Self-Representation difference matrix graph convolutional network for hyperspectral image classification

Hyperspectral image (HSI) classification identifies surface objects using spectral information from multiple bands. Since adjacent bands are highly correlated, redundant spectral information can reduce efficiency and affect classification. This paper proposes a Self-Representation Difference Matrix Graph Convolutional Network (SDM-GCN) for spectral band selection. First, entropy rate segmentation is used to divide HSIs into homogeneous patches and extract target profile information. Within each patch, a spectral difference matrix is constructed from the reflectance differences among different targets, where smaller differences indicate bands with weaker discriminative ability. The difference matrices are then treated as nodes in a graph convolutional network, and node relationships describe correlations among spectral bands. Finally, a sparse band self-representation optimization model is built to learn band weights and select an optimal subset of informative and non-redundant bands. Experiments on multiple hyperspectral datasets demonstrate that the proposed method outperforms several comparative approaches and produces classification maps that better distinguish different land-cover categories.