Cone-based joint sparse modelling for hyperspectral image classification

Abstract

Joint sparse model (JSM) is being extensively investigated on hyperspectral images (HSIs) and has achieved promising performance for classification. In JSM, it is assumed that neighbouring hyperspectral pixels can share sparse representations. However, the coefficients of the endmembers used to reconstruct a test HSI pixel is desirable to be non-negative for the sake of physical interpretation. Hence in this paper, we introduce the non-negativity constraint into JSM. The non-negativity constraint implies a cone-shaped space instead of the infinite sample space for pixel representation. This leads us to propose a new model called cone-based joint sparse model (C-JSM), to install the non-negativity on top of the sparse and joint modelling. To solve the C-JSM problem, we also propose a new algorithm through introducing the non-negativity constraint into the simultaneous orthogonal matching pursuit (SOMP) algorithm. The new algorithm is called non-negative simultaneous orthogonal matching pursuit (NN-SOMP). Experiments and investigations show that the proposed C-JSM can produce a more stable, sparse representation and a superior classification than other methods which only ensure the sparsity, non-negativity or spatial coherence.