Performance comparison of YOLOv5 and YOLOv8 for Plumeria Leaf Gall detection

Abstract

Leaf galls are abnormal plant growths resulting from insect parasitism, which can cause major agricultural losses if not managed in a timely manner. Precise early detection of leaf galls is crucial for enabling targeted treatment of affected plants and precision crop management. However, manual monitoring and identification of leaf galls across large cultivated areas can be extremely labor-intensive, slow and error-prone. This necessitates the development of automated computer vision techniques using deep learning to accurately detect leaf galls at scale for crop health monitoring. This work develops and systematically compares two state-of-the-art convolutional neural network architectures - YOLOv8 and YOLOv5 for automated detection of leaf galls on plumeria leaves. A dataset of 489 high resolution images of plumeria leaves exhibiting leaf galls of various shapes, sizes, textures and colors was collected through extensive field surveys. Each image was annotated by an experienced researcher using bounding boxes demarcating each gall instance. 73% of the images were utilized for training, 12% for validation, while the remaining 15% were held-out for testing model performance. Both YOLOv8 and YOLOv5 models were optimized by tuning key hyperparameters and leveraging data augmentation techniques to minimize overfitting. On the 142-image test set, YOLOv8 achieved a higher mean Average Precision (mAP) of 92.1%, compared to 89.1% attained by YOLOv5, demonstrating YOLOv8's superior accuracy. YOLOv8 also attained higher precision of 90.3% and recall of 88.3%, versus 89.4% and 84.8% for YOLOv5, indicating improved classification and localization capabilities. However, YOLOv5 exhibited slightly faster inference time versus YOLOv8. Overall, this rigorous comparative evaluation highlights YOLOv8 as a more robust and accurate solution for automated leaf gall detection, while YOLOv5 may be more suitable for real-time analysis. The findings provide meaningful insights on deep learning advancements for agriculture applications.