the surface vector. It accounts for cross-polarised
as well as depolarisation effects. The application of the model to experimental data
demonstrate a high inversion accuracy shows a good agreement between inverted values and
ground measurements (Hajnsek, 2001).
However, the main limitation for surface parameters estimation from polarimetric SAR data is
the presence of vegetation. This combined with the fact that most natural surfaces are
temporarily or permanently covered by vegetation restricts significantly the importance of
radar remote sensing for a wide spectrum of geophysical and environmental applications.
However, the evolution of radar technology and techniques allows optimism concerning the
vitiation of this limitation.

3.2 The Extended or X-Bragg model

Based on these experimental observations, a new model for the investigation of surface
parameters from polarimetric SAR data is introduced in this paper. The proposed model is a
two component model including a Bragg scattering term and a roughness induced rotation
symmetric disturbance. In order to decouple the real part of the dielectric constant ε’ from
surface roughness s, the model is addressed in terms of the polarimetric scattering entropy
(H), scattering anisotropy (A) and alpha angle (α) which are derived from the eigenvalues and
eigenvectors of the polarimetric coherency matrix (Cloude et al. 1999). This allows the
implementation of a simple inversion algorithm for both surface roughness and soil moisture.

3.2.1 Second Order Scattering Description of Surface Scatterers

In order to describe correlation properties of natural scatterers, the polarimetric coherency


matrix which contains the second order moments of the scattering r
process is introduced. Its
formation is based on the introduction of a scattering vector k P as the vectorisation of the
scattering matrix [S] using the Pauli spin matrices basis set (Cloude et al. 1996, Cloude et al.
2001).
S S HV  v 1
[S HH + SVV 2 S HV ]
T
[ S ] =  HH → k P := S HH − SVV (6)
 SVH SVV  2
r
The coherency matrix [T] is formed by averaging the outer product of k P as
 S HH + SVV
2
( S HH + SVV )(S HH − SVV )* 2 ( S HH + SVV )S *HV 
 
v v 1 
=  ( S HH − SVV )(S HH + SVV )* 2 (S HH − SVV )S ∗ HV 
2
[ T ] := k P ⋅ k P+ S HH − SVV (7)
2
 
2 S HV (S HH + SVV )∗ 2 S HV (S HH − SVV ) ∗ 2
 4 S HV 
 

The diagonal elements of [T] are given by the real backscattered power values while
r
the off-
diagonal elements contain the complex cross correlations between the elements of k P . [T] is
by definition hermitian positive semi-definite, which implies that it has real non-negative
eigenvalues and orthogonal eigenvectors, and can always be diagonalised by an unitary
similarity transformation of the form (Cloude et al. 1997)
λ1 0 0
[T ] = [U 3 ][Λ ][U 3 ] where [Λ ] =  0 0  , [U 3 ] = [e1 , e2 , e3 ]
−1 v v v T
 λ2 (8)
 0 0 λ3 

[Λ] is a diagonal matrix with elements the real non-negative eigenvalues, 0 ≤ λ1 ≤ λ 2 ≤ λ3 of


[T], and [U3] is the unitary eigenvector matrix with columns corresponding to the orthonormal
eigenvectors ev1 , ev2 and ev3 . The diagonalisation of the coherency matrix [T] may be
interpreted as its decomposition into a non-coherent sum of three independent scattering
contributions
(v v
) (
v v
) v v
(
[T ] = [U 3 ][Λ][U 3 ]−1 = λ1 e1 ⋅ e1+ + λ2 e2 ⋅ e2+ + λ3 e3 ⋅ e3+ ) (9)

The contribution of each scattering mechanism in terms of power is given by the appropriate
eigenvalue. The information about which ‘kind’ of scattering mechanism is obtained is
contained in the corresponding eigenvectors.
There are three important physical features arising directly from the diagonalisation of the
coherency matrix. The first two are obtained from the eigenvalues of [T] which - normalised
by the absolute scattering magnitudes - can be interpreted as scattering probabilities pi
(Cloude et al. 2001)
λi λi
pi := = → p1 + p2 + p3 = 1 (10)
∑ λ1 + λ2 + λ3
3
j =1
λj

Using the probabilities pi two ratios for the description of an arbitrary rough surface can be
defined: the polarimetric scattering entropy H and the scattering anisotropy A, which are
defined as
3
H = −∑ pi log 3 pi (11)
i=1
p2 − p3
A = (12)
p2 + p3
Both parameters vary between 0 and 1. The entropy can be interpreted as a measure of the
randomness of the scattering process. The anisotropy defines the relation between the second
and the third eigenvalues, and is a measure for the difference of the secondary scattering
mechanisms. For smooth surfaces, H becomes zero implying a non-depolarising scattering
process described by a single scattering matrix and increases with surface roughness.
Depolarising surfaces are characterised by non-zero entropy values. A can be zero even for
rough surfaces. For surfaces characterised by intermediate entropy values, a high anisotropy
indicates the presence of only one strong secondary scattering process, while a low anisotropy
indicates the appearance of two equally strong scattering processes. For azimuthaly
symmetric surfaces λ2=λ3 by definition and A becomes 0 (Cloude et al. 1996, Cloude et al.
2001). In this sense, the anisotropy provides complementary information to the entropy and
facilitates the interpretation of the surface scatterer. The great advantage of these two
parameters arises from the invariance of the eigenvalue problem under unitary
transformations: The same surface leads to the same eigenvalues and consequently to the
same entropy and anisotropy values independently on the basis used to measure the
corresponding scattering matrix.
The third important parameter is obtained from the eigenvectors of [T]. Each eigenvector evi
can be expressed in terms of five angles as (Cloude et al. 1997)
v
ei = [cos α i exp (i φ 1 i ) sin α i cos β i exp (i φ 2 i ) sin α i sin β i exp (i φ 3 i )]
T
(13)
The β i angles can be interpreted as a rotation of the corresponding eigenvector evi in the
plane perpendicular to the scattering plane, while φ 1 i , φ 2 i and φ 3 i are accounting for the phase
relations between the elements of evi . More important in the context of this work is the mean
scattering angle α defined as
α = p 1α 1 + p 2α 2 + p 3α 3 (14)

The alpha angle α indicates the “type” of the mean scattering process occurring (Cloude et al.
1999). Its interpretation in terms of the surface scattering problem will be given in the
following sections.
r
Returning now to the SPM, the corresponding scattering vector k P is given by
 S HH + SVV   Rs + RP  cos α exp(iφ 1 )
→ 1 
kP = S − S  = m  R − R  = m  sin α exp(iφ )
(15)
2 
HH VV  s S P  2 

 2 S HV   0   0 

where m denotes the absolute scattering amplitude. The coherency matrix of a Bragg scatterer
follows as


RS + RP
2
(R
S + R p )(RS − RP )
*
0

r r+
[T ] = k P kP = ms  (RS − RP )(RS + RP )
2


*
RS − RP
2
0
 (16)
 0 0 0
 
Two of the three off-diagonal elements disappear as a consequence of zero cross-polarisation,
while the third one depends only on the surface dielectric constant and the radar incidence
angle. The corresponding normalised correlation coefficient is one
| ( S HH + SVV )( S HH − SVV )* | | ( RS + R p )( RS − RP )* |
γ ( HH +VV )( HH −VV ) := = =1 (17)
| S HH + SVV |2 | S HH − SVV |2 | RS + RP |2 | RS − RP |2

as a consequence of the inability of the SPM to describe depolarisation effects.


Regarding now the interpretation of the alpha angle, from (15) follows that α − similar to the
co-polarised ratio Rs / Rp − is independent of roughness, and therefore, can be used for the
estimation of the dielectric constant if the local incidence angle is known
 
 RS + RP 
α = arccos  (18)
 RS + RP
2
+ RS − RP
2

 
As shown in Figure 1, for a dry surface the alpha angle α is small and increases with
increasing soil moisture. The highest sensitivity occurs between 0 and 20 mv [vol. %] and
saturates with further increasing of mv.

60 degrees

25 degrees

Figure 1 The relation of the alpha angle to the volumetric soil moisture
Summarising, the main limitations of the SPM are its small surface roughness validity range –
below 0.3 ks and the saturation of its sensitivity to soil moisture content above mv 20 [vol. %].
Its inability to describe depolarisation effects restricts its usefulness regarding the
interpretation and inversion of experimental data from natural surfaces. However, the
robustness of SPM inside its validity range and its relevant physical background leaded to
several investigations to use it as a valuable starting point for an extended model.
As mentioned in the previous section, two main effects have to be introduced in order to
extend the Bragg scattering model for a wider range of roughness conditions: non-zero cross-
polarised backscattering and depolarisation. In fact, conventional two- or multiple-scale
extensions introduce cross-polarisation but fail to express depolarisation effects. An
alternative way to introduce roughness disturbance is to model the surface as a reflection
symmetric depolariser by rotating the Bragg coherency matrix [T] about an angle ß in the
plane perpendicular to the scattering plane (Lee et al. 2000) as

1 0 0 
 RS + RP
2
(R S − R p )(RS + R P )
*
0 1
 0 0 
[T ( β )] = 0 cos 2β sin 2β   (RS + R P )(RS − R P )

*
RS − R P
2
0 0 cos 2 β

− sin 2β  (19)
0 − sin 2 β cos 2 β   0 0 0 0 sin 2 β cos 2β 
 

and performing a configurational averaging over a given distribution P(ß) of ß:



[T ] = ∫ [T ( β )]P( β )dβ (20)
0

The width of the assumed distribution corresponds to the amount of roughness disturbance of
the modelled surface. Assuming P(ß) to be a uniform distribution about zero with width ß1
(see Figure 2)
 1
 β ≤ β1
P (β ) =  2 β 1 (21)
 0 ≤ β1 ≤ π
 2
the coherency matrix for the rough surface becomes
T11 T12 T13   C1 C 2 sin c (2 β 1 ) 0 
[T ] = T21 T22  
T23  = C 2 sin c (2 β 1 ) C 3 (1 + sin c (4 β 1 )) 0 
 (22)
T31 T32 T33   0 0 C 3 (1 − sin c (4 β 1 ))

(11.50)
Sensor

Scattering Plane

Azimuthal Oriented Surface

β
Surface

k
P(β)

β
−β1 +β1
Figure 2 Schematic representation of the uniform distribution of the slope
with sinc(x) = sin(x) / x. The coefficients C1, C2 , and C3 describing the Bragg components of
the surface, and are given by
C1 = R S + R P
2
(
C 2 = (R S + R P ) R S* − R P* ) C3 =
1
2
RS − RP
2
(23)

Equation (22) represents the coherency matrix of an extended Bragg surface characterised by
cross-polarised energy and at the same time a polarimetric coherence less than one
T12 sin c(2 β 1 )
γ ( HH +VV )( HH −VV ) := = ≤1 (24)
T11 T22 (1 + sin c(4 β 1 )) / 2.
The width of the distribution ß1 describes the roughness component and controls both, the
level of cross-polarised power and the (HH+VV)(HH-VV) coherence. Fig. (11.14) show the
variation of (HH+VV)(HH-VV) coherence (Fig. (11.14)) and normalised cross-polarised
power according to (22). In the limit of a smooth surface (i.e. ß1 = 0), the (HH+VV)(HH-VV)
coherence is one and the HV backscattered power zero. In this case, the coherency matrix
obtains the form of the “pure” Bragg coherency matrix (see (16)). With increasing roughness
(i.e. increasing ß1), the HV power increases, while the (HH+VV)(HH-VV) coherence decreases
monotonically from 1 to zero. In this high roughness limit, the surface becomes azimuthally
symmetric.
Figure 3 Variation of the coherence and the cross polarised power with increasing β1

Furthermore, it is important to realise that, while the increase of cross polarised power with
roughness depends on the dielectric constant of the surface (and the incidence angle) the
decrease of the (HH+VV)(HH-VV) coherence is independent from both.
On the other hand, according to (22), the ratio
2
T22 + T33 R − RP
= S 2 (25)
T11 RS + RP
is independent on the surface roughness and depends only on the dielectric constant of the
surface and the incidence angle as shown in Figure 4. Thus, the extended Bragg model is
characterised by an inherent separation of roughness and moisture effects obtained in form of
ratios of the elements of the coherency matrix allowing an independent estimation of these
parameters.

Figure 4 . The polarisation ratio as a function of the β1 parameter.


http://earth.esa.int/polsarpro/Manuals/4_Estimation_Of_Surface_Characteristics.pdf

Embedded Secure Document


The file http://earth.esa.int/polsarpro/Manuals/4_Estimation_Of_Surface_Characteristics.pdf is a
secure document that has been embedded in this document. Double click the pushpin to view.

http://earth.esa.int/polsarpro/Manuals/4_Estimation_Of_Surface_Characteristics.pdf4/14/2007 2:18:32 PM