A COMPARISON OF TREE SEGMENTATION METHODS USING VERY HIGH DENSITY AIRBORNE LASER SCANNER DATA
- 1CIRGEO, Interdepartmental Research Center of Geomatics, University of Padua, Viale dell'Università 16, 35020 Legnaro, Italy
- 2TESAF Department, University of Padua, Viale dell'Università 16, 35020 Legnaro, Italy
- 3Department of Forestry and Renewable Forest Resources, Biotechnical Faculty, University of Ljubljana, Večna pot 83, 1000 Ljubljana, Slovenia
- 4Centre de recherche sur les matériaux renouvelables, Département des sciences du bois et de la forêt, Pavillon Gene-H.-Kruger, 2425 rue de la Terrasse, Université Laval, Québec, QC G1V 0A6, Canada
Keywords: Airborne laser scanner, LiDAR, tree segmentation, dense point cloud, forestry
Abstract. Developments of LiDAR technology are decreasing the unit cost per single point (e.g. single-photo counting). This brings to the possibility of future LiDAR datasets having very dense point clouds. In this work, we process a very dense point cloud (~200 points per square meter), using three different methods for segmenting single trees and extracting tree positions and other metrics of interest in forestry, such as tree height distribution and canopy area distribution. The three algorithms are tested at decreasing densities, up to a lowest density of ~5 point per square meter.
Accuracy assessment is done using Kappa, recall, precision and F-Score metrics comparing results with tree positions from groundtruth measurements in six ground plots where tree positions and heights were surveyed manually. Results show that one method provides better Kappa and recall accuracy results for all cases, and that different point densities, in the range used in this study, do not affect accuracy significantly. Processing time is also considered; the method with better accuracy is several times slower than the other two methods and increases exponentially with point density. Best performer gave Kappa = 0.7. The implications of metrics for determining the accuracy of results of point positions’ detection is reported. Motives for the different performances of the three methods is discussed and further research direction is proposed.