A Comparative Deep Learning Approach for Classifying Oil Palm Fruit Ripeness Levels Using YOLOv8s and Faster R-CNN

Abstract

Assessing oil palm fruit ripeness is essential for optimizing harvest timing and maximizing market value. In many developing regions, harvesting is still performed every 10–15 days through manual visual inspection, a process prone to human error that often causes premature harvesting and reduces selling value by up to 50%. This study explores deep learning-based object detection for automatic classification of oil palm fruit bunches. A dataset of 4,578 annotated high-resolution images was prepared and categorized into six ripeness classes: Empty, Immature, Underripe, Abnormal, Ripe, and Overripe. Two advanced detection models, YOLOv8s and Faster R-CNN with a ResNet-50 backbone, were evaluated under identical conditions using precision, recall, and mean Average Precision (mAP) metrics. YOLOv8s achieved precision and recall above 99%, with a mAP 0.5:0.95 of 0.9254, demonstrating strong reliability and efficiency for real-time use. Faster R-CNN achieved a higher mAP 0.5 of 0.9964, indicating superior localization accuracy but slower computation. Overall, YOLOv8s provides a better trade-off between accuracy and speed, making it more practical for automated harvesting. This research supports precision agriculture by emphasizing AI-driven solutions that improve productivity, minimize losses, and promote sustainable palm oil management.