Synthetic 3D SPREAD Dataset for Forest Biomass Analysis

• Synthetic 3D SPREAD Dataset is a synthetic collection of photorealistic forest scenes with multimodal annotations designed for aboveground biomass estimation.
• It leverages advanced procedural generation in Unreal Engine 5 to create diverse, high-resolution views with detailed per-tree metadata and environmental variability.
• The dataset supports pixelwise biomass regression, depth estimation, and instance segmentation, enabling scalable, machine learning–driven forest monitoring.

The Synthetic 3D SPREAD Dataset is a highly detailed collection of photorealistic, procedurally generated forest scenes designed for research on aboveground biomass (AGB) estimation and related dense-prediction tasks in forestry and ecological monitoring. SPREAD offers multimodal data assets—including high-resolution RGB frames, depth maps, and comprehensive per-tree annotations—rendered from a diverse range of virtual environments using the Unreal Engine 5 platform. This dataset enables the training and benchmarking of computer vision models for pixelwise biomass regression, depth estimation, instance segmentation, and AGB density mapping, providing a controlled and scalable resource for advancing machine learning-based forest monitoring methodologies (Zuffi, 28 Nov 2025).

1. Dataset Structure and Scene Composition

SPREAD comprises 13 photorealistic 3D “virtual plots,” each assembled and rendered in Unreal Engine 5. The virtual environments encompass diverse ecological and anthropogenic settings, including temperate deciduous woods (e.g., Birch Forest), mixed broadleaf stands, redwood groves, tropical rainforests, plantation rows, meadows, and urban parks. Each plot is rendered under multiple environmental conditions—such as clear sky, overcast, and rain—that are modulated with physically based lighting, varying sun angles, cloud cover, and HDRI-driven ambient illumination to simulate temporal and weather variability.

For each scene, hundreds to thousands of unique camera viewpoints are sampled, producing high-resolution RGB images (originally up to 4,000 × 3,000 pixels, resized to 224 × 224 for network training) and associated auxiliary render passes. Tree assets derive from a large species-specific 3D model library, capturing variation in trunk, branch, and leaf geometry to support a broad spectrum of crown architectures, sizes, and morphologies. Scene assembly leverages physically based rendering (PBR) and procedural-placement tools in the engine, ensuring structural and radiometric realism across scenes.

2. Annotations and Metadata

Each frame in the SPREAD dataset is furnished with rich, multimodal annotations:

• RGB image (PNG format)
• Depth buffer (in meters)
• Semantic segmentation maps (per-pixel class labels: trunk, foliage, ground, sky)
• Instance segmentation maps (per-pixel tree identifiers via unique 24-bit color masks)

For every scene, a JSON metadata file provides per-tree instance information, including:

• Diameter at breast height (DBH, in centimeters)
• Total height (H, in meters)
• Canopy diameter (in meters)
• Ground-plane coordinates (x, y) of the trunk base
• Wood specific density ($\rho$, in g cm$^{-3}$), selected per species

These comprehensive annotations facilitate the use of SPREAD for a wide range of machine perception and forestry analytics tasks. The combination of semantic and instance-level segmentations alongside depth maps and metadata supports supervised training and rigorous evaluation of dense-prediction algorithms.

3. Generation Pipeline and Rendering Methodology

Forest plot assembly, species placement, and environmental effect scripting are fully executed within Unreal Engine 5 using its foliage-painting and procedural placement tools. Surface materials adhere to a PBR standard, with lighting realized through a combination of directional "sun" sources and skydome HDRI for ambient effects. Weather variations—including fog, rain, and cloud cover—are procedurally scripted to increase visual diversity and robustness to real-world environmental noise.

Camera intrinsics, such as focal length and sensor size, are configurable, enabling emulation of various real-world camera systems. Camera extrinsics, like height, pan, and tilt, are randomized within user-specified ranges to produce extensive viewpoint diversity. At each camera pose, the engine outputs synchronized render passes (RGB, depth, semantic labels, and instance IDs) to ensure consistent pixel correspondences across modalities.

4. Biomass Computation and Allometric Equations

To derive per-pixel aboveground biomass ground truth, SPREAD employs a multi-step process grounded in established forestry allometric principles:

• Single-tree Biomass: For each tree instance $i$, AGB (in kilograms) is computed following the allometric model of Chave et al.:

$$\mathrm{AGB}_i = 0.0673 \times (\rho \cdot \mathrm{DBH}^2 \cdot H)^{0.967}$$

where $\mathrm{DBH}$ is in cm, $H$ in m, and $\rho$ in g cm$^{-3}$.

• AGB Density Mapping: Let $A_p$ denote the plot area in m$^2$ (ground-plane bounding rectangle of all visible tree bases), and $A_i$ the pixel count of tree $i$’s instance mask. Each pixel $(h, k)$ associated with tree $i$ is assigned:

$$d\mathrm{AGB}[h,k] = \frac{\mathrm{AGB}_i}{A_p \cdot A_i}$$

giving a per-pixel AGB density in kg m$^{-2}$.

• Total Plot-Level AGB: Integrating the pixelwise density map across all scene pixels yields

$$\mathrm{AGB} = \sum_{h, k} d\mathrm{AGB}[h,k] = \sum_i \frac{\mathrm{AGB}_i}{A_p}$$

This framework supplies dense, spatially explicit AGB ground truth, enabling supervised learning for pixelwise AGB regression from RGB inputs (Zuffi, 28 Nov 2025).

5. Dataset Statistics and Practical Applications

Of the 13 virtual plots, two temperate forest scenes—Birch Forest and Broadleaf Forest—are prioritized for model development due to their complete instance annotations and DBH records. After filtering (removing tree masks covering <2% of image area and outlier viewpoints), the training set consists of approximately 5,660 image pairs (RGB plus derived AGB density map), split 80%/20% into train/test partitions.

AGB distributions for these scenes span roughly 0–10 kg/m$^2$ (Birch) and 0–15 kg/m$^2$ (Broadleaf). Typical plots contain tens to a few hundred unique trees per image. The structure and annotation richness of SPREAD natively support tasks including:

• Pixelwise biomass regression (AGB density mapping)
• Depth estimation from RGB imagery
• Instance segmentation (tree identification)
• Algorithmic benchmarking for synthetic-to-real transfer

A key application is end-to-end training of neural networks that map RGB images to 1-channel AGB density outputs. This enables, for the first time, estimation of AGB directly from single ground-based RGB images without reliance on field measurements or airborne remote sensing, facilitating scalable forest monitoring and citizen science engagement (Zuffi, 28 Nov 2025).

6. Limitations, Domain Gaps, and Considerations

Despite its photorealism, SPREAD remains a synthetic resource with several limitations:

• Fine-grained understory detail and leaf texture variability may not match real forest floors or reflect phenological (seasonal) dynamics such as foliage color and density changes.
• Tree species diversity is restricted by available Unreal Engine assets: rare or ecologically unique species are omitted, and all biomass calculations exclude belowground/root carbon pools (AGB only).
• Plot-area $A_p$ is computed assuming a perfectly flat terrain; real-world sloped or uneven terrains require remeasurement, or else density estimates risk systematic error.
• Camera models must be carefully calibrated to align synthetic data with real instrument intrinsics; field deployment involving non-standard optics (e.g., wide-angle, fisheye lenses) may necessitate re-rendering scenes with matching parameters.
• Hybrid training approaches—such as pretraining on SPREAD followed by fine-tuning with a smaller real-image dataset, or explicit domain adaptation—are recommended to address the synthetic-to-real gap and maximize ecological monitoring accuracy.

A plausible implication is that, while SPREAD offers scalable and precise pretraining, operational deployment in heterogeneous real forests requires careful calibration and adaptation to local ecological and imaging conditions (Zuffi, 28 Nov 2025).