Abstract:To address the difficulty of jointly modeling local structures, multiscale spatial dependencies, and observation deviations in deep learning methods for hyperspectral unmixing, an unmixing method based on dual-stream feature fusion of a convolutional neural network (CNN) and Mamba is proposed. A residual-projection-based band reduction method is used to obtain a low-dimensional representation of the original hyperspectral image. A local branch combining CNN with the fast Fourier transform (FFT) and a multiscale window-based Mamba branch are constructed in parallel, and joint gating is used to fuse local textures, frequency-domain information, and multiscale contextual information. A bidirectional spectral Mamba module is employed to update the endmembers, while illumination modulation and bounded residual compensation are incorporated into linear mixture reconstruction. Experimental results on real-world datasets show that local features and multiscale spatial context are complementary, and that their integration with constrained reconstruction compensation enhances joint spatial–spectral representations, demonstrating the effectiveness of the proposed method for blind hyperspectral unmixing.