Abstract: To address the problems of high spectral redundancy and insufficient spatial-spectral information fusion in hyperspectral images of rice diseases and pests, which limited segmentation accuracy, a spatial-spectral attention U-Net model (SSA U-Net) was proposed. A band gating module was introduced at the input stage to adaptively weight spectral bands, thereby highlighting disease- and pest-sensitive bands while suppressing redundant spectral information. In the skip connections, a spectral-guided residual attention gate module was designed to generate spatial attention weights by integrating high-level decoder semantics with spectral priors, which improved the representation of weak lesions, small-scale regions, and blurred boundaries. Meanwhile, a joint loss function combining cross-entropy loss, Dice loss, and band sparsity regularization was constructed to alleviate class imbalance and enhance the learning ability for key spectral features. Experiments on 1 270 of 34-band hyperspectral images of rice diseases and pests showed that SSA U-Net achieved a mean Dice coefficient of 89.6% and a mean intersection over union of 81.2% on the test set, which were 5.4 and 8.5 percentage points higher than those of the baseline U-Net, respectively. The overall performance metrics were higher than those of DeepLabV3+, U-Net, U-Net++, and SegNet, while the parameter size (3.16×10⁷) was at the same order of magnitude as that of U-Net. SSA U-Net achieved pixel-level segmentation of rice disease and pest regions, providing technical support for intelligent monitoring and precision plant protection.

Key words: rice diseases and pests, hyperspectral imaging, semantic segmentation, spatial-spectral attention, U-Net