A unified spatial-spectral pyramid mamba for hyperspectral image classification

Abstract

Hyperspectral image (HSI) classification is vital in remote sensing for land cover mapping, environmental surveillance, and precision agriculture. However, achieving high accuracy with minimal samples is a major challenge. Mamba’s superior long-range modeling and linear complexity offer a promising solution. Furthermore, leveraging the 3D structure of HSIs, effective multiscale spatial-spectral feature extraction is key to improving performance. Towards this end, we propose the Pyramid-based Spatial-Spectral Mamba (PS²Mamba). It has a dual-branch architecture within a unified framework with three key modules. In the spectral branch, the Pyramid-based Spectral Mamba (PSpeM) incrementally captures hierarchical spectral representations through spatial neighborhood weighting and a bidirectional scanning strategy to enhance spectral discrimination. The spatial branch’s Pyramid-based Spatial Mamba (PSpaM) captures multi-scale, multi-directional spatial features, incorporating an innovative eight-direction scanning mechanism to encode both fine and coarse patterns. In addition, a Spatial-Spectral Interaction Fusion (SSIF) module enables efficient cross-domain interaction for seamless fusion. Experimental results demonstrate that the proposed method achieves overall accuracy (OA) scores of 90.90%, 95.62%, 93.09%, 95.78% and 91.83% on five benchmark HSI datasets, under a limited training regime with only 10 labeled samples per class. Notably, these results represent significant improvements over state-of-the-art algorithms, with relative gains of 1.20%, 1.11%, 2.61%, 1.14% and 1.87% respectively. The code will be available at: https://github.com/Duppy0509/PS2Mamba.