Comparative of YOLOv5 and YOLOv8 for rice leaf disease detection on diverse image datasets

Abstract

Rice (Oryza sativa) is Indonesia’s primary food crop, yet its productivity is often threatened by leaf diseases such as Brownspot, Hispa, and Sheath Blight. To address the limitations of manual inspection, this study proposes an automated detection and classification framework based on deep learning, with a comparative evaluation of the YOLOv5 and YOLOv8 models. This study is novel in that it assesses the robustness of models across a variety of data sources, such as a public dataset collected under controlled conditions and a private dataset collected in the field that replicates real-world agricultural contexts. The experimental results suggest that YOLOv8 consistently outperforms YOLOv5 in a variety of evaluation metrics. YOLOv8 performed best on the private dataset, with a precision of 0.907, recall of 0.886, F1-score of 0.896, Intersection over Union (IoU) of 0.71, and mAP50 of 0.924 under the 90:5:5 data split configuration. It shows that it can detect things well even in difficult field conditions. Both models performed about the same on the public dataset; however, YOLOv8 was better at finding objects, as shown by higher mAP50–95 values. Both models also did a great job of classifying; however, YOLOv8 was better at generalising across different dataset distributions. These results demonstrate that YOLOv8, which operates without anchors, is a superior and more dependable method for the real-time detection of rice leaf disease. This study offers pragmatic insights for implementing advanced computer vision models in precision agriculture systems, particularly in resource-constrained, dynamic agricultural environments.