Annotation Refinement and Minority-Class Augmentation for Coffee Leaf Disease Detection Using YOLOv8

Abstract

Coffee leaf diseases negatively affect crop productivity, making early detection an essential task in precision agriculture. The widely used BRACOL dataset presents critical object detection challenges related to spatial localization consistency and extreme class imbalance. Rather than proposing a new detection architecture, this study evaluates data quality and class representation by investigating the incremental effects of manual annotation refinement and targeted minority-class augmentation on detection performance using YOLOv8s. Three experimental scenarios were systematically compared: a baseline dataset directly reproduced from previous work, an annotation-refined dataset emphasizing bounding box tightness, and a dataset combining refinement with offline augmentation for the minority Cercospora class. Experimental results demonstrate a meaningful but modest incremental improvement. Annotation refinement increased mAP50-95 from 0.301 to 0.321, while combining it with augmentation achieved the highest score of 0.328. Class-wise, the severely underrepresented Cercospora showed the largest mAP50-95 improvement, increasing from 0.155 to 0.214. However, the augmentation scenario exhibited a precision-recall trade-off, improving precision but decreasing overall recall compared to refinement alone. Ultimately, while these data-centric interventions positively improve detection under small-lesion and imbalanced conditions, the absolute detection performance remains moderate. Therefore, the current model is not yet ready for reliable field-level disease monitoring and requires broader validation before practical deployment.