CMHLFG: A Cross-Layer Multihop Graph Convolution and High-Low-Frequency Graph Attention Fusion Network for Hyperspectral Image Classification

Graph convolutional networks (GCNs) have emerged as a prominent research focus for hyperspectral image classification (HSIC). However, existing GCN-based HSIC methods still face the following challenges: the graph structure has difficulty fully capturing complex neighborhood relationships, and the adaptability to heterogeneous topological structures is limited. To address the aforementioned issues, we integrate multiscale information, cross-layer multihop graphs, and high-low-frequency graph attention into a unified framework and propose a cross-layer multihop graph convolution and high-low-frequency graph attention fusion network for HSIC, namely, CMHLFG. The CMHLFG employs a dual-branch network, including three core components: a multiscale dilated convolution module (MDC-M), a cross-layer multihop graph convolution branch (CMGC-B), and a high-low frequency graph attention branch (HLFGA-B). The MDC-M extracts hierarchical features from hyperspectral images while preserving both detailed spatial information and contextual semantic representations, thereby enhancing feature discriminability. In the CMGC-B, diversified node relationships are comprehensively modeled to improve feature separability. Cross-layer connections are utilized to strengthen the model's expressive capacity for node features and broaden its learning scope, effectively capturing graph structural information across different scales. The HLFGA-B distinguishes frequency components in graph signals and uses a gating mechanism to dynamically allocate weights between high-frequency and low frequency components, with the residual channel fusing original input features, improving the model's adaptability to heterogeneous topological structures in complex land-cover distributions of hyperspectral images. The efficacy of the CMHLFG was confirmed through experiments on some real public HSI datasets, and the CMHLFG achieved the best classification performance compared with recently related state of-the-art HSIC approaches.