Publication: Fine-tuning and comparative analysis Of yolo and transformer models for UAV-based palm oil tree detection.
Loading...
Date
2026-07
Authors
Mohammad Farhan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Oil palm plantation monitoring requires accurate tree detection for yield estimation, resource allocation, and sustainability compliance, yet manual counting methods remain labour-intensive and error-prone at scale. Despite growing adoption of UAV-based deep learning detection, no study has systematically compared the latest CNN-based and transformer-based paradigms under identical conditions for oil palm tree detection. This study fine-tunes and comparatively evaluates YOLOv11 (CNN-based) and D-FINE (transformer-based) using 6,080 high-resolution UAV images containing 111,810 annotated trees from Penang, Malaysia, across standardised GPU and CPU platforms. YOLOv11 achieved higher precision (0.953), mAP@0.5 (0.988), and faster inference (20.61 FPS on GPU) with a compact 109.1 MB footprint, favouring real time and resource-constrained deployments. D-FINE achieved higher recall (0.977) and superior robustness under adverse conditions: in compound scenarios combining dense canopy layouts with motion blur, YOLOv11 suffered a 34% reduction in
detections while D-FINE maintained performance with only a 5% reduction. These findings establish that YOLOv11 is preferable for real-time monitoring under favourable imaging conditions, while D-FINE better serves scenarios demanding detection consistency under dense canopy overlap and motion-degraded imagery, providing empirical guidance for architecture selection in UAV-based oil palm plantation monitoring.