MCFRNet: Multi-path contextual feature refinement network for hyperspectral image classification

The key challenge in hyperspectral image (HSI) classification is to obtain spectral-spatial representations that can capture sufficient contextual information without sacrificing fine-grained local details or introducing excessive computational cost. Existing CNN-based methods are generally efficient at extracting local features; however, their limited receptive fields restrict broader contextual modeling. Transformer-based methods can capture long-range dependencies, but they often introduce high computational cost and may be less effective in preserving local spatial details. To address this challenge, this paper proposes a lightweight multi-path contextual feature refinement network, termed MCFRNet. Specifically, the HEFE module is first designed to enlarge the effective receptive field with low computational cost by using depthwise convolution and depthwise separable convolution. Then, the MCFR module integrates multi-layer feature reconstruction (MFR) and multi-group contextual feature aggregation (MCFA) to refine contextual features and preserve discriminative details in a unified process. Finally, the SACA module combines channel attention and global spatial attention to adaptively enhance informative spectral channels and strengthen spatial correlations. Experimental results on four public HSI datasets demonstrate that MCFRNet achieves competitive classification accuracy with low parameter count and computational complexity. The code of this work is available at https://github.com/sunziqi-study/MCFRNet.