Efficient Progressive Mamba Model for Hyperspectral Sequence Unmixing

In recent years, deep learning-based hyperspectral unmixing (HU) has increasingly incorporated spatial information to improve performance. However, the extent of spatial information introduced involves a complex trade-off: too little offers limited gains, while too much leads to spectral mixing from neighboring pixels (e.g., CNNs and Vision Transformers), which contradicts the goal of recovering pure endmembers. Meanwhile, pixel-level unmixing requires per-pixel modeling with high computational cost, limiting the application of powerful sequence models. To address this, we propose an efficient progressive hyperspectral sequence unmixing model (ProMU), introducing a multi-stage progressive context selection (CVPS) strategy that ensures computational efficiency while alleviating the performance degradation caused by the length mismatch between spectral and abundance sequences. We further design a stage-aware Mamba module to capture dependencies across stages, positions, and dimensions, and propose an autoregressive mechanism for fine-grained abundance prediction. Besides comparisons with image-level methods, we systematically evaluate mainstream sequence models for pixel-level unmixing on three real hyperspectral datasets. Experiments show that ProMU achieves state-of-the-art (SOTA) performance with extremely low parameter count and computational cost, rivaling image-level methods. This validates the potential of sequence models for pixel-level unmixing and offers a novel solution for hyperspectral unmixing research. Code is available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Liujehong/ProMU</uri>.