Mamba-based spatial-spectral fusion network for hyperspectral unmixing

• Published in: Journal of King Saud University: Computer and Information Sciences
• Main Authors: Yuquan Gan, Jingtao Wei, Mengmeng Xu
• Format: Article
• Language: English
• Published: Springer 2025-05-01
• Subjects: Hyperspectral unmixing; Hyperspectral image; Mamba model; Fusion mechanism; Multi-Scale analysis
• Online Access: https://doi.org/10.1007/s44443-025-00033-2

Abstract Hyperspectral unmixing (HU) is a critical technique in hyperspectral image (HSI) analysis, aimed at decomposing mixed pixels into a set of spectral signatures (endmembers) and their corresponding abundance values. Recently, the Mamba model has gained significant attention for its exceptional performance in natural language processing and has been extended to vision research. With its strong capability for long-range modeling and linear computational complexity, Mamba demonstrates substantial potential in hyperspectral image processing. However, due to the intrinsic requirement of HU tasks for comprehensive integration of spatial and spectral information, challenges remain in effectively leveraging Mamba for hyperspectral representation. To address these issues, we propose a novel Mamba-based spatial-spectral fusion network for hyperspectral unmixing (Mamba-SSFN). This network introduces a fusion mechanism to jointly learn spectral and spatial feature representations, enabling more efficient extraction of critical HSI features. Specifically, in the spatial feature extraction module, we integrate multi-scale analysis with the Mamba module, enabling the capturing of both local and global spatial information. In the spectral feature extraction module, the Mamba module is employed in a grouped manner to process spectral vectors, exploring the correlations among different spectral groups. Finally, an effective fusion mechanism is implemented to integrate spatial and spectral features.Experimental results demonstrate that Mamba-SSFN achieves outstanding performance across multiple benchmark datasets, significantly surpassing existing state-of-the-art methods in terms of unmixing accuracy, model robustness, and computational efficiency.