A hybrid method for segmenting individual trees from airborne lidar data

Abstract

Segmentation of individual trees from airborne lidar data uses either the point cloud directly or canopy height models (CHMs) derived from the point cloud. Point-based segmentation is able to detect understorey trees but is sensitive to the point density and often demands a high overhead cost of computing. Conversely, CHM-based segmentation can be easily implemented but it is impractical for the detection of understorey trees. To identify highly accurate treetops as well as understorey trees, this paper presents a hybrid method by modifying a CHM-based individual tree crown delineation (ITCD) algorithm and integrating it into a point-based algorithm. A multiscale local maxima (LM) algorithm is developed to improve the accuracy of LM obtained from CHMs in different spatial resolutions. The improved LM are used as seeds to segment the lidar point cloud into individual trees. For each tree, histogram analysis is applied to investigate the presence of understorey trees. Field measurements of tree heights and crown widths are used as ground truth to evaluate how well the proposed method is performing. The mean errors of tree heights and crown widths are 0.147 m and −0.004 m, respectively. The proposed method is also compared with five conventional methods of individual tree segmentation, namely ITCD, fixed window local maxima, Popescu and Wynne’s local maxima, variable area local maxima, and Li’s point-based segmentation. The comparison results indicate that the proposed hybrid method outperforms the conventional methods in terms of detection rate, omission error, commission error, mean absolute error of tree heights and root-mean-squared-error of tree heights.