Dual Graph Convolutional Network for Hyperspectral Images With Spatial Graph and Spectral Multigraph

To accurately represent the graph structure of the pixel nodes in the hyperspectral remote sensing image classification based on graph convolutional networks (GCNs), a spectral multi-graph adjacency matrix is presented by the weighted fusion of four single spectral adjacency matrices constructed from four different similarity measures; a strategy of extracting pixel neighborhood spatial features as pixel nodes to construct spatial adjacency matrices is presented, i.e., using Gabor wavelet transform to extract the shallow pixel spatial texture feature to construct the shallow spatial feature adjacency matrix and using a 2-dimensional convolutional neural network (2D-CNN) to extract the deep pixel spatial feature to construct the deep spatial feature adjacency matrix. By combining a shallow texture or deep spatial graph branch with a spectral multi-graph branch, a spatial texture feature spectral multi-graph dual interactive GCN (STSM-DGCN) and a spatial deep feature spectral multi-graph dual interactive GCN (SDSM-DGCN) are designed. The experimental results on three real datasets show that the presented methods can improve the classification accuracy compared to support vector machine and K nearest neighbor classifiers, 3D-CNN, SSRN, HybridSN, GCN, FuNet-C, CEGCN, and WFCG models, especially under small-size training samples.