A lightweight multidirectional Mamba architecture for hyperspectral image classification

In the paper, we propose a novel network architecture—Lightweight and MultiDirectional Mamba (LMD-Mamba)— specifically designed for Hyperspectral image classification. LMD-Mamba utilizes a multi-directional scanning mechanism with shared parameters, encoding pixel and band sequences along spatial dimensions (both row-wise and column-wise) as well as along the spectral dimension (in both forward and reverse orders). This comprehensive approach effectively captures global dependencies across dimensions and substantially enhances feature extraction capabilities. Experimental results demonstrate that LMD-Mamba not only outperforms existing mainstream methods in classification accuracy on several public benchmark datasets but also exhibits significant advantages in parameter efficiency and computational performance, thereby offering a robust and effective solution for Hyperspectral image classification.