Rice Leaf Disease Classification Based on Convolutional Neural Network Transfer Learning Using MobileNetV2 and EfficientNetB0

Abstract

Rice is a strategic commodity that plays a vital role in Indonesia’s national food security. Conventional identification of rice diseases, which relies on direct visual inspection, has several limitations; therefore, developing an automated identification system using Transfer Learning–based Convolutional Neural Networks (CNNs) has become increasingly important. This study employs MobileNetV2 and EfficientNetB0 architectures to compare the performance of both models in classifying rice leaf diseases. The dataset consists of 2,756 images obtained from Kaggle, with 690 images used for validation and 383 images for testing, classified into six categories of rice leaf conditions. These categories include five disease classes—Bacterial Leaf Blight, Brown Spot, Leaf Blast, Leaf Scald, and Sheath Blight—and one healthy leaf class. Preprocessing stages include pixel normalization, data augmentation, and resizing to 224×224 pixels. Model training was conducted for 20 epochs using a learning rate of 0.001 with the Adam optimizer. Evaluation results show that MobileNetV2 achieved an accuracy of 83.29%, precision of 87.20%, recall of 76.50%, and an F1-score of 82.87%. In comparison, EfficientNetB0 achieved an accuracy of 86.95%, precision of 90.00%, recall of 80.15%, and an F1-score of 86.78%. These findings highlight that although both models perform effectively for rice disease classification, EfficientNetB0 consistently provides a superior margin of approximately 3–4% across all metrics, making it a more stable choice for accurate early detection of rice leaf diseases. This study contributes to accelerating the identification process and supporting decision-making in the agricultural sector.