SSA-Mamba: Spatial-Spectral Attentive State Space Model for Hyperspectral Image Classification

Hyperspectral image(HSI) classification faces critical challenges in effectively modeling long-range dependencies while maintaining computational efficiency and synergistically exploiting fields, spatial-spectral information. Convolutional neural networks(CNNs) are constrained by local receptive transformers computational complexity, suffer from quadratic and existing state space model(SSM)-based methods lack sophisticated cross-domain interaction mechanisms. This paper proposes SSA-Mamba (Spatial-Spectral Attentive Mamba), a novel classification approach addressing these limitations through 3 synergistic innovations. First, a dual-branch independent modeling strategy allocates separate parameter spaces for spatial and spectral feature extraction via parallel SSMs, preventing feature coupling while enabling domain-specific learning. Second, an asymmetric cross-domain attention mechanism allows spatial features to actively query spectral information through multi-head attention, establishing adaptive fusion via gating mechanisms and channel attention. Third, a multi-scale residual architecture operating at module-internal, block-internal, and global pathway levels achieves hierarchical feature fusion while maintaining numerical stability through exponential parameterization. The recursive computation mechanism of SSMs enables each position to aggregate global historical information through compact hidden states, achieving O(L) linear complexity compared to transformers' O(L²) quadratic complexity. Extensive experiments on three benchmark datasets—Houston2013, WHU-Hi-HongHu, and XiongAn—validate the effectiveness of these innovations. SSA-Mamba achieves overall accuracies of 93.98%, 93.58%, and 96.06%, surpassing state-of-the-art approaches by 1.27%, 0.25%, and 1.27%, respectively. The dual-branch design enables effective discrimination of spectrally similar categories, improving Brassica variety classification by 19.21-23.33 percentage points over coupled-feature approaches. The cross-domain attention mechanism enhances urban land cover classification, with Commercial and Highway categories improving by 1.74% and 15.66%. On the large-scale XiongAn dataset (5.92 million pixels), SSA-Mamba demonstrates exceptional scalability with peak GPU memory of only 317.89 MB and per-sample inference time of 0.646 ms, providing an efficient solution for real-time HSI processing. The source code for SSA-Mamba will be made publicly available at: https://github.com/Jason20155/.