Locality Constraint Joint-Sparse and Weighted Low-Rank Based Hyperspectral Image Classification

In this paper, a spectral-spatial classification method based on locality constrained joint-sparse and weighted low-rank (LCJS-WLR) for homogeneous regions in hyperspectral image (HSI) which are extracted by superpixel segmentation is proposed. Pixels in a superpixel have similar spectral signatures and they are represented by joint atoms in a training dictionary, which means that coefficient matrix has joint-sparse and low-rank structure. In order to give higher weights to atoms whose spectral characteristics are similar to pixels inside superpixel, locality constraint is adopted to the joint-sparse term. Weighted low-rank representation is used to enhance low-rankness property of the data. The possible sparse noise terms are eliminated by adding a sparse noise regularization term. Experiments on a real hyperspectral data set point out that LCJS-WLR provides better classification performance than the methods under comparison.