**I.1 SAR Polarimetry**

**I.1.2 Polarimetric decomposition**

Target decomposition (TD), introduced in the ﬁrst place in [9], aims to interpret polarimet-ric data by assessing and analysing the components involved in the scattering process [10].

Roughly speaking, the assessing assumes the derivation of the involved backscattering com-ponents, while the analysis dominantly concerns their parametrization. The former can be deﬁned as:

X= XN

i=1

k_{i}X_{i}, (I.7)

with X being the scattering matrix (S) in case of a coherent target, or the coherence matrix
(T) in case of an incoherent one. Both the means of deriving the components (X_{i}) and their
parametrizations, diﬀer for diﬀerent types of decompositions. Here, we present the ones we
found to be the most notable with respect to their historical and practical relevance.
After-wards, in the Chapter III, we will elaborate a new decomposition, the principal contribution
of this thesis.

I.1.2.1 Coherent decompositions

The most elementary approach in decomposing a scattering matrix is based on a set of Pauli matrices, originally used by Wolfgang Pauli in his theory of quantum-mechanical spin. Due to their numerous interesting mathematical properties (e.g. hermitian, unitary and commutation properties), these matrices found their applications in many domains, among which, the SAR polarimetry. In case of monostatic conﬁguration, the scattering matrix is decomposed into standard mechanisms as:

12 Chapter I. POLSAR image and BSS
making this approach a "model based" one. The ﬁrst term in Eq. I.8 represent the
odd-bounce scattering: ﬂat surface, sphere, trihedral. The second one is related to the even-bounce
scattering without polarization change - a dihedral with the axe of symmetry being parallel to
the incident horizontally polarized wave. The third one represent the scatterer which favours
the cross-polarized channel - a dihedral with the axe of symmetry rotated45^{◦} with respect to
the incident horizontally polarized wave [6].

A slightly diﬀerent model based approach is proposed by Krogager in [11]. The second component in this case rather represents a dihedral with any orientation, while the third one represents a helix. The decomposition takes form of a:

S_{hh} S_{hv}

with the angles φ,φ_{s}, and ✓ being respectively the absolute phase, the single (odd) bounce
component phase, the orientation of the dihedral. The real coeﬃcients k_{s}, k_{d}, k_{h} are the
contributions of single bounce, double bounce and helix scatterer. The latter one represents
the non-symmetrical scattering i.e. the case when the target axe of symmetry doesn’t lie in
the plane perpendicular to the line of sight.

However, this decomposition is usually employed in the circular basis making this a suitable point for introducing the scattering matrix projected on the circular polarization basis. In the same way the projection onto the Pauli basis leads to the target vector, the projection onto the circular polarization ones, gives us the circular scattering matrix, which for the monostatic conﬁguration takes the following form:

S_{C} =

The lexicographic ordering of the these elements leads to the circular target vector:

k_{C} =h
S_{ll} p

2S_{lr} S_{rr}iT

. (I.12)

The most representative algebraic coherent decomposition would be the Cameron decom-position [12]. In this approach we cannot actually assume reciprocity, meaning that we have to consider all four elements of the scattering matrix. The decomposing process can be roughly divided on two steps:

• The ﬁrst step is related to the target geometrical properties. More precisely, we are initially trying to isolate the symmetric scattering from the non-symmetric one. By maximizing the former and consequently, minimizing the latter one, we can express the scattering matrix as:

I.1. SAR Polarimetry 13

The ↵ parameters correspond to the ones introduced in Eq. I.9 and is the rotation angle with respect to the reference basis.

• The second step would be algebraic, given that it represents one sort of the parametri-sation of the symmetric part. Namely, if we express the matrix given in Eq. I.14 as:

S^{max}_{sym} =

1 0 1 z

�

. (I.16)

we can deﬁne a vector:

λ(z) = 1 corre-spond respectively to sphere, dihedral, cylinder, narrow dihedral), the symmetric target can be categorized by calculating the closeness to these elementary reﬂectors:

λ(z) = 1
p1 +|z|^{2}

p 1

1 +|z_{ref}|^{2}|1 +z^{⇤}z_{ref}|. (I.18)
I.1.2.2 Incoherent decompositions

Unlike it was the case with the coherent decompositions, here we are rather concentrated
either on the already introduced coherence matrix, or on the covariance matrix, derived as
the spatial covariance of the lexicographic target vector k_{c}:

C = E⇥

14 Chapter I. POLSAR image and BSS Theﬁrst incoherent decomposition was proposed by Huynen in [9], and it assumes decom-posing the scattering as the incoherent sum of the fully polarized mechanisms and the fully unpolarized ones. Originally, the decomposition is based on Mueller matrix formalism:

M=M_{pol}+M_{unpol} (I.20)

but the same reasoning can be equally applied on either covariance or coherence matrix.

The polarized matrix is then parametrized using nine parameters, among which only two are invariant with respect to the rotation around the line of sight.

The counterpart of formerly presented decompositions on standard mechanisms, in case of incoherent targets, are Freeman [13] and Yamaguchi [14] decompositions.

Freeman decomposition assumes decomposing the covariance matrix into the sum of ﬁrst-order Bragg, double bounce and volume scattering:

C = C_{s}+C_{d}+C_{v} = (I.21)

Bragg scattering would be a more elaborated odd-bounce backscattering, which can provide
us with some details about the surface dynamics. The real parametersβ,f_{s},f_{d}andf_{v}, and the
complex parameter↵,ﬁguring in Eq. I.22 are supposed to be derived from a set of equations:

h|S_{hh}|^{2}i = f_{s}β^{2}+f_{d}|↵|^{2}+f_{v} (I.22)
h|Svv|^{2}i = fs+f_{d}+fv

hS_{hh}S_{vv}^{⇤} i = f_{s}β+f_{d}↵+f_{v}/3
h|S_{hv}|^{2}i = f_{v}/3.

However, there is one more unknown variable than there are equations. Therefore, in order
to make this a well-posed problem, we need to consider the sign of hS_{hh}S_{vv}^{⇤} i−h|S_{hv}|^{2}i. If it
is negative, we take↵=−1, if not β= 1.

In case of Yamaguchi, there is an additional fourth component, representing non-symmetric, helix backscattering:

I.1. SAR Polarimetry 15
Another diﬀerence would be a complexβ parameter, makingC_{s}_{31} to be rather β^{⇤}. In this
case, we have a system ofﬁve equations:

h|S_{hh}|^{2}i = f_{s}|β|^{2}+f_{d}|↵|^{2}+ 8

Two parameters can be derived from the second and the last equation (Eq. I.25), while the rest of the system can be solved using the same hypothesis as in the case of Freeman decomposition.

The incoherent equivalents of the Cameron algebraic method would be the Cloude and Pottier [15] and the Touzi [16] decompositions. The particular emphasise will be put on introducing these two, given that the highlight of this thesis concern one incoherent target decomposition.

The ﬁrst step in both of these two decompositions is the eigenvector decomposition of the target coherence matrix, allowing us to represent the total backscattering as a sum of three backscattering mechanism. Namely, given the Hermitian nature of positive semi-deﬁnite coherence matrix, the derived eigenvectors are mutually orthogonal and characterized by real eigenvalues. Each of the derived eigenvectors forms a coherence matrix with unity rank and therefore happens to be a fully polarized target vector. The corresponding eigenvalue (λ) represents its contribution to the total backscattering.

T=λ_{1}k_{1}k^{H}_{1} +λ_{2}k_{2}k^{H}_{2} +λ_{3}k_{3}k^{H}_{3} (I.25)
The principal diﬀerence between two decompositions would be, the second step - the
parametrisation of the estimated target vectors i.e. backscattering mechanisms.

The Cloude and Pottier decomposition is based on the↵−β−γ−δ parametrisation of a target vector [15]:

Among the obtained parameters, the very central place has the angle ↵_{p}, used in the

16 Chapter I. POLSAR image and BSS

(a) (b)

Figure I.4: Poincaré sphere representation of TSVM parameters: (a) symmetric scattering
(⌧_{m} = 0), (b) non-symmetric scattering (Φ↵s = 0).

derivation of the mean backscattering mechanism (↵_{p}), conditioned by the probabilities of↵_{pi}
to occur in the random sequence of parameters (Pi):

↵_{p}=P_{1}↵_{p1}+P_{2}↵_{p2}+P_{3}↵_{p3} (I.27)
Aside from this one, very important polarimetric descriptor would be the entropy, deﬁning
the appropriateness of using polarimetry [6]:

H= X3

i=1

−P_{i}log_{3}P_{i}, P_{i}= λ_{i}
P_{3}

i=1λ_{i}. (I.28)

The value zero indicates the absolute dominance of one fully polarized mechanism, while the value one points to the completely depolarizing target.

The third polarimetric descriptor we ought to mention would be the anisotropy:

A= λ_{2}−λ_{3}

λ_{2}+λ_{3} = P_{2}−P_{3}

P_{2}+P_{3}, (I.29)

describing the relation of the second and the third estimated mechanisms.

The Touzi decomposition is rather based on the Target Scattering Vector Model (TSVM) [16].

I.2. SAR images statistics 17