Authors - Soumyadeep Basak, Shubham Sahu, Sankur Kundu, Ankita Ray Chowdhury Abstract - Hyperspectral image (HSI) classification requires effective modeling of high-dimensional spectral signatures and fine-grained spa tial structures while maintaining computational efficiency for real-world deployment. Although recent Transformer- and state-space-based ap proaches enhance long-range dependency modeling, they often introduce substantial architectural complexity and computational overhead. To ad dress these challenges, we propose MF-HSINet, a lightweight dual branch framework that enables adaptive spectral–spatial fusion via se lective state-space modeling. The spectral branch captures inter-band de pendencies, the spatial branch extracts local structural patterns, and the proposed Mamba-Enhanced Attention Fusion (MAF) module integrates these complementary representations through selective state updates, cross-attention, and adaptive gating to achieve pixel-wise feature balanc ing. This design preserves discriminative local details while strengthen ing global contextual modeling with reduced parameter complexity. Ex tensive experiments on nine benchmark hyperspectral datasets demon strate that MF-HSINet achieves competitive and consistent performance in terms of Overall Accuracy, Average Accuracy, and Kappa coefficient, while offering improved efficiency and inference speed, making it suitable for practical and resource-constrained HSI applications.
Open Zoom