Tree crown prediction of spruce tree using a machine-learning-based hybrid image processing method

This paper aims to estimate the crown of spruce trees using aerial top-view images captured by drones. In order to obtain real tree crown areas/diameter, we collected more than 2000 spruce trees in Regina, Saskatchewan, Canada, and each tree was labeled that corresponding to its location in the image. The measurements of crown were done using a Haglof Vertex V Hypsometer. This paper proposes aerial images by a machine-learning-based hybrid image processing approach, which combines YOLOv5 and Watershed Segmentation Technique (WST). Where, YOLOv5 is used for tree identification by a bounding box and WST is used to segment each tree crown based on the given bounding box. Meanwhile, we use Non-Maximum Suppression (NMS) to avoid the overlapping of bounding boxes. The aerial images were taken by a DJI Mavic 3, at altitudes of 50 m and 70 m above ground. The prediction accuracy is calculated via comparing with the real measured crown value. The results show that the developed model yielded better performance compared with other methods like YOLOv5 + Canny Edge Detection, with an accuracy of 89.1 % on tree crown area estimation.