A new clustering algorithm for PolSAR images segmentation

Abstract

This paper deals with polarimetric synthetic aperture radar (PolSAR) image segmentation. More precisely, we present a new robust clustering algorithm designed for non-Gaussian data. The algorithm is based on an expectation-maximization approach. Its novelty is that, in addition to the estimation of each cluster center and covariance matrix, it also provides for each observation an estimation of the scale parameter, allowing a better flexibility when assigning each observation in one cluster. The method performances are evaluated on both simulated and real multi-looked PolSAR data. It is demonstrated that the algorithm outperforms the classical clustering algorithms such as k-means and GMM (Gaussian-based EM algorithm) in various scenarios.