Conv-J2SMamba: A New Hybrid Convolutional and Joint Scan Merge Structured Mamba Network for Hyperspectral Image Classification

Recently, Mamba has shown great potential in longsequence modeling and begun to be applied to hyperspectral image classification (HSIC). However, as a causal sequence model, Mamba can only obtain a global receptive field at the end of the sequence, which limits its ability to capture complex spatial structures in hyperspectral images (HSIs). To address this problem, a new Hybrid Convolution and Joint Scan Merge Structured Mamba Network (Conv-J2SMamba) is proposed for HSIC, which establishes neighborhood connectivity within the state space model to extract global-local features from HSIs. Specifically, our Joint Scan Merge Structured Mamba (J2SMamba) is designed to extract global features, while convolution captures local features. The core module of J2SMamba is the Joint Scan Merge Structure-Aware State Space Model (J2-SASSM), which introduces two key innovations. First, the Joint Scan Merge strategy concatenates dual HSIs and constructs four stride-based scanning paths for joint scanning and merging. This strategy enables omnidirectional global feature extraction. Second, the Structure-Aware State Space Model (SASSM) utilizes the Structure-Aware State Fusion (SASF) equation to capture spatial structural dependencies in the HSI, significantly enhancing the flow of contextual information. To better integrate the generated features, a Contrastive Guided Feature Fusion (CGFF) module is also introduced, which employs a multi-level attention mechanism to guide key feature enhancement and multi-level feature fusion. Experimental results on three public HSI datasets demonstrate that the proposed Conv-J2SMamba model outperforms state-of-the-art methods.