Spectral–Spatial Hypergraph Convolutional Network for Hyperspectral Image Classification

Abstract

In recent years, due to the high-order nonlinear modeling capability of hypergraph convolutional network (HGCN), it has been introduced into the field of hyperspectral image (HSI) classification. However, the inherent spectral redundancy and variability directly affect the quality of the hypergraph constructed from HSI. In addition, single spectral hypergraph convolution or spatial hypergraph convolution has limited ability to jointly learn global and local features, limiting the performance of HGCN in HSI classification. In this article, we propose a novel spectral–spatial hypergraphs convolutional network (S2HGCN) that simultaneously models long-range relationships and local spectral–spatial features. Specifically, HSI is first processed through an attention feature enhancement module to reduce interference noise caused by spectral information redundancy and variability. Then, spectral and spatial hypergraphs are constructed to model high-order correlations of land cover. Finally, it combines spectral and spatial hypergraphs convolution for learning high-order correlations. Experiments on three public HSI datasets demonstrate that the proposed S2HGCN effectively models global high-order relationships and local detailed features, achieving better classification results than a few state-of-the-art HSI classification methods.