Rice Leaf Disease Detection Using Transfer Learning with YOLOv8 and the Ultralytics Validation Engine

Abstract

Rice leaf diseases are difficult to distinguish visually when symptoms overlap, and image classification alone does not localize symptomatic regions. This study investigated whether a transfer-learned YOLOv8m detector, evaluated through the Ultralytics validation engine and integrated into a web prototype, could provide lesion-level detection for bacterial leaf blight, brown spot, and leaf blast. The dataset combined 220 field images from Sukabumi, 792 disease images from Roboflow, and 759 null images. The 1,771 original images were split before augmentation into 1,239 training, 357 validation, and 175 testing images, and augmentation was applied only to the training partition. The model was fine-tuned from COCO-pretrained weights, while the prototype was examined through six black-box functional scenarios on one desktop and two Android devices. On the internal validation partition, the model achieved 86.9% precision, 72.9% recall, 84.6% mAP@50, and 47.5% mAP@50–95. Bacterial leaf blight obtained the highest class AP@50 at 92.5%, whereas brown spot obtained the lowest at 79.0%. The lower recall indicates that some annotated lesions remained undetected, while the gap between mAP@50 and mAP@50–95 indicates reduced bounding-box stability under stricter localization criteria. All six software scenarios returned the expected interface outputs, demonstrating functional prototype operation rather than independent diagnostic generalization. The study contributes an end-to-end proof of concept linking lesion localization, internal validation, and web deployment. Independent test-set evaluation, expert-confirmed labels, leakage screening, external field validation, and measured inference latency remain necessary before operational use.