Silvilaser 2019 - Poster Presentations »
Tree detection and height measurement from aerial laser scanning point cloud
Forest stand variables such as number of trees, height and diameter are important for forest management. Traditionally, these variables are estimated by forest inventory plots which can be time consuming and costly. LiDAR aerial data have been an efficient technique to measurement tree heights and detect individual trees. The most of conventional methods convert LiDAR data point to Digital Elevation Model (DEM) and Digital Surface Model (DSM) to obtain the Canopy Height Model (CHM). In this work, normalize point cloud was applied to detect the individual trees and measuring their heights. The experimental area is a Eucalyptus urograndis plantation located in Telêmaco Borba, Paraná State, Brazil. The stand age is five years old and the initial stand density was 1.111 trees ha-1. Four plots (512 m2 each) were selected to measurement of trees height and obtaining the numbers of trees. The LiDAR data were collected with a Trimble® Harrier 68i sensor mounted in a Cessna 206, flying height of 666.17 meters above sea level and density of five points per square meters. The FUSION 3.70 software was used to generate a normalized point cloud and the tree_detection function in the package lidR (2.0.3 version) in R software was used to detect trees and heights. This function implements an algorithm based on a local maximum filter. The arguments of the function were: the window size (numeric or a function), the minimum height of a tree and the shape of window (square or circular). In this study the window size was coincident with the average crown area (empirical) and the tree heights minimum was based on forest inventory data. The accuracy of number of trees was evaluated in terms of recall, precision and F-score. The statistical significance of the mean differences between the tree height observed in ground (plots) and the LiDAR height was evaluated by t-test. The algorithm detected correctly 91% of trees (recall), the value of precision was 97% and the F-score was 94%. The t-test (p-value = 0.37) indicated there was no statistical difference between the tree heights. The bias of height tree was overestimated in 0.34 m (RMSE = 3.5%, R2 = 0.85). The method proposed can be used to support the forest inventory planning with high accuracy.