Abstract
<title>Abstract</title> <p>Harvesting decisions for oil palm fresh fruit bunches (FFBs) are difficult in natural plantations because candidate regions are affected by grass attachment, trunk fibers, fallen fruits, occlusion, shadows, and complex backgrounds. Conventional maturity classifiers usually assume that each detected candidate is a valid FFB, which can lead to unsafe decisions when false candidates or heavily occluded bunches are present. This study proposes a lightweight two-stage harvesting decision support framework. Stage 1 uses YOLO11s-seg to generate instance-level FFB candidate masks. Stage 2 uses YOLOv8n-cls to classify each candidate as non_FFB, occluded_FFB, unripe, underripe, or ripe, and maps the outputs to four image-level decisions: Harvest available, Review, Wait, and No target. On a grouped test set of 99 samples, the five-class model achieved 90.91% classification accuracy and 93.94% harvesting decision accuracy. It also achieved 100.00% Harvest precision, 93.75% Harvest recall, 96.77% Harvest F1-score, 0.00% false harvest rate, and 6.25% missed harvest rate. The results show that explicit hard-negative rejection and occlusion-aware assessment can improve the safety and interpretability of FFB harvesting decision support.</p>