DualMamba is a lightweight spectral-spatial Mamba-convolution network designed for hyperspectral image (HSI) classification. The method addresses the challenges of efficiently capturing global and local spectral-spatial features in HSI classification, which is crucial for accurate material and object identification. Existing methods based on CNNs and transformers suffer from high computational costs and limited effectiveness in capturing complex spectral-spatial relationships. DualMamba introduces a parallel lightweight design combining Mamba and CNN blocks to extract global and local features. The cross-attention spectral-spatial Mamba module enhances spatial location information through dynamic positional embedding and efficiently captures global spectral-spatial features using lightweight Mamba blocks. The lightweight spectral-spatial residual convolution module extracts local features using residual learning. An adaptive global-local fusion dynamically combines global Mamba features and local convolution features for a comprehensive spectral-spatial representation. DualMamba achieves significant classification accuracy on three public HSI datasets with a substantial reduction in model parameters and floating point operations (FLOPs) compared to state-of-the-art methods. The method's lightweight design and efficient computation make it suitable for HSI classification tasks. The architecture includes a cross-attention spectral-spatial Mamba module, lightweight spectral-spatial residual convolution module, and adaptive global-local fusion. Experimental results demonstrate that DualMamba outperforms existing methods in terms of classification accuracy and computational efficiency.DualMamba is a lightweight spectral-spatial Mamba-convolution network designed for hyperspectral image (HSI) classification. The method addresses the challenges of efficiently capturing global and local spectral-spatial features in HSI classification, which is crucial for accurate material and object identification. Existing methods based on CNNs and transformers suffer from high computational costs and limited effectiveness in capturing complex spectral-spatial relationships. DualMamba introduces a parallel lightweight design combining Mamba and CNN blocks to extract global and local features. The cross-attention spectral-spatial Mamba module enhances spatial location information through dynamic positional embedding and efficiently captures global spectral-spatial features using lightweight Mamba blocks. The lightweight spectral-spatial residual convolution module extracts local features using residual learning. An adaptive global-local fusion dynamically combines global Mamba features and local convolution features for a comprehensive spectral-spatial representation. DualMamba achieves significant classification accuracy on three public HSI datasets with a substantial reduction in model parameters and floating point operations (FLOPs) compared to state-of-the-art methods. The method's lightweight design and efficient computation make it suitable for HSI classification tasks. The architecture includes a cross-attention spectral-spatial Mamba module, lightweight spectral-spatial residual convolution module, and adaptive global-local fusion. Experimental results demonstrate that DualMamba outperforms existing methods in terms of classification accuracy and computational efficiency.