Multiscale sparse representation classification for robust hyperspectral image analysis

Sparse representation of signals has been recently applied for hyperspectral imagery classification. It relies on the assumption that a test pixel can be linearly and sparsely represented as a combination of all training samples. Although recent work reported in the literature has exploited the sparsity of hyperspectral images, its use in effective multiscale representation of hyperspectral imagery is an unexplored area. In this work, a multiscale sparse representation algorithm is proposed for robust hyperspectral image classification. An automatic and adaptive weight assignment scheme based on the spectral angle ratio has been incorporated into the proposed multiclassifier framework to fuse sparse representation information across all scales. Experimental results based on a real-world benchmarking hyperspectral dataset show that utilizing multiscale information and adaptive weight assignment results in much higher classification accuracies than merely using the spectral information in hyperspectral image classification tasks.