Nguyen Thanh Tuan, Phan Van Tuan, Nguyen Thanh Hung, Duong Tien Duc
Individual tree detection and crown segmentation for Pinus kesiya using UAV–LiDAR and deep learning models
Dendrobiology 2026, vol. 95: 156-169
https://doi.org/10.12657/denbio.095.011
Abstract:
Unmanned aerial vehicle (UAV) is increasingly used for individual tree crown detection and segmentation; however, performance remains limited in structurally heterogeneous stands due to crown overlap and variable crown geometry. This study evaluates and compares three advanced deep learning approaches for individual tree crown detection and segmentation from orthophoto RGB imagery in Pinus kesiya plantations, including YOLOv8m-Seg, YOLOv11m-Seg, and an integrated YOLOv12m–SAM2 framework. The optimal model is subsequently selected to derive individual tree attributes by integrating LiDAR data and orthophoto RGB imagery. The results revealed significant differences in performance among the evaluated models. YOLOv11m-Seg framework achieved the highest detection performance, with mAP50 = 0.62 (0.59–0.64), mAP50–95 = 0.290 (0.287–0.294), and an F1-score of 0.61 (0.58–0.63), indicating accurate localization of tree crowns. For the segmentation task, the model achieved F1-score = 0.59 (0.56–0.61), mAP50 = 0.60 (0.58–0.62), and mAP50–95 = 0.26 (0.25–0.27), outperforming YOLOv12m + SAM2 and YOLOv8m-Seg at the IoU > 0.5 threshold. Finally, the accuracy of individual tree attribute extraction from LiDAR data and RGB imagery using YOLOv11m-Seg was validated through comparison with ground measurements. The results showed a higher agreement between reference data and UAV–LiDAR data for tree height than for crown diameter with R² values ranging from 0.85 to 0.93 for tree height and from 0.58 to 0.86 for tree crown diameter. Overall, the findings suggest that YOLOv11m-Seg provides a reliable approach for extracting individual tree attributes from UAV–LiDAR data in Pinus kesiya plantations.
Keywords: drone, forest inventory, Segment Anything Model (SAM), You Only Look Once (YOLO)