The paper introduces DualMamba, a lightweight dual-stream Mamba-convolution network designed for hyperspectral image (HSI) classification. DualMamba aims to efficiently model complex spectral-spatial relationships by combining the global context modeling capabilities of Mamba with the local feature extraction capabilities of a lightweight spectral-spatial CNN. The key contributions of DualMamba include: 1. **Cross-Attention Spectral-Spatial Mamba Module**: This module leverages dynamic positional embedding to enhance spatial location information and employs cross-attention fusion to integrate global spectral and spatial features. It captures global spectral-spatial relations with linear complexity and minimal parameters. 2. **Lightweight Spectral-Spatial Residual Convolution Module**: This module uses residual learning to extract local spectral-spatial features through lightweight spectral and spatial branches, efficiently capturing local spectral-spatial relations. 3. **Adaptive Global-Local Fusion**: This module dynamically adjusts the balance between global and local features to achieve a comprehensive spectral-spatial representation, enhancing the model's adaptability to different HSI contents. Experimental results on three public HSI datasets (Indian Pines, WHU-Hi-Longkou, and Houston 2018) demonstrate that DualMamba outperforms state-of-the-art methods in terms of classification accuracy while significantly reducing model parameters and floating-point operations (FLOPs). The method's effectiveness is further validated through ablation studies, which show that each component of DualMamba contributes to its overall performance.The paper introduces DualMamba, a lightweight dual-stream Mamba-convolution network designed for hyperspectral image (HSI) classification. DualMamba aims to efficiently model complex spectral-spatial relationships by combining the global context modeling capabilities of Mamba with the local feature extraction capabilities of a lightweight spectral-spatial CNN. The key contributions of DualMamba include: 1. **Cross-Attention Spectral-Spatial Mamba Module**: This module leverages dynamic positional embedding to enhance spatial location information and employs cross-attention fusion to integrate global spectral and spatial features. It captures global spectral-spatial relations with linear complexity and minimal parameters. 2. **Lightweight Spectral-Spatial Residual Convolution Module**: This module uses residual learning to extract local spectral-spatial features through lightweight spectral and spatial branches, efficiently capturing local spectral-spatial relations. 3. **Adaptive Global-Local Fusion**: This module dynamically adjusts the balance between global and local features to achieve a comprehensive spectral-spatial representation, enhancing the model's adaptability to different HSI contents. Experimental results on three public HSI datasets (Indian Pines, WHU-Hi-Longkou, and Houston 2018) demonstrate that DualMamba outperforms state-of-the-art methods in terms of classification accuracy while significantly reducing model parameters and floating-point operations (FLOPs). The method's effectiveness is further validated through ablation studies, which show that each component of DualMamba contributes to its overall performance.