

Statistical Segmentation of Polarimetric SAR Data
Dr. Laurent FerroFamil ^{(1)}, Dr. Douglas Corr^{(2)}
, Pr. Eric Pottier^{(1)}
, and Mr. Alex Rodrigues^{(2)}
^{(1)}
University of Rennes 1,
Campus de Beaulieu, Bat 11.D,
35042 Rennes Cedex,
France
^{(2)} QinetiQ, Cody technology park, Farnborough, Hampshire, United Kingdom
Abstract
The polarimetric response of a medium is highly related to intrinsic parameters (such as its geometrical structure, its dielectric properties etc.) and observation parameters (such as incidence angle, centre frequency etc.). Segmentation techniques in polarimetric applications have significantly improved the interpretation of the scattering from natural media. However, a wide number of schemes have been proposed, each having particular advantages and disadvantages. Maximum Likelihood kmean clustering procedures using the coherency matrix and Wishart statistics were proposed by Lee and Cloude (Wishart Halpha classifier) and Pottier and Lee (Wishart HalphaA classifier). These procedures are initialised using the data distribution in the Halpha plane (Cloude and Pottier). They have the advantage of being highly related to a physical interpretation of the scattering phenomenon. However, they are limited by the high sensitivity of the kmean procedure to the number of classes and to the initial distribution of the pixels into these classes. To overcome these limitations, Lee et al. proposed an initialisation scheme with a large number of classes followed by a reduction procedure that preserves the scattering mechanisms. This approach used a decomposition theorem (Freeman et al.) that presents an underdetermination and is not roll invariant. Skriver et al. also proposed a class number reduction procedure, uniquely based on a Wishart test. Schou et al. proposed a classification scheme using Gaussian Hidden Markov Random Fields (HMRF) while D'hondt, FerroFamil and Pottier introduced a less restrictive HMRF classifier using Potts model and Wishart statistics. The main limitation of these classification schemes is high computing time . We propose a classification procedure that gathers the main advantages of all of these approaches . A large number of initial classes is obtained from the HalphaA domain and the polarimetric span. The number of classes is then is reduced to a value defined by the data by using test statistics. Finally, the clusters are identified using a Freeman decomposition (Lee et al.) or a Halpha A decomposition (FerroFamil et al.)
Full paper
Full paper
