Multi-Defect Classification in Oranges Using YOLOv8, YOLO11, and YOLOv8m: A Comparative Study

Abstract

Orange farming is among the most economically significant agricultural activities in the world, with the crop feeding global markets and supporting millions of smallholder livelihoods. Yet despite that importance, quality grading at the post-harvest stage still depends heavily on manual inspection, a slow and inconsistent process that struggles to keep pace with the demands of modern supply chains. Surface diseases such as Blackspot, Canker, and Greening are among the leading causes of post-harvest losses, and catching them early is critical to protecting fruit quality and commercial value. This paper presents an automated classification framework for orange fruit defects that evaluates three members of the You Only Look Once (YOLO) model family: YOLOv8n, YOLO11n, and an enhanced YOLOv8m trained with an extended preprocessing and augmentation pipeline. The system classifies orange images into four categories—Blackspot, Canker, Fresh, and Greening—enabling defect-specific grading decisions rather than simple pass/fail quality labelling. All models were trained and tested on a publicly available orange disease dataset under identical experimental conditions, providing a controlled multi-model benchmark. Evaluation metrics include Top-1 Accuracy, Precision, Recall, and F1-Score per class. The enhanced YOLOv8m achieved the best overall performance with 97.98% accuracy and a macro F1-Score of 0.98, followed by YOLO11n at 96.97% and YOLOv8n at 93.94%. These results show that targeted augmentation and regularization strategies can produce meaningful gains in classification performance even on small, class-imbalanced agricultural datasets.