A Multiscale Synergistic Attention Network With Initialized Endmembers for Hyperspectral Unmixing

Autoencoder-based approaches have achieved remarkable performance in hyperspectral unmixing (HU). However, most existing methods constrain endmember estimation by directly deriving endmembers from decoder weights, limiting their flexibility and representational capacity. Moreover, the feature extraction process often fails to fully exploit the spatial-spectral information inherent in hyperspectral images (HSIs). To overcome these limitations, we propose a multiscale synergistic attention network with initialized endmembers (MSSA-IEm) for HU. A physically interpretable dual-branch structure is developed to decouple abundance and endmember estimation, enabling more adaptive unmixing in complex scenarios. In the abundance extraction branch, a multiscale synergistic attention (MSSA) module is designed to hierarchically integrate spatial contextual cues and spectral correlations, enhancing the network's capability to balance spectral heterogeneity and spatial locality. In the endmember estimation branch, a learnable endmember matrix is incorporated to preserve physical interpretability and facilitate flexible adaptation to complex spectral variations. Extensive experiments on synthetic and real-world hyperspectral datasets demonstrate that MSSA-IEm significantly outperforms several state-of-the-art (SOTA) unmixing methods, validating its effectiveness and robustness. The code is available at https://github.com/Octopus-Squidward/MSSA-IEm