Toward Structural Understanding of Complex Forest Environments using Terrestrial point clouds

Terrestrial laser scanning (TLS) captures forest structure at centimeter resolution, producing dense 3D point clouds that can inform fine-scale inventory and ecological assessment. This project develops an end-to-end pipeline that converts raw TLS data into actionable structural information. Our field sites span contrasting biomesÑHarvard Forest (Massachusetts) and mangrove roots in PalauÑproviding diverse canopy architectures and acquisition conditions.

The pipeline integrates: (1) preprocessing and registration, (2) semantic segmentation of tree components (stems, branches, foliage, ground), (3) individual-tree delineation, (4) 3D tree-structure reconstruction, and (5) quantitative analysis (e.g., diameter at breast height, tree height, crown volume, and branch architecture). We combine established geometric/graph methods with state-of-the-art deep learning to balance accuracy, interpretability, and scalability for operational forest management.

We curated a >30 million-point TLS mangrove dataset (Josie Clapp, Fei Zhang) for semantic segmentation; built a spherical-projection, feature-enriched, uncertainty-aware semi-automated annotation pipeline; and achieved up to 89% overall accuracy on 2D spherical feature maps, with a 3D PointNet++ baseline at 87%. Aisha and Fei implemented TreeISO (graph-based individual-tree segmentation) and AdTree (tree-mesh reconstruction via skeletonization and branch fitting) on Harvard Forest data.

In MarÐApr 2024, Jan van Aardt and Rob Chancia led a USFS-supported field campaign in Palau, collecting SET measurements at 17 sites (pin, TLS, and pilot iPhone-LiDAR), training local partners for ongoing monitoring, and conducting community seminars/demos that generated strong engagement and media coverage (Island Times cover, Apr 5, 2024).

Fei presented project highlights at the 2025 Finger Lakes Science & Technology Showcase (Rochester, NY, April 24, 2025; poster link: https://github.com/fz-rit/TLS-showcase/blob/main/Fei_mangrove_roots_pos…).

Figure 1: Semantic segmentation of a Harvard Forest plot.

Figure 2: Semantic segmentation of a mangrove stand in Palau.

Figure 3: Spherical projection of a mangrove stand showing a PCA composite of multiple feature channels: intensity, range, height (Z), curvature, anisotropy, planarity, and normal-based pseudo-coloring.

Figure 4: More Spherical projection of a mangrove stand showing a PCA composite of multiple feature channels: intensity, range, height (Z), curvature, anisotropy, planarity, and normal-based pseudo-coloring.

Figure 5: Virtual-sphere visualization of TLS scan coverage, colored by the PCA composite in Figure 3.

Figure 6: Individual-tree segmentation in a Harvard Forest plot using TreeISO.

Figure 7: Single-tree mesh reconstruction from TLS point clouds using AdTree.

Reference

[1] Chancia, R. O., van Aardt, J. A., and MacKenzie, R. A. Enhanced surface elevation change assessment in mangrove forests using a lightweight, low-cost, and rapid-scanning terrestrial lidar. AGU23 (2023).
[2] Zhang, F. Tls-showcase. https://github.com/fz-rit/TLS-showcase, 2025. Accessed: 2025-08-01.