MD 2 F-Mamba: Multidirectional Depthwise Convolution and Dual-Branch Mamba Feature Fusion Networks for Hyperspectral Image Classification

Hyperspectral image (HSI) classification necessitates adept modeling of both intricate local variations and long-range spectral-spatial dependencies, while maintaining computational efficiency. Conventional methods frequently prioritize on either local or global features, neglecting directional information, or employ simplistic fusion techniques, which leads to inadequate feature representations and reduced discriminative ability. To address these challenges, this paper presents MD2F-Mamba, a novel dual-branch architecture that integrates a multi-directional depthwise convolution (MDConv) module to capture spatial features from multiple orientations-namely square, horizontal, and vertical convolutions, enriching local representations. The architecture comprises a local branch, featuring a multi-scale local feature enhancement module (MS-LFEM) with positional encoding, which effectively captures diverse spatial-spectral patterns. Simultaneously, the global branch utilizes a hierarchical state-space Mamba for global feature extraction (HSM-GFE) that models multiscale long-range dependencies with linear complexity. A cosine similarity feature fusion (CSFF) module adaptively merges local and global features, optimizing discriminability by reducing redundancy. Experimental results on the Pavia University, Houston2013, WHU-Hi-LongKou, and WHU-Hi-HanChuan datasets demonstrate that MD2F-Mamba achieves competitive classification results while maintaining a minimal parameter count compared with several state-of-the-art methods, underscoring its innovative efficiency and robustness in HSI classification.