To address the limitations of traditional convolutional neural networks in hyperspectral image classification under sample-limited scenarios, including insufficient feature extraction and information redundancy, this paper proposes a novel network model called CGMSANet. This model replaces traditional batch normalization with group normalization (GN) and integrates it with attention mechanisms to effectively mitigate statistical bias in small-sample training, thereby enhancing the model’s generalization capability. Additionally, a dual-path attention module is designed, combining spatial dependency capture (SDC) and spectral feature weighting (SFW). SDC models cross-band global dependencies in the spatial dimension, while SFW dynamically suppresses redundant channels in the spectral dimension, thereby enhancing focus on discriminative features. Experiments on three public data sets—HongHu, LongKou, and HanChuan—demonstrate that CGMSANet achieves superior performance in terms of F1 score and classification accuracy. Specifically, on the HongHu data set, the average F1 score reaches 96.93%, with a 57% reduction in computational complexity and a 2.45× improvement in inference speed compared to the transformer-based SpectralFormer model. Ablation studies further validate the collaborative effectiveness of the GN, SDC, and SFW modules. This model provides an efficient and robust solution for few-shot hyperspectral classification tasks and shows significant potential for practical applications in agricultural monitoring and resource management. The data are available at https://github.com/YMY666yy/CGMASNet.
Ye et al. (Wed,) studied this question.