Texture modelling for land cover classification of fully polarimetric SAR images

Classification based on polarimetric data alone does not provide sufficient sensitivity for classes' separation such as forests. The use of other characteristics such as texture provides a powerful tool to get better class discrimination. Thus, this article outlines an approach for texture modelling adapted to land identification using polarimetric synthetic aperture radar PolSAR) images. As radar images may contain a myriad of textures and to take into account their nonstationarity feature, the model is non-parametric and is based on a likeness measure between neighbourhoods. This measure is incorporated in probability formulation of the textural model. The latter is validated through texture synthesis applied to Brodatz' natural textures. While having multiple applications, we focused here on the issue of terrain mapping of PolSAR images. Using Bayesian approach and some basic assumptions, we defined a classification scheme based on the textural model. It was tested on original Pauli decomposed images and filtered ones. The test area used is the Oberpfaffenhofen in Munich and the PolSAR images were acquired in the P band. For evaluation purposes, the classifications results obtained were compared with the Wishart classifier result performed on H-alpha partition showing interesting results and reaching classification rates approximate to 90%. Such an analysis has allowed us to assess the importance of texture considering and proved that the proposed texture model is able to describe the target's physical properties of PolSAR data.