Tree crown segmentation and estimation of metrics from pointclouds with improved local maximum method

On the Earth's surface, spatial information and volume of natural and human-made elements are importance for countries' governance policies in terms of economic, ecological, and sociocultural aspects. 3D spatial information is needed for control and analysis of these elements. LiDAR and photogrammetric point clouds are used in many applications such as urban modeling, disaster management, and monitoring forested areas. In recent years, to prevent problems like climate change and global warming, key role of trees needs to be preserved and their current status needs to be recorded. However, since GPS cannot receive a continuous signal, especially in densely forested areas, it is time-consuming and expensive to do field studies with terrestrial measurement techniques in these regions. In accordance with forestry principles adopted by countries' it is of great importance for sustainable development to process tree inventory information with minimal manpower. In this thesis study, a novel approach based on a two-stage developed local maxima method for individual tree crown segmentation from point clouds and the estimation of four different tree metrics (tree height, crown width, breast diameter, crown base height) is proposed. The study utilizes aerial LiDAR and/or photogrammetric point clouds of forested areas covered with Stone Pine (Pinus pinea) at the Mediterranean University, Larch (Picea orientalis) in Maçka, and Oak (Quercus alba- L.) in Geneva as the dataset. The proposed workflow consists of detecting individual trees using the existing Canopy Height Model in the MATLAB software and then providing an alternative solution by improving the erroneous tree clusters that occur in adjacent trees using a two-plane approach supported by Bézier Curves. Existing and improved allometric equations were used for the metric estimation of the obtained tree clusters. According to the object-based accuracy analysis, the segmentation accuracy was calculated as an average of 79%, and the positional error in tree centers was ±0.785 m.