Illumination-Decoupled Transformer Learning for Shadow-Robust Crop Disease Diagnosis Under Structured Cast Shadows

Crop disease diagnosis can be degraded by structured cast shadows, including panel-like strip shadows that motivate applications in agrivoltaic-style farming. This paper presents ShadowFormer-AV, a transformer-based framework that adapts general shadow-robust visual learning to crop disease classification by separating disease evidence from illumination interference. The proposed approach combines a soft shadow-prior extractor, an illumination-decoupled dual-stream token encoder, lesion-preserving adaptive attention, and a cross-view consistency objective between original and shadow-perturbed images. The method uses only standard RGB inputs and does not require shadow-free reference images, multispectral sensing, or pixel-level shadow annotation. We evaluated the framework on publicly available plant disease datasets using calibrated panel-like synthetic shadows and a naturally shadowed PlantDoc subset. Because no on-site agrivoltaic disease dataset was used, the conclusions were limited to shadow robustness under these simulated and naturally shadowed test conditions rather than verified performance under real photovoltaic-panel shadows. Within this validation boundary, ShadowFormer-AV improved accuracy, Macro-F1, and calibration over representative convolutional and transformer baselines, suggesting that illumination-aware token learning is useful for crop disease recognition under structured shadow interference.