Location
MC 5479
Candidate
Roman Kaharlytskyi| Applied Mathematics, University of Waterloo
Title
The Role of Radiometric and Skeleton-Derived Features in Cross-Platform LiDAR Leaf-Wood Segmentation
Abstract
Leaf-wood segmentation, the per-point classification of tree LiDAR data into wood and leaf material, is a foundational preprocessing step for above-ground biomass estimation via quantitative structure models (QSMs). Existing methods have been developed and evaluated predominantly on terrestrial laser scanning (TLS) data, leaving cross-platform generalisation to aerial acquisitions largely unaddressed. This thesis investigates the contributions of geometric, radiometric, and skeleton-derived features to leaf-wood segmentation across two platforms: ground-based TLS (Heidelberg dataset, 11 trees, 1550 nm) and remotely piloted aircraft laser scanning (RPA-LS; 16 individually labelled deciduous trees, Ontario, Canada,905 nm), the latter constituting a novel benchmark introduced here.Controlled ablation experiments across three scenarios (within-platform RPA-LS, within- platform TLS, and cross-platform TLS→RPA-LS transfer) reveal a systematic inversion of feature family importance between platforms. On RPA-LS, radiometric features are the dominant contributor: three intensity-based features alone outperform 41 geometric features, and their addition to a geometric baseline increases Wood F1 by 0.190. On TLS,skeleton-derived features extracted from QSMs fitted to leaf-on point clouds are dominant,contributing approximately twice the Wood F1 improvement of radiometric features. Under TLS→RPA-LS transfer, radiometric features are necessary for competitive performance, while the reverse direction (RPA-LS→TLS) is actively harmed by their inclusion, an asymmetry explained by a sign inversion of the wood-leaf intensity contrast between 905 nm and 1550 nm, documented here directly in operational point cloud data.The best cross-platform model (CatBoost, F1 = 0.657) outperforms published baselines including LeWoS, ForestFormer3D, and PointsToWood by margins of 0.100–0.306 F1 on the RPA-LS evaluation set. Biomass analysis demonstrates that a 0.047 F1 difference between cross-platform models corresponds to a 24.6 percentage-point difference in biomass MAPE and a sign change in R2, establishing Wood Recall (not aggregate F1) as the diagnostically relevant metric for segmentation quality in downstream biomass applications.