An Abundance-Guided Mamba Network for Hyperspectral Unmixing

Hyperspectral unmixing aims to recover endmembers and their corresponding abundances from mixed pixels, which is a key task in quantitative hyperspectral remote sensing. Although convolutional neural networks (CNNs) are effective in modeling local spatial–spectral features, their ability to capture long-range spatial dependencies remains limited. Transformers can enhance global feature interactions but typically incur high computational complexity. To address these issues, this study proposes an abundance-guided Mamba network for hyperspectral unmixing, termed AGMNet. Built upon an interpretable autoencoder-based unmixing framework, the proposed method uses CNNs to extract local spatial–spectral features and introduces a lightweight spatial Mamba branch to incorporate whole-image contextual information. Meanwhile, an abundance estimation mechanism with physical constraints and a linear decoder consistent with the linear mixing model are employed to achieve physically interpretable endmember-abundance decomposition. Experiments on three public datasets, namely Samson, Jasper Ridge, and Urban, show that AGMNet provides accurate and stable abundance recovery, achieving the lowest average root mean square error (RMSE) across all three datasets. Its spectral angle distance (SAD) performance is more dataset-dependent: AGMNet obtains the best average SAD on Urban and competitive SAD values on Samson and Jasper Ridge, but it does not consistently achieve the lowest SAD for every endmember. These results indicate that AGMNet is particularly effective for abundance recovery while retaining physically interpretable spectral reconstruction, although further improvement of endmember spectral-shape consistency remains necessary.