Answers are grounded in this paper and cite their sources.
Hyperspectral unmixing (HU) aims to extract the pure material spectra (endmember) and their corresponding fractions (abundances) from the mixed pixels. Convolutional neural networks are limited by local receptive fields and fail to capture long-range dependencies, while graph convolutional networks rely on pairwise association modeling and struggle to capture high-order interactions among spectral features. To address these issues, we propose a hypergraph high-order association and collaborative interactive attention network for HU. The model employs a multiscale hypergraph high-order association modeling module to construct fine, medium, and coarse hypergraphs via a spatial–spectral dual-strategy scheme, capturing high-order group relationships. Within this module, the intrahyperedge node dynamic aggregation attention adaptively aggregates node features to mitigate information dilution, while the multiscale hyperedge cooperative attention enables cross-scale interaction, overcoming static-scale fusion limitations. A local spatial sparse perception module with enhanced native sparse attention preserves fine spatial structures while reducing computational redundancy. A local spectral perception module performs spectral adaptive calibration to strengthen discriminability. A collaborative interactive attention module with bidirectional cross-attention fuses global high-order association features with local spatial–spectral details. Physical priors, including abundance nonnegativity and sum-to-one constraints, are embedded in feature modeling and loss design to ensure unmixing consistency. Experiments on one synthetic and three real datasets show that our method outperforms state-of-the-art algorithms in both endmember and abundance estimation accuracy, with superior robustness in complex scenes.
No takes yet. Share an insight, caveat, or question.
Qu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: