Hyperspectral images band selection using multi-dictionary based sparse representation

Abstract

Hyperspectral images have more spectral bands than the multispectral images. Rich band information provide more favorable conditions for the application of hyperspectral images. Whereas there are a large amount of redundant information among hyperspectral image bands. Therefore, band selection is a useful operation to reduce the dimensionality of hyperspectral image bands for decreasing computational complexity and avoiding Hugh phenomenon. In this paper, we present a novel algorithm for band selection based on a sparse representation group of the hyperspectral image. If each band data can be represented approximately by the linear combination of some band data group, the group are the features we select. For the original band data, the linear combination's weights are sparse. The Orthogonal Matching Pursuit (OMP) algorithm is adopt to obtain the weights. For every band data we get a corresponding weight vector, and the coefficient weights matrix will be obtained for full bands. Experimental results show that our algorithm has good performance in hyperspectral image classification applications than random band selection and Principle Component Analysis (PCA) dimension reduction.