💻 computer science

Counting Trees from Satellite Imagery with Noisy Supervision

This paper introduces a novel tree counting framework for satellite imagery that addresses the challenges of ambiguous crown boundaries and noisy supervision by formulating the task as a spatial density matching problem using Unbalanced Optimal Transport with a self-correction mechanism, achieving state-of-the-art performance on the new, large-scale TinyTrees benchmark.

Original authors: Dimitri Gominski, Maurice Mugabowindekwe, Qiue Xu, Xiaowei Tong, Martin Brandt, Hieu Le, Rasmus Fensholt, Dimitris Samaras, Loic Landrieu

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dimitri Gominski, Maurice Mugabowindekwe, Qiue Xu, Xiaowei Tong, Martin Brandt, Hieu Le, Rasmus Fensholt, Dimitris Samaras, Loic Landrieu

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 looking at a massive forest from a satellite high above the ground. Your goal is to count every single tree.

If the trees are far apart, like people standing in a park, it's easy: you just point and say, "That's one, that's two." But in a dense forest, the trees are packed so tightly that their leaves and branches overlap. From space, it looks like one giant green blob. You can't tell where one tree ends and another begins. Trying to draw a line around each tree is like trying to count individual grains of sand in a bucket of wet sand—you just can't do it precisely.

This is the problem the paper tackles: How do we count trees from space when the "rules" for what counts as a single tree are blurry, and we don't have enough human experts to draw perfect maps for every forest on Earth?

Here is how the authors solved it, using simple analogies:

1. The Problem: The "Noisy" Map

To teach a computer to count trees, you usually need a "teacher" (data) to show it the right answers.

The Perfect Teacher (Strong Labels): These are humans manually drawing dots on every tree. It's incredibly accurate, but it takes forever and costs a fortune. We only have a few of these maps.
The Noisy Teacher (Weak Labels): These are maps created automatically by lasers (LiDAR) flying over the forest. They cover huge areas, but they are messy. Sometimes they miss a tree; sometimes they think a bush is a tree. It's like having a student who knows the general area but keeps making small mistakes.

The challenge is: How do you train a smart AI using a messy teacher without letting the mistakes ruin the lesson?

2. The Solution: "Unbalanced Optimal Transport"

The authors invented a new way to teach the AI called Unbalanced Optimal Transport (UOT). Let's break that down with an analogy.

The "Moving Boxes" Analogy:
Imagine you have a pile of boxes (the trees the AI thinks it sees) and a pile of target spots on the floor (the trees the teacher says are there).

Old Way (Balanced Transport): You must move every single box to a target spot perfectly. If the teacher says there are 100 trees but the AI sees 105, the system panics because it can't match them 1-to-1. It forces the AI to be perfect, even when the teacher's map is wrong.
The New Way (Unbalanced Transport): The authors say, "It's okay if the numbers don't match exactly."
- If the AI sees a tree where the teacher didn't mark one, the system can say, "Okay, let's just create a little bit of mass there" (add a tree).
- If the teacher marked a tree but the AI doesn't see it, the system can say, "Let's destroy that mark" (remove the tree).
- The Cost: You can create or destroy trees, but it costs a little "penalty fee." The goal is to move the boxes to the right spots while paying the least amount of penalty fees possible.

This allows the AI to be flexible. It can handle the blurry, overlapping trees in dense forests without getting confused by the teacher's mistakes.

3. The "Self-Correction" Mechanism

The paper introduces a clever trick to make the "Noisy Teacher" better over time.

Imagine the AI is taking a test.

First Pass: The AI looks at the messy map and makes its best guess.
The Check: The system compares the AI's guess with the messy map. It finds the "residuals" (the differences).
- Example: The messy map says "Tree here," but the AI says "No tree there." The system flags this as a likely error in the map.
- Example: The messy map says "No tree," but the AI sees a clear tree shape. The system flags this as a likely missing tree.
The Fix: The system quietly updates the messy map to fix these specific errors before the AI learns from it again.
Repeat: The AI learns from this slightly cleaner map, gets better, and the cycle continues.

It's like a student who keeps correcting their own homework based on what they know, gradually turning a messy draft into a clean final copy.

4. The Result: TINYTREES

To prove this works, the authors built a massive new dataset called TINYTREES.

Scale: It covers 23,000 square kilometers (roughly the size of New Jersey) across three continents (China, Rwanda, France).
Data: It contains over 215 million tree annotations.
Performance: Their method (called TREEMATCH) beat all other existing methods. Whether the trees were sparse or packed tight, and whether the satellite images were high-resolution or a bit blurry, their system was the most accurate.

Summary

The paper says: Counting trees from space is hard because forests are messy and perfect data is expensive.

They solved this by using a math tool (Unbalanced Optimal Transport) that allows the AI to "create" or "delete" trees to match the map, rather than forcing a perfect 1-to-1 match.
They added a self-correcting loop that fixes the messy data as the AI learns.
They proved it works on a massive, new global dataset, showing that this approach is the most reliable way to count trees from space today.

The paper does not claim this can be used for medical diagnosis, predicting future climate change models, or specific commercial products yet; it strictly focuses on the technical achievement of counting trees accurately from satellite images.

Technical Summary: Counting Trees from Satellite Imagery with Noisy Supervision

Problem Definition

Counting individual trees from satellite imagery is a critical task for environmental monitoring, reforetracking, biomass estimation, and biodiversity assessment. However, this task faces significant technical hurdles at satellite resolutions:

Ambiguity in Dense Forests: While isolated trees are identifiable, tree crown boundaries in dense forests become ambiguous, making the concept of an "individual tree" ill-defined. Sensor blur and overlapping crowns further complicate delineation.
Annotation Scarcity and Cost: Large-scale manual annotation of individual tree centers is prohibitively expensive, limiting the availability of high-quality training data.
Noisy Supervision: Scalable supervision can be derived from Airborne Laser Scanning (ALS), but these annotations are inherently noisy, containing spurious detections and missed trees, particularly in dense canopies.

Existing methods, such as detection-based approaches (bounding boxes/points) or standard regression, struggle under these conditions. Detection methods fail when objects overlap heavily, while standard regression often degrades when trained on noisy, heterogeneous annotations.

Methodology: TREEMATCH

The authors propose TREEMATCH, a framework that formulates tree counting as a spatial density matching problem supervised via Unbalanced Optimal Transport (UOT).

1. Unbalanced Optimal Transport (UOT)

Unlike classical Optimal Transport (OT) which enforces strict mass conservation, UOT relaxes this constraint. This is crucial for tree counting where:

Annotations may be incomplete (missing trees).
Annotations may be noisy (spurious detections).
The total predicted count may differ from the annotated count.

The loss function minimizes the cost of transporting mass between the predicted density map $z$ and the annotated measure $y$ , while allowing for controlled mass creation or destruction penalized by a parameter $\tau$ . The cost matrix $C$ is based on squared Euclidean distances between pixel locations, encouraging the predicted density to align geometrically with annotations while tolerating small localization shifts.

2. Self-Correction Mechanism

To address the noise in large-scale weak labels (derived from ALS), the authors introduce a bootstrapping mechanism:

Marginal Residuals: During training, the UOT solution yields a transport plan $\gamma$ . The marginal residuals ( $r = \gamma^\top \mathbf{1} - y$ ) reveal discrepancies between the annotation and the model's prediction. Positive residuals suggest missed trees; negative residuals suggest over-annotation.
Progressive Refinement: These residuals are used to generate corrected annotations $\tilde{y} = y + \lambda(t)r$ . The correction weight $\lambda(t)$ increases from 0 to a maximum $\alpha$ during training. This allows the model to initially rely on raw annotations and progressively incorporate transport-consistent corrections without backpropagating through the correction step itself.

3. Hybrid Supervision Strategy

The framework leverages two data sources:

Strong Labels: Sparse, high-quality manual annotations (point centers). These supervise both global count consistency and spatial alignment.
Weak Labels: Large-scale, noisy annotations (tree-cover masks or ALS-derived points). These are corrected via the marginal residual mechanism and used to train the model on a massive scale.

Key Contributions

TINYTREES Benchmark: The authors introduce the first large-scale benchmark for satellite tree counting. It spans three continents (China, Rwanda, France) and three satellite sensors (Gaofen-2, PlanetScope, SPOT-6), covering over 23,000 km² with more than 215 million tree annotations (including 773K manually verified instances).
TREEMATCH Framework: A density-matching approach based on UOT that naturally handles heterogeneous counting regimes, enabling precise localization of isolated trees while remaining robust in dense canopies.
Noise-Robust Learning: A principled method to automatically correct weak annotations during training using transport residuals, enabling effective learning from large-scale noisy supervision alongside sparse high-quality labels.

Results

Evaluated on TINYTREES, TREEMATCH consistently outperforms:

Regression-based baselines: (Density regression, Count regression, Uncertainty-aware regression).
Detection-based baselines: (YOLOv8, CenterNet, P2P, TreeFormer).
Distribution-matching baselines: (DM-Count, VRSNet, Bayesian approaches).

Key Performance Indicators:

TREEMATCH achieves the lowest Root Mean Squared Error (RMSE) and Normalized Mean Absolute Error (nMAE) across all regions and sensors.
It demonstrates superior robustness to weak supervision compared to standard regression, which degrades significantly when trained on noisy ALS data.
Detection-based methods fail to converge or perform poorly in dense forest regions due to the inability to separate overlapping crowns.
The model generalizes reasonably well across regions when trained jointly, though performance varies based on sensor resolution and annotation quality (e.g., higher performance on Gaofen-2 data compared to PlanetScope).

Significance and Claims

The paper claims that unbalanced optimal transport provides a principled and flexible foundation for large-scale tree counting under real-world annotation constraints. By decoupling geometric alignment from exact mass conservation, the method effectively handles the inherent ambiguity of tree boundaries in dense forests and the noise inherent in automated labeling pipelines.

The work highlights that while detection-based methods are standard for distinct objects, they are ill-suited for satellite-based tree counting in heterogeneous forest regimes. Instead, formulating the problem as a distribution matching task with noise-aware supervision offers a scalable solution for environmental monitoring where precise manual annotation is impossible at the required scale. The authors position TINYTREES as a necessary resource to advance systematic evaluation and method development in this under-explored domain.