IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO.

4, APRIL 2010 1809

Coherency Matrix Estimation of Heterogeneous

Clutter in High-Resolution Polarimetric SAR Images
Gabriel Vasile, Member, IEEE, Jean-Philippe Ovarlez, Member, IEEE,
Frederic Pascal, Member, IEEE, and Céline Tison

Abstract—This paper presents an application of the recent ad- ML Maximum likelihood.

vances in the field of spherically invariant random vector (SIRV) MPWF Multilook polarimetric whitening filter.
modeling for coherency matrix estimation in heterogeneous clut- P Generic span.
ter. The complete description of the polarimetric synthetic aper-
ture radar (POLSAR) data set is achieved by estimating the span PDF Probability density function.
and the normalized coherency independently. The normalized PWF Polarimetric whitening filter.
coherency describes the polarimetric diversity, while the span SCM Sample covariance matrix.
indicates the total received power. The main advantages of the SDAN Span-driven adaptive neighborhood.
proposed fixed-point (FP) estimator are that it does not require
SIRP Spherically invariant random process.
any a priori information about the probability density function of
the texture (or span) and that it can directly be applied on adaptive SIRVs Spherically invariant random vectors.
neighborhoods. Interesting results are obtained when coupling this [T ] Generic polarimetric coherency matrix.
FP estimator with an adaptive spatial support based on the scalar
span information. Based on the SIRV model, a new maximum-
likelihood distance measure is introduced for unsupervised I. I NTRODUCTION
POLSAR classification. The proposed method is tested with both
simulated POLSAR data and airborne POLSAR images provided
by the Radar Aéroporté Multi-Spectral d’Etude des Signatures
system. Results of entropy/alpha/anisotropy decomposition, fol-
A SYNTHETIC aperture radar (SAR) measures both the
amplitude and phase of a backscattered signal, producing
one complex image for each recording. With the sensors being
lowed by unsupervised classification, allow discussing the use of
the normalized coherency and the span as two separate descriptors able to emit or receive two orthogonal polarizations, fully
of POLSAR data sets. polarimetric SAR (POLSAR) systems describe the interactions
Index Terms—Estimation, heterogeneous clutter, polarimetry, between the electromagnetic wave and the target area by means
segmentation, synthetic aperture radar (SAR). of the Sinclair matrix [1]. Among the difficulties encountered
when using POLSAR imagery, one important feature is the
N OMENCLATURE presence of speckle. Occurring in all types of coherent imagery,
the speckle is due to the random interference of the waves scat-
BN Boxcar neighborhood. tered by the elementary targets belonging to one resolution cell
[C] Generic covariance matrix. [2]. In general, POLSAR data are locally modeled by a multi-
FP Fixed point. variate zero-mean circular Gaussian PDF, which is completely
i.i.d. Independent and identically distributed. determined by the covariance matrix [3].
LLMMSE Locally linear minimum mean-squared error. The recently launched POLSAR systems are now capable
[M ] Generic normalized polarimetric coherency of producing high-quality images of the Earth’s surface with
matrix. meter resolution. The decrease of the resolution cell offers the
opportunity to observe much thinner spatial features than the
Manuscript received July 21, 2008; revised December 8, 2008, June 12, decametric resolution of the up-to-now available SAR images.
2009, and September 21, 2009. First published December 15, 2009; current
version published March 24, 2010. This work was supported by the Centre
Recent studies [4] show that the higher scene heterogeneity
National d’Etudes Spatiales, the French space agency. leads to non-Gaussian clutter modeling, particularly for urban
G. Vasile is with the Grenoble Image Speech Signal Automatics Laboratory, areas. One commonly used fully polarimetric non-Gaussian
National Council for Scientific Research (CNRS), Grenoble Cedex 38402,
France (e-mail: [email protected]).
clutter model is the product model [5]: The spatial non-
J.-P. Ovarlez is with the French Aerospace Laboratory (ONERA), Palaiseau homogeneity is incorporated by modeling the clutter as a prod-
Cedex 91761, France, and also with the SONDRA Laboratory, Supelec, uct between the square root of a scalar random variable (texture)
National University of Singapore, Defence Science and Technology Agency
Research Alliance, Gif-sur-Yvette Cedex 91192, France (e-mail: ovarlez@
and an independent zero-mean complex circular Gaussian
onera.fr). random vector (speckle). If the texture random variable is sup-
F. Pascal is with the SONDRA Laboratory, Supelec, National University posed to be a Gamma spatial distributed intensity, the product
of Singapore, Defence Science and Technology Agency Research Alliance,
Gif-sur-Yvette Cedex 91192, France (e-mail: [email protected]). model is equivalent to the well-known K-distributed clutter
C. Tison is with the Centre National d’Etudes Spatiales, Toulouse Cedex 9 model [6], [7].
31401, France (e-mail: [email protected]). For a Gaussian polarimetric clutter model, the estimation
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. of the polarimetric coherency matrix is treated in the context
Digital Object Identifier 10.1109/TGRS.2009.2035496 of POLSAR speckle filtering. The POLSAR adaptive filtering
0196-2892/$26.00 © 2009 IEEE
1810 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

techniques can roughly be divided into two main classes [8]: II. H ETEROGENEOUS M ODEL FOR P OLARIMETRIC
based on the optimization of the spatial support and based T EXTURED S CENES
on the use of the local statistics to derive adaptive estimators.
The goal of the estimation process is to derive the scene
These two directions are not exclusive since both of them
signature from the observed data set. In the case of spatially
can be applied simultaneously [9], [10]. For example, the
changing surfaces (“heterogeneous” or “textured” scenes), the
refined Lee filter couples eight edge-aligned directional neigh-
first step is to define an appropriate model describing the depen-
borhoods with an adaptive estimator based on the LLMMSE
dence between the polarimetric signature and the observable
criterion [9].
as a function of the speckle. In general, the multiplicative
In the context of the non-Gaussian polarimetric clutter mod-
model [5] has been employed for SAR data processing as a
els, several studies tackled POLSAR parameter estimation us-
product between the square root of a scalar positive quantity
ing the product model. For deterministic texture, Novak et al.
(texture) and the description of an equivalent homogeneous
derived the PWF by optimally combining the elements of the
surface (speckle) by means of the following:
polarimetric covariance matrix to produce a single scalar image
[11], [12]. Using the complex Wishart distribution, the PWF 1) the intensity descriptor for single-polarization SAR im-
for homogeneous surfaces has been generalized to an MPWF ages [22], [23];
[13], [14]. In general, the texture random variable is speci- 2) the complex SAR signal descriptor for single-polarization
fied by the PDF. For Gamma-distributed texture, Lopes and SAR data [24];
Sery [13] derived the ML estimator of the covariance matrix. 3) the polarimetric target vector descriptor in lexicographic
Moreover, the vector spatial LLMMSE filter applied on the basis for monostatic POLSAR images [11], [25], [26];
scalar ML texture estimator has also been introduced when the 4) the normalized polarimetric target vector descriptor in
texture variance and spatial correlation functions are a priori lexicographic basis [6], [27], [28];
known [13]. In [15], DeGrandi et al. performed an exten- 5) the polarimetric covariance matrix descriptor for
sive study on the dependence of the normalized second-order POLSAR data [13], [29].
moment of intensity on polarization state for a K-distributed In this paper, the polarimetric descriptors used are the tar-
clutter model. This dependence was condensed in a graphical get vectors k = [k1 , k2 , k3 ]T in the Pauli basis (monostatic
form by a formalism called the polarimetric texture signature. acquisition). The following section presents an application of
This study has been applied for target detection and texture the recent advances, in the field of SIRV modeling [20], for
segmentation using the discrete wavelet transform generated estimating span and normalized coherency matrices of high-
with the first derivative of a B-spline of order three as mother resolution POLSAR data.
wavelet [16].
The POLSAR information allows the discrimination of dif-
ferent scattering mechanisms. In [17], Cloude and Pottier in- A. Gaussian Model
troduced the target entropy and the entropy–alpha–anisotropy The elements of a vector are generally modeled by a multi-
(H−α−A) model by assigning to each eigenvector the cor- variate zero-mean complex Gaussian random process. The PDF
responding coherent single scattering mechanism. Based on is given by the following expression [2]:
this decomposition, unsupervised classification for land appli-
1  
cations was performed by an iterative algorithm based on the pm (k) = exp −k† [T ]−1 k (1)
complex Wishart density function [18], [19]. πm det {[T ]}
The objective of this paper is to present a new coherency
estimation technique [20] based on the SIRV model [21] and where [T ] = E{kk† } is the polarimetric coherency matrix,
to analyze the consequences that this model has on the con- det{. . .} denotes the matrix determinant, † is the conjugate
ventional POLSAR processing chain. This paper is organized transpose operator, m is the dimension of the target vector
as follows. Section II is dedicated to the presentation of the (m = 3 for monostatic POLSAR acquisitions), and E{. . .}
proposed estimation scheme. The heterogeneity of the polari- denotes the statistical mean over the polarimetric channels.
metric textured scenes is taken into account by coupling the ML According to (1), a Gaussian stochastic process is completely
normalized coherency estimator with adaptive neighborhoods characterized by the coherency matrix. In this case, the ML
(ANs) driven on the scalar ML span estimators. A new ML estimator of the polarimetric coherency matrix is the SCM
distance measure is also introduced for classifying normalized obtained by replacing the statistical mean by spatial averaging
coherency matrices under the SIRV model. In Section III, the N
1 
results obtained using the proposed approach are presented [T]SCM = ki k†i (2)
and compared to those given by the Gaussian ML estimator. N i=1
Results of the H−α−A decomposition, followed by unsuper-
vised POLSAR classification, allow discussing the use of the where N is the number of samples. The SCM is statistically
normalized coherency and the span as two separate descriptors determined by the Wishart PDF [2].
of POLSAR data sets. Detailed discussion on the advantages Another POLSAR parameter is the span (or total power) P
and the limitations of the SIRV model is given in Section IV. generally defined for each pixel as [1]
Eventually, in Section V, some conclusions and perspectives are
presented. PSLC = k† k. (3)
VASILE et al.: COHERENCY MATRIX ESTIMATION OF HETEROGENEOUS CLUTTER 1811

The corresponding multilook span can be estimated within a 1) Model Identification: When using the product model, an
local neighborhood according to identification problem can be pointed out: The SIRV model is
  uniquely defined with respect to the covariance matrix param-
P = E k† k = Tr {[T ]} (4) eter up to a multiplicative constant. Let [M1 ] and [M2 ] be
where Tr{[T ]} denotes the trace of the matrix [T ]. Hence, the two covariance matrices such that [M1 ] = κ · [M2 ] ∀κ ∈ R∗+ .
common span estimator for the Gaussian case can directly be Notice that √the two sets of parameters defined as {τ1 , [M1 ]} and
obtained from the SCM as {τ2 = (τ1 / κ), [M2 ]} describe the same SIRV. For solving
  this identification problem, the covariance matrix has to be
PSCM = Tr [T]SCM . (5) normalized. In the following, the covariance matrix [M ] is nor-
malized such that Tr{[M ]} = m, with m being the dimension
of the target vector.
B. SIRV Model
One important consequence of the imposed normalization
SIRVs and their applications to estimation and detection in condition is that the resulting normalized polarimetric co-
communication theory were first introduced by Yao [21]. The herency matrix reveals information concerning the polarimetric
SIRV is a class of nonhomogeneous Gaussian processes with diversity only: The total power information is transferred into
random variance. The complex m-dimensional measurement the texture random variable. The POLSAR data can fully be
k is defined as the product between the independent complex characterized by coupling the normalized coherency matrix
circular Gaussian vector z (speckle) with zero mean and covari- with the span descriptor
ance matrix [M ] = E{zz† } and the square root of the positive
random variable τ (representing the texture) PSLC = k† k = τ (z† z). (8)
√
k = τ z. (6) When operating on the polarimetric statistical axis, the span for
the SIRV case is given by
It is important to notice that, in the SIRV definition, the PDF
   
of the texture random variable is not explicitly specified. As P = E τ (z† z); τ = τ · E z† z = τ · Tr {[M ]} = τ · m.
a consequence, SIRVs describe a whole class of stochastic (9)
processes defined by (6). This class includes the conventional
clutter models having Gaussian, K-distributed, Chi, Rayleigh, An estimate of P can be obtained when considering τ as an
Weibull, or Rician PDFs [30]. unknown deterministic parameter from cell to cell.
For POLSAR data, the SIRV product model is the product of 2) Stationarity Definition: In the following, several generic
two separate random processes operating across two different concepts are recalled. Given a SIRP, this process is wide-sense
statistical axes. stationary if and only if both the texture random variable and
1) The polarimetric diversity is modeled by a multi- the speckle random vector are wide-sense stationary. As the
dimensional Gaussian kernel characterized by its covari- speckle is a zero-mean complex Gaussian vector, the latter
ance matrix [M ]. means that the statistical samples ki used in the estimation
2) The randomness of spatial variations in the radar back- process must have the same theoretical covariance matrix [M ].
scattering from cell to cell is characterized by τ . The cor- This condition is called “matrix stationarity.”
responding random process operates along the spatial axis However, as the results presented in this section can be
given by the image support. Relatively to the polarimetric applied whatever the texture PDF (∀p(τ )), the previous prop-
axis, the texture random variable τ can be viewed as an erties can be reformulated using the SIRV class of stochastic
unknown deterministic parameter from cell to cell. processes. Given a “matrix stationary” stochastic process, this
One major advantage of the SIRV clutter model is the high process is “SIRV homogeneous” if and only if the texture
degree of generality with respect to other texture-aware models random variable is “texture homogeneous,” where texture ho-
employed in the literature [4], [30]. Nevertheless, this model is mogeneous means that it is possible to define a texture PDF
founded on the validity of three basic assumptions: The texture (∃p(τ )) such that the stochastic process can be described by
random variable affects the backscattered power only; it is the product model from (6). We illustrate these properties using
multiplicative and spatially uncorrelated. When applied to high- four local populations which often occur in practical POLSAR
resolution POLSAR clutter, the SIRV model postulates that the applications.
texture descriptor τ from (6) is identical for all polarization 1) One zero-mean Gaussian process with covariance matrix
channels. [M ]: N (0, [M ]). Being a “Gaussian stationary”1 process,
Now, let p(τ ) be the texture PDF associated to the SIRV it is also “SIRP stationary” and SIRV homogeneous. This
model. The SIRP corresponding to (6) has the following model is widely used for POLSAR data analysis [32].
PDF [31]: 2) Two adjacent Gaussian processes with different covari-
ance matrix: N = {N (1) (0, [M ]1 ), N (2) (0, [M ]2 )}. The
+∞
1 Gaussian mixture N is neither SIRP stationary nor SIRV
F {p(τ ), [M ]} = pm (k) = homogeneous as the matrix stationarity condition is not
(πτ )m det {[M ]}
0
k† [M ]−1 k 1 A “Gaussian stationary” process is a stochastic process whose Gaussian
× exp − p(τ ) dτ. (7)
τ PDF does not change when shifted in time or space.
1812 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

respected. Generally, such cases are treated by employing The ML estimator of the normalized covariance matrix in the
adaptive estimation schemes [8], [9] in order to approxi- deterministic texture case is obtained by canceling the gradient
mate the local “Gaussian stationarity” condition. of Lk with respect to [M ] as the solution of the following
3) One K-distributed process [33] with Gamma-distributed recursive equation:
texture pG (τ ; τ , ν) and covariance matrix [M ]:  
FK {pG (τ ; τ , ν), [M ]}. This process is SIRP stationary [M ]FP = f [M ]FP
as it is “K stationary”,2 but obviously, it is not Gaussian
N
stationary. m ki k†i
4) Two adjacent K-distributed processes with two different = †
(1) (2) N i=1 ki [M ]−1 ki
Gamma texture PDFs pG (τ ; τ 1 , ν1 ), pG (τ ; τ 2 , ν2 ) FP

and the same covariance matrix [M ]: FK = N

(1) (1) (2) (2) m zi z†i
{FK {pG (τ ; τ 1 , ν1 ), [M ]}, FK {pG (τ ; τ 2 , ν2 ), [M ]}}. = . (13)
(1) (2) N i=1 z†i [M ]−1 zi
The K-distributed processes FK and FK are SIRP FP
stationary and K stationary, but the mixture FK is not
K stationary. Despite this, the process FK is SIRV This approach has been used in [38] by Conte et al. to derive
homogeneous as it is possible to define a texture PDF a recursive algorithm for estimating the matrix [M ]. This al-
which models the Gamma mixture. As a consequence, gorithm consists in computing the FP of f using the sequence
the results presented in this section can still be applied in ([M ]i )i≥0 defined by
this case.
[M ]i+1 = f ([M ]i ) . (14)
In conclusion, the two properties to be verified in order to
apply the SIRV model are the matrix stationarity and the “tex- This study has been completed by the work of Pascal et al.
ture homogeneity.” Moreover, the latter considerably relaxes [20], [39], which recently established the existence and the
the “texture stationarity” condition required when using explicit uniqueness, up to a scalar factor, of the FP estimator of the
texture models such as the Gamma or the Fisher PDF. normalized covariance matrix, as well as the convergence of
3) SIRV Parameter Estimation: In the field of target detec- the recursive algorithm whatever the initialization. The al-
tion for radar applications, the SIRV model led to many inves- gorithm can therefore be initialized with the identity matrix
tigations [34]–[37]. In (6) and (7), the normalized covariance [M ]0 = [Im ]. One way to analyze the convergence of the
matrix is an unknown parameter which can be estimated from FP estimator consists in evaluating the following criterion:
the ML theory. In [31], Gini and Greco derived the exact ML  
 
estimate [M ] of the normalized covariance matrix when τi are [M ](i + 1) − [M ](i)
assumed to be unknown deterministic parameters. For N i.i.d. C(i) =   F
(15)
 
data, the likelihood function to maximize with respect to [M ] [M ](i)
F
and τi is given by
where . . . F represents the Frobenius norm. When comput-
Lk (k1 , . . . , kN ; [M ], τ1 , . . . , τN ) ing the FP estimator, (14) is iterated until C becomes smaller
N
 than a predefined lower limit. Note that only a few iterations
1 1 k†i [M ]−1 ki
= × exp − . (10) suffice to reach an error that is less than 10−15 [20].
π mN det {[M ]}N τm
i=1 i
τi It has also been shown in [31] and [38] that the recursive
estimation scheme from (14) can be applied to derive an exact
For a given [M ], maximizing Lk (k1 , . . . , kN ; [M ], ML estimator of the normalized covariance matrix
τ1 , . . . , τN ) with respect to τi yields the texture ML estimator  
† −1
N h
m+1 k i [M ] k
ML i
k†i [M ]−1 ki [M ]ML =
m
ki k†i
τi = . (11) N i=1 hm (k†i [M ]−1 ki )
m ML

Replacing τi in (10) by their ML estimates, the generalized ∞ q

likelihood is obtained as with hm (q) = τ p exp p(τ ) dτ. (16)
τ
0
1
Lk (k1 , . . . , kN ; [M ]) =
π mN det {[M ]}N In the previous equation, the exact ML estimator depends on
N
the texture PDF through the SIRV density-generating func-
mm exp(−m) tion hm (q). Chitour and Pascal [40] have recently demonstrated
×  m . (12)
† −1 that (16) admits a unique solution and that its corresponding
i=1 ki [M ] ki
iterative algorithm converges to the FP solution for every
admissible initial condition. Pascal et al. [20], [39] have also
demonstrated that the normalized covariance ML estimator de-
2 A “K stationary” process is a stochastic process whose K PDF does not veloped under the deterministic texture case (13) yields also an
change when shifted in time or space. approximate ML estimator under stochastic texture hypothesis.
VASILE et al.: COHERENCY MATRIX ESTIMATION OF HETEROGENEOUS CLUTTER 1813

We propose to apply these results in estimating normal- MPWF is that it takes into account the correlation between
ized coherency matrices for high-resolution POLSAR data. the different polarization channels (speckle) in the whitening
The main advantage of this approach is that the local “scene process.
heterogeneity” can be taken into account without any a priori
hypothesis regarding the texture random variable τ [(14) does
C. Spatial Support
not depend on τ ]. The obtained FP is the approximate ML esti-
mate under the stochastic τ assumption and the exact ML under In the estimation process, a certain number of samples must
the deterministic τ assumption. Moreover, the normalized po- be gathered for deriving the observation vector. In this purpose,
larimetric coherency matrix estimated using the FP method is the boxcar sliding neighborhood is usually employed. The
unbiased and asymptotically Gaussian distributed [20], [39]. main drawback of nonadaptive BN is that the available number
Note also that the texture estimator from (11) can directly be of samples is directly proportional with the loss of spatial
linked to the total scattered power (span) according to (9). By resolution. In order to deal with this undesired effect, several
estimating the normalized coherency as the FP solution of (13), strategies to obtain locally ANs were proposed for POLSAR
the derived estimate is independent of the total power, and it data processing. In [8], three local neighborhoods are analyzed,
contains polarimetric information only. Using this matrix, it is and their performances are discussed with respect to different
possible to compute the SIRV span ML estimator for unknown end-user applications (visual interpretation, classification, etc.).
deterministic τ as Experiments on real data sets have shown that the intensity-
driven adaptive neighborhood (IDAN) represents, on the whole,
PPWF = k† [M ]−1
FP k. (17) a good tradeoff between preserving signal characteristics and
gathering a significant number of samples for coherency and
One can observe that the span estimator from (17) has the H−α−A parameter estimation [8], [43].
same form as the PWF introduced by Novak and Burl in [11]. Recent studies have revealed that the original IDAN algo-
The only difference is the use of the normalized coherency rithm tends to introduce a bias with respect to the radiometry
estimate given by the FP estimator instead of the conven- information [44]. The main reasons are the use of a symmetric
tional SCM. confidence interval around the mean for the Gamma-distributed
Finally, it is possible to derive an estimate of the conventional intensity and the estimation of the initial seed by the median
polarimetric coherency matrix according to (6) computed within a 3 × 3 neighborhood. In order to deal
PPWF with these problems, the SDAN algorithm has been introduced
[T]FP = [M ]FP . (18) in [45]. It allows using heterogeneous scene models, such as
m
SIRV, in the estimation step. Note that this approach is not
4) Gaussian Model in the SIRV Context: The multivariate optimal as the resulting AN is driven on the texture (span)
Gaussian distribution presented in Section II-A is obviously a information only. One may use other existing locally ANs (e.g.,
member of the SIRV class. Let us assume N i.i.d. realizations directional neighborhoods [9]), but, up to now, the existing
of the target vector k. The SCM from (2) is the ML estimator AN algorithms are also tributary to the span information.
of [T ] in the Gaussian clutter but not in the clutter described SDAN successively truncates the texture PDF using two
by the product model [41]. In the specific case of completely symmetric confidence intervals around the mean. The trunca-
correlated texture (τ = τi ∀i ∈ {1, . . . , N }), Richmond [42] tion thresholds are expressed with Gamma-distributed texture.
proved that the SCM is, again, the exact ML estimator of However, different PDFs can be truncated according to the
[M ] provided that the M normalization is respected. In fact, same thresholds (initially set using a Gamma prior). In this
the completely correlated τ case is equivalent to the Gaussian paper, the SDAN is employed to eliminate eventual outliers
model for a given realization of data across all resolution from the local neighborhood. The main advantage of this ap-
cells [31]. Consequently, it is possible to define the normalized proach consists in selecting spatially connected pixels within a
SCM as certain confidence interval. Its main inconvenience is the esti-
mation bias which can be induced by truncating the significant
[T]
[M ]SCM = m  SCM  . (19) part of the unknown texture PDF.
Tr [T]SCM Within the SIRV context, the SDAN algorithm operates
under deterministic texture hypothesis: If τ is deterministic,
In other words, in Gaussian clutter, the local power P is no the span statistics over matrix stationary areas is given by the
more random in (9), but m · τ = P with probability one [31]. Gamma PDF resulting from the complex Gaussian kernel. This
Based on this consideration and according to (17), the MPWF is coherent with the general hypothesis adopted for POLSAR
can be defined as speckle filtering, stating that the local matrix stationarity prop-
erty is revealed by changes in the span image when texture is
N
1  † absent [9].
PMPWF = k [Mi ]−1
FP ki . (20)
N i=1 i
D. Application to POLSAR Parameter Estimation
The MPWF is the span ML estimator for Gaussian clutter with
known power P , and it is unbiased [13], [14]. When compared One way to derive the normalized coherency matrix is the
to the span estimator from (5), the main advantage of the normalized sample covariance estimator, obtained by locally
1814 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

replacing the statistical mean by spatial average within the

sliding neighborhood W
m
[M ]SCM (i, j) =   [T]SCM

Tr [T ]SCM

1 
with [T]SCM = k(p, q)k† (p, q)
card {W(i, j)}
(p,q)∈W(i,j)
Fig. 1. Proposed estimation scheme.
(21)
this tradeoff by applying the LLMMSE criterion for the span
estimation [9]
where (i, j) represents the current range/azimuth position and
card{W} denotes the cardinal of W. The main advantage of PLLMMSE = PMPWF +αLLMMSE (PP W F − PMPWF )
the [M ]SCM estimator consists in deriving the polarimetric  
covariance matrix independently of the span for the Gaussian σP2 W F 1+σn2
with αLLMMSE = 2 (23)
case. The normalized SCM estimator presents also one major σMPWF −μ2MPWF σn2
disadvantage: It is not SIRP stationary, and, as a consequence,
where μMPWF , σPWF , and σMPWF are the signal mean and
this estimator is not consistent over textured areas. Although
standard deviations computed inside the local estimation neigh-
the derivation of the normalized SCM estimator from the stan-
borhood, respectively, and σn is the noise standard deviation
dard SCM estimator is straightforward, we could not find any
(a priori known). In (23), the two span estimators can be
specific paper to report its use for POLSAR data.
computed according to (17) and (20).
In this paper, we propose to extend the estimation of the
In the last stage, it is also possible to unify these two descrip-
normalized polarimetric coherency matrix by using a hetero-
tors by multiplying them according to the SIRV model from (6).
geneous scene model over the sliding neighborhood. The FP
An important remark is that, by multiplying the two descriptors,
estimator of the normalized covariance matrix for the SIRV
the separation between the total received power (span) and
model is applied using the procedure described in Section II-B.
the polarimetric information (speckle normalized coherency) is
More precisely, the FP normalized coherency matrix is com-
lost. Finally, the resulting coherency matrix [T] does not obey
puted iteratively as
the Wishart PDF as it depends on the estimated span PDF.
m In summary, this section introduces a novel estimation
[Ml ]FP (i, j) = scheme (see Fig. 1) for deriving normalized polarimetric co-
card {W(i, j)}
herency matrices and resulting estimated span. The proposed
 k(p, q)k† (p, q) algorithm couples span-driven multiresolution techniques [45]
× with heterogeneous SIRV scene models [20] to deal with
(p,q)∈W(i,j) k† (p, q)[Ml−1 ]−1
FP (i, j)k(p, q)
the polarimetric texture inside the estimation neighborhood.
It is important to notice that the proposed FP estimator uses
with [M0 ]FP = [Im ] (22) normalized coherency matrix inversion, and thus, it works
only with Hermitian positive definite normalized coherency
where l is the iteration index. Equation (22) gives the covari- matrices. This constraint is still acceptable since, in practice,
ance matrix estimate of the SIRV complex Gaussian kernel image coherency matrices are generally of full rank (three for
without imposing any statistical constraint over the texture monostatic POLSAR data) [46]. However, in the specific case
random variable τ . The resulting matrix [M ]FP is asymptot- of a noninvertible matrix, which can correspond to a strongly
ically Gaussian distributed. The proposed procedure (SDAN- polarized scattered signal, the SIRV model can be applied by
FP) starts by computing the AN using the SDAN algorithm [45] using only the nonzero signal subspace.
at each range/azimuth position. The resulting AN is supposed to
respect the matrix stationarity condition. Finally, the FP estima-
E. Distance Measures for POLSAR Segmentation
tor is applied to derive the normalized polarimetric coherency
matrix estimate under a compound Gaussian polarimetric clut- Classification of ground cover with POLSAR data is an
ter model (22). important application [17]–[19], [47]–[49]. Generally, one has
Another physical parameter to be estimated is the total to find a distance between the pixel covariance matrix [C] and
power. For the SIRV model, the PWF span estimator is the the class center [C]ω . Based on this distance, conventional clus-
ML estimator; hence, it should be applied for textured areas. tering methods have already been introduced with POLSAR
However, on Gaussian textureless areas, a stronger speckle re- data: “naive” Bayesian ML classifier or K-means [18], fuzzy
duction can be obtained using the MPWF estimator. In practical K-means, or expectation maximization [47].
applications, the PWF and the MPWF estimators should be When the POLSAR data are modeled by a stochastic process
applied as follows: On Gaussian stationary regions, the best with a known PDF, it is possible to derive optimal ML distance
span estimator is the MPWF, while on SIRV homogeneous measures (e.g., the Wishart distance for Gaussian processes).
areas only, the PWF should be applied. We propose to deal with In [27], Yueh et al. derived an optimal ML distance measure
VASILE et al.: COHERENCY MATRIX ESTIMATION OF HETEROGENEOUS CLUTTER 1815

for terrain cover classification using the normalized target the test given by (27). This procedure is here simplified, con-
vector in the lexicographical basis. The adopted normalization sidering a particular SIRV process with a texture characterized
condition was the Euclidian norm, and the distance measure by an unknown deterministic parameter. Consequently, each
was computed applying the Bayesian ML classifier with the resolution cell is now associated with its own p(τ ) = δ(τ −
PDF of the normalized polarimetric data. Note that, in Yueh’s τn ) in (7), where τn are the unknown deterministic texture
approach, the covariance matrix is estimated using the SCM variables. This way, the texture descriptor can be discarded for
(ML estimator only with Gaussian clutter). In consequence, each pixel independently.
the derived optimal distance is a generalized ML distance for By taking the natural logarithm in (28), one obtains
Gaussian clutter only. N  
We propose the following general binary hypothesis test for det{[M ]ω}  k†n [M ]−1 ω −[M ]
−1
kn
ln(ΛSIRV ) = −N ln − .
a given class ω: det{[M ]} n=1 τn
 (29)
H0 : [C] = [C]ω
(24)
H1 : [C] = [C]ω . Now, since τn ’s and [M ] are unknown, they are replaced by
their ML estimates from (11) and (13). The resulting general-
According to the Neyman–Pearson Lemma, the likelihood ratio ized LRT ΛSIRV is given by
test (LRT) provides the most powerful test [50]
det {[M ]ω }
pm (k1 , . . . , kN /H1 ) ln(ΛSIRV ) = − N ln  
Λ= . (25) det [M ]FP
pm (k1 , . . . , kN /H0 )
N
 k†n [M ]−1
ω kn
For Gaussian clutter, maximizing the LRT from (25) and re- −m + N m. (30)
placing the pixel coherency matrix [T ] with the ML estimate n=1 k†n [M ]−1
FP kn
[T]SCM are equivalent to minimizing the conventional Wishart Maximizing the generalized LRT over all classes is equivalent
distance to minimizing the following SIRV distance:
  det {[T ]ω }  
DWishart [T]SCM , [T ]ω = ln   det {[M ]ω }
DSIRV [M ]FP , [M ]ω = ln  
det [T]SCM det [M ]FP
  N
+ Tr [T ]−1 
ω [T ]SCM . (26)
m  k†n [M ]−1
ω kn
+ † −1
. (31)
N n=1 kn [M ] kn
FP
This distance has been widely used for supervised and unsuper-
vised POLSAR data clustering [18], [19], [47]. Notice that computing the distance from (31) needs the original
In the case of the SIRV model, one can rewrite the hypothesis scattering vectors kn .
test as In this paper, the distance measure from (31) is used as a dis-
 √ similarity measure in the conventional K-means clustering for
H0 : [M ] = [M ]ω ⇔ k = √τ z, with z ∼ N (0, [M ]) POLSAR data. The full description of the K-means algorithm
H1 : [M ] = [M ]ω ⇔ k = τ z, with z ∼ N (0, [M ]ω ) can be found in [18].
(27) In summary, this section introduces a new distance measure
where τ is the unknown deterministic texture. between normalized coherency matrices. The resulting approx-
For a given class [M ]ω , the LRT with respect to the texture τ imate generalized ML distance is optimal for POLSAR data
and the normalized coherency matrix [M ] is given by characterized by the SIRV model.
 † −1  An interesting remark concerning the SIRV distance can be
N 1
n=1 π τn det{[M ]ω }
m m exp −kn [Mτn]ω kn observed in (31). On the one hand, when the texture (span
ΛSIRV =   † −1  . (28) information) is high, the second term of the SIRV distance
N kn [M ] kn
n=1 π m τ m det{[M ]} exp −
1
n τn DSIRV becomes small, and the distance measure is dominated
by the determinant ratio. This usually corresponds to strongly
Notice that the likelihood function in (28) does not use polarized targets with a dominant scattering mechanism (e.g.,
the stochastic texture description as the PDF p(τ ) is sup- dihedral, trihedral, etc.). On the other hand, with smaller span
posed unknown in the SIRV model. As previously stated in values, the distance is dominated by the second term which
Section II-B3, the texture parameter τ can be considered ei- takes into account the N observed samples. This second case
ther as a random variable with unknown PDF p(τ ) or as an often corresponds to distributed targets.
unknown deterministic parameter with PDF p(τ ) = δ(τ − τn )
which characterizes yet a particular SIRV process. It can be
III. R ESULTS AND D ISCUSSION
shown that the ML estimation of the coherency matrix yields a
good approximate ML estimate in the first case and the true ML This section has two main objectives. The first one consists
estimate in the second case [31], [38]. With the general PDF in evaluating the performance of the normalized coherency
being unknown, it is therefore impossible to derive a texture- estimation techniques presented in Section II. The second
independent closed-form expression for the likelihood ratio of objective is to show the improvement in the conventional
1816 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

coherency are computed over the SE quadrant. A global error

measure for the normalized coherency matrix is also intro-
duced as
 
 
1 
N [M i ] − [M ref 
]
F
= (32)
N i=1 [Mref ] F

where [Mref ] is the reference normalized coherency matrix used

for data simulation. As observed in Table I, the best results are
obtained using the BN-SCM estimator. Being the ML estimator
for Gaussian stationary regions, the SCM is used as a bench-
Fig. 2. Simulated POLSAR data, Gaussian case (200 × 200 pixels). (a) Initial
one-look span estimated using (3). (b) Amplitude color composition of the
mark for the SDAN and the FP estimator. The mean value is
target vector elements k1 − k3 − k2 . well preserved for both BN-FP and SDAN-FP estimates, while
the variance of the BN-FP is smaller than the measured variance
POLSAR processing chain brought by introducing the normal- of the SDAN-FP. The latter observation is explained by the fact
ized coherence matrix related to the SIRV model. that the BN is optimal on such SIRV homogeneous regions.
Three different estimation techniques are analyzed: the nor- One can also note that, despite the mean of each element of
malized SCM coupled with the 7 × 7 BN (BN-SCM) and the the normalized coherency being quite similar, a better error
FP estimator coupled either with the 7 × 7 BN (BN-FP) or measure is provided by the parameter. Using the Frobenius
with the SDAN (SDAN-FP). In all three cases, the correspond- norm, which is a norm associated to the inner product on the
ing span image is estimated using the LLMMSE estimator ring of all complex matrices, the corresponding error shows
from (23). The parameters used for the SDAN algorithm are that the smallest error is obtained for the optimal BN-SCM
Leq = 3 and Nmax = 50. estimator. When introducing the FP estimator, increases, and
it increases even more by using the SDAN adaptive spatial
support. This behavior corresponds to the expected theoretical
A. Simulated POLSAR Data
observations. However, it is important to notice that, for both
As, for real data, it is impossible to find reference regions BN-FP and SDAN-FP, the error is not increased by more than
with known coherency matrix, the effectiveness of the esti- 7% with respect to the ML estimator. This is acceptable for the
mation schemes is demonstrated using simulated POLSAR POLSAR applications where the clutter is characterized by a
data [30]. Gaussian stationary stochastic process.
1) Gaussian Case: The first POLSAR data set consists of A similar objective performance assessment is carried out
four adjacent Gaussian regions as presented in Fig. 2. Each for the estimation of the span image. Table II shows the span
of the four quadrants is associated with a known deterministic mean ratio and the speckle coefficient of variation computed
texture value and a known theoretical covariance matrix. Using for the same Gaussian stationary region. As for the normalized
these parameters, each component of the polarimetric target coherency, the bias in the estimated radiometry is less than
vector is simulated accordingly. Fig. 2(a) shows the initial 7% for all three estimation techniques. An interesting remark
span image computed using (3), and Fig. 2(b) shows the re- consists in the fact that, when using the FP estimator, the
sulting amplitude color composition of the three target vector bias from Table II is also linked with the average computed
components. over the corresponding homogeneous region. Even if the mean
The LLMMSE span P and the normalized coherency mat- ratio is a standard parameter for evaluating the speckle filter
rix [M ] are estimated using the three different estimation radiometric bias, this parameter is not so well adapted for
schemes. Note that, in the Gaussian case, the optimal ML esti- the FP estimation. Although being asymptotically Gaussian
mation technique is the BN-SCM from Fig. 3(a) and (d). Inside distributed, the FP estimator is outperformed by the SCM
each quadrant, the stochastic process characterizing the data is (ML estimator for Gaussian clutter) with a fixed number of
Gaussian stationary; hence, it is also SIRV homogeneous. The samples. Consequently, the average over a homogeneous area
BN-FP estimation yields quite similar results, from the visual should be coupled with the estimation of the FP normalized
point of view, as shown in Fig. 3(b) and (e). However, the coherency over the same homogeneous population for optimal
use of the BN is associated with the well-known edge-blurring performances. This effect can be noticed with the SDAN-FP
effect as the matrix stationarity condition is not respected over span, where the local SDAN can gather more than 49 samples
the transitions between the quadrants. The SDAN-FP estima- over Gaussian stationary areas. The resulting SDAN-FP span
tion reduces this undesired effect as presented in Fig. 3(c) estimator exhibits a radiometric bias that is less than 1% (same
and (f). As a general remark, the blurring is more present within as for the BN-SCM).
the normalized coherency diagonal elements than within the In summary, the subjective and objective performance as-
span images due to the use of the adaptive LLMMSE span sessment carried out for Gaussian POLSAR clutter shows that,
estimator. despite being suboptimal, the proposed FP estimator and the
In order to objectively assess the estimation performances, SDANs give good overall performances. The corresponding
the mean and the variance for each element of the normalized error measure is less than 7% for all estimation schemes.
VASILE et al.: COHERENCY MATRIX ESTIMATION OF HETEROGENEOUS CLUTTER 1817

Fig. 3. Simulated POLSAR data, Gaussian case (200 × 200 pixels). Square root of the LLMMSE span image using the normalized coherency estimated by
(a) BN-SCM, (b) BN-FP, and (c) SDAN-FP. Color composition of the normalized coherency diagonal elements [M ]11 − [M ]33 − [M ]22 estimated by
(d) BN-SCM, (e) BN-FP, and (f) SDAN-FP.

TABLE I
S IMULATED POLSAR DATA , G AUSSIAN C ASE : M EAN AND S TANDARD D EVIATION OF THE
N ORMALIZED C OHERENCY E LEMENTS OVER G AUSSIAN S TATIONARY A REAS

TABLE II
/μref AND C OEFFICIENT OF VARIATION OVER G AUSSIAN S TATIONARY A REAS
S IMULATED POLSAR DATA , G AUSSIAN C ASE : S PAN M EAN R ATIO μ

2) SIRV Case: The second simulated POLSAR data set shows the LLMMSE span P and the normalized coherency ma-
proposes the same four quadrants but with Gamma-distributed trix [M ] estimated using the three different estimation schemes.
texture [Fig. 4(a)]: Each quadrant is K distributed. The tex- As the data set is not Gaussian, the PWF span estimator
ture coefficient of variation used for simulation is equal to is dominant in the LLMMSE criterion, and the correspond-
three, which corresponds to a highly non-Gaussian clutter ing speckle reduction is performed using only three samples.
(urban areas). Fig. 4(b) shows the initial span image. Fig. 4(c) Hence, concerning the LLMMSE span estimation, BN-SCM,
shows the corresponding amplitude color composition of the BN-FP, and SDAN-FP [Fig. 5(a)–(c)] look similar from the
three target vector components. visual point of view.
The overall data set is not SIRV homogeneous as the ma- The effectiveness of the FP estimator in compound Gaussian
trix stationarity condition is not respected on the boundaries; clutter can be observed in Fig. 5(e) and (f). While the
however, each quadrant is SIRP stationary. In the following, we BN-SCM normalized coherency [Fig. 5(d)] presents a “patchy”
shall use only the “SIRV homogeneity” assumption over each appearance, the BN-FP estimation [Fig. 5(e)] provides better
quadrant, namely, the texture PDF is supposed unknown. Fig. 5 visual homogeneity within each quadrant. The adaptive SDAN
1818 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

Fig. 4. Simulated POLSAR data, SIRV case (200 × 200 pixels). (a) Texture image. (b) Initial one-look span estimated using (3). (c) Amplitude color composition
of the target vector elements k1 − k3 − k2 .

Fig. 5. Simulated POLSAR data, SIRV case (200 × 200 pixels). Square root of the LLMMSE span image using the normalized coherency estimated by
(a) BN-SCM, (b) BN-FP, and (c) SDAN-FP. Color composition of the normalized coherency diagonal elements [M ]11 − [M ]33 − [M ]22 estimated by
(d) BN-SCM, (e) BN-FP, and (f) SDAN-FP.

spatial support [Fig. 5(f)] assures better edge preservation for Kolmogorov–Smirnov (KS) test with respect to the reference
the transitions between quadrants. One important issue is that span used for simulation. The resulting KS values, com-
the diagonal elements of the BN-FP normalized coherency for puted over the entire span image, indicate that BN-FP outper-
the SIRV case [Fig. 5(e)] have the same visual aspect as for the forms BN-SCM. The best results are reported when using the
previous Gaussian POLSAR data set [Fig. 3(e)]. This shows SDAN-FP estimator. Note that the KS distance is rather small
that the FP estimate of the covariance matrix does not depend ∈ (0.07, 0.12) in all three cases.
on the texture PDF.
Using the same reference region as for the Gaussian case,
B. Airborne POLSAR Data
Table III presents the mean and the variance for each element of
the normalized coherency and also the overall error measure To illustrate the improvements in the standard POLSAR
computed for the three estimation schemes. BN-FP and SDAN- processing chain, the results obtained with high- and very high
FP outperform BN-SCM in retrieving the reference value and resolution airborne data are reported. Both data sets were ac-
also in terms of variance reduction. Since the matrix stationarity quired by the airborne French Aerospace Laboratory (ONERA)
is always assured within the reference region, BN-FP outper- RAMSES system [51].
forms the SDAN-FP also. Finally, another interesting result 1) High-Resolution POLSAR Data: The first POLSAR data
consists in the fact that Table III indicates the same BN-FP set was acquired in Brétigny, France. The mean incidence angle
value for the error parameter as in the Gaussian case (Table I). is 30◦ . It represents a fully polarimetric (monostatic mode)
This objective issue confirms the visual comparison mentioned X-band acquisition with a spatial resolution of approximately
in the previous paragraph. 1.5 m in range and azimuth.
Objective performance assessment has been carried out for Fig. 6(a) shows the color composition of the target vector
the LLMMSE span estimation also. Table IV presents the amplitudes. The target area is composed of three buildings, a
VASILE et al.: COHERENCY MATRIX ESTIMATION OF HETEROGENEOUS CLUTTER 1819

TABLE III
S IMULATED POLSAR DATA , SIRV C ASE : M EAN AND S TANDARD D EVIATION OF THE
N ORMALIZED C OHERENCY E LEMENTS OVER SIRP S TATIONARY A REAS

TABLE IV mation techniques [Fig. 7(d)–(i)]. Both parameters exhibit the

S IMULATED POLSAR DATA , SIRV C ASE : T HE KS T EST
KSn = maxx |Fn (x) − Fref (x)| FOR THE S PAN D ISTRIBUTION same behavior:
1) BN-SCM: patchy appearance mainly due to the texture;
2) BN-FP: blurring as the matrix stationarity condition is not
respected;
3) SDAN-FP: higher spatial feature preservation but more
variance.
As the target area is highly heterogeneous, the SDAN-FP
estimation is a good tradeoff between robust estimation and
spatial resolution preservation.
Finally, it is possible to derive the SDAN-FP polarimetric co-
herency matrix as a product between the span image [Fig. 7(c)]
and the corresponding normalized coherency [Fig. 7(f)]. In
Fig. 8(a)–(c), the SDAN-FP coherency is compared with the
conventional coherency matrices obtained by the SCM estima-
tor coupled with two spatial supports: the BN and the IDAN [8].
Subjective visual assessment can be expressed in terms of
the hue–saturation–lightness color space [53] by associating
Fig. 6. Brétigny, RAMSES POLSAR data, X-band (501 × 501 pixels). the lightness to the span and the saturation to the polari-
(a) Amplitude color composition of the target vector elements k1 − k3 − k2 . metric diversity. The SDAN-FP coherency from Fig. 8(c) ex-
(b) Optical image (137 × 137 pixel zoom of the initial span superposed for hibits better performances in terms of lightness and saturation,
illustrating the region of interest).
which means that both the span and the normalized coherency
are better estimated. The corresponding H−α classification
parking lot, and the surrounding agricultural areas. For further maps [17] are shown in Fig. 8(d)–(f). For H−α classification
illustration, a non-Gaussian urban (building) region has been also, the SDAN-FP coherency provides better performances as
selected, namely, the span image superposed over the airborne it achieves stronger noise reduction than the IDAN filter.
photograph from Fig. 6(b). One key issue to be discussed is whether the normalized
The LLMMSE span and the normalized coherency matrix coherency matrix and the span should be aggregated in the final
are estimated using the three different estimation schemes estimation step or not (the question mark from Fig. 1). Most of
(Fig. 7). The BN-FP span illustrated in Fig. 7(b) exhibits better the existing processing chains use the conventional coherency
whitening in the estimation process than the BN-SCM span matrix for representing POLSAR data for unsupervised land
from Fig. 7(a). This can be observed on the isolated brilliant cover classification [18], [19], [43], [47] and for target detection
pointwise structures surrounding the building. However, both applications [12], [33]. Due to the SIRV model identification
BN-SCM and BN-FP are tributary to the “ring effect” (two problem discussed in Section II-B, the complete description
large dips on a spatial profile near the boundaries of a pointwise of the POLSAR data set is achieved by estimating the span
target) induced by coupling the BN spatial support with the and the normalized coherency independently. The latter de-
LLMMSE estimator [52]. This effect is reduced in the SDAN- scribes the polarimetric diversity, while the span indicates the
FP span image as it can be observed over the metallic structures total received power. Moreover, the FP estimation of the nor-
that are present on the roof of the building from Fig. 7(c). malized coherence does not depend on the span information.
Visual assessment is carried out also with the normalized co- Given these facts, we propose to investigate this problem in
herency [M ] estimates. Color compositions, constructed from the framework of unsupervised POLSAR classification. The
either the diagonal elements of [M ] or the corresponding classification scheme discussed in the following is the stan-
H−α−A parameters [17], are computed for the three esti- dard Wishart H−α segmentation [18]. For segmenting the
1820 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

Fig. 7. Brétigny, RAMSES POLSAR data, X-band (137 × 137 pixels). Square root of the LLMMSE span image using the normalized coherency estimated
by (a) BN-SCM, (b) BN-FP, and (c) SDAN-FP. Color composition of the normalized coherency diagonal elements [M ]11 − [M ]33 − [M ]22 estimated by
(d) BN-SCM, (e) BN-FP, and (f) SDAN-FP. Color composition of the polarimetric H−α−A parameters estimated by (g) BN-SCM, (h) BN-FP, and (i) SDAN-FP.

Fig. 8. Brétigny, RAMSES POLSAR data, X-band (137 × 137 pixels). Color composition of the coherency diagonal elements [T ]11 − [T ]33 − [T ]22 estimated
by (a) BN, (b) IDAN, and (c) SDAN-FP after multiplication with the LLMMSE span from Fig. 7(c). H−α classification results using (d) BN, (e) IDAN, and
(f) SDAN-FP.
VASILE et al.: COHERENCY MATRIX ESTIMATION OF HETEROGENEOUS CLUTTER 1821

Fig. 9. Brétigny, RAMSES POLSAR data, X-band (501 × 501 pixels). LLMMSE span using the normalized coherency estimated by SDAN-FP: (a) Span
image, (d) Gamma unsupervised classification, and (g) physical mechanism identification (odd-bounce classes, even-bounce classes, and volume class) using the
SDAN-FP normalized coherency. SDAN-FP coherency matrix after span multiplication: (b) Color composition of the diagonal elements [T ]11 − [T ]33 − [T ]22 ,
(e) Wishart unsupervised classification, and (h) physical mechanism identification. SDAN-FP normalized coherency matrix: (c) Color composition of the diagonal
elements [M ]11 − [M ]33 − [M ]22 , (f) SIRV unsupervised classification, and (i) physical mechanism identification.

normalized coherency, we have modified the Wishart H−α

algorithm by replacing the Wishart distance with the SIRV ML
distance discussed in Section II-E. For comparison, we have
also used the scalar Gamma K-means classification with H−α
initialization.
Fig. 9 shows the POLSAR unsupervised classification re-
sults using three descriptors estimated by SDAN-FP: span
[Fig. 9(a)], coherency [Fig. 9(b)], and normalized coherency
[Fig. 9(c)]. The selected scene is composed of both Gaussian
(agricultural fields) and non-Gaussian (urban) areas. This case
is encountered in many practical POLSAR classification ap-
plications. Fig. 9(e) shows the eight-class segmentation map
obtained using the SDAN-FP coherency matrix. When com- Fig. 10. Toulouse, RAMSES POLSAR data, X-band (500 × 500 pixels),
resolution azimuth and range of 50 cm. (a) Amplitude color composition of
pared to the scalar unsupervised classification map [Fig. 9(d)] the target vector elements k1 − k3 − k2 . (b) Optical image.
obtained using the span only, one can observe the high degree
of similarity between them. This leads to the following con- sociated SIRV distance. The visual assessment in Fig. 9(e) and
clusion concerning the Brétigny data set: The Wishart H−α (f) reveals that a significant part of the polarimetric information
classification is mainly influenced by the information contained is lost when using the standard coherency matrix: The building
in the span image. The same behavior has been reported over class separation is lost, as well as the natural canonical targets
alpine glaciers, with L-band E-SAR data [45]. Regarding the (trihedral, dihedral, etc.) that are present over different “field”
polarimetric information, Fig. 9(f) shows the classification map classes. One important remark concerning the Wishart H−α
computed using the normalized coherency matrix and the as- classification is that a large number of samples are usually
1822 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

Fig. 11. Toulouse, RAMSES POLSAR data, X-band (500 × 500 pixels), resolution azimuth and range of 50 cm. Color composition of the diagonal elements
[T ]11 − [T ]33 − [T ]22 : (a) Lee refined filter and (b) SDAN-FP coherency matrix after span multiplication. (c) Color composition of the diagonal elements
[M ]11 − [M ]33 − [M ]22 estimated by SDAN-FP. Wishart unsupervised classification: (d) Coherency estimated by the Lee refined filter and (e) SDAN-FP
coherency matrix after span multiplication. (f) Unsupervised classification of the SDAN-FP normalized coherency based on the SIRV distance measure. Physical
mechanism identification (odd-bounce classes, even-bounce classes, and volume class) using (g) coherency estimated by the Lee refined filter, (h) SDAN-FP
coherency matrix after span multiplication, and (i) SDAN-FP normalized coherency.

assigned to the class feature vector when iterating the K-means Fig. 11 shows the visual assessment of the proposed esti-
clustering algorithm. Due to this, locally Gaussian areas (agri- mation scheme applied to very high resolution POLSAR data
cultural fields) may become heterogeneous regions as neither acquired in an urban environment. The obtained results are
the matrix stationarity nor the texture homogeneity conditions visually compared to those obtained by the refined Lee filter
are respected. operating under Gaussian clutter hypothesis [9]. With a 50-cm
The same behavior can also be observed in Fig. 9(g)–(i) spatial resolution, the SDAN-FP normalized coherency from
with the classification maps obtained after basic scattering Fig. 11(c) reveals higher variability in polarimetric signatures
mechanism identification [54]. The use of polarimetric indi- than with a 1.5-m spatial resolution [Fig. 7(f)]. Fig. 11(b) shows
cators, derived from the eigenvector–eigenvalue decomposition the color composition of the diagonal elements of the SDAN-
of the normalized coherency matrix, allows the interpretation FP coherency matrix after multiplication by the corresponding
of each cluster scattering mechanism from Fig. 9(d)–(f). In all LLMMSE span. When compared to the polarimetric coherency
three cases, the POLSAR parameters were computed using the derived by the refined Lee filter [Fig. 11(a)], the SDAN-FP
SDAN-FP normalized coherency from Fig. 9(f). The observed coherency better preserves the polarimetric and radiometric
scene is then classified into three canonical scattering types: signatures over thin spatial features (brilliant points and edges),
even bounce (blue or cyan), odd bounce (red or dark red), and while over larger structures (buildings, fields, and roads), the
volume scattering (green) [55]. two images look similar. This can also be observed in the eight-
2) Very High Resolution POLSAR Data: The second class segmentation maps obtained by Wishart H−α clustering
POLSAR data (Fig. 10) set was acquired in Toulouse, France, [Fig. 11(d) and (e)]. It is important to stress that, for very high
with a mean incidence angle of 50◦ . It represents a fully po- resolution urban POLSAR data, the polarimetric coherency
larimetric (monostatic mode) X-band acquisition with a spatial matrix is not Wishart distributed. Hence, the unsupervised
resolution of approximately 50 cm in range and azimuth. H−α classification based on the SIRV distance measure can
VASILE et al.: COHERENCY MATRIX ESTIMATION OF HETEROGENEOUS CLUTTER 1823

properly be applied using the SDAN-FP normalized coherency. covariance matrix parameter is replaced by the SCM, or the
The result is shown in Fig. 11(f). Finally, the three classifi- SCM depends on the texture PDF p(τ ), and it is not the ML
cation maps are interpreted according to the basic scattering estimator of the covariance matrix in the K-distributed clutter.
mechanism identification procedure [54]. The subjective visual The exact ML normalized covariance estimator can be derived
assessment indicates that quite realistic results are obtained us- using Yao’s representation theorem for SIRVs, and its exact
ing the SDAN-FP normalized coherency descriptor [Fig. 11(i)]: expression is given in (16).
Buildings and cars are mainly retrieved in the red “odd-bounce” The product model has also been used by Novak et al. [11],
class, while “even-bounce” scattering mechanism (cyan class) [12] for deriving the PWF. Based on this result, Lopes and Sery
appears on the flat regions (roads). [13] derived the MPWF as well as the adaptive LLMMSE filter
In conclusion, the joint analysis of the span and the nor- for Gaussian and K-distributed clutter. The SIRV representation
malized coherency presents several advantages with respect theorem allows the derivation of the PWF as an ML estimator
to the coherency matrix descriptor: separation between the of the deterministic texture. Once the texture parameter is
total received power and the polarimetric information, esti- obtained for every resolution cell, further statistical processing
mation of the normalized coherency matrix independently of can be applied over a population of texture parameters (e.g., the
the span, and the existence of the SIRV distance measure proposed LLMMSE span filter).
for unsupervised ML classification of normalized coheren- For Gaussian clutter, Lee et al. [18], [19], [47] introduced
cies. However, the span–normalized-coherency description of optimal polarimetric covariance matrix classification schemes
POLSAR images raises new problems which still remain under based on the Wishart distance. The proposed methods can
investigation. The first issue concerns the use of span for be extended to the SIRV model by using the SIRV distance
testing the matrix stationarity condition for the normalized presented in Section II-E. Moreover, the asymptotic distribution
coherency estimation. This test is currently used for POLSAR of the FP estimator from (13) has been derived in [39]. The
data speckle filtering, and it is founded on the basic principle FP estimator computed with N samples (secondary data) con-
that changes within the polarimetric signature are revealed by verges in distribution to the normalized SCM computed with
changes in the total received power. Consequently, one may N [m/(m + 1)] secondary data. Since the normalized SCM is
envisage other estimation schemes dedicated to the SIRV model the SCM up to a scale factor, we may conclude that, in prob-
with stochastic texture by considering external estimators of lems invariant with respect to a scale factor on the covariance
matrix stationarity. The second important remark concerns matrix, the FP estimate is asymptotically equivalent to the SCM
the Wishart unsupervised classification scheme. Although all computed with N [m/(m + 1)] secondary data.
statistical requirements employed for unsupervised classifica- We can conclude that Yao’s representation theorem allows
tion are met, the polarimetric information is quite difficult to optimal multivariate signal processing of POLSAR data in a
extract using the K-means clustering. As it can be noticed general framework. The SIRV model provides the methodology
in Fig. 9(c), the polarimetric signatures are strongly mixed, for retrieving the conventional cases (multivariate Gaussian
and the class boundaries are smoothed within high-resolution and multivariate K distribution). This methodology can also be
POLSAR images (even for highly heterogeneous target areas). generalized to other heterogeneous clutter models defined by
Therefore, other clustering strategies should be better suited to explicit texture PDFs (inverse Gamma, Fisher, etc.) [59].
capture the spatial distribution of different polarimetric signa- More recent studies have revealed the presence of different
tures. One starting point could be the POLSAR segmentation scattering characteristics between the cross- and copolar terms
by likelihood approximation [56], spectral clustering ensemble of the Sinclair matrix [16], [60]. In consequence, the POLSAR
[57], or the support vector machines kernel-based nonlinear clutter could be modeled by different texture random variables
classification [58]. for each polarization channel. Such a stochastic model already
Finally, one can observe that span information does also, in exists in the literature, and it is known as the generalized SIRV
some cases, contribute to classification quality (e.g., discrim- model [61]. Unfortunately, the covariance matrix generalized
ination of roads in Fig. 9 and buildings in Fig. 11), although SIRV estimator of the Gaussian kernel could not be found,
the polarimetric signature clustering suffers. Based on the without taking into account any a priori information about
SIRV model, the separation span/polarimetric signature is the texture multivariate PDF. Future work should investigate
achieved. Future work is needed to objectively assess the clas- the coupling between SIRVs and multiple single-channel spa-
sification potential of these two descriptors separately. tial texture descriptors, such as the nonstationary anisotropic
Gaussian-kernel model [62].
Despite being quite general, the SIRV clutter model sup-
IV. G ENERAL R EMARKS
poses the matrix stationarity condition to be verified over the
One critical point of the SIRV model is linked to the scalar observation vector. We proposed the use of an adaptive spatial
texture (span) descriptor τ . The validity of the product model support based on the scalar span information. The resulting
for POLSAR data has been investigated in many papers over SDAN operates under deterministic texture hypothesis, and it
the last decades [7], [11]–[13], [28]. states that the local matrix stationarity property is revealed by
Yueh et al. [27] derived the generalized likelihood of the changes in the span image.
normalized polarimetric target vector in Gaussian clutter. This One limitation of the proposed estimation scheme concerns
approach has been extended to the K-distributed clutter in [6] the determination of the SIRV homogeneous neighborhood
and [28]. Note that this extension is not optimal since the surrounding a pixel. The strategy adopted for this paper consists
1824 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

in testing the matrix stationarity condition using the span, under R EFERENCES
deterministic texture assumption. Despite not being optimal in [1] W. M. Boerner, “Basic concepts of radar polarimetry and its applica-
the context of the SIRV model, the proposed approach does tions to target discrimination, classification, imaging, and identification,”
not require additional a priori information regarding the local Commun. Lab., Univ. Illinois, Chicago, IL, 1982.
[2] J. W. Goodman, “Some fundamental properties of speckle,” J. Opt. Soc.
clutter statistics. Amer., vol. 66, no. 11, pp. 1145–1149, 1976.
Finally, the SDAN-FP algorithm is more computation inten- [3] C. Lopez-Martinez and X. Fabregas, “Polarimetric SAR speckle
sive than other existing POLSAR speckle filtering algorithms noise model,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 10,
pp. 2232–2242, Oct. 2003.
developed for Gaussian clutter [8]–[10], and it handles single- [4] M. S. Greco and F. Gini, “Statistical analysis of high-resolution SAR
look complex data only. Further work should address the exten- ground clutter data,” IEEE Trans. Geosci. Remote Sens., vol. 45, no. 3,
sion of the proposed approach to adaptive nonlinear filtering of pp. 566–575, Mar. 2007.
[5] F. T. Ulaby, F. Kouyate, B. Brisco, and T. H. L. Williams, “Textural infor-
multilook POLSAR data. mation in SAR images,” IEEE Trans. Geosci. Remote Sens., vol. GRS-24,
no. 2, pp. 235–245, Mar. 1986.
[6] S. H. Yueh, J. A. Kong, J. K. Jao, R. T. Shin, and L. M. Novak,
V. C ONCLUSION AND P ERSPECTIVES “K-distribution and polarimetric terrain radar clutter,” J. Electromagn.
Waves Appl., vol. 3, no. 8, pp. 747–768, 1989.
This paper has presented a new estimation scheme for deriv- [7] J. S. Lee, D. L. Schuler, R. H. Lang, and K. J. Ranson, “K-distribution
ing normalized coherency matrices and the resulting estimated for multi-look processed polarimetric SAR imagery,” in Proc.
IEEE Int. Geosci. Remote Sens. Symp., Pasadena, CA, 1994, vol. 4,
span with high-resolution POLSAR images. The proposed ap- pp. 2179–2182.
proach couples nonlinear ML estimators with SDANs for taking [8] G. Vasile, E. Trouvé, J. S. Lee, and V. Buzuloiu, “Intensity-driven
the local scene heterogeneity into account. adaptive-neighborhood technique for polarimetric and interferometric
SAR parameters estimation,” IEEE Trans. Geosci. Remote Sens., vol. 44,
The heterogeneous clutter in POLSAR data was described by no. 6, pp. 1609–1621, Jun. 2006.
the SIRV model. Two estimators were introduced for describing [9] J. S. Lee, M. R. Grunes, and G. DeGrandi, “Polarimetric SAR speckle
the POLSAR data set: the FP estimator of the normalized filtering and its impact on terrain classification,” IEEE Trans. Geosci.
Remote Sens., vol. 37, no. 5, pp. 2363–2373, Sep. 1999.
coherency matrix and the corresponding LLMMSE span. The [10] J. S. Lee, J. H. Wen, T. L. Ainsworth, K. S. Chen, and A. J. Chen,
FP estimation is independent on the span PDF and represents “Improved sigma filter for speckle filtering of SAR imagery,” IEEE Trans.
an approximate ML estimator for a large class of stochastic Geosci. Remote Sens., vol. 47, no. 1, pp. 202–213, Jan. 2009.
[11] L. M. Novak and M. C. Burl, “Optimal speckle reduction in polarimet-
processes obeying the SIRV model. Moreover, the derived ric SAR imagery,” IEEE Trans. Aerosp. Electron. Syst., vol. 26, no. 2,
normalized coherency is asymptotically Gaussian distributed. pp. 293–305, Mar. 1990.
For the SIRV clutter, a new ML distance measure was [12] L. M. Novak, M. C. Burl, and W. W. Irving, “Optimal polarimetric
processing for enhanced target detection,” IEEE Trans. Aerosp. Electron.
introduced for unsupervised POLSAR classification. This dis- Syst., vol. 29, no. 1, pp. 234–244, Jan. 1993.
tance was used in conventional K-means clustering initialized [13] A. Lopes and F. Sery, “Optimal speckle reduction for the product model in
by the H−α polarimetric decomposition. Other extensions of multilook polarimetric SAR imagery and the Wishart distribution,” IEEE
Trans. Geosci. Remote Sens., vol. 35, no. 3, pp. 632–647, May 1997.
the existing unsupervised or supervised POLSAR clustering [14] G. Liu, S. Huang, A. Torre, and F. Rubertone, “The multilook polarimetric
methods (e.g., Bayes ML or fuzzy K-means) can be derived by whitening filter (MPWF) for intensity speckle reduction in polarimetric
replacing the conventional Wishart distance with the proposed SAR images,” IEEE Trans. Geosci. Remote Sens., vol. 36, no. 3, pp. 1016–
1020, May 1998.
SIRV distance. [15] G. DeGrandi, J. S. Lee, D. Schuler, and E. Nezry, “Texture and speckle
The effectiveness of the proposed estimation scheme was statistics in polarimetric SAR synthesized images,” IEEE Trans. Geosci.
illustrated by high- and very high resolution ONERA RAMSES Remote Sens., vol. 41, no. 9, pp. 2070–2088, Sep. 2003.
[16] G. DeGrandi, J. S. Lee, and D. Schuler, “Target detection and texture seg-
X-band POLSAR data. The reliability of the obtained results mentation in polarimetric SAR images using a wavelet frame: Theoretical
was demonstrated by quantitative performance assessments aspects,” IEEE Trans. Geosci. Remote Sens., vol. 45, no. 11, pp. 3437–
using simulated POLSAR data. 3453, Nov. 2007.
[17] S. R. Cloude and E. Pottier, “An entropy based classification scheme
This work has many interesting perspectives. We believe that for land applications of polarimetric SAR,” IEEE Trans. Geosci. Remote
this paper contributes toward the description and the analysis Sens., vol. 35, no. 1, pp. 68–78, Jan. 1997.
of heterogeneous clutter over scenes exhibiting complex po- [18] J. S. Lee, M. R. Grunes, T. L. Ainsworth, D. Li-Jen, D. L. Schuler, and
S. R. Cloude, “Unsupervised classification using polarimetric decompo-
larimetric signatures. The proposed approach presents a high
sition and the complex Wishart classifier,” IEEE Trans. Geosci. Remote
degree of generality as no explicit stochastic texture model is Sens., vol. 37, no. 5, pp. 2249–2258, Sep. 1999.
needed. Finally, the proposed estimation scheme can be ex- [19] J. S. Lee, M. R. Grunes, E. Pottier, and L. Ferro-Famil, “Unsuper-
tended to other multidimensional SAR techniques using the co- vised terrain classification preserving polarimetric scattering character-
istics,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 4, pp. 722–731,
variance matrix descriptor, such as the following: multibaseline Apr. 2004.
interferometry, polarimetric interferometry, or multifrequency [20] F. Pascal, Y. Chitour, J. P. Ovarlez, P. Forster, and P. Larzabal, “Covariance
polarimetry. Future work should address the quantitative perfor- structure maximum-likelihood estimates in compound Gaussian noise:
Existence and algorithm analysis,” IEEE Trans. Signal Process., vol. 56,
mance analysis of classification and target detection algorithms no. 1, pp. 34–48, Jan. 2008.
based on these estimators. [21] K. Yao, “A representation theorem and its applications to spherically-
invariant random processes,” IEEE Trans. Inf. Theory, vol. IT-19, no. 5,
pp. 600–608, Sep. 1973.
ACKNOWLEDGMENT [22] J. S. Lee, “Speckle analysis and smoothing of synthetic aperture radar
images,” Comput. Graph. Image Process., vol. 17, no. 1, pp. 24–32,
The authors would like to thank the anonymous review- Sep. 1981.
[23] R. Touzi, “A review of speckle filtering in the context of estimation
ers and the associate editor for the helpful comments and theory,” IEEE Trans. Geosci. Remote Sens., vol. 40, no. 11, pp. 2392–
suggestions. 2404, Nov. 2002.
VASILE et al.: COHERENCY MATRIX ESTIMATION OF HETEROGENEOUS CLUTTER 1825

[24] R. Fjortoft and A. Lopes, “Estimation of the mean radar reflectivity from a the complex Wishart clustering for fully polarimetric SAR data analy-
finite number of correlated samples,” IEEE Trans. Geosci. Remote Sens., sis,” IEEE Trans. Geosci. Remote Sens., vol. 45, no. 11, pp. 3454–3467,
vol. 39, no. 1, pp. 196–199, Jan. 2001. Nov. 2007.
[25] J. S. Lee, M. R. Grunes, and S. A. Mango, “Speckle reduction in multi- [49] A. P. Doulgeris, S. N. Anfinsen, and T. Eltoft, “Classification with a non-
polarization, multifrequency SAR imagery,” IEEE Trans. Geosci. Remote Gaussian model for POLSAR data,” IEEE Trans. Geosci. Remote Sens.,
Sens., vol. 29, no. 4, pp. 535–544, Jul. 1991. vol. 46, no. 10, pp. 2999–3009, Oct. 2008.
[26] A. DeMaio, G. Alfano, and E. Conte, “Polarization diversity detection in [50] E. L. Lehmann, Testing Statistical Hypotheses., 2nd ed. New York:
compound-Gaussian clutter,” IEEE Trans. Aerosp. Electron. Syst., vol. 40, Springer-Verlag, 1986.
no. 1, pp. 114–131, Jan. 2004. [51] P. Dreuillet, H. Cantalloube, E. Colin, P. Dubois-Fernandez, X. Dupuis,
[27] H. A. Yueh, A. A. Swartz, J. A. Kong, R. T. Shin, and L. M. Novak, P. Fromage, F. Garestier, D. Heuze, H. Oriot, J. L. Peron, J. Peyret,
“Bayes classification of terrain cover using normalized polarimetric data,” G. Bonin, O. R. duPlessis, J. F. Nouvel, and B. Vaizan, “The
J. Geophys. Res., vol. 93, no. B12, pp. 15 261–15 267, Dec. 1988. ONERA RAMSES SAR: Latest significant results and future develop-
[28] H. A. Yueh, J. A. Kong, J. K. Jao, R. T. Shin, H. A. Zebker, T. LeToan, and ments,” in Proc. IEEE Int. Radar Conf., Verona, NJ, 2006, 7 pp., DOI:
L. M. Novak, “Polarimetric remote sensing,” in PIER 3, J. A. Kong, Ed. 10.1109/RADAR.2006.1631849.
Amsterdam, The Netherlands: Elsevier, 1990, ch. 4. [52] J. S. Lee, S. R. Cloude, K. P. Papathanassiou, M. R. Grunes, and
[29] R. Touzi and A. Lopes, “The principle of speckle filtering in polarimet- I. H. Woodhouse, “Speckle filtering and coherence estimation of po-
ric SAR imagery,” IEEE Trans. Geosci. Remote Sens., vol. 32, no. 5, larimetric SAR interferometry data for forest applications,” IEEE Trans.
pp. 1110–1114, Sep. 1994. Geosci. Remote Sens., vol. 41, no. 10, pp. 2254–2263, Oct. 2003.
[30] M. Rangaswamy, D. Weiner, and A. Ozturk, “Computer generation of [53] P. Imbo and J. C. Souyris, “Visualization of the polarimetric information:
correlated non-Gaussian radar clutter,” IEEE Trans. Aerosp. Electron. Comparison between partial versus full polarimetry architectures,” in
Syst., vol. 31, no. 1, pp. 106–116, Jan. 1995. Proc. IEEE Int. Geosci. Remote Sens. Symp., Honolulu, HI, 2000, vol. 3,
[31] F. Gini and M. V. Greco, “Covariance matrix estimation for CFAR detec- pp. 1116–1118.
tion in correlated heavy tailed clutter,” Signal Process., vol. 82, no. 12, [54] L. Ferro-Famil, E. Pottier, and J. S. Lee, “Classification and interpretation
pp. 1847–1859, Dec. 2002. of polarimetric interferometric SAR data,” in Proc. IEEE Int. Geosci.
[32] N. R. Goodman, “Statistical analysis based on a certain multivariate com- Remote Sens. Symp., Toronto, ON, Canada, 2002, vol. 1, pp. 635–637.
plex Gaussian distribution (an introduction),” Ann. Math. Stat., vol. 34, [55] E. Pottier, L. Ferro-Famil, S. Allain, S. Cloude, I. Hajnsek,
no. 1, pp. 152–177, 1963. K. Papathanassiou, A. Moreira, M. Williams, T. Pearson, and Y. Desnos,
[33] L. M. Novak, M. B. Sechtin, and M. J. Cardullo, “Studies of target de- “An overview of the PolSARpro v2.0 software. The educational toolbox
tection algorithms that use polarimetric radar data,” IEEE Trans. Aerosp. for polarimetric and interferometric polarimetric SAR data processing,”
Electron. Syst., vol. 25, no. 2, pp. 150–165, Mar. 1989. in Proc. POLinSAR Workshop, 2007.
[34] J. Goldman, “Detection in the presence of spherically symmetric ran- [56] J. M. Beaulieu and R. Touzi, “Segmentation of textured polarimetric SAR
dom vectors,” IEEE Trans. Inf. Theory, vol. IT-22, no. 1, pp. 52–59, scenes by likelihood approximation,” IEEE Trans. Geosci. Remote Sens.,
Jan. 1976. vol. 42, no. 10, pp. 2063–2072, Oct. 2004.
[35] E. Conte, M. Lops, and G. Ricci, “Asymptotically optimum radar detec- [57] X. Zhang, L. Jiao, F. Liu, L. Bo, and M. Gong, “Spectral clustering ensem-
tion in compound-Gaussian clutter,” IEEE Trans. Aerosp. Electron. Syst., ble applied to SAR image segmentation,” IEEE Trans. Geosci. Remote
vol. 31, no. 2, pp. 617–625, Apr. 1995. Sens., vol. 46, no. 7, pp. 2126–2136, Jul. 2008.
[36] F. Gini, “Sub-optimum coherent radar detection in a mixture of [58] S. Fukuda and H. Hirosawa, “Support vector machine classifica-
K-distributed and Gaussian clutter,” Proc. Inst. Elect. Eng.—Radar, Sonar tion of land cover: Application to polarimetric SAR data,” in Proc.
Navig., vol. 144, no. 1, pp. 39–48, Feb. 1997. IEEE Int. Geosci. Remote Sens. Symp., Sydney, Australia, 2001, vol. 1,
[37] M. Rangaswamy, J. H. Michels, and D. D. Weiner, “Multichannel de- pp. 187–189.
tection for correlated non-Gaussian random processes based on inno- [59] L. Bombrun, J. M. Beaulieu, G. Vasile, J. P. Ovarlez, F. Pascal, and
vations,” IEEE Trans. Signal Process., vol. 43, no. 8, pp. 1915–1922, M. Gay, “Segmentation of polarimetric SAR images using heterogeneous
Aug. 2005. clutter models,” in Proc. IEEE Int. Geosci. Remote Sens. Symp., Cape
[38] E. Conte, A. DeMaio, and G. Ricci, “Recursive estimation of the co- Town, South Africa, 2009, to be published.
variance matrix of a compound-Gaussian process and its application to [60] O. D’Hondt, L. Ferro-Famil, and E. Pottier, “The gradient structure tensor
adaptive CFAR detection,” IEEE Trans. Image Process., vol. 50, no. 8, as an efficient descriptor of spatial texture in polarimetric SAR data,” in
pp. 1908–1915, Aug. 2002. Proc. IEEE Int. Geosci. Remote Sens. Symp., Denver, CO, 2006, vol. 1,
[39] F. Pascal, P. Forster, J. P. Ovarlez, and P. Larzabal, “Performance analysis pp. 164–167.
of covariance matrix estimates in impulsive noise,” IEEE Trans. Signal [61] T. J. Barnard and D. D. Weiner, “Non-Gaussian clutter modeling with
Process., vol. 56, no. 6, pp. 2206–2216, Jun. 2008. generalized spherically invariant random vectors,” IEEE Trans. Signal
[40] Y. Chitour and F. Pascal, “Exact maximum likelihood estimates for SIRV Process., vol. 44, no. 10, pp. 2384–2390, Oct. 1996.
covariance matrix: Existence and algorithm analysis,” IEEE Trans. Signal [62] O. D’Hondt, C. López-Martínez, L. Ferro-Famil, and E. Pottier, “Spatially
Process., vol. 56, no. 10, pp. 4563–4573, Oct. 2008. nonstationary anisotropic texture analysis in SAR images,” IEEE Trans.
[41] E. J. Kelly, “An adaptive detection algorithm,” IEEE Trans. Aerosp. Geosci. Remote Sens., vol. 45, no. 12, pp. 3905–3918, Dec. 2007.
Electron. Syst., vol. AES-22, no. 1, pp. 115–127, Mar. 1986.
[42] C. D. Richmond, “A note on non-Gaussian adaptive array detection and
signal parameter estimation,” IEEE Signal Process. Lett., vol. 3, no. 8,
pp. 251–252, Aug. 1996.
[43] M. Jager, M. Neumann, S. Guillaso, and A. Reigber, “A self-initializing
PolInSAR classifier using interferometric phase differences,” IEEE Trans.
Geosci. Remote Sens., vol. 45, no. 11, pp. 3503–3518, Nov. 2007. Gabriel Vasile (S’06–M’07) received the Mas-
[44] G. Vasile, “Imagerie radar à synthèse d’ouverture interférométrique et po- ter of Engineering degree in electrical engineer-
larimétrique. Application au suivi des glaciers alpins,” Ph.D. dissertation, ing and computer science and the M.S. degree
Univ. Savoie, Chambery, France, 2007. in image, shapes, and artificial intelligence from
[45] G. Vasile, J. P. Ovarlez, F. Pascal, C. Tison, L. Bombrun, M. Gay, and the Politehnica University of Bucharest, Bucharest,
E. Trouvé, “Normalized coherency matrix estimation under the SIRV Romania, in 2003 and 2004, respectively, and the
model. Alpine glacier POLSAR data analysis,” in Proc. IEEE Int. Geosci. Ph.D. degree in signal and image processing from
Remote Sens. Symp., Boston, MA, 2008, vol. 1, pp. I-74–I-77. Université de Savoie, Savoie, France, in 2007.
[46] S. R. Cloude and K. P. Papathanassiou, “Polarimetric SAR interferome- From 2007 to 2008, he was a Postdoctoral Fel-
try,” IEEE Trans. Geosci. Remote Sens., vol. 36, no. 5, pp. 1551–1565, low with the Centre National d’Etudes Spatiales
Sep. 1998. (the French space agency) and was with the French
[47] P. R. Kersten, J. S. Lee, and T. L. Ainsworth, “Unsupervised classification Aerospace Laboratory (ONERA), Palaiseau, France. In 2008, he joined the
of polarimetric synthetic aperture radar images using fuzzy clustering French National Council for Scientific Research (CNRS), where he is currently
and EM clustering,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 3, a Research Fellow and a member of the Grenoble Image Speech Signal
pp. 519–527, Mar. 2005. Automatics Laboratory, Grenoble, France. His research interests include signal
[48] F. Cao, W. Hong, Y. Wu, and E. Pottier, “An unsupervised segmentation and image processing, synthetic aperture radar remote sensing, polarimetry, and
with an adaptive number of clusters using the span/H/α/A space and interferometry.
1826 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010

Jean-Philippe Ovarlez (M’06) was born in Céline Tison received the Master of Engineering and
Denain, France, in 1963. He received the Master Ph.D. degrees from the Ecole Nationale Supérieure
of Engineering degree from the Ecole Supérieure des Télécommunications (Télécom Paris), Paris,
d’Electronique Automatique et Informatique France, in 2001 and 2004, respectively.
(ESIEA), Paris, France, the Diplôme d’Etudes She is currently with the Altimetry and Radar
Approfondies degree in signal processing from the Department, Centre National d’Etudes Spatiales (the
University of Paris XI, Paris, in 1987, and the Ph.D. French space agency), Toulouse, France. She works
degree in physics from the University of Paris VI, on radar instrument design and radar signal process-
Paris, in 1992. ing. Her main research interests, in the framework
In 1992, he joined the Electromagnetic and Radar of synthetic aperture radar (SAR) image processing,
Division, French Aerospace Laboratory (ONERA), include SAR image information extraction, in par-
Palaiseau, France, where he is currently the Chief Scientist and a member of ticular, for high-resolution urban imagery, high-resolution interferometry, and
the Scientific Committee, ONERA Physics Branch. Since January 2008, he polarimetry.
has been with the French–Singaporean SONDRA Laboratory, Supelec, Gif-sur-
Yvette, France, on a part-time basis to supervise signal processing activities.
His current activities of research are centered in the topic of signal processing
for radar and synthetic aperture radar applications such as time–frequency
analysis, imaging, detection, and parameter estimation.

Frederic Pascal (M’06) was born in Sallanches,

France, in 1979. He received the Master’s Degree
(“Probabilities, Statistics and Applications: Signal,
Image et Networks”) with merit in applied statistics
from the University of Paris VII, Paris, France, in
2003 and the Ph.D. degree in signal processing from
the University of Paris X, Paris, in 2006, where
it was advised by Pr. Philippe Forster (“Detection
and Estimation in Compound Gaussian Noise”). This
Ph.D. thesis was in collaboration with the French
Aerospace Laboratory (ONERA), Palaiseau, France.
From November 2006 to February 2008, he was a Postdoctoral Fellow with
the Signal Processing and Information Team, Laboratory Système et Applica-
tions des Technologies de l’Information et de l’Energie, National Council for
Scientific Research (CNRS), Ecole Normale Supérieure, Cachan, France. Since
March 2008, he has been an Assistant Professor with the SONDRA Laboratory,
Supelec, Gif-sur-Yvette, France. His research interests are estimation in statis-
tical signal processing and radar detection.