Adaptive Improved GCNs and SAM Superpixels for Hyperspectral Image Classification

Hyperspectral image (HSI) classification with limited training samples is a challenging problem. According to recent results, effectively exploiting the spatial-spectral information of the HSI is crucial for HSI classification, even when the dataset is highly unbalanced with very few training samples per class. For instance, the graph convolutional networks (GCNs) and the convolutional neural networks (CNNs) can be jointly used to learn the feature representations from the HSI data. However, most existing GCNs and CNNs inherently suffer from the limitations of over-smoothing, over-fitting, and huge computational burden. In this paper, we propose a novel HSI classification algorithm by combining the adaptive improved graph convolutional networks (AIGCNs) and the CNNs, named AIGCN-CNN. First, for the AIGCNs, we propose the AIGCN model to improve the conventional GCN model and provide its theoretical benefits. With the proposed AIGCN model, two AIGCN subnetworks from spatial and spectral domains are constructed to extract spatial-spectral information from the superpixels of the HSI. The superpixels are obtained by using the segment anything model (SAM) and can effectively capture the homogeneous information among the pixels of the HSI and reduce the computational burden. Second, the CNNs are formed by using two simple 1-D and 2-D kernels, separated from the 3-D convolutional kernel, to extract the heterogeneous information of the pixels and to reduce the computational burden. Competitive compared with other state-of-the-art methods, extensive experiments on three HSI datasets and a new HSI dataset captured by Chinese hyperspectral satellite, demonstrate the outstanding performances of the proposed AIGCN-CNN, especially in the case of lacking labeled samples.