Deep Learning in Forestry Applications: A Comparative Analysis of YOLOv8 and YOLOv10 for Individual Tree Detection from UAV Images

Creative Commons License

Şevik U.

Düzce Üniversitesi Orman Fakültesi Ormancılık Dergisi, cilt.22, sa.1, ss.62-75, 2026 (Hakemli Dergi)

Özet

Accurate and efficient monitoring of forest ecosystems at the individual-tree level is essential for assessing forest dynamics and biodiversity, and for estimating carbon stocks, particularly under changing climate conditions. Using RGB images readily obtained with Unmanned Aerial Vehicles (UAVs), cost-effective solutions for large-scale Individual Tree Crown (ITC) detection with Deep Learning (DL) approaches can be developed. This study presents a comparative analysis of two new versions of the You Only Look Once (YOLO) algorithm, YOLOv8n and YOLOv10n, for ITC detection in temperate forest environments using a publicly available high-resolution UAV RGB dataset obtained from Roboflow. We used a transfer-learning approach, fine-tuning pre-trained models on a dataset relabeled with a single 'Tree' class. The models were trained and evaluated using a T4 GPU in Google Colab; performance was assessed using standard object detection metrics, including mean Average Precision (mAP50, mAP50-95), precision, and recall. The results show that both models achieved high performance, effectively detecting individual tree crowns. YOLOv8n performed slightly superior with 0.913 mAP50 and 0.588 mAP50-95, compared to YOLOv10n's 0.902 mAP50 and 0.584 mAP50-95. Furthermore, YOLOv8n demonstrated significantly faster training times (2650.6 seconds versus 3322.5 seconds for YOLOv10n under the same conditions). Both models reached high F1-scores at their optimal confidence thresholds (0.85 for YOLOv8n and 0.84 for YOLOv10n), and inference benchmarking on an NVIDIA A100 GPU confirmed real-time speeds, with YOLOv8n reaching 87.0 FPS compared to 71.6 FPS for YOLOv10n. A canopy-density-stratified analysis further showed that detection performance was highest in medium- and high-density canopies (F1 up to 0.79) and lowest in sparse canopies. Qualitative analysis of true positive (TP), false positive (FP), and false negative (FN) predictions provides insights into model behavior in complex crown structures. This study demonstrates the high potential of current YOLO architectures for efficient and accurate ITC detection from UAV RGB imagery. It highlights the suitable balance between accuracy and computational efficiency of YOLOv8n for practical forestry applications.