Wheatleafnet: A Novel Explainable Ai-Based Hybrid DeepLearning Approach for Wheat Leaf Disease Detection and Classification

Agriculture is the second-largest sector in India, contributing 20.2% to the national GDP, with wheat ‎cultivation playing a crucial role in global food security. However, wheat is highly susceptible to fungal, ‎bacterial, viral, and nutrient deficiency-related diseases that can severely reduce yield. Early and accurate ‎disease detection is therefore vital for sustainable crop management. In this study, we introduce ‎WheatLeafNet, a deep learning framework for multi-class wheat leaf disease classification. The dataset ‎comprises 6,247 images across 24 categories (23 disease types and healthy samples). To enhance model ‎robustness and mitigate overfitting, data augmentation and regularization strategies were employed, ‎alongside ablation studies on their effectiveness. We evaluated a custom CNN and transfer learning ‎architectures (MobileNetV2, ResNet50, EfficientNet-B0) under four optimizers (SGD, RMSProp, ‎Adam, and AdaGrad), with stratified 5-fold cross-validation ensuring reliable assessment. Among the ‎tested models, MobileNetV2 with the AdaGrad optimizer achieved the best performance, reaching an ‎accuracy of 97.91% without augmentation and 97.84% with augmentation. Comprehensive evaluation ‎metrics, including per-class precision, recall, F1-score, ROC-AUC, Expected Calibration Error (ECE), ‎and Grad-CAM visualizations, confirm the reliability and interpretability of the framework. The ‎integration of augmentation, optimization, and explainability strengthens the model’s generalizability for ‎real-world applications. The proposed system offers an effective and scalable solution for early disease ‎identification in wheat, enabling timely interventions, reducing crop loss, and promoting sustainable ‎agricultural practices. Future work will incorporate segmentation-driven severity estimation and multi-‎label detection to enable real-world deployment.