Hyperspectral Image Classification With Online Structured Dictionary Learning

Abstract

In this study, the spectral and spatial redundancies of hyperspectral images are used for designing a sparse representation-based classification approach. The spectral redundancy is used to define spectral blocks and they are used to adaptively recognize the distinctive bands. The most distinctive blocks are identified as active blocks in a block sparse representation approach. Then the sparse coefficients within each spatial group are imposed to share a common subspace. To achieve this hierarchical sparsity pattern a sparse coding algorithm is proposed. This sparse coding is done over a block-structured dictionary, which is learned from the image data using the online dictionary learning algorithm. The obtained sparse coefficients are then classified using a support vector machine classifier. This structured sparsity pattern alleviates the instability of the sparse coefficients. Experiments on two standard datasets namely, Indian Pines and Pavia University, verify the effectiveness of the proposed approach for the classification of hyperspectral images.