Tree Detection using Genus-Specific RetinaNet from Orthophoto for Segmentation Access of Airborne LiDAR Data

Abstract

Understanding rainforest habitats are made difficult by accessibility, the density of foliage, and coverage; yet it is vital for ecologist and conservationists to understand the variety and diversity of these ecosystems. An airborne light detection and ranging (LiDAR) technology has carried a breakthrough in the acquisition and management of forest data specifically on the aspect of forest inventory. LiDAR can accurately measure tree-level features such as tree elevation, crown dimension as well as derivative estimates such as trunk diameter (diameter at breast height - DBH). In order to do this, an accurate tree segmentation approaches within LiDAR data are required. Tree segmentation within LiDAR data always starts with by locating treetops through local maxima (LM). Wide-ranging efforts has been used to segment individual trees from LiDAR data by starting to localize treetops through LM within LiDAR data. In this study, a demonstration of a pipeline for new tree segmentation within LiDAR Point Cloud by integrating a new approach of detecting trees using deep learning-based object detection. Tree detection has been done using RetinaNet with a mean average precision (mAP) score of 90.73 % with a classification loss of 0.7732 and a regression loss of 0.1322. Tree segmentation within the LiDAR point cloud has been effective and possible using the coordinates of the bounding-box produced by the tree-detection.