Effects of imaging angle and field of view on detection and tracking of bareroot loblolly pine seedlings using computer vision

Abstract

An efficient, accurate inventory of bareroot loblolly pine seedlings is critical for effective nursery management and large-scale reforestation efforts. Traditional sampling-based counting can be labor-intensive, time-consuming, and prone to errors, highlighting the need for an automated solution. This study developed an automated spring inventory system using a tracking-by-detection framework to detect, track, and count loblolly pine seedlings in RGB imagery. You Only Look Once (YOLO): YOLOv8, YOLOv9, and YOLOv10 detection models were employed and trained on 480 images captured at a commercial forest nursery and evaluated on eight unique videos (480 frames total). The YOLOv10-balanced model achieved a high mean Average Precision (mAP) of 95.5%, while the BoT-SORT tracking algorithm reached a Multi-Object Tracking Accuracy (MOTA) of 85.33%. Results revealed that seedlings closer to the image periphery exhibited lower detection accuracy, underscoring the impact of imaging angles and field of view. Detection accuracy peaked at a nadir (top-down) perspective and declined as the camera’s horizontal and vertical viewing angles became more oblique. Increasing the vertical field of view improved counting accuracy up to a certain threshold, after which it plateaued, whereas a wider horizontal angle negatively influenced accuracy due to factors such as overlapping and occlusion. These findings provide practical insights into optimizing camera placement and the number of cameras for accurate automated seedling counts during spring inventory in forest nurseries.