WheatNet: A Domain-Aware Attention-Driven Framework for Robust Wheat Disease Detection in Real-World Field Conditions

Wheat diseases pose a persistent threat to global food security, as their visual symptoms vary widely across growth stages, environmental conditions, and plant organs, making reliable field-level detection particularly challenging. To address these limitations, this paper proposes WheatNet, a domain-aware deep learning framework for accurate and robust wheat disease detection under real-world agricultural conditions. The framework integrates leaf- and head-region imagery through an attention-enhanced DenseNet–FPN backbone to capture subtle lesion textures and spatial patterns. A disease-aware augmentation strategy models occlusions, illumination variability, and sensor noise, while an adaptive loss formulation mitigates severe class imbalance. Extensive experiments on enriched multi-source datasets show that WheatNet consistently outperforms classical CNNs and state-of-the-art detectors in mAP, F1-score, and robustness. Interpretability analyses further confirm that the model attends to biologically meaningful disease regions, supporting practical deployment in precision agriculture.