KERNELIZED SPARSE GRAPH-EMBEDDED DIMENSIONALITY REDUCTION FOR HYPERSPECTRAL IMAGE CLASSIFICATION

A new sparse graph-embedded dimensionality reduction (DR) method for hyperspectral image classification is proposed in this paper. The proposed method incorporates the contextual information and the class information to address the supervised transform-based DR problems. On one hand, a class-oriented sparse graph construction method is proposed, where the contextual information is integrated via a simultaneous sparsity model, forcing the spectrally similar samples to have similar sparse representations. On the other hand, the discriminative power is further enhanced by kernelizing a sparse graph-embedded DR method. In this approach, the sparse representation is conducted class-wisely, which is very efficient compared with the tranditional pixel-wisely-based sparse graph construction methods. The proposed method is evaluated by using a multinomial logistic regression classifier. Experimental results with real hyperspectral data sets indicate that the proposed method can yield superior classification performance compared to other related approaches.