Discriminative Dictionary Learning for Polarimetric SAR Image Classification

Polarimetric SAR image classification is a high-dimensional nonlinear mapping problem, sparse representation (CS) has shown great potential to tackle such problems. Dictionary training based-CS plays an important role in SAR image classification. In this paper, we propose a novel dictionary learning model to make dictionary more discriminative, thus such dictionary is very fit for SAR image classification. According to the role of two kinds of sub-dictionaries, different sparsity regularizations are imposed on their corresponding coefficients. Specifically, the coefficients corresponding a common sub-dictionary are imposed with sparse regularization, so that the common sub-dictionary obtains features shared by all classes; the coefficients corresponding to class-specific sub-dictionaries are imposed with both sparse and low rank regularization, so that the class-specific sub-dictionaries capture intra-class intrinsic local and globe structure features. Due to the common sub-dictionary represents features shared by all classes, we use the reconstruction error based on each class-specific sub-dictionary for classification. The results obtained with AIRSAR Flevoland data set confirm that the proposed method is validated.