An Explainable Artificial Intelligence Approach to Rice Leaf Disease Classification Using MobileNetV3-Large

Abstract

In Indonesia, rice is a vital component of food security, but persistent leaf diseases threaten productivity. Currently, field identification relies on labor-intensive, human-error-prone manual observation. To address this, we developed an accurate and interpretable classification model integrating the MobileNetV3-Large with Gradient-weighted Class Activation Mapping (Grad-CAM). The model obtained 98.82% validation accuracy after being trained via transfer learning on a public dataset with four classes: healthy, bacterial leaf blight, brown spot, and blast. The dataset was partitioned using a 90:10 data split (5,018 training and 612 validation samples). The model achieved a validation accuracy of 98.82% alongside robust multi-metric performance, obtaining a weighted average precision of 99.00%, a recall of 98.80%, Micro AUC and Macro AUC both reached 0.9977, and an F1-score of 98.90%. Furthermore, Grad-CAM analysis confirmed that the model's predictions are interpretable. The resulting heatmaps consistently concentrated on leaf regions exhibiting disease symptoms, demonstrating that the network learned visually meaningful features despite minor background activations. These findings validate that combining MobileNetV3-Large with Grad-CAM yields an effective, transparent classification system, offering strong potential for robust, user-verifiable rice disease diagnosis in agricultural deployments.