Dictionary learning using novel multiscale context sensitive spectral features for classification of hyperspectral imagery

Abstract

Sparse representation models for the classification of hyperspectral images have been greatly enhanced by dictionary learning techniques. The effectiveness of these techniques depends on the discriminative power of the patterns used to learn the dictionaries. In this research, to learn quality, discriminative and comprehensive dictionaries, we propose novel features extracted by exploiting singular value decomposition (SVD). Here, SVD is exploited to extract context-sensitive spectral features (CSSF) of the pixel by taking into account its appropriate spatial neighbor pixels. In the proposed technique, multiple CSSFs are extracted by considering spatial neighborhood of the pixel at different scales to learn dictionaries for classification. The effectiveness of the proposed technique is evaluated by comparing it with several state-of-the-art techniques.