CFO-RetinaNet: Convolutional Feature Optimization for Oil Palm Ripeness Assessment in Precision Agriculture

Accurate ripeness assessment of oil palm fruit bunches (FFB) is critical for optimizing yield and quality in the palm oil industry, yet manual grading remains subjective and labor-intensive. This study proposes CFO-RetinaNet, an enhanced RetinaNet framework integrating deformable convolutions and hybrid attention mechanisms to optimize multi-scale convolutional features for robust ripeness classification under variable field conditions. Our key contribution is threefold: (1) a novel dataset of 4,728 high-resolution, expert-annotated FFB images spanning five ripeness stages (Immature to Decayed), collected under diverse lighting and occlusion scenarios in Central Kalimantan, Indonesia; (2) a feature optimization pipeline combining adaptive feature fusion and dynamic focal loss to improve discriminative capability for nuanced inter-class distinctions; and (3) a scalable deep learning solution validated through rigorous field testing. The model achieves a mean average precision (mAP) of 83.6% and an F1-score of 98.3%, outperforming YOLOv5 (82.5% mAP) and Faster R-CNN (76.4% mAP), with 18.5% fewer misclassifications than standard RetinaNet. It retains 99% accuracy in low-light conditions and reduces labor costs by automating error-prone grading tasks. By publicly releasing the dataset and framework, this work advances precision agriculture standards, offering a transferable solution for ordinal maturity classification in perennial crops while supporting sustainable palm oil production through optimized harvesting decisions.