SPECGAN: Extracting sensitive bands from plant disease spectra based on generative adversarial network

Hyperspectral imaging provides detailed spectral information for non-destructive plant disease diagnosis, yet its use is limited by high dimensionality of the original spectra, as well as insufficient and imbalanced data records. These issues hinder the extraction of weak pathological signals and ultimately reduce model applicability. To overcome these challenges, this study proposes SPECGAN, a Generative Adversarial Network with a Temporal-Domain Feature Pyramid Fusion (TD-FPNF) and a residual attention mechanism. SPECGAN forms a general framework for both sensitive band extraction and data augmentation. A multi-scale convolutional module captures local narrow-band features related to biochemical changes, as well as global broadband trends linked to physiological structure. The residual attention mechanism further enhances subtle disease cues by adaptively reweighting multi-level fused features and suppressing background noise. SPECGAN accurately identifies key discriminatory bands based on gradient saliency analysis of the discriminator, while generating high-quality synthetic samples to alleviate data scarcity. Experiment results demonstrate that the sensitive bands concentrate in the green peak (520-550 nm) and red-edge (680-720 nm) regions, consistent with disease-induced physiological changes. By only inputting the top 20 bands (8% of the spectrum), the MLP classifier achieves 96.22% accuracy. Under a 14.6:1 imbalance scenario, generating 1500 synthetic samples boosts performance by 6%-13%. Overall, SPECGAN provides an efficient and interpretable approach for early diagnosis of rice bacterial leaf blight.