Polarimetric SAR image classification based on contextual sparse representation

A CSR-Based (Contextual Sparse Representation) classification method for PolSAR image is proposed based on the idea of sparse representation and spatial correlation, which incorporates the intrinsic polarimetric information and the spatial contextual information in the sparse representation procedure. Firstly, multiple useful features are extracted to describe PolSAR images at various aspects. Then, the feature vectors of training samples construct an over-complete dictionary. Then sparsely represent the training samples using the over-complete dictionary and obtain the corresponding coefficients. In this step, the spatial neighboring feature-vectors are assumed to have a similar sparse representation way. Specifically, they can be linearly represented by the same atoms while the weights are different. That is the kernel of CSR. In this way, the efficiency of sparse classification can be highly raised and the result can also be improved by adding the contextual information. The proposed method is validated by the Danish EMISAR L-band fully polarimetric SAR data and the experimental results confirm the performance of the proposed method in PolSAR image classification.