Comparative evaluation of YOLO algorithms for detection and classification of rice leaf diseases using field-collected datasets

Abstract

Rice is the staple food for more than half of the global population. However, the productivity and quality of rice or grain dropped significantly due to leaf diseases. These diseases are difficult to identify through manual processes, which are time-consuming, labour-intensive and often inaccurate, particularly in rural farming communities. With recent advances in computer vision technology, object detection algorithms, namely the you only look once (YOLO) family, can provide high-speed, high-accuracy solutions for automated plant disease detection. This study evaluates four YOLO variants, such as YOLOv5, YOLOv7, YOLOv8 and YOLOv11, using about 1500 field images collected in Bangladesh. The data represent four major rice leaf diseases, such as bacterial leaf blight, brown spot, leaf blast and sheath blight. Data pre-processing, including image annotation and data augmentation, was conducted before model training and was followed by the training of the YOLO models. All the models were trained with the same hyperparameters and their performance was evaluated using standard metrics, such as F1 scores, precision, recall and mean average precision (mAP). According to experimental findings, the YOLOv7 recorded the highest performance based on F1 score of 0.77 and mAP of 0.85 in comparison with the rest of the variants. The results suggest that YOLOv7 will be the most appropriate to use instead of other models in the detection of rice leaf disease in real-time, which can be utilised in precision agriculture and mobile-based disease management systems.

Plant phenotyping relevance

イネ葉の病害状態を画像から検出・分類するYOLO手法を複数比較し、フィールド画像データセット上で性能検証しているため、植物フェノタイピング手法が中心である。

Code and data availability

The supplied article blocks describe YOLO evaluation on ~1500 field-collected rice leaf disease images from Bangladesh, but contain no data availability statement, no public dataset deposit, no code repository, and no supplement with images, annotations, or trained models. The only links are ORCID profiles, the article

No evidence-backed public reproduction asset is currently recorded.