HDCGNet: A Hypergraph Dual-Branch CNN-GCN Network for Mars Hyperspectral Image Classification

The hyperspectral image (HSI) classification, as an important research direction in the field of remote sensing, has made significant progress in Earth observation. However, its application in Mars exploration missions is still in the exploratory stage. While convolutional neural networks (CNNs) are effective in extracting local features, their inherent local receptive field restricts the ability to capture long-range spatial dependencies. To overcome this limitation, we propose a dual-branch fusion architecture, termed HDCGNet, which seamlessly integrates CNNs and graph convolutional networks (GCNs), and further incorporates hypergraph learning into the GCN branch. This design enables the extraction of both local and global information from HSI and facilitates the effective modeling of high-order nonlinear relationships among multiple nodes. First, a cascade processing structure is employed in the CNN feature extraction branch to perform spectral denoising and transformation. By leveraging depthwise separable convolution (DSC) and cross multiscale convolution modules, the model effectively captures multiscale spatial-spectral joint features and enhances pixel-level feature representation with differentiated receptive fields. Second, to enrich the topological features of the HSI segmentation area, a hypergraph is added to the GCN feature modeling branch, while the ContraNorm normalization layer ensures a uniform distribution of the representation space and facilitates hyperpixel-level feature extraction. Finally, the adaptive cross-attention fusion module (ACAFM) is employed to fuse the features of the two branches, thereby ensuring complementarity between global and local information. Experimental results on three Mars HSI datasets demonstrate that HDCGNet achieves better performance than some existing methods. Our codes and data will be made publicly available at: https://github.com/Ctao0820/HDCGNet.git