MCF-SSGN: A Multiscale Spectral–Spatial Graph Fusion Network for Hyperspectral Image Classification

Graph Convolutional Networks (GCNs) have been widely applied to hyperspectral image classification (HSIC). However, HSIC still faces challenges such as spectral redundancy, spatial heterogeneity, and limited labeled samples. Although recent studies have combined GCNs with Convolutional Neural Networks (CNNs) to exploit complementary strengths, these models remain vulnerable to noise under limited supervision and often incur high computational cost. To address these issues, we propose a Multi-scale Spectral-Spatial Graph Fusion Network (MCF-SSGN). First, a Spectral-Attention Superpixel Graph Convolution Submodule is introduced to perform band-aware weighted aggregation on superpixel graphs, high-lighting informative spectral bands and suppressing noise for improved global representation. Second, a Multi-scale Expert Aggregation Module is designed to dynamically generate input-dependent kernels, enhancing adaptability to diverse spatial patterns while reducing computational complexity. Finally, a dual-branch adaptive fusion mechanism is developed to learn channel-wise mixing weights for balancing and integrating the CNN and GCN branches, thereby overcoming the limitations of static fusion strategies. Extensive experiments on four widely used benchmark datasets—Indian Pines, Pavia University, Salinas, and WHU-Hi-HongHu—demonstrate that MCF-SSGN achieves overall accuracies of 95.53%, 99.26%, 99.20%, and 95.65%, significantly surpassing 11 state-of-the-art base-lines, including CNN-based, graph-based, Mamba-based, and hybrid models. The core code for MCF-SSGN is published at https://github.com/Majunyi310321/MCF-SSGN.