Hyperspectral Image Classification Via Shape-Adaptive Joint Sparse Representation

A new shape-adaptive joint sparse representation classification (SAJSRC) method is proposed for hyperspectral images (HSIs) classification. The proposed method adaptively explores the spatial information and incorporates it into a joint sparse representation classifier. First, the HSI is transformed with the principal component analysis (PCA) algorithm. Then, the first principal component (PC), which represents the most spatial variation in the HSI, is used in the shape-adaptive algorithm to construct a shape-adaptive local smooth region for each test pixel. Unlike the fixed-sized window used in other sparse representation-based methods, the shape-adaptive regions have adaptive sizes and shapes, and conform to the spatial structure of the HSI as far as possible. Finally, the label of the test pixel is determined by applying the joint sparse representation classifier to the first several PCs of pixels within the corresponding SA region. According to the experiments performed on several HSIs, the proposed SAJSRC method outperforms some widely used HSIs classification approaches.