Radiometric Features Enable Cross-Platform (TLS-to-ULS) Generalization for Forest Structure Assessment

Leaf-wood segmentation is a fundamental prerequisite for generating Quantitative Structure Models (QSMs) used in non-destructive biomass estimation. However, state-of-the-art segmentation models, typically trained on terrestrial laser scanning (TLS) data, often exclude radiometric features to ensure sensor-agnostic applicability. We challenge this design choice by investigating whether excluding radiometric data limits cross-platform generalization when transferring models from ground-based scans to remotely piloted aircraft (RPA) platforms. The RPA platform offers the ability to acquire data across much larger spatial extents relative to TLS data acquisition. 

We utilized a gradient-boosting framework to evaluate domain generalization, training on the public Heidelberg TLS dataset (European mixed forest; RIEGL VZ-400) and testing on a novel manually labeled RPA-LS dataset from a mixed deciduous forest in Southern Ontario, Canada. The testing data were acquired with a RIEGL Ultra120 at a density of approximately 10,000–12,000 pts/m². We compared a geometry-only model (utilizing 26 descriptors including eigenvalue features, verticality, and neighbor counts) against a radiometrically augmented variant (incorporating normalized intensity, return number and number of returns) and benchmarked these against established methods (LeWoS, ForestFormer3D, PointsToWood). 

Results indicate that geometry-only approaches fail to generalize to the aerial viewpoint, achieving F1 scores ≤ 0.56 and producing fragmented predictions. The inclusion of radiometric features increased the F1 score to 0.61 and more than doubled wood recall from 0.16 to 0.35. Crucially, the integration of radiometric data substantially enhanced structural coherence, reducing disconnections between stem and branch components observed in geometry-only predictions. 

Our results suggest that geometric descriptors are limited by their dependence on the scanner's viewpoint, while radiometric features rely on physical material properties that persist regardless of the sensor used. For operational forest inventory, leveraging these consistent radiometric signatures is essential for preserving the topological continuity required for downstream QSM reconstruction.