关键词: 激光雷达, 单木分割, 分水岭算法, 冠层高度模型, 数字表面模型

Abstract:

Objective: Considering that the unmanned aerial vehicle(UAV) LiDAR-derived canopy height model (CHM) is prone to distortion in areas with complex terrain, which significantly limits the accuracy of the individual tree segmentation, this study aims to propose a new method that utilizes the combination of CHM and digital surface model (DSM) to segment individual tree. Method: By using UAV-LiDAR data and choosing 3 standard plots of middle-aged forest and young forest in steep areas of Cunninghamia lanceolata plantation with middle and high crown density in Yangkou forest farm, Shunchang country, Fujian Province, this paper used the field-measured and visual interpretation data to calculate and analyze the accuracy of treetop detection and individual tree segmentation which were acquired by using local maximum algorithm with 4 fixed window size and extremum controlling watershed algorithm based on the combination of CHM and DSM and compared them with the traditional method of using CHM alone. Result: In the matter of treetop detection, number of detecting treetops and detection percentage for each plot showed a downward trend with window size increasing; The optimal window size of 3 middle-aged forest plots is 0.3 m and the optimal window size of 3 young forest plots is 0.2 m. In this configuration, the individual tree of one-to-one correspondence and the producer’s accuracy are the maxima. By using the CHM alone, the detection percentage are higher but the accuracy of treetop detection are lower than using the combination of CHM and DSM because the CHM-based local maximum method is prone to multiple detecting treetops in the individual tree on the optimal window size. The accuracy of treetop detection in the young forest plots are higher than in the middle-aged forest plots. The reason is that the crown breadth and adjacent distance of individual trees in the young forest plots are more consistent which is more adapted to the local maximum algorithm. In the matter of individual tree segmentation, on the optimal window size, by using the combination of CHM and DSM, the accuracy of individual tree segmentation are higher than using the CHM alone and the accuracy of individual tree segmentation in the young forest plots are higher than the middle-aged forest plots. Conclusion: Because the direct data source of treetop detection and individual tree segmentation is DSM which reflects the crown surface veritably with no distortion, the result represents the real individual tree crown boundary and reaches the high accuracy in the young and middle ages of Cunninghamia lanceolata plantation with middle and high crown density (the detection rate r is more than 88%, the accuracy p is more than 92% and the F-score is more than 91% for each plot). Moreover, in this study, the method integrates into the ArcGIS Model Builder which is helpful to build an accurate, automatic, and integrated platform for individual tree segmentation of UAV-LiDAR.

Key words: light detection and ranging (LiDAR), individual tree segmentation, watershed algorithm, canopy height model (CHM), digital surface model (DSM)