<![CDATA[Purpose – Manual assessment of oil palm fresh fruit bunch (FFB) ripeness remains subjective and may reduce harvest quality consistency. This study aims to evaluate YOLOv11n for six-class FFB ripeness detection using instance-level object detection metrics. Methods – A Roboflow dataset of 17,437 augmented images was split into training, validation, and held-out test subsets across six classes: Empty Bunch, Less Ripe, Abnormal FFB, Ripe FFB, Unripe FFB, and Overripe. A qualitative consistency check was conducted by one harvest foreman. Findings – Evaluation on 1,756 held-out test images containing 6,372 FFB instances achieved precision of 0.968, recall of 0.980, F1-score of 0.974, mAP50 of 0.988, and mAP50-95 of 0.901. Overripe was the weakest class, with mAP50-95 of 0.846. Research implications – The results indicate that YOLOv11n has strong potential to support automated FFB ripeness grading. However, broader field validation, multi-annotator agreement analysis, and testing under more diverse plantation conditions are still required. Originality – This study contributes a six-class YOLOv11n-based FFB ripeness detection evaluation using held-out test-set instance-level metrics and strict localization assessment through mAP50-95.]]>
Copyrights © 2026
