Accurate detection of oil palm fruit maturity levels plays a crucial role in improving harvesting efficiency and maintaining the quality of palm oil production. In practice, this task remains challenging due to the presence of severe class imbalance in real-world field datasets, where certain classes have far fewer samples than others, often leading to biased model learning and reduced detection accuracy. This study investigates the performance of several Class-Balanced Loss Function variants integrated into the YOLOv11-nano framework using a publicly available oil-palm fruit dataset for harvest estimation, which presents a significantly imbalanced class ratio. Four training configurations were evaluated: the baseline Binary Cross-Entropy (BCE), Class-Balanced Focal Loss (CB-Focal), Class-Balanced Sigmoid Loss (CB-Sigmoid), and Class-Balanced Softmax Loss (CB-Softmax). The experimental results indicate that CB-Focal achieved the highest performance with an mAP@50 of 0.783, approximately 0.5 percent higher than the BCE baseline (0.778) and 4 to 5 percent greater than YOLOv8-n and YOLOv8-s models trained on the same dataset. CB-Focal also demonstrated smoother convergence and more balanced per-class performance compared to the other loss functions. These findings suggest that integrating CB-Focal into the YOLOv11-nano framework not only improves accuracy for minority classes but also holds strong potential for supporting more accurate, efficient, and scalable automated harvest monitoring systems in real plantation environments.
Suparto et al. (2026) studied this question.