Spectral-spatial discriminative broad graph convolution networks for hyperspectral image classification

Graph convolutional networks (GCN) can provide excellent performance in hyperspectral image classification due to their ability to capture feature representations. However, the loss function of GCN model only uses labeled data for model training, and does not consider the relationship between inter-class spacing and intra-class spacing of sample features. It is difficult to ensure effective separation between samples and sufficient aggregation within samples, which limits the classification performance of GCN. To address these issues, we proposed a discriminative broad graph convolution network for hyperspectral image classification (DBGCN). Firstly, we use multiple edge preserving filters to extract spatial spectral features, and then use PCA to fuse the spatial spectral joint features obtained by edge preserving filters. Secondly, graph convolution was used to obtain the deep-level features of the hyperspectral image in the non-Euclidean domain. Finally, the intra-class divergence and inter-class divergence matrix were calculated according to the obtained features, and the weights of the fully connected layer were then trained such that DBGCN exhibited stronger discriminative ability and achieved better classification results. The experimental results show that the proposed model was superior to the state-of-the-art results.