An Efficient Technique for Determining Tree Coordinates using LiDAR Data via Deep Learning

The forest is a complex ecosystem, consisting of a wide variety of plants, such as trees, shrubs, grass, small vegetation, and tree sprouts. The use of high-resolution LiDAR terrain surveying leads to difficulties in identifying forest components. This makes it challenging to detect and classify forest constituents. Therefore, the development of new accounting methods is an important task for more efficient study and management of forest resources. In this work, the authors described and formalized the main stages of tree trunk detection and fixation of coordinates. The main dependencies of data distribution were described, and equations were obtained, according to which one can predict in advance the accuracy of the system and the probability of tree detection. Optimal values for point filtering were considered and found, which allow obtaining the greatest information content of the data. Reliable detection of only the trees of interest was achieved using deep learning methods based on the PointNet++ architecture.