Physics-guided spatial-spectral Mamba network for hyperspectral nonlinear blind unmixing

Hyperspectral blind unmixing aims to recover endmember spectra and abundance fractions from mixed pixels. In complex real scenes, nonlinear interactions, illumination variations, and other unmodeled effects make the linear mixing assumption insufficient. Existing deep unmixing methods still struggle to jointly model long-range spatial-spectral dependencies, preserve physical interpretability in reconstruction, and provide stable endmember initialization. To address these issues, we propose a physics-guided spatial-spectral Mamba network for hyperspectral nonlinear blind unmixing. The proposed framework consists of a cascaded Mamba-based spatial-spectral feature extractor, a physics-guided nonlinear reconstruction module, and a superpixel-based endmember initialization strategy. Specifically, each Mamba block integrates an SS2D-based spatial branch and a latent spectral branch, and the reconstruction module is decomposed into a dominant linear branch, a scale compensation branch, and a lightweight nonlinear residual compensation branch. Experimental results on one synthetic dataset, one ray-tracing-based orchard scene dataset, and three real datasets show that the proposed method achieves the best overall performance on synthetic, orchard, and Urban datasets, while remaining competitive on Jasper Ridge and Samson. These results demonstrate the effectiveness and robustness of the proposed method for hyperspectral nonlinear blind unmixing across diverse scenarios.