Polarimetric SAR image classification using collaborative representation based nearest subspace

Polarimetric synthetic radar (PolSAR) images contain a huge volume of polarimetric and spatial features, which can be useful for class discrimination. A nonparametric PolSAR classification method called collaborative representation based nearest subspace (CRNS) is proposed in this paper. CRNS simultaneously removes speckle noise and extracts polarimetric-spatial features in two successive stages. At first, it obtains the collaborative representation of the polarimetric cube. Then, it extracts more continuity information from the neighboring pixels through spatial averaging. The extracted polarimetric-spatial feature cube is then classified by using a regularized version of the nearest subspace classifier. The experimental results on two simulated and real PolSAR images show the superior performance of CRNS compared to other state-of-the-art classifiers in both small and large training sets.