💻 computer science

TreeDGS: Aerial Gaussian Splatting for Distant DBH Measurement

TreeDGS is an aerial reconstruction method that leverages 3D Gaussian Splatting and opacity-weighted circle fitting to achieve accurate, low-cost tree diameter at breast height (DBH) measurements in complex forest scenes, outperforming state-of-the-art LiDAR baselines despite the challenges of distant and sparsely observed trunks.

Original authors: Belal Shaheen, Minh-Hieu Nguyen, Bach-Thuan Bui, Shubham, Tim Wu, Michael Fairley, Matthew David Zane, Michael Wu, James Tompkin

Published 2026-03-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Belal Shaheen, Minh-Hieu Nguyen, Bach-Thuan Bui, Shubham, Tim Wu, Michael Fairley, Matthew David Zane, Michael Wu, James Tompkin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

🌲 The Big Problem: Measuring Trees from a Helicopter

Imagine you are a forester trying to measure the thickness of a tree trunk (specifically at "breast height," about 4.5 feet off the ground). Usually, you'd walk up to the tree and wrap a tape measure around it.

But what if you have to measure thousands of trees in a dense forest from a drone flying 70 meters (230 feet) in the air?

This is the problem the authors solved.

The Challenge: From that high up, tree trunks look tiny—sometimes just a few pixels wide in a photo.
The Obstacle: Leaves, branches, and other trees block the view. It's like trying to measure a pipe through a window covered in fog and hanging vines.
The Old Way: Previous methods used LiDAR (laser scanners from drones). While good, lasers often bounce off leaves or miss the thin trunk entirely, leaving "holes" in the data. It's like trying to guess the shape of a statue by throwing sand at it; sometimes the sand misses the statue and hits the background.

✨ The Solution: TreeDGS (The "Digital Fog" Trick)

The authors created a new method called TreeDGS. Instead of trying to build a solid 3D model (like a clay statue), they use a technology called 3D Gaussian Splatting.

Here is the best way to visualize it:

1. The "Cloud of Paint" Analogy

Imagine you are trying to recreate a tree trunk using a spray paint can, but instead of solid paint, you are spraying glowing, fuzzy clouds of mist.

Traditional 3D models try to build a solid wall. If a leaf blocks the view, the wall has a hole.
TreeDGS sprays thousands of tiny, fuzzy, glowing clouds (called "Gaussians") that overlap.
Where the tree trunk is, the clouds are dense and bright (high opacity). Where there is empty air or just a few leaves, the clouds are thin and faint.

2. The "Trust Score" (Opacity)

This is the magic part. The system doesn't just look at where the clouds are; it looks at how confident the clouds are.

If a specific spot on the trunk is seen clearly by 10 different drone cameras, the "cloud" there becomes bright and solid.
If a spot is only seen through a gap in the leaves by 1 camera, the "cloud" there stays dim and fuzzy.

The system uses this "brightness" as a Trust Score. It says, "I'm 90% sure this part is the tree trunk, but I'm only 20% sure about that leafy bit."

3. The "Slicing" and "Fitting"

Once the system has this "cloud of trust," it does two things:

Slicing: It virtually slices the tree at breast height (like slicing a loaf of bread).
Fitting: It looks at the slice and tries to fit a circle around the trunk.
- The Smart Move: It ignores the dim, fuzzy clouds (the leaves and noise) and focuses only on the bright, solid clouds (the trunk). It essentially draws a circle through the "high-confidence" parts of the tree.

🏆 Why is this better than the old ways?

The authors tested this against LiDAR (the expensive laser scanner) and found that TreeDGS was actually more accurate.

LiDAR's Flaw: Lasers are great, but if a leaf is in the way, the laser hits the leaf and thinks the trunk is there. It gets confused by the "noise."
TreeDGS's Win: Because TreeDGS uses the "Trust Score" (opacity), it can tell the difference between a solid trunk and a flimsy leaf. It effectively filters out the noise.

The Results:

Old LiDAR method: Made mistakes of about 8 cm (3 inches) on average.
TreeDGS method: Made mistakes of only 4.8 cm (2 inches) on average.
Cost: TreeDGS only needs a standard camera on a cheap drone, whereas LiDAR requires expensive, heavy equipment.

🚀 The Takeaway

Think of TreeDGS as a smart detective looking at a crime scene through a foggy window.

Old methods just guess the shape based on the blurry outline.
TreeDGS looks at every clue from every angle, figures out which clues are reliable (the trunk) and which are red herrings (the leaves), and then draws a perfect circle around the truth.

This means we can now measure forests, calculate carbon storage, and manage timber resources much more accurately and cheaply, using nothing but a drone and a camera.

1. Problem Statement

The paper addresses the challenge of accurately measuring Diameter at Breast Height (DBH) for trees using aerial remote sensing, specifically from UAV (drone) RGB imagery.

The Core Difficulty: In typical UAV surveys (operating at ~~70m altitude), tree trunks are distant and occupy very few pixels (often only ~13 pixels wide). The "breast-height" band (~~1.37m above ground) is frequently occluded by the canopy, branches, and understory vegetation.
Limitations of Existing Methods:
- Traditional SfM/MVS: Struggles with weak textures and intermittent visibility, often failing to reconstruct coherent trunk surfaces.
- Aerial LiDAR: While effective for canopy structure, downward-looking LiDAR often yields sparse or uneven returns on the trunk at breast height due to occlusion, making direct DBH measurement difficult.
- Radiance Fields (NeRF/3DGS): While these improve reconstruction fidelity, converting "fluffy" volumetric reconstructions into precise geometric measurements (like diameter) is non-trivial, especially when the geometry is only weakly observed.

2. Methodology: The TreeDGS Pipeline

TreeDGS proposes a pipeline that leverages 3D Gaussian Splatting (3DGS) not just for rendering, but as a continuous, densifiable scene representation for geometric measurement. The workflow consists of four main stages:

A. Initialization and Optimization

SfM/MVS: The pipeline starts with Structure-from-Motion (SfM) to estimate camera poses and a sparse point cloud, followed by Multi-View Stereo (MVS) to create a denser initial point set.
3D Gaussian Optimization: Using RaDe-GS (a variant of 3DGS that explicitly rasterizes depth), the system optimizes a continuous field of anisotropic 3D Gaussians. This field is trained using multi-view photometric consistency to refine the scene geometry and appearance.

B. Opacity-Aware Surface Sampling

Directly exporting Gaussian centers yields a sparse, fragmented cloud. Instead, TreeDGS extracts a dense point set using a depth-aware cumulative-opacity integration strategy:

Sampling: Points are drawn from the volume of each Gaussian based on its learned opacity ( $\alpha$ ).
Reliability Scoring: For each sampled point, the system computes a multi-view opacity support score. It integrates the cumulative opacity along the viewing ray for every calibrated camera.
Filtering: Points are retained only if they project to a foreground pixel in multiple views and have a high accumulated opacity. This effectively filters out "floaters" and semi-transparent vegetation, isolating the solid trunk surface.

C. Semantic Trunk Extraction

The extracted point cloud contains a mix of trunks, foliage, and noise.

Segmentation: A 3D semantic segmentation model (ForestFormer3D) is applied to label points as "trunk" or "vegetation."
Isolation: Only points classified as trunks are retained for each tree instance, forming a clean trunk-only point cloud.

D. Opacity-Weighted DBH Estimation

The final step estimates the diameter using a robust fitting strategy tailored to the volumetric nature of the data:

Slicing: The trunk is sliced horizontally at 10cm intervals around the breast height (1.37m).
Solid-Circle Fitting: Unlike traditional methods that fit circles to surface rings, TreeDGS treats the trunk cross-section as a filled disk (since sampling occurs inside the volume).
RANSAC with Weighting: A RANSAC algorithm fits a circle to the 2D slice points. Crucially, it uses opacity weights ( $w = \bar{\alpha}$ ) derived from the multi-view consistency test. Points with low confidence (likely noise or occlusion artifacts) are down-weighted during hypothesis generation and scoring.
Taper Regression: A robust linear regression with a negative slope constraint is applied across slices to model the tree's taper, ensuring the final DBH estimate at 1.37m is physically plausible and robust to local noise.

3. Key Contributions

TreeDGS Framework: The first method to utilize 3D Gaussian Splatting specifically for direct, object-level DBH measurement from distant, pixel-limited UAV RGB imagery.
Opacity-Guided Sampling: A novel extraction technique that converts the volumetric Gaussian field into a dense, surface-consistent point cloud while attaching a multi-view reliability score to every point.
Weighted Geometric Fitting: A measurement strategy that combines solid-circle RANSAC with opacity-based weighting, significantly reducing the impact of occlusion and noise compared to unweighted or cylinder-fitting baselines.
Performance Validation: Demonstrated that RGB-only methods can outperform state-of-the-art aerial LiDAR for DBH estimation in complex, dense forest environments.

4. Experimental Results

The method was evaluated on 10 circular plots (0.2 acres each) in a managed loblolly pine forest in Arkansas, containing 210 trees with ground-truth DBH measurements.

Accuracy:
- TreeDGS (Opacity-Weighted Circle Fit): Achieved an RMSE of 4.79 cm and MAE of 3.67 cm.
- Comparison: This outperformed the best aerial LiDAR baseline (Cylinder Fit) which had an RMSE of 7.91 cm. TreeDGS reduced the error by approximately 39%.
- Bias: TreeDGS showed negligible systematic bias (ME = -0.38 cm), whereas other methods exhibited significant underestimation or overestimation.
Robustness: The opacity-weighted approach significantly reduced variance compared to unweighted circle fitting or standard cylinder fitting, particularly in plots with heavy occlusion (e.g., Plot 3, where LiDAR error was 17.22 cm vs. TreeDGS 5.93 cm).
Ablation: The paper confirms that using standard cylinder fitting on TreeDGS points performs poorly (RMSE 13.29 cm), highlighting the necessity of the "solid-circle" fitting model and opacity weighting.

5. Significance and Limitations

Significance:

Cost-Effectiveness: Demonstrates that high-accuracy DBH measurement can be achieved using commodity RGB cameras on standard UAVs, eliminating the need for expensive LiDAR sensors.
Overcoming Occlusion: The opacity-weighting mechanism effectively handles the "few pixels, few views" regime typical of aerial forest surveys, where traditional photogrammetry fails.
New Paradigm: Shifts the use of 3DGS from purely visual rendering (NeRF-like) to a geometric measurement tool, bridging the gap between dense reconstruction and forestry metrics.

Limitations:

Visibility Dependency: The method requires the breast-height band to be visible in at least 2–3 informative views. If the trunk is fully occluded in all views, the geometry cannot be recovered.
Segmentation Sensitivity: The accuracy relies on the quality of the 3D semantic segmentation. False-positive trunk labels from nearby vegetation can bias the diameter estimate, though opacity weighting mitigates this.
Future Work: The authors suggest future research into jointly optimizing explicit trunk primitives (e.g., tapered cylinders) within the Gaussian field to reduce reliance on post-processing segmentation.