From Canopy to Ground via ForestGen3D: Learning Cross-Domain Generation of 3D Forest Structure from Aerial-to-Terrestrial LiDAR

TL;DR

ForestGen3D addresses the challenge of recovering sub-canopy and ground-level forest structure from aerial LiDAR by learning a cross-domain generative model conditioned on ALS inputs. It uses a conditional denoising diffusion probabilistic framework (DDPM) trained on co-registered ALS/TLS data to produce TLS-like 3D point clouds that align with ALS canopy geometry, enabling scalable, high-fidelity reconstructions across tree, plot, and landscape scales. The approach introduces the Expected Point Containment (EPC) as a practical proxy for generation quality when TLS ground truth is unavailable and demonstrates that ALS+ForestGen3D biometrics closely match TLS-derived distributions, with substantial improvements over ALS-only methods in DBH and crown volume estimations. At regional scales, ForestGen3D maintains spatial coherence with ALS geometry while enriching sub-canopy detail, supporting ecological analysis, fuel characterization, and wildfire modeling in ALS-dominant environments. The framework thus provides a practical, extensible solution for generating detailed 3D forest structure from readily available ALS data, with potential for broader ecosystem coverage through expanded training data and temporal conditioning.

Abstract

The 3D structure of living and non-living components in ecosystems plays a critical role in determining ecological processes and feedbacks from both natural and human-driven disturbances. Anticipating the effects of wildfire, drought, disease, or atmospheric deposition depends on accurate characterization of 3D vegetation structure, yet widespread measurement remains prohibitively expensive and often infeasible. We present ForestGen3D, a cross-domain generative framework that preserves aerial LiDAR (ALS) observed 3D forest structure while inferring missing sub-canopy detail. ForestGen3D is based on conditional denoising diffusion probabilistic models trained on co-registered ALS and terrestrial LiDAR (TLS) data. The model generates realistic TLS-like point clouds that remain spatially consistent with ALS geometry, enabling landscape-scalable reconstruction of full vertical forest structure. We evaluate ForestGen3D at tree, plot, and landscape scales using real-world data from mixed conifer ecosystems, and show through qualitative and quantitative geometric and distributional analyses that it produces high-fidelity reconstructions closely matching TLS reference data in terms of 3D structural similarity and downstream biophysical metrics, including tree height, DBH, crown diameter, and crown volume. We further introduce and demonstrate the expected point containment (EPC) metric which serves as a practical proxy for generation quality in settings where TLS ground truth is unavailable. Our results demonstrate that ForestGen3D enhances the utility of ALS only environments by inferring ecologically plausible sub-canopy structure while faithfully preserving the landscape heterogeneity encoded in ALS observations, thereby providing a richer 3D representation for ecological analysis, structural fuel characterization and related remote sensing applications.

Figures (18)

• Figure 1: Characteristics of ALS (in yellow) and TLS data at a surveying forest site, covering 2.5 km in diameter with 31 TLS scans randomly distributed (a color-coded blob per scan). Overlapping ALS in top-right shows a side-view point cloud with sparse/missing information below the canopy. Co-registered TLS in the bottom-right, on the other hand shows high detail along the tree-stand direction but is limited in the lat-long direction due to physical range limitations of TLS (as shown in the top-down images).
• Figure 2: ForestGen3D based on denoising diffusion probabilistic model (DDPM) formulation. Model iteratively denoises a sample guided by ALS measurements and generates a sample of forest tree 3D structure from the learned training distribution. Note that the 3D structural information provided by ALS is effective for constraining 3D structural generation in the vertical tree stand direction.
• Figure 3: 9 examples of the 3D training dataset consisting of 2900 examples of tree-segmented and co-registered aerial lidar scanning (ALS)/ terrestrial lidar scanning (TLS) trees. ALS trees in yellow in the top row, corresponding co-registered TLS trees in aqua color in the bottom row.
• Figure 5: Schematic of ALS-to-TLS cross-domain ForestGen3D. The forward process encodes a TLS sample into Gaussian noise, while the reverse process learns to denoise a sample conditioned on ALS input following a denoising diffusion probabilistic model (DDPM) formulation, reconstructing high-resolution 3D vegetation structure.
• Figure 6: Three examples of ForestGen3D generation at the tree-scale conditioned by ALS inputs in the training set. This qualitative examples showcase the architecture's capacity to encode and represent the 3D structural details of trees.

Theorems & Definitions (1)

• Remark : Containment Consistency under ELBO