The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
132
Application I: The determinant of the covariance matrix is the
generalized variance of polarimetric data, and the ratio of two
determinants is an important parameter in applications as in edge
detection (Skriver et al., 2001), change detection (Conradsen2003
et al., 2003) and SAR image tracking (Erten et al., 2008). An un
biased estimator of the ratio of two covariance matrices determi
nants j^j is given by . The expectation of is derived
from (3) as a function of the canonical correlation coefficients
r = [p\ p2 p 3 ] T . For that, a same procedure as in (Lliopoulos,
2006) is applied into (3), and the unbiased estimate of jfj^j for
n > 2q results
lAn[(n-p)(n-g) 1
|A 22 |(n-ç- 1)
^2(n — p — 1) +
+
n — 2q J
(18)
To obtain (18) from (3), integrations over An and A 22 are nec
essary. These integrations are only valid if the variance of obser
vations are finite, or in other words, the condition of | An | < oo
and IA 22 1 < oo which are always satisfied in the polarimetric
case. Figure 2 presents the evaluation of the unbiased estimate of
related to the number of looks n and the canonical corre
lation coefficients r. It is clear that the generalized variance ratio
is asymptotically unbiased for a large number of looks and the
o
-P
Figure 2: The expected value of the polarimetric variance ratio
estimator with r = [p\ p 2 p$\ T
Application II: As indicated in (Martinez et al., 2005), the proba
bility of the maximum eigenvalue related to the dominant scatter
ing is a important parameter for target detection and its analysis.
The probability of the detection that there is just one dominant
scattering mechanism is a function of the detection threshold T,
that can be obtained from (15) as
F\max{T) =P(Amax < T) = / px max (x)dx (19)
Jo
where Fx (x) indicates the cumulative density function. In addi
tion, having the closed form of the eigenvalue pdf (15), the prob
ability of false alarm can be also computed, as well as receiver
operating characteristic (ROC) curves, allowing a complete de
tection problem analysis.
Application III: Another application of (3) has been detailed ex
plained in (Erten et al., 2008) with the aim of multidimensional
SAR tracking. The joint distribution of polarimetric covariance
matrices over time has been used to perform maximum likelihood
tracking into amplitude data without making any assumption of
their independence.
4 CONCLUSIONS AND FUTURE WORK
The statistical description of two (possible) correlated Wishart
distributions has been presented. Closed forms for the general
distribution have been derived, as well as for the joint distribu
tions of the eigenvalues of the two Wishart matrices and for the
joint distribution of their maximum eigenvalue. This analysis can
be applied to a wide field of aplications, whenever the applica
tion in question follows the statistical assumptions. Examples of
applications have been given in the paper, as the assessment of
different aspects in polrimetric statistical analysis over time. It
has been showed that the performance of analysis may increase
with high canonical correlation coefficients, the number of look
and number of polarimetric channel or may decrease due to the
presence of speckle that effects the calculation of true values of
the parameters.
REFERENCES
Conradsen, K., Nielsen, A. A., Schou, J. and Skriver, H.,2003. A
test statistic in the complex Wishart distribution and its applica
tion to change detection in polarimetric SAR data IEEE Trans,
on Geoscience and Remote Sensing, 41, 1, pp. 4—19.
Erten, E., Reigber, A., Hellwich, O. and Prats, P., 2007. Indepen
dency preserving dependent maximum likelihood texture track
ing model EUSAR 2008, in procedings.
Gross, K.I, Richards, D.S.P., 1989. Total positivity, spherical se
ries, and hypergeoometric functions of matrix argument Journal
of Approx. Theory, 59, pp. 224-246.
James, A.T., 1964. Distribution of matrix variates and latent roots
derived from normal samples The Annals of Mathematical Statis
tics, 35, 2, pp. 475-501.
Kuo, P. H., Smith, P.J. and Garth, L.M., 2007. Joint den
sity for eigenvalues of two correlated complex wishart matri-
cesxharacterization of MIMO systems IEEE Trans, on Wireless
Communications, 6, 11, pp. 3902-3906.
Famil, L. F., Pottier, E. and Lee, J. S., 2001. Unsupervised clas
sification of multifrequency and fully polarimetric SAR images
based on the H/A/Alpha-Wishart Classifier IEEE Trans, on Geo
science and Remote Sensing, 39, 11, pp. 2332-2342.
Lliopoulos, G., 2006. UMVU estimation of the ratio of pow
ers of normal generalized variances under correlation Journal of
Multivariate Analysis, available online.
Martinez, C. L., Pottier, E. and Cloude S. R., 2005. Statistical
assessment of eigenvector-based target decomposition theorems
in radar polarimetry IEEE Trans, on Geoscience and Remote
Sensing, 43, pp. 2058-2074.
McKay, M.R., Grant, A.J. and Collings, I.B., 2007. Performance
analysis of MIMO-MRC in Double-Correlated Rayleigh Envi
ronments IEEE Trans, on Communications, 55, 3, pp. 497-507.
Muirhead, R. J., 1982. Aspects of Multivariate Statistical Theory
New York: Wiley.
Skriver, H., Schou, J., Nielsen, A. A. and Conradsen, K., 2001.
Polarimetric edge detector based on the complex Wishartdistri-
bution Geoscience and Remote Sensing Symposium, 2001, 7,
pp. 3149-3151.
Smith, P.J. and Garth, L.M., 2007. Distribution and characteris
tic functions for correlated complex Wishart matrices Journal of
Multivariate Analysis, 98, pp. 661-667.