Extraction of individual tree parameters by combining terrestrial and UAV LiDAR

Abstract: Lightning Detection and Ranging (LiDAR) can be expected to measure the forest's spatial structure. It is probable to detect the position structure inside the vegetation. However, the information in the upper part of the forest tends to lose using the ground-based LiDAR alone, whereas, that in the middle and lower part of the forest using UAV LiDAR alone. It is a high demand to completely describe the vertical structure of the forest. In this study, the point clouds of terrestrial and UAV LiDAR were fused to extract the individual tree parameters, fully considering the different LiDAR data. The Comparative Shortest-Path (CSP) and the point cloud region growing algorithm were firstly used to identify the individual tree from the terrestrial and UAV LiDAR point clouds, respectively. The point cloud was then roughly matched, according to the tree position of terrestrial and UAV LiDAR, as well as the ground measured. The ground point cloud was transformed into the UAV point cloud coordinate system. Then, the Iterative Closest Point (ICP) algorithm was used to perform the fine matching of point clouds. The highest value and Density-based spatial clustering of applications with noise (DBSCAN) algorithm were used to extract the tree height and the Diameter at Breast Height (DBH). The extraction accuracy of individual tree parameters was evaluated in the terrestrial, UAV, and fusion point clouds. The results showed that the consistent detection rate of individual trees was achieved using the terrestrial and fusion point clouds. The detection rates of individual trees in easy, medium, and difficult plots were 98%, 94%, and 91%, respectively. There was a low detection rate of individual trees using the UAV point cloud, where the overall detection rate was 42%, and the highest detection rate was only 58% in the three types of plots. There was all the same in the DBH extraction accuracy using the terrestrial and fusion point cloud. Specifically, the R2 values of easy, medium, and difficult plots were around 0.98, 0.97, and 0.96, respectively, and the Root Mean Square Error (RMSE) were 1.20, 1.46, and 1.50 cm, respectively. The highest accuracy was achieved in the tree height extraction using a fusion point cloud, where the R2 values of easy, medium, and difficult plots were 0.98, 0.94, and 0.73, respectively, and the RMSE was between 1.38 and 4.19 m. The fusion point cloud greatly improved the extraction accuracy of tree height in the medium plots, where the RMSE was reduced by 0.34 m, compared with the terrestrial point cloud, and the R2 value increased by 3%. Nevertheless, there was a small improvement for the easy and difficult plots. Among them, the accuracy of parameter extraction was higher for the Cunninghamia lanceolata than that for the eucalyptus, where the R2 values of DBH were 0.99 and 0.96, respectively, and the RMSEs were 1.35 and 1.49 cm, respectively. The R2 values of tree height were 0.89 and 0.73, and the RMSEs were 1.96 and 3.47 m, respectively. The extraction accuracy of tree parameters decreased gradually, as the plot was much more complex. In general, the fusion point cloud of terrestrial and UAV LiDAR can be applied to more precisely measure the forest spatial structure, in order to improve the parameter extraction accuracy of the easy and medium plots for the better application of forest resources.