Palm Fruit Bunch Ripeness Detection and Maturity Time Prediction

Reliable, plantation-scale estimation of palm fruit bunch(PFB) ripeness and harvest timing is restricted by labour-intensive manual checks, while the indoor-trained detectors underperformed outdoors. We built a 3,018-image unharvested PFB dataset plus 71-day time-series sequences. The proposed OcclusionCut reached 89.8 % mAP for YOLOv12s while remaining lightweight, aiming for edge device deployment. The proposed hybrid color-correction method lifted mAP for overall and each ripeness class at different weather conditions. Finally, a regressive Mixture-of-Experts with NAS-tuned experts predicted days-to-maturity with 0.96 R², enabling proactive harvest planning. The integration of detection and forecast advances autonomous, field-ready management of Malaysian oil-palm plantations.