Automatic tree reconstruction for CFD simulations
H.J. Gujda (TU Delft - Architecture and the Built Environment)
S.J. Deuten (TU Delft - Architecture and the Built Environment)
A.L. Scherrenburg (TU Delft - Architecture and the Built Environment)
C. Garcia Sanchez – Mentor (TU Delft - Architecture and the Built Environment)
T. Vargiemezis – Mentor (TU Delft - Architecture and the Built Environment)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Urban computational fluid dynamics (CFD) models used for pedestrian wind comfort assessment typically omit vegetation, despite trees having a measurable effect on wind flow. With the Dutch NEN8100 pedestrian wind comfort guideline under review, there is growing interest in automated, scalable methods for incorporating tree geometry into urban CFD workflows. However, reconstructing individual trees from airborne LiDAR remains challenging: point clouds are collected top-down and frequently contain incomplete or overlapping canopy returns, while existing segmentation and reconstruction pipelines have not been systematically evaluated for CFD compatibility or scalability.
This research extends an existing tree reconstruction pipeline (Tree4CFD) along three stages: individual tree segmentation, tree reconstruction, and scalability through an accessible interface. For segmentation, a hybrid approach combining a 2D canopy-height-model (CHM) coarse pass with 3D multi-directional analysis of under-segmented clusters was implemented and benchmarked against the original watershed-only method, using a sensitivity analysis and grid search to tune parameters across densely, moderately, and sparsely built-up areas. This is later combined with a variable window for finding tree peaks in the CHM. For reconstruction, LoD2 and LoD3 tree geometries were compared on computational efficiency, geometric validity (via val3dity), and mesh simplification using CGAL alpha wrapping. Finally, an interactive web interface was developed, allowing users to define an area of interest, select desired output parameters, and automatically trigger data retrieval, segmentation, and reconstruction, producing georeferenced OBJ geometries ready for integration into frameworks such as City4CFD.
Results indicate that cell size, smooth sigma, and minimum peak distance are the most influential segmentation parameters. The multi-directional refinement resulted in a minimal increase in F1-score (average of 1.68%), showing that improvement opportunities are possible but need further evaluation. Reconstruction time was found not to be a bottleneck, with negligible differences between LoD2 and LoD3 processing times, and CGAL alpha wrapping produced watertight, substantially simplified meshes suitable for downstream simulation. The developed interface demonstrates that the full pipeline, from AOI selection to downloadable 3D geometry, can be executed without manual intervention, supporting its practical use by stakeholders such as Haskoning.