Hypergraph convolutional network for hyperspectral image classification

Recently, graph-based neural networks have been investigated in hyperspectral image (HSI) classification to address the limited global feature representation capability issue of HSI classification methods based on convolutional neural networks (CNN). However, most of the existing graph-based neural networks for HSI classification methods either characterize the relation information by the pair-wise modeling or rely on the CNNs to extract the local spectral–spatial features. To solve this problem, in this paper, a concise hypergraph convolutional network (HGCN) is proposed for semi-supervised HSI classification. To effectively and efficiently capture the global and local features of HSI, the hypergraph model is established on superpixel level which characterizes the spectral affinities rather than the spatial distance. The designed hypergraph model not only incorporates the local homogeneity and complex correlations of HSI but also consumes little computation. Two hypergraph convolution layers are designed to propagate and update the features of nodes. To construct an end-to-end architecture, a mapping matrix is defined for pixels encoding and superpixels decoding. The proposed method is hinged on the goodness the clustering algorithm used in superpixel segmentation and the experiments has shown that the clustering algorithm affects the effectiveness of proposed method. Thus, we give a strategy for selecting the segmentation parameter. The comparison experiments conducted on four real-world benchmark HSI data sets demonstrate that the proposed method provides more stable and effective classification performance than some state-of-the-art deep approaches with very limited training samples. The overall accuracies are 95.42% on Indian Pines, 98.48% on Kennedy Space center, 98.23% on Salinas Valley and 96.91% on Pavia University.