ForestGen3D: 3D Forest Structure Generation

• ForestGen3D is a generative modeling framework that reconstructs dense 3D forest structures, including sub-canopy and understory details, from sparse aerial LiDAR data.
• It employs a conditional denoising diffusion probabilistic model with PointNet++ layers and feature modulation to synthesize TLS-like point clouds from ALS observations.
• A novel geometric containment prior based on the ALS convex hull ensures spatial consistency, as demonstrated by improved Chamfer Distance, EMD, and EPC scores.

ForestGen3D is a generative modeling framework for reconstructing terrestrial-scale 3D forest structure from aerial LiDAR alone. Introduced in “From Canopy to Ground via ForestGen3D: Learning Cross-Domain Generation of 3D Forest Structure from Aerial-to-Terrestrial LiDAR” (Castorena et al., 19 Sep 2025), it learns to synthesize TLS-like point clouds conditioned on ALS observations by means of a conditional denoising diffusion probabilistic model trained on co-registered ALS/TLS pairs. Its stated purpose is to recover sub-canopy, stem, understory, and near-ground vegetation that ALS sparsely samples or misses entirely, while maintaining spatial consistency through a geometric containment prior based on the convex hull of ALS observations.

1. Problem formulation and scope

ForestGen3D addresses a cross-domain generation problem in forest remote sensing: the input domain is aerial laser scanning, which provides broad landscape coverage and strong canopy observations, while the target domain is terrestrial laser scanning, which provides rich local detail of stems, understory, branches, internal canopy structure, and ground-adjacent vegetation (Castorena et al., 19 Sep 2025). The motivating asymmetry is operational as much as geometric. ALS is scalable but top-down and occlusion-limited; TLS is structurally informative but expensive, slow, and spatially constrained.

The framework is designed around the claim that many ecological and operational tasks depend on the full vertical organization of vegetation rather than canopy-top summaries alone. The paper explicitly links this to wildfire simulation, structural fuel characterization, and ecological modeling. In that sense, ForestGen3D is neither a procedural visualization engine nor a classical forest simulator based on simplified cylinders or field-summary statistics. It is a sensor-conditioned generative model whose output is intended to approximate the conditional distribution of detailed terrestrial forest structure given sparse aerial observations.

A notable aspect of the formulation is its multi-scale deployment strategy. The model is trained at tree scale, where the containment prior is empirically most defensible, but is then applied compositionally at plot and landscape scales through detection, local generation, and aggregation. The basic modeling unit is not always a single botanical tree: the detector may yield a tightly packed vegetation group or small cluster when neighboring crowns cannot be cleanly separated. This suggests that ForestGen3D is best understood as operating on localized vegetation objects extracted from ALS, rather than on globally coherent forest scenes in a single generative pass.

2. Data, sensing domains, and preprocessing

The input ALS point cloud is denoted

$\mathbf{Y} \in \mathbb{R}^{M \times 3},$

and the target TLS-like point cloud is denoted

$\mathbf{X} \in \mathbb{R}^{N \times 3}.$

Training uses co-registered ALS/TLS pairs from the CoLiDAR-Forest3D (ALS/TLS) dataset (Castorena et al., 19 Sep 2025). The reported corpus contains 2,900 co-registered ALS/TLS tree scans for training, 300 additional non-overlapping tree samples for validation, and 50 TLS scans from geographically separate regions for testing, yielding 1,457 test trees after detection. The study area is a mixed conifer forest at Fort Stewart, USA, including longleaf pine, slash pine, loblolly pine, turkey oak, and understory species such as wiregrass, gallberry, and saw palmetto.

The ALS platform is a Galaxy T2000 operating at 1064 nm, with up to 8 returns per pulse, a 28° field of view, flight altitude of about 1675 m, density of about 15 points/m$^2$, and 0.03–0.25 m RMSE over operational range. The TLS platform is a Leica BLK360 operating at 830 nm with 300° vertical and 360° horizontal field of view. The TLS campaign includes 240 scans over a region about 2.5 km in diameter, each with about 8 million points, of which about 4.6 million are ground points. Because most TLS returns lie within 25 m, points beyond 25 m are removed to reduce density nonuniformity.

Co-registration proceeds from initial GPS placement of TLS scans into global coordinates, followed by extraction of a 50 m radius ALS subset around each TLS scan center and automatic alignment using prior work, with both translation and rotation refinement. Pairs are then manually checked by an expert. Tree extraction is performed by a deep-learning detector operating from a bird’s-eye view, which produces 3D bounding boxes spanning from the highest vegetation point to the lowest ground point.

Preprocessing standardizes each tree cloud by centering and normalizing it to

$[0,1]^3.$

During training, each ALS/TLS pair is uniformly subsampled to

$N = 2048$

points, with batch size 16. This fixed-size subsampling functions simultaneously as a computational constraint and as a stochastic augmentation mechanism, since different point subsets are exposed across epochs.

3. Conditional diffusion architecture

ForestGen3D models the conditional distribution

$p(\mathbf{X}\mid \mathbf{Y}),$

where $\mathbf{Y}$ is the ALS observation and $\mathbf{X}$ is the TLS-like forest structure to be generated (Castorena et al., 19 Sep 2025). The architecture is described as an encoder-decoder U-Net adapted to point clouds, using PointNet++ layers as its basic hierarchical operators. ALS conditioning is injected through feature-wise modulation blocks at each hierarchy level so that aerial structure influences denoising at every scale.

The clean TLS point cloud is written as

$\mathbf{X}^0 \in \mathbb{R}^{N \times 3},$

and the forward diffusion trajectory is

$$\mathbf{X}^0 \rightarrow \mathbf{X}^1 \rightarrow \cdots \rightarrow \mathbf{X}^T \sim \mathcal{N}(0,\mathbf{I}).$$

The paper states that the forward process progressively corrupts the sample over $\mathbf{X} \in \mathbb{R}^{N \times 3}.$0 steps by multiplying with a decaying factor and adding Gaussian noise. The reverse process learns to approximate

$\mathbf{X} \in \mathbb{R}^{N \times 3}.$1

thereby denoising the latent variable back to a plausible TLS point cloud. Training is described as ELBO minimization following DDPM theory.

The reported diffusion schedules are ablated over

$\mathbf{X} \in \mathbb{R}^{N \times 3}.$2

with noise schedule endpoints

$\mathbf{X} \in \mathbb{R}^{N \times 3}.$3

At inference, the model begins from Gaussian noise,

$\mathbf{X} \in \mathbb{R}^{N \times 3}.$4

and iteratively samples $\mathbf{X} \in \mathbb{R}^{N \times 3}.$5 while conditioning on the ALS input. The reported runtime is approximately 2 seconds per tree on a Tesla V100S-PCIE-32GB GPU.

This architecture reflects a deliberate modeling choice. Rather than predicting a deterministic completion, ForestGen3D treats hidden sub-canopy structure as a conditional distribution shaped by ALS evidence and by the empirical statistics of co-registered ALS/TLS pairs. A plausible implication is that the framework is intended to reconstruct not merely missing geometry, but the distribution of ecologically plausible hidden structure consistent with an observed canopy envelope.

4. Geometric containment prior

A central contribution of ForestGen3D is a geometric containment prior defined from the ALS convex hull (Castorena et al., 19 Sep 2025). For an ALS point cloud $\mathbf{X} \in \mathbb{R}^{N \times 3}.$6, the hull is

$\mathbf{X} \in \mathbb{R}^{N \times 3}.$7

The prior is motivated by the empirical observation that, at tree scale, TLS structure usually lies spatially beneath or within the support implied by the ALS crown envelope.

The paper formalizes an empirical containment condition as

$\mathbf{X} \in \mathbb{R}^{N \times 3}.$8

for small $\mathbf{X} \in \mathbb{R}^{N \times 3}.$9. Empirically, about 3.9% of TLS points fall outside the ALS convex hull, so the paper interprets this as

$^2$0

The main theoretical statement in the paper is

$^2$1

where $^2$2 denotes model approximation error. In the appendix, this is sharpened through a KL-based bound: $^2$3

The operational metric derived from this prior is Expected Point Containment (EPC), defined as the fraction of generated points lying inside the ALS convex hull. The paper treats EPC both as a spatial plausibility diagnostic and as a proxy quality metric when TLS ground truth is unavailable. Importantly, the prior is not implemented as a hard geometric projection or clipping rule, and the paper does not describe it as an explicit additional training loss. It is therefore a soft prior: data-derived, implicitly inherited through ELBO-based approximation of the conditional distribution, and evaluated empirically rather than enforced by construction.

5. Evaluation across tree, plot, and landscape scales

Tree-scale evaluation compares ForestGen3D against PCN, 3D-GAN, latent-GAN, and PointFlow using Chamfer Distance (CD), Earth Mover’s Distance (EMD), and EPC (Castorena et al., 19 Sep 2025). On the reported 1,457-example test set, ForestGen3D achieves the best score on all three metrics.

Method	CD	EMD	EPC
PCN	$^2$4	$^2$5	0.921
3D-GAN	$^2$6	$^2$7	0.926
latent-GAN	$^2$8	$^2$9	0.862
PointFlow	$[0,1]^3.$0	$[0,1]^3.$1	0.939
ForestGen3D	$[0,1]^3.$2	$[0,1]^3.$3	0.945

At plot scale, the framework aggregates treewise generations over 25 m radius plots and evaluates height, DBH, crown diameter, and crown volume distributions using Wasserstein distance. The stated result is that ALS+ForestGen3D is consistently closest to ALS+TLS, which the paper treats as the most complete structural reference. This is central to the paper’s ecological claim: the generated sub-canopy structure improves biometric distributions relative to ALS-only observations.

Landscape-scale deployment is evaluated on 200 m radius ALS-only regions. Here the paper emphasizes containment behavior rather than direct TLS comparison. Across 10 landscapes, the out-of-envelope fraction remains around 1.4%–1.6%, with mean distances of outside points typically below 0.2 m. The authors use this as evidence that the containment prior remains practically informative under ALS-only deployment.

The paper also reports a strong empirical association between containment and geometric fidelity: as EPC rises during training, CD and EMD fall. This makes containment more than a theoretical appendix result. It becomes a practical model-selection and quality-monitoring signal in environments where terrestrial ground truth is unavailable or prohibitively expensive to obtain.

6. Position in the literature and stated limitations

ForestGen3D occupies a distinct position within recent forest-3D research. “ForestFormer3D” (Xiang et al., 20 Jun 2025) addresses unified instance and semantic segmentation of forest LiDAR point clouds; it is a forest-specific transformer for parsing structure rather than generating missing structure. “Multi-Layer Modeling of Dense Vegetation from Aerial LiDAR Scans” (Kalinicheva et al., 2022) provides a layered occupancy-and-height representation for aerial LiDAR reconstruction, emphasizing semantic points, raster occupancy maps, thickness estimation, and watertight meshes rather than cross-domain synthesis. “Scaling Up Forest Vision with Synthetic Data” (She et al., 14 Sep 2025) contributes a scalable synthetic-data pipeline for forest LiDAR segmentation, emphasizing physics, diversity, and scale as prerequisites for transfer. ForestGen3D differs from all three by treating ALS-to-TLS translation itself as a generative modeling problem.

The paper is also explicit about its limitations (Castorena et al., 19 Sep 2025). Training data come primarily from mixed conifer ecosystems, so generalization to boreal, tropical, hardwood-dominated, or structurally novel forests is not demonstrated. Large-scale deployment depends on a tree detection and bounding-box extraction stage; if neighboring trees are merged, the generator inherits that ambiguity. The containment prior is approximate: the convex hull may under- or over-estimate occupied space and does not by itself guarantee realism. The method depends on ALS providing enough canopy structure to condition generation, assumes reasonably accurate ALS/TLS co-registration during training, and does not provide calibrated uncertainty intervals for generated structures. It also has no temporal component: seasonality, thinning, prescribed fire, wildfire effects, disease, and recovery dynamics fall outside the reported formulation.

A further limitation appears in tightly packed groups. The paper describes cases in which ground-truth TLS shows four tightly packed trees while the generated output contains fewer distinct trunks. The resulting structure may remain plausible at the level of local vegetation mass and vertical profile, but not botanically exact at the individual-tree level. This suggests that ForestGen3D is strongest when the task is structural reconstruction under ALS-only conditions, rather than exact recovery of every hidden stem in dense multi-tree aggregates.

7. Scientific significance and likely uses

Within forest remote sensing, ForestGen3D is best interpreted as a method for statistical reconstruction of hidden vertical structure from aerial evidence, rather than as a generic point-cloud generator. Its scientific importance lies in linking a conditional point-cloud diffusion model to a sensor-specific geometric prior that reflects how ALS and TLS observe forests differently (Castorena et al., 19 Sep 2025). That combination makes it relevant to structural fuel mapping, wildfire modeling, and ecological analysis in ALS-only settings where direct terrestrial surveying is impractical.

The framework also provides a methodological template. It shows that cross-domain forest generation can be formulated with real co-registered sensor pairs, point-based conditional diffusion, and a soft geometric prior that is both theoretically analyzable and operationally useful. This suggests a broader research direction in which forest reconstruction, quality control, and ecological plausibility are treated jointly rather than as separate problems.

At the same time, ForestGen3D is not a complete forest-world model. It does not simulate temporal growth, stand dynamics, or species interactions, and it does not replace tree detection, co-registration, or domain adaptation. Its contribution is narrower and more specific: given ALS observations of a local vegetation object, generate a TLS-like 3D point cloud that reconstructs hidden sub-canopy detail while remaining spatially consistent with the observed canopy envelope. Within that scope, it represents one of the clearest recent attempts to turn aerial LiDAR into a generative substrate for terrestrial forest structure.