Using machine learning methods to estimate spruce tree crown and DBH from aerial imagery

Abstract

Spruce trees play a vital role in Canada’s forest ecosystems, which is widely used in construction, paper production, and other industries. However, spruce trees are particularly susceptible to wildfires, which pose a major risk to both natural landscapes and human settlements. Therefore, to evaluate the spruce forest biomass volume is an important step to estimate its yield and combustibility. This paper aims to use Machine Learning (ML) approaches to estimate the biomass volume of spruce trees from aerial top-view images. Since the aerial images are only show the tree crown shapes, we set up the relationship between tree crown diameter (TCD) and Diameter of tree Breast Height (DBH), and this DBH can be further used to estimate the tree biomass volume. Here, Spruce trees top-view images were taken by a DJI Mavic 3, at an altitude of 50m above ground. We measured the actual TCD and DBH in the field. 2,155 spruce trees were labelled in our dataset according to its location in the aerial images. The actual TCD values of 2155 samples were measured with a Hypsometer device that uses ultrasound at the extremities of the tree branches, which will be further used to compare and calculate the accuracy of the TCD values that are measured from top-view images after our model training. After experimenting with tree detection methods, we conclude that YOLO performed better than MaskRCNN by 4%. And then we proposed two methods that use YOLOs: First method, a combination of YOLOv5 bounding box to identify the trees and watershed technique to segment tree crowns from aerial images. Compared to the second method YOLOv11 that uses instance segmentation to segment the trees. A study is conducted to showcase a relationship between TCD and DBH of the field measurements. This linear relationship can be used to estimate DBH out of TCD and then could futher calculate tree biomass volume with the estimated DBH.