A Hybrid Approach for Handling Occlusion in UAV Images of Fresh Oil Palm Fruit Bunches Based on Computer Vision

Abstract

This paper presents a method that integrates Loss Rank Mining (LRM) into the YOLOv11 object detection framework to enhance training efficiency by focusing on harder examples. A key issue in detecting Fresh Fruit Bunches (FFB) in UAV imagery is occlusion, caused by overlapping fronds, clustered bunches, and inconsistent viewing angles, which often leads to undetected or misclassified objects. By prioritizing instances with elevated loss values, the model is guided to learn from complex and informative cases, particularly those related to partial occlusions a prevalent challenge in oil palm plantations. We evaluated the proposed YOLOv11-LRM model against other YOLOv11 variants using standard object detection metrics, specifically focusing on mean Average Precision (mAP). Experimental results show that YOLOv11LRM outperformed the other models, achieving an mAP50 of 99.2 % and mAP50-95 of 85.6 %. These findings demonstrate that LRM is effective in improving object detection under challenging agricultural conditions, making it especially suitable for UAV-based FFB monitoring systems.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint