From individual trees to virtual forests: a comprehensive 3D structural dataset of Central European tree species

Abstract

This dataset provides a comprehensive, multi-scale collection of forest structural data for Central European tree species. The data were mainly derived from extensive terrestrial laser scanning campaigns conducted between 2009 and 2025 across 47 forest sites in the Czech Republic, capturing 12 tree species in both leaf-on and leaf-off conditions. The repository is comprised of three components. First, it includes a library of 3618 individual segmented terrestrial laser scanning tree point clouds, accompanied by detailed metadata coupling field-measured traits with structural parameters derived from a quantitative structure model (QSM). Second, the dataset contains 15,718 virtual airborne laser scanning scenes (50 × 50 m plots) simulated using the HELIOS++ radiative transfer model. These synthetic plots were procedurally assembled using the Forest Factory growth model and provide multi-resolution point clouds (8 points/m² and 16 points/m² densities) alongside precise spatial and structural metadata for every tree. Third, the repository features 273 fully reconstructed three-dimensional (3D) tree representations, combining a QSM for woody architecture with biologically realistic foliage distribution based on empirical traits. This dataset serves as a robust ground-truth resource for training, testing, and validating machine learning algorithms aimed at individual tree segmentation, species classification, and structural parameter extraction from airborne laser scanning data. Furthermore, the ready-to-use 3D representations and virtual scenes facilitate advanced 3D radiative transfer modelling and the development of a virtual forest.

Specifications Table

Subject	Earth & Environmental Sciences
Specific subject area	Terrestrial laser scanning; Virtual scenes; 3D tree representations; Tree mechanical stability; Quantitative Structure Model; RayCloudTools, Virtual airborne laser scanning
Type of data	Point cloud (.laz), Photos (.png), Tree metadata (.csv), 3D model (.obj)
Data collection	Terrestrial laser scanning (TLS) point clouds (leaf-on and leaf-off) were collected across the Czech Republic between 2009 and 2025 using OPTECH Ilris-36D, Riegl VZ-400, Riegl VZ-400i, Leica RTC360, and Hovermap ST-X, and subsequently segmented into individual trees. From a selected subset of these scans, complete 3D tree representations were reconstructed. These consist of quantitative structure models representing the woody structure, and biologically realistic foliage distribution driven by leaf area index. Virtual airborne laser scanning data were simulated in HELIOS++ using scenes based on acquired TLS trees positioned on generated coordinates with detailed attribute metadata.
Data source location	Country: Czech Republic Localisation: Various forest sites throughout the country Latitude and longitude: 49° 44′ 37″ N, 15° 30′ 20″ E (center of the Czech Republic)
Data accessibility	Repository name: Czech National Repository Data identification number: 10.48700/datst.xkadt-0cb92 Direct URL to data: https://datarepo.eosc.cz/datasets/records/datst.xkadt-0cb92
Related research article	None

1. Value of the Data

• •The dataset provides 3618 segmented terrestrial laser scanning (TLS) individual tree point clouds across 12 Central European species, serving as a ground-truth resource for training and validating machine learning algorithms for tree segmentation, species classification, and structural parameter extraction.
• •The repository features 15,718 virtual airborne laser scanning (ALS) forest scenes simulated via the HELIOS++ radiative transfer model, providing an annotated dataset for the training and testing of machine learning algorithms designed to retrieve forest structural variables.
• •A subset of 273 highly detailed three-dimensional (3D) tree representations, combining quantitative structure models (QSM) for woody architecture with biologically realistic distribution of foliage. The reconstruction workflow utilises PointNet++ deep-learning semantic separation to isolate wood and foliage components; the wood points are subsequently used to construct the QSM geometry, while foliage placement is driven by empirical traits (target leaf area index, leaf angle distribution, and specific leaf dimensions). Provided as ready-to-use .obj meshes and .xyz point clouds, these representations can serve as structural inputs directly compatible with 3D radiative transfer models.
• •The virtual ALS scenes are generated directly from TLS point clouds of the same trees using the Forest Factory growth model and simulated via the Helios++ radiative transfer model. This setup provides a tightly coupled ground-truth reference that avoids the spatial and temporal mismatches inherent in independently collected field and airborne datasets, replicating a Riegl LMS Q-780 sensor at standard (8 points/m²) and high-density (16 points/m²) resolutions.
• •The dataset uniquely combines critical wind speed parameters for stem breakage derived from QSM with species-specific tree mechanical properties for 12 Central European species. These values are explicitly coupled in the metadata (.csv files) with variables such as maximum bending stress and modulus of elasticity, providing a dedicated resource for biomechanical modelling and assessment of forest structural vulnerability to wind disturbance.

2. Background

This dataset [1] was compiled to train and validate machine-learning approaches for retrieving forest structural properties, such as diameter at breast height (DBH) or biomass, from ALS data. While the optimal amount of training data for these algorithms is not clearly established, these methods benefit from large volumes [2]. To overcome the human and financial costs of acquiring large-scale datasets in the field, simulated virtual forest scenes are frequently used [3]. Nevertheless, these scenes often lack the architectural complexity of real trees or are optimised for one specific site and sensor, limiting their suitability for broader implementation. Our approach addresses this gap by combining individual tree point clouds with airborne simulations designed for specific regional ecosystems. The focus was therefore on Central European tree species, representing the dominant forest composition in the region. To capture this complexity, a library of 3618 individual tree point clouds (leaf-on and leaf-off) was compiled. From this dataset and publicly available sources, 15,718 virtual ALS plots (50 × 50 m) were generated with tree metadata to directly support machine learning applications. Furthermore, detailed 3D representations were reconstructed for a subset of trees, providing ready-to-use structural inputs for radiative transfer models such as the Discrete Anisotropic Radiative Transfer (DART) model or HELIOS++. The dataset establishes a baseline for developing robust retrievals of forest structural traits.

3. Data Description

The dataset is structured into two primary zip archives: the complete database (“Tree_dataset.zip”) and a representative sample (“Tree_dataset_subset.zip”). In the “Tree_dataset.zip” file, there are three folders and a text file. The overall folder structure of the repository is shown in Fig. 1. The “ReadMe.txt” contains a description of the folder structure, file naming conventions, and species list. The folder “SEGMENTED_TREES”, includes the comprehensive library of individual segmented tree point clouds acquired via TLS. The data are organised into two subfolders: “LAZ”, containing the compressed 3D point clouds (.laz), and “PNG”, providing image previews of the segmented clouds. The .laz files follow a naming convention denoting the plot ID, Latin species abbreviation, tree ID, and leaf status (e.g., DN20_FASY_15_LON.laz for a leaf-on Fagus sylvatica L. tree with ID 15 from site DN20). This folder is accompanied by two metadata files. The “SEGMENTED_TREES_METADATA.csv” file provides a detailed table of individual tree traits, coupling field-measured metrics (e.g., DBH, tree height, and partitioned biomass components such as leaf, branch, stem, and root) with structural traits extracted directly from the QSM (e.g., QSM-derived DBH, stem and branch volumes, crown base height, and mechanical stability parameters). The “SPECIES_ATTRIBUTES.csv” file supplies physical and mechanical tree properties for all 12 represented species, directly underpinning the mechanical stability parameters included in the “SEGMENTED_TREES_METADATA.csv”.

Fig. 1.

Structure of the Tree_dataset.zip and Tree_dataset_subset.zip data repository.

The second folder, “VIRTUAL_SCENES”, contains the virtual ALS scenes (50 × 50 m). To account for seasonal variability, the data are divided into two primary scenario folders: “ON” (simulated leaf-on conditions with full foliage) and “OFF” (simulated leaf-off conditions). Within each scenario, the content is further separated into a “DATA” subfolder containing the actual simulated files (.laz), and a “CSV” subfolder containing metadata (.csv) for each virtual plot. Each virtual scene was simulated from two different flight altitudes, resulting in two distinct point cloud densities: 8 points/m² (denoted as 8p) and 16 points/m² (16p). This is reflected in the file naming convention (e.g., 30_18_2_8p.laz, where the numeric prefix serves as a unique scene identifier and the 8p suffix denotes the point cloud density). The accompanying metadata files detail the exact spatial arrangement and structural attributes of every tree within the simulated scene.

The third folder, “3D_TREES” contains 3D representations of individual trees. They are divided into folders by Latin species names (e.g., ACER_PSEUDOPLATANUS, FAGUS_SYLVATICA). Accompanying these is a “3D_TREE_METADATA.csv” file containing parameters for each 3D tree representation. Within each species directory, individual trees are stored in separate subfolders, each containing five files: a complete 3D representation (.obj), a QSM of the woody structure (.obj), separate point clouds for wood and foliage (.xyz), which were used to generate the previously mentioned objects, and a preview image (.png).

The file “Tree_dataset_subset.zip” has the same structure as the file “Tree_dataset.zip” but contains only a sample of the whole dataset. Sample files include 10 randomly selected 3D representations from the folder “3D_TREES”, 100 segmented point clouds from “SEGMENTED_TREES”, and 50 virtual scenes from “VIRTUAL_SCENES”, comprising both 8p and 16p density versions.

4. Experimental Design, Materials and Methods

4.1. Field data acquisition and TLS data processing

TLS campaigns were conducted between years 2009 and 2025 across various forest sites in the Czech Republic (47 in total), and the majority of the plots were part of DendroNetwork infrastructure (www.dendronet.cz). Although the plots are predominantly characterised by four species (Fagus sylvatica L., Quercus robur L., Pinus sylvestris L., Picea abies (L.) H.Karst.), the dataset includes a total of 12 Central European tree species. The spatial distribution of the sampled forest plots is illustrated in Fig. 2. Data were acquired during both leaf-on and leaf-off seasons using several laser scanners, including OPTECH Ilris-36D, Riegl VZ-400, Riegl VZ-400i, Leica RTC360, and Hovermap ST-X. Some plots dominated by Pinus sylvestris were scanned by colleagues from the Freie Universität Berlin, and the scans are available as part of their dataset [4]. Each tree on plots was also measured using FieldMap (www.field-map.com) to obtain height and DBH, which were compared with TLS-derived data metrics and used to calculate aboveground biomass using allometric equations [5].

Fig. 2.

Spatial distribution of forest inventory plots within the Czech Republic. The map shows the geographical locations where terrestrial laser scanning data were acquired, with plot symbols categorised by the dominant tree species.

Following data acquisition, the raw point clouds were co-registered using the software provided by the scanner manufacturer (RiSCAN PRO 2.1.1 - Reigl VZ400/i; Leica Cyclone REGISTER 360 2025.0.1 - Leica RTC360; Emesent Aura 1.10.2 - Hovermap ST-X). Subsequent processing steps, including noise filtering, classification into ground and non-ground points, height normalisation, and voxelisation at a 1 cm voxel size, were performed using LAStools [6]. Individual trees were then extracted and segmented utilising the RayExtract method within RayCloudTools [7,8]. Because automated segmentation often struggles in dense forest conditions (resulting in artefacts such as detached trunk bases or merged crowns from neighbouring trees), manual corrections were applied using CloudCompare software (version 2.14) [9]. These cleaned and refined tree point clouds were subsequently reprocessed in RayExtract to generate QSM. Finally, a custom script was employed to calculate numerous structural parameters directly from these models (e.g., tree volume, surface area, convex projection areas, and centroid height), all of which are included in the dataset's metadata table. Height and DBH distribution are shown in Table 1.

Table 1.

Species distribution and summary statistics of the sampled trees. Values for tree height and diameter at breast height (DBH) are presented as means ± standard deviation.

Species Latin name	Number of trees	Tree height [m]	Diameter at breast height [cm]
Fagus sylvatica	1028	22.4 ± 6.6	22.1 ± 10.9
Picea abies	975	22.1 ± 5.7	25.1 ± 9.1
Pinus sylvestris	852	20.9 ± 2.9	20.1 ± 5.6
Quercus robur	682	19.1 ± 4.2	20.8 ± 6.6
Acer pseudoplatanus	29	21.4 ± 5.2	27.5 ± 9.9
Larix decidua	22	22.7 ± 6.6	29.0 ± 9.9
Carpinus betulus	15	19.0 ± 7.4	19.7 ± 8.3
Tilia cordata	6	13.5 ± 3.3	15.2 ± 3.0
Betula pendula	5	25.5 ± 8.6	29.3 ± 9.2
Ulmus minor	2	10.5 ± 0.1	16.4 ± 4.8
Abies alba	1	19.8	13.1
Acer platanoides	1	8	7.2

To assess mechanical stability, the critical wind speed for stem breakage was determined based on the principle of moment equilibrium, balancing external bending moments against the mechanical resistance of the stem. Two main sources of loading are considered: i) the wind-induced moment M_wind and ii) the gravitational moment of the crown M_crown. The wind moment is expressed as

$${\mathrm{M}}_{\text{wind}}\left(\mathrm{x}\right)=\frac{1}{2}·{\rho }_{\text{air}}·\mathrm{c}·{\mathrm{A}}_{\text{crown}}·{\mathrm{V}}^2.\left(\mathrm{L}-\mathrm{x}\right)·\cos \left({\theta }_{\text{trunk}}\right)$$ (1)

where ρ_air is air density set as 1.2 [10], c is the drag coefficient of the crown [11], A_crown is the projected crown area, V is wind speed, L is the height of the crown centre of gravity, x is the height of the potential break point, and θ_trunk is the trunk tilt. The gravitational contribution is given by

$${\mathrm{M}}_{\text{crown}}\left(\mathrm{x}\right)={\mathrm{F}}_{\mathrm{e}}·\sqrt{{\left(\left(\mathrm{L}-\mathrm{x}\right)·\sin \left({\theta }_{\text{trunk}}\right)\right)}^2+{\mathrm{L}}_{\text{exc}}^2}$$ (2)

where F_e is the weight force of the crown, and L_exc represents the horizontal eccentricity of the crown’s centre of gravity. These applied moments are resisted by the stem’s bending capacity (M_break), defined as

$${\mathrm{M}}_{\text{break}}=\frac{\pi }{32}·{\sigma }_{\max }·{\left(2\mathrm{r}\left(\mathrm{x}\right)\right)}^3$$ (3)

where σ_max is the maximum bending stress of the wood [12], and r(x) is the stem radius at the break point at height x. The condition for failure is reached when

$${\mathrm{M}}_{\text{break}}={\mathrm{M}}_{\text{wind}}+{\mathrm{M}}_{\text{crown}}$$ (4)

By rearranging this balance, the critical wind speed for breakage is obtained as

$$\mathrm{V}\left(\mathrm{x}\right)=\sqrt{\frac{{\mathrm{M}}_{\text{break}}-{\mathrm{M}}_{\text{crown}}}{\frac{1}{2}·{\rho }_{\text{air}}·\mathrm{c}·{\mathrm{A}}_{\text{crown}}·\left(\mathrm{L}-\mathrm{X}\right)·\cos \left({\theta }_{\text{trunk}}\right)}}$$ (5)

This formulation integrates aerodynamic loading, tree architecture, and material strength, providing a mechanistic estimate of the wind speed at which the stem is expected to fail.

4.2. Virtual scene generation and ALS simulation

Seasonally variable training data for machine learning algorithms was generated by simulating 6576 leaf-on and 9142 leaf-off synthetic ALS forest plots (50 × 50 m) to capture seasonal structural variability. Virtual scene generation relied on the Forest Factory growth model [13,14]. The model was parameterised for Central European forests using settings provided by Schäfer et al. [2], and initially simulated four dominant species: Fagus sylvatica, Quercus robur, Picea abies, and Pinus sylvestris. The Forest Factory outputs provide specifically tree position, species, height, and DBH, and provide the spatial framework for the assembly of each plot.

The assignment of segmented TLS tree data to the simulated positions followed a matching algorithm. This process utilised both the proprietary TLS database compiled in Section 4.1 and supplementary open-source datasets (specifically, subsets from FOR-species20K [2], Wytham forest [15], and PyTreeDB [16]). First, the databases were filtered by the required species and the target phenological state. For deciduous species, the algorithm strictly selected TLS point clouds representing the corresponding seasonal status (leaf-on or leaf-off state). For coniferous species, seasonal differences were not distinguished, and the same point clouds were used in both scenarios. Next, the 10 trees with a measured height closest to the Forest Factory output requirement were pre-selected. From this subset, three trees with the DBH closest to the simulated value were chosen, and from this selection, one tree was randomly selected. The corresponding TLS point cloud was then placed at the calculated coordinates and randomly rotated around the Z-axis to ensure geometric variability. This iterative selection and placement procedure was applied to every tree within the scene. Finally, a synthetic ground point cloud was incorporated beneath the canopy, resulting in the creation of a complete point cloud for the given forest plot (Fig. 3).

Fig. 3.

Generalised workflow for virtual forest scene assembly. (A) 2D output from the Forest Factory model depicting tree spatial distribution and parameters (shown is tree ID); (B) assignment of a structurally compatible tree from the terrestrial laser scanning (TLS) database; (C) placement of the selected tree into the virtual scene; (D) complete artificial TLS scene, visualised by height (blue to red gradient).

ALS data simulation was performed by processing the fully assembled 3D point cloud scenes (Fig. 3D) within the HELIOS++ radiative transfer model [17]. The simulations were parameterised using the predefined Riegl LMS Q-780 scanner model, with the maximum pulse repetition frequency set to 400 kHz, to replicate the real airborne LiDAR system and standard operational mapping workflows operated by the Global Change Research Institute CAS (CzechGlobe) [18]. To convert the input TLS point clouds into the HELIOS++ framework, a voxel edge length of 10 cm was applied. This resolution serves as an optimal compromise; reducing the voxel size to 5 cm would increase computational demands, whereas a standard low-altitude ALS acquisition yielding 16 points/m² corresponds to a point spacing of roughly 25 cm. Thus, the 10 cm voxel size ensures high structural fidelity without prohibitive computational costs.

Diverse acquisition strategies and multi-resolution training data were captured by configuring virtual flight lines for two distinct scanning scenarios (Fig. 4). A standard survey configuration uses parallel, overlapping flight lines at 1030 m above ground level (AGL) to produce a point cloud density of 8 points/m² (8p). Conversely, a detailed mapping setup employs a cross-hatch pattern with intersecting flight lines at 500 m AGL to maximise canopy penetration, yielding 16 points/m² (16p).

Fig. 4.

Schematic of virtual flight lines for airborne laser scanning simulation in HELIOS++. The solid black line demarcates the extent of the input terrestrial laser scanned (TLS) virtual scene. The blue lines illustrate the scanning configuration for standard surveys, using parallel, overlapping flight lines at 1030 m above ground level (AGL) to yield a point cloud density of 8 points/m² (8p). The red lines depict the detailed scanning configuration, employing a cross-hatch pattern with intersecting flight lines at 500 m AGL to maximise canopy penetration and yield a higher density of 16 points/m² (16p).

4.3. 3D tree representations reconstruction

Advanced 3D radiative transfer modelling requires highly realistic virtual assets; therefore, a selected subset of 273 segmented TLS point clouds (Section 4.1) was reconstructed into complete 3D representations. The reconstruction workflow was executed according to the methodologies detailed by Janoutová et al. [19] and Hanousek et al. [20]. This multi-step process initially included the semantic separation of wood and foliage components via a custom-trained PointNet++ deep learning model. The wood points were then utilised to construct QSM by progressively fitting cylinders to the data, thereby capturing the precise topological and geometrical traits of the branching architecture. Finally, the procedurally realistic placement of biologically foliage was driven by empirical traits, namely target leaf area index, leaf angle distribution, and specific leaf dimensions, resulting in high-fidelity meshes. A visual example illustrating the transition from the original segmented TLS point cloud to the finalised complete 3D representation is provided in Fig. 5. The distribution of the generated 3D representations across the tree species is summarised in Table 2. The remaining six species represented in the “SEGMENTED_TREES” library do not have corresponding 3D representations in this dataset.

Fig. 5.

Generalised workflow of the three-dimensional (3D) tree representation reconstruction. (A) Original segmented terrestrial laser scanning point cloud (PC) of an individual tree; (B) semantically isolated wood PC; (C) semantically isolated foliage PC; (D) reconstructed quantitative structure model representing the woody architecture; (E) 3D foliage mesh; (F) finalised complete 3D representation.

Table 2.

Species distribution and summary statistics of the reconstructed three-dimensional tree representations. Values for leaf area index and tree height are presented as means ± standard deviation.

Species Latin name	Number of trees	Leaf area index [m2/m2]	Tree height [m]
Fagus sylvatica	168	3.49 ± 1.67	28.82 ± 6.04
Picea abies	90	5.68 ± 3.28	30.09 ± 7.29
Acer pseudoplatanus	8	4.44 ± 0.97	29.73 ± 0.72
Quercus robur	3	2.44 ± 0.75	26.87 ± 4.22
Betula pendula	2	2.75 ± 0.35	28.29 ± 0.33
Carpinus betulus	2	2.08 ± 0.87	32.30 ± 2.74

Limitations

The source TLS data and subsequent 3D representations come from geographically restricted Central European forest ecosystems. Although 12 species are present, the dataset is heavily biased towards four dominant species (Fagus sylvatica, Quercus robur, Picea abies, and Pinus sylvestris). This restricted sampling does not fully represent the structural variability that may occur across distinct biomes. The methodological design introduces several limitations that should be acknowledged. The virtual scenes assembled in Forest Factory and simulated via HELIOS++ utilise a synthetic flat ground point cloud. This simplification does not fully capture the structural complexity of understory vegetation, resulting in less terrain-level noise compared to real-world ALS acquisitions. Additionally, structural occlusion during the initial TLS data acquisition inevitably limits the complete spatial capture of dense inner-crown branching architectures. When combined with the foliation algorithm, this inherent occlusion introduces an underestimation of leaf area index and lower overall foliage density in the canopy sections of the final 3D representations in comparison with real trees.

Ethics Statement

The authors declare that this work did not involve human subjects nor animal experiment nor data collected from social media platforms. They have read and followed the ethical requirements for publication in the Data in Brief journal.

Declaration of Generative AI and AI-assisted Technologies in the Manuscript Preparation Process

During the preparation of this work, the author(s) used the Gemini 3.1 Pro model and Grammarly in order to improve language and polish text to a scientific level. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article.

CRediT authorship contribution statement

Tomáš Hanousek: Conceptualization, Methodology, Data curation, Visualization, Writing – original draft. Barbora Pavelková Navrátilová: Methodology, Validation, Data curation. Jan Novotný: Methodology. Barbora Schneider: Data curation. Růžena Janoutová: Methodology, Resources, Supervision, Writing – review & editing.

Data Availability

• Czech National RepositoryDatabase of TLS Point Clouds, 3D Models, and Virtual ALS for Central European Trees (Original data).