Explainable Lightweight Deep-Learning Framework for Field-Based Java Apple Leaf-Disease Classification Using a Custom Dataset from Bangladesh

Java Apple (Syzygium samarangense), an economically important tropical fruit, is highly susceptible to foliar diseases that can significantly diminish yield and quality. Manual disease diagnosis is time-consuming and dependent on expert availability, highlighting the need for an automated, reliable solution. In this study, we present a lightweight deeplearning framework for classifying five common Java Apple leaf conditions—Pest Damage, Downy Mildew, Bacterial Leaf Spot, Healthy, and Dried—using a custom dataset of 3,812 field-acquired images collected across Bangladesh and enriched through targeted augmentation techniques. We evaluated four state-of-the-art convolutional neural networks (ResNet50, EfficientNet-B0, DenseNet121, and MobileNetV2) with rigorous hyperparameter tuning and early stopping to balance performance and efficiency. MobileNetV2 emerged as the optimal model, achieving a peak accuracy of 99.10%, precision of 0.97, recall of 0.94, and F1-score of 0.97, while maintaining a compact architecture suitable for edge deployment. To ensure transparency, we integrated explainable AI techniques—GradCAM and Ablation-CAM—to visualize the class-discriminative regions driving model predictions, confirming alignment with biologically meaningful leaf features. The proposed system offers a practical tool for rapid, in-field disease monitoring, with potential extensions to mobile and UAV-based surveillance for scalable smart-agriculture applications.