TOPOLOGY-BASED INDIVIDUAL TREE MAPPING FROM LIDAR POINT CLOUDS

Abstract

Light Detection and Ranging (LiDAR) techniques have dramatically enhanced our ability to characterize forest structures remotely by acquiring 3D point cloud samplings of forest shapes. Extracting individual trees from the forests plays a critical role in the automated processing pipeline of forest point cloud analysis. However, there is still a lack of automated, efficient, and easy-to-use approaches available to identify and extract individual trees in a forest point cloud. This is mainly due to inconsistent point cloud quality, diverse forest structure, and complicated plant morphology. Most existing methods require intensive parameter tuning, time-consuming user interactions, and external information (i.e., allometric function). In this dissertation, we consider the problem of extracting single-tree point clouds from input forest point clouds. We propose two novel Topology-based Tree Segmentation (TTS) approaches, namely TTS-ALS and TTS-TLS, for airborne and terrestrial laser scanning data analysis, respectively. TTS algorithms are plug-and-play by nature and controlled by at most one parameter, ensuring user-friendliness. The implemented TTS software tools can extract single trees from 3D point clouds on various forest types, including conifer trees, broadleaf deciduous forests, and evergreen subtropical trees. Compared to state-of-the-art software tools, TTS tools achieve more accurate stem localization and tree extraction results on a broad set of forest types and point densities. Further experiments show that point normalization, one preprocessing step before TTS, slightly affects the TTS-ALS's performance of detected tree locations while strongly influencing tree crowns. Compared to TTS-ALS, TTS-TLS segmentation accuracy is more sensitive to normalized points. However, TTS-TLS can effectively limit errors introduced by the preprocessing step in local regions and maintain consistent results across entire areas. Because of their reliability and generality, TTS approaches are promising for ample usage of forest LiDAR point clouds in forestry and ecology studies, such as automated forest inventory generation from point clouds. Additionally, our extra research includes a novel building footprint delineation method for ALS point clouds and a comprehensive review of tree reconstruction methods tailored to single-tree point clouds, enhancing the breadth and depth of our contribution.