Spatial Multifeature and Dual-Layer Multihop Graph Convolution Networks for Hyperspectral Image Classification

Hyperspectral image (HSI) classification consti-tutes a crucial research direction within the domain of remote sensing. Convolutional neural networks (CNN) and graph convo-lutional neural networks (GCN) have exhibited outstanding clas-sification performance in this field, emerging as current research focuses. Nevertheless, GCN possesses certain limitations in cap-turing the neighborhood features of images, while traditional 2D CNNs are incapable of fully extracting the spatial information of HSI. To address these problems, we propose a novel architecture dubbed spatial multi-feature and dual-layer multi-hop graph convolutional network (SMTGCN). This network is capable of concurrently extracting pixel-level spatial features and superpix-el-level spectral features. Specifically, a dual-layer multi-hop graph convolutional network (DMGCN) is constructed within the GCN branch, which can take the features of superpixel at differ-ent segmentation scales as network nodes to effectively capture and fuse the superpixel features in HSI. In the CNN branch, a multi-scale spatial structure is constructed for feature extraction and fusion, and a hybrid attention mechanism model is proposed to enhance the feature capture ability, a multi-layer pooling structure is added to retain more detailed information while sup-pressing excessive redundant data. Finally, the features extracted by the GCN branch and the CNN branch are fused to realize HSI classification. Experimental results conducted on four bench-mark HSI datasets indicate that, in comparison with existing classification methods, SMTGCN achieves remarkable improve-ments in classification performance when using a small number of training samples.