Polarimetric radar backscattering is classically modeled as a multivariate Gaussian process. However, PolSAR data sets acquired by recent sensors with high resolution performance show some clutter heterogeneity, whose handling requires higher order representations. In this paper, the statistics of textured polarimetric images are modeled as a SIRV process. The comparison of the performance of the SIRV with respect to the classical Wishart distribution for the classification of real polarimetric SAR data clearly illustrates the strong sensitivity of the Wishart classification to the texture information, i.e. the backscattered intensity. In order to overcome this dependency, a new algorithm, based on the M-Box test, is proposed to classify pixels solely from their polarimetric properties.
展开▼
