Automated Segmentation of Leaves From Deciduous Trees in Terrestrial Laser Scanning Point Clouds

Abstract

Recent improvements in topographic LiDAR technology allow the detailed characterization of individual trees at both branch and leaf scale, providing more accurate information to support phenological and ecological research. However, an effective methodology to map single leaves in 3-D is still missing. This letter presents a point cloud segmentation approach for single leaf detection and the derivation of selected morphological features (i.e., leaf area (LA), maximal leaf length, width, and slope) using terrestrial laser scanning. The developed approach consists of 1) filtering noise points; 2) region growing segmentation; 3) separating leaf and nonleaf segments; and 4) calculating leaf-morphological features. For the evaluation of the workflow, two deciduous trees were scanned. A Selection of leaves of the specified trees was randomly harvested during the field campaign for comparison. A qualitative comparison analysis was carried out between the area of the harvested leaves and the leaf area (LA) derived from 3-D point cloud segmentation. In addition, a sensitivity analysis investigated the effect of the segmentation parameterization. This step revealed that the proposed segmentation algorithm is robust when using an optimum subset of parameter values. However, the determination of leaf outlines is limited due to the orientation of leaves to the scanner, shadow effects, and the inhomogeneity of the point cloud. The results underline the potential of region growing segmentation of point clouds for providing accurate information on single leaves and vegetation structure in more detail. This facilitates improvements in applications such as estimating water balance, biomass, or leaf area (LA) index.