M2FNet: a multiscale graph-convolution collaborative network with node interaction attention for hyperspectral change detection

Abstract

Hyperspectral image change detection faces challenges from spectral redundancy and spatial-scale diversity, which hinder the extraction of discriminative features and lead to pseudo-change noise. To address these issues, we propose M2FNet: a multiscale graph-convolution collaborative network with node interaction attention (NIA). The network employs a dual-branch architecture, where a multi-scale decomposition perception module extracts local spatial–spectral features, and a graph convolutional network (GCN) branch models global contextual relationships. Two modules are introduced: the NIA module enhances feature discriminability by enabling cross-channel and spatial interactions among superpixel nodes, and the multi-scale adaptive fusion (MSAF) module achieves deep alignment and fusion of convolutional neural network-based local features and GCN-based global features through an adaptive multi-scale pyramid and local window attention. Experiments on three public hyperspectral image change detection datasets (Farmland, Santa Barbara, and USA) demonstrate that M2FNet outperforms state-of-the-art methods, achieving superior overall accuracy (95.48%, 99.05%, and 98.71%), F1 scores (91.18%, 98.91%, and 95.10%), and kappa coefficients (0.8815, 0.9807, and 0.9436). Ablation studies confirm the individual and synergistic contributions of the NIA and MSAF modules. The model also maintains competitive efficiency, offering a balanced solution for accurate and robust hyperspectral change detection.