Segmentation of Individual Trees From TLS and MLS Data

Abstract

Terrestrial laser scanning (TLS) and mobile laser scanning (MLS) data can be used to obtain abundant and precise side information of trees. Therefore, it can enable extracting individual tree parameters, such as the tree height, crown size, crown base height, and diameter at breast height, and it can provide basic data for forest research and management. This study proposes a technical framework for segmenting individual trees from TLS and MLS data. This framework contains six steps: 1) data preprocessing, 2) octree construction, 3) spatial clustering, 4) stem detection, 5) initial segmentation, and 6) overlapped canopy segmentation. This framework makes two main contributions: 1) a top-down hierarchical segmentation approach, including connectivity-based spatial clustering (regional scale), stem-based initial segmentation (individual tree scale), and fine segmentation of overlapped canopy (canopy scale), is proposed to reduce technical difficulties and improve process efficiency; and 2) a modified node similarity calculation for normalized cut method aiming at segmenting overlapped canopy, which can effectively separate neighboring trees even if their canopies are overlapped, is proposed. The proposed framework was tested on a leaves-off terrestrial LiDAR dataset and a leaves-on mobile LiDAR dataset. For terrestrial LiDAR data, our framework achieved completeness of 92.4%, correctness of 95.4%, and F-score of 0.94. For mobile LiDAR data, the corresponding values were 94.0%, 93.7%, and 0.94.

Constrained Spectral Clustering of Individual Trees in Dense Forest Using Terrestrial Laser Scanning Data

“…However, Raumonen et al [10] did not provide accuracy metrics for the segmentation and used single trees segmentation in simple forest stands as pre-clustering stage for subsequent tree modelling. The good detection rates of Zhong et al [30] of more than 90% are difficult to compare and interpret, since the authors sparsely described forest characteristics of their test sites and omitted a description of the reference data. Since they explained that their algorithm does not work in vertically structured forests and because they partly used MLS data close to urban locations, it can be assumed that their accuracies related to single layered and relatively simple forest structure or even urban trees.…”

“…Raumonen et al [10] and Zhong et al [30] are among the only authors. However, Raumonen et al [10] did not provide accuracy metrics for the segmentation and used single trees segmentation in simple forest stands as pre-clustering stage for subsequent tree modelling.…”

“…While some initial studies already worked with partly overlapping tree crowns, the overall datasets still appeared to be characterized by well defined mature trees and with little or no understory [10,20,29]. Zhong et al [30] significantly improved the separation of overlapping tree crowns in the upper canopy layer, but could not distinguish individual trees in vertically layered forest structures. Zou et al [31] did not extract the complete trees, but only point cloud segments within a defined radius around the stem position that were intended to be used for species classification.…”

“…Hu et al [11] and Lee et al [37] likewise applied it on airborne data. For TLS data, this technique was only used by Zhong et al [30], but their comparatively more simple approach proved to be not capable to separate trees in vertically structured forests. Integrating spectral clustering into a processing framework which manages to deal with complex forest structures captured with high point densities and the different perspective of TLS data is a new challenge of this study.…”

An Automated Hierarchical Approach for Three-Dimensional Segmentation of Single Trees Using UAV LiDAR Data

“…To delineate small trees under canopies from airborne LiDAR data, Reitberger et al [30] introduced normalized cut (NCut), originally proposed for 2D image segmentation, and achieved a recognition rate of 12%, which is higher than conventional watershed segmentation methods. Zhong et al [31] improved the algorithm to segment overlapped trees with a correctness of above 90% by means of terrestrial and mobile LiDAR data. NCut has shown great potential in 3D segmentation and works very well when separating two overlapping objects [32,33].…”