A Lightweight Hybrid CNN-Mamba Network for Hyperspectral Image Classification

In hyperspectral image (HSI) classification, the effectiveness and efficiency of modeling complex spectral-spatial relationships are critical. However, most of CNN and Transformer-based methods face challenges in modeling local and global spectral-spatial information while demanding high computational resources. To address these issues, we propose a lightweight hybrid CNN-Mamba network for hyperspectral image classification (HSIC), which combines the strengths of CNN and Mamba models, and effectively captures both global and local spectral-spatial features. First, we introduce a wavelet separable convolution (WSC) module to enhance spatial feature extraction. The WSC module integrates wavelet transform with depth-wise separable convolution, expanding the receptive field and improving the model's ability to capture complex spatial structures while maintaining computational efficiency. Second, we propose a lightweight spectral convolution (LSC) module, which focuses on extracting local spectral features through 3-D convolution, balancing accuracy and computational burden. Additionally, the network incorporates the Mamba module for modeling long-range dependencies via a state-space model (SSM). The experimental results show that the classification performance of our method on Pavia University datasets outstands several state-of-the-art methods.