Real-time tree-crown detection in Christmas tree nurseries using YOLOv11 and edge computing

Abstract

Advanced technologies, including unmanned aerial vehicles, artificial intelligence, and agricultural robotics, are transforming agricultural systems toward greater sustainability and efficiency, enabling automated solutions for labor-intensive manual processes in horticulture. Accurate tree crown detection from UAV imagery is a key enabling step for autonomous screening processes in Christmas tree production. This paper investigates real-time tree-crown detection of transplanted Christmas trees (Abies nordmanniana) under standard production conditions, using object detection based on deep learning and UAV-acquired RGB imagery. Different size variants of the YOLOv11 deep learning architecture (n, s, m, and l) and input image sizes (320, 640, 960, and 1280 pixels) are evaluated on a custom dataset of 322 images containing 7,082 annotated instances, using a DJI Mini 3 Pro. Models are trained using a five-fold cross-validation with stratification at flight altitude level. To assess the feasibility of edge-based real-time processing, the models are deployed on three NVIDIA Jetson edge devices (Orin Nano, Orin NX, and AGX Orin) using FP32 and FP16 precision. The ANOVA results confirmed that the input image size, architecture size, and numeric precision have statistically significant effects on the detection performance measured by mAP50:95, while the choice of edge device had no significant influence. On the held-out test set, the three selected deployment configurations achieved mAP50:95 values between 87.00% and 90.97%, with YOLOv11l at 1280 pixels reaching the highest detection performance.