💻 computer science

Estimating the Diameter at Breast Height of Trees in a Forest from RGB

This paper presents a low-cost, semi-automated pipeline using consumer-grade 360° video and SfM photogrammetry to estimate tree diameter at breast height with 5–9% median absolute relative error, achieving accuracy comparable to expensive LiDAR methods while significantly reducing cost and operational complexity.

Original authors: Siming He, Zachary Osman, Fernando Cladera, Dexter Ong, Nitant Rai, Patrick Corey Green, Vijay Kumar, Pratik Chaudhari

Published 2026-04-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Siming He, Zachary Osman, Fernando Cladera, Dexter Ong, Nitant Rai, Patrick Corey Green, Vijay Kumar, Pratik Chaudhari

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

Imagine you are a forest manager trying to figure out how much carbon a forest is storing or how much timber it holds. To do this, you need to measure the thickness of the tree trunks at a specific height (about chest height, or 1.3 meters). This measurement is called DBH (Diameter at Breast Height).

Traditionally, foresters have to walk into the woods, find every tree, and wrap a tape measure around it. It's slow, tiring, and hard to do for thousands of trees.

This paper introduces a low-cost, high-tech shortcut that uses a simple 360-degree camera (like the ones used for virtual reality tours) to measure trees automatically. Here is how it works, broken down into simple steps with some creative analogies.

1. The Problem: The Expensive "Laser Scanner" vs. The Cheap "Camera"

The Old Way (LiDAR): Think of LiDAR as a high-end, military-grade laser scanner. It shoots lasers at trees and builds a perfect 3D model. It's incredibly accurate (like a master carpenter's ruler), but it costs thousands of dollars and requires a specialist to operate.
The New Way (360° Video): The authors say, "Why spend a fortune when we can use a consumer camera?" They use a cheap 360-degree camera (like an Insta360) mounted on a drone or a backpack. It's like using a smartphone instead of a professional DSLR camera. It's cheap, easy to use, and available to everyone.

2. The Process: How They Turn Video into Measurements

The team built a "recipe" (pipeline) to turn that video into tree measurements. Here are the three main ingredients:

Step A: Building the 3D World (The "Puzzle Solver")

First, they take the 360-degree video and feed it into software called Agisoft Metashape.

The Analogy: Imagine taking hundreds of photos of a statue from every angle and using a computer to glue them together into a 3D model. This is called Structure from Motion (SfM). The software looks at how the pixels move between frames to figure out where the camera was and builds a "point cloud" (a digital cloud of dots representing the forest).
The Catch: The computer builds the forest, but it doesn't know the real size yet. It might think a tree is 10 meters tall when it's actually 5. So, the team places a known-sized marker (a 1-meter square) in the scene to tell the computer, "Hey, this square is exactly 1 meter. Scale everything else to match."

Step B: Finding the Trunks (The "Digital Highlighter")

Now they have a 3D forest, but it's a messy mix of leaves, branches, ground, and trunks. They need to isolate just the trunks.

The Analogy: Imagine you have a giant photo of a forest, and you want to highlight only the tree trunks in red. You ask a super-smart AI (called Grounded SAM) to "find the tree trunks." The AI draws a mask around the trunks in the 2D photos.
The Magic: The system projects those 2D red masks onto the 3D point cloud. Suddenly, the computer knows exactly which 3D dots belong to a tree trunk and which belong to a leaf. It's like using a digital highlighter to color-code the trunks in a 3D space.

Step C: Measuring the Diameter (The "Shape Fitter")

Once the trunk is isolated, the computer slices the 3D model at "chest height" (1.3 meters) to get a cross-section. It then tries to guess the shape of that slice.

The Challenge: Tree trunks aren't perfect circles. They are lumpy, bumpy, and sometimes only partially visible because of leaves or other trees.
The Solution: The team tried three ways to fit a shape to the slice:
1. Ellipse: Trying to fit a stretched circle.
2. B-Spline: Trying to draw a smooth, flexible line through the dots.
3. Fourier Series: This is the winner. Think of this as a mathematical "wiggle" that can bend and twist to fit the lumpy, imperfect shape of a real tree trunk, even if parts of the trunk are missing or hidden.
The Result: The "Fourier" method was the most robust, acting like a flexible rubber band that snaps perfectly around the irregular shape of the tree, ignoring the noise and gaps.

3. The Results: How Good Is It?

The team tested this on 43 trees with 61 different scans.

Accuracy: Their method was off by about 5% to 9% compared to a human measuring with a tape measure.
Comparison: The expensive LiDAR method is slightly more accurate (off by only 2-4%), but the difference is small.
The Trade-off: You are trading a tiny bit of precision for a massive reduction in cost and complexity. Instead of needing a $20,000 laser scanner and a PhD to operate it, you can use a $500 camera and a backpack.

4. The "Secret Sauce" Visualization

They also built a cool interactive tool (like a video game viewer) where you can click on a tree in the 3D forest, and it instantly tells you the tree's ID and its estimated thickness. This helps humans double-check the computer's work.

Summary: Why Does This Matter?

This paper is about democratizing forestry.

Before: Only rich governments or big companies could afford to measure forests accurately.
Now: Anyone with a 360-degree camera can map a forest, estimate its carbon storage, and manage its health.

It's like going from needing a professional film crew to make a movie, to just using an iPhone. The quality isn't quite Hollywood-level, but it's good enough for most people, and it's accessible to everyone. This allows us to monitor our planet's forests more frequently, cheaply, and effectively.

1. Problem Statement

Forest inventories require accurate measurements of Diameter at Breast Height (DBH) (measured at 1.3m above ground) for carbon accounting, biomass estimation, and resource management.

Current Limitations: Traditional manual methods (tape/calipers) are labor-intensive and inefficient at scale. LiDAR-based techniques offer high precision (often <2 cm RMSE) but are cost-prohibitive, require specialized processing software, and need trained personnel, limiting their widespread adoption.
Goal: Develop a low-cost, scalable, and accessible alternative using consumer-grade hardware that achieves accuracy comparable to LiDAR without the associated costs or complexity.

2. Methodology

The authors propose a semi-automated pipeline that converts 360° video data into DBH estimates. The workflow consists of six main stages:

A. Data Acquisition

Hardware: Consumer-grade Insta360 360° camera mounted on either a drone (Falcon UAV) or a backpack-mounted tower.
Ground Truth: Manual measurements using diameter tapes and LiDAR data (Ouster sensor) collected simultaneously for baseline comparison.
Scaling: A 1m × 1m ArUco marker is placed in the scene to manually scale the 3D reconstruction to real-world metric units.

B. 3D Reconstruction (SfM)

Software: Agisoft Metashape is used to process the 360° video.
Process:
1. Alignment: Structure from Motion (SfM) estimates camera poses and intrinsic parameters from overlapping equirectangular images.
2. Dense Cloud: Multi-View Stereo (MVS) generates a dense 3D point cloud and a textured mesh.
3. Scaling: The model is scaled using the known dimensions of the ArUco marker.

C. Automated Tree Trunk Segmentation

To isolate tree trunks from the noisy forest environment (ground, canopy, understory):

Semantic Projection: Grounded SAM (Segment Anything Model) is used with the text prompt "tree trunk" to generate 2D segmentation masks on the equirectangular images.
3D Projection: These 2D masks are projected onto the 3D point cloud. Points receiving a "trunk" label from multiple views are retained.
Filtering:
- Ground Removal: RANSAC identifies and removes the ground plane.
- Normal Filtering: Points not parallel to the ground (assuming trees tilt <45°) are removed.
- Clustering: DBSCAN clusters the remaining points. Small clusters (noise) are discarded using Otsu's thresholding.
- Expansion: Clusters are expanded to recover points missed by initial projection.

D. DBH Estimation (Geometric Fitting)

Once a trunk is segmented, a cross-section is extracted at the 1.3m height (specifically a 30cm band from 1.22m to 1.52m). The authors compare three fitting methods within a RANSAC framework to handle outliers:

Least-squares Ellipse Fitting.
Periodic B-spline Fitting.
Truncated Fourier Series (Degree 2): Represents the curve parametrically in polar coordinates ( $r(t) = a_0 + \sum a_i \cos(it) + b_i \sin(it)$ ).

Selection: The Fourier Series method consistently yielded the lowest error and variance. The final DBH is calculated by dividing the fitted curve's perimeter by $\pi$ .

E. Visualization

A custom interactive tool (based on viser) allows users to inspect the 3D point cloud, view segmented trunks (highlighted in red), and click on specific trees to view their unique ID and estimated DBH.

3. Key Contributions

End-to-End Pipeline: A complete system transforming 360° video into DBH estimates using SfM, Grounded SAM for segmentation, and robust geometric fitting.
Cost-Effective Hardware: Demonstrates that a single consumer 360° camera (orders of magnitude cheaper than LiDAR) can achieve near-LiDAR accuracy.
Robust Fitting Algorithm: Identification of Truncated Fourier Series as the superior fitting method for noisy, partially observed cross-sections compared to ellipses or B-splines.
Interactive Tool: Development of a visualization interface for validation and analysis by forestry researchers.

4. Experimental Results

The system was evaluated on 61 samples of 43 trees across various species in Stadium Woods, Virginia Tech.

Accuracy:
- The method achieved median absolute relative errors of 5–9% compared to manual ground truth.
- This is only 2–4% higher than state-of-the-art LiDAR-based estimates (which typically achieve ~2–5% error).
Fitting Performance:
- Fourier Fitting: Consistently outperformed others, with a median error of 4.8% for fully observed trees and 8.1% for partially observed trees.
- Ellipse/B-Spline: Showed higher sensitivity to partial occlusion and noise, with errors rising significantly (up to 19.5% for ellipses in partial views).
Platform Comparison:
- Backpack (Walking): Lower error (6.2% median) due to better coverage and less motion blur.
- Drone: Higher error (11.5% median) due to motion blur and inability to fly close to foliage/branches.
LiDAR Comparison:
- When compared to a LiDAR-inertial odometry system using the same pipeline, the photogrammetry approach actually performed slightly better in some metrics, though the LiDAR "per-scan" method (cleaner data, no drift) remains the gold standard for sub-centimeter precision.

5. Significance and Conclusion

Accessibility: The proposed method democratizes forest inventory by replacing expensive, specialized LiDAR systems with widely available, low-cost 360° cameras.
Trade-off: While it sacrifices sub-centimeter precision (offering ~5–9% error vs. LiDAR's ~2–5%), the two orders of magnitude reduction in sensor cost and ease of deployment make it ideal for large-scale or resource-constrained forestry applications where sub-centimeter precision is not critical.
Future Work: The authors aim to automate the manual scaling step, validate on larger diverse forest stands, and extend the system to evaluate tree species and health.

In summary, this paper presents a viable, low-cost alternative to LiDAR for forest monitoring, proving that advanced computer vision techniques (SfM + Foundation Models + Robust Fitting) can bridge the gap between consumer hardware and professional forestry requirements.