⚡ electrical engineering

Self-Supervised Tree-level Biomass Estimation in Urban Environments From Airborne LiDAR and Optical Observations

This paper presents a self-supervised deep learning framework that integrates airborne LiDAR and optical imagery to generate a high-resolution, bitemporal database of individual urban tree biomass across an 810 km² landscape in Ontario, achieving accurate crown-level estimation without manual annotation and revealing a net carbon gain of 39 Gg C over five years.

Original authors: Jose Bermudez (McMaster University, Hamilton, Ontario, Canada), Zilong Zhong (McMaster University, Hamilton, Ontario, Canada), Dominic Cyr (, Environment and Climate Change Canada, Montreal, Quebec, C

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jose Bermudez (McMaster University, Hamilton, Ontario, Canada), Zilong Zhong (McMaster University, Hamilton, Ontario, Canada), Dominic Cyr (, Environment and Climate Change Canada, Montreal, Quebec, Canada), Camile Sothe (Planet Labs PBC, San Francisco, California, USA), Alemu Gonsamo (McMaster University, Hamilton, Ontario, Canada)

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 a giant, invisible forest growing not in the wild, but right in the middle of our cities, suburbs, and towns. These are the trees lining our streets, sitting in our parks, and dotting our neighborhoods. For a long time, counting how much "wood" (biomass) and carbon these urban trees hold has been like trying to count grains of sand on a beach while wearing thick gloves: we could guess the total, but we couldn't see the individual grains, and our maps were often blurry.

This paper presents a new, high-tech way to count and weigh these city trees, one by one, across a massive area of Southern Ontario (about 810 square kilometers). Here is how they did it, explained simply:

1. The "Winter Photo" Strategy

Usually, when you take a photo of a tree in winter (when it has no leaves), it looks like a bare, gray skeleton. This makes it hard for computers to tell the difference between a tree branch and a building roof.

The Paper's Trick: The researchers used two types of "eyes" to look at the city at the same time:
- Eyes that see height (LiDAR): This is like a 3D laser scanner that bounces light off the ground and trees to build a precise 3D model. It sees the shape of the tree, even without leaves.
- Eyes that see color (Optical): These are high-resolution photos taken in winter.
The Analogy: Think of it like trying to identify a person in a dark room. If you only have a flashlight (the photo), you might miss them. But if you also have a sonar device (the laser) that tells you exactly where their body is, you can find them easily. By combining the "shape" from the laser and the "color" from the photo, the computer can spot trees even when they are bare.

2. Teaching the Computer Without a Teacher

Normally, to teach a computer to recognize trees, humans have to spend years drawing outlines around thousands of trees in photos. This is slow and expensive.

The Paper's Trick: They used a "Self-Supervised" approach. Instead of hiring a teacher, they gave the computer a set of rough, rule-based guesses (like "if it's flat and high, it's a building; if it's bumpy and green, it's a tree"). These guesses were imperfect, like a student's first draft.
The Analogy: Imagine a student learning to draw by looking at a messy sketchbook. The computer started with these messy, rule-based sketches and then practiced millions of times, correcting its own mistakes until it learned to draw the trees perfectly on its own. It didn't need a human to draw every single tree for it.

3. Separating the "Forest" from the "City"

One of the hardest parts of urban mapping is telling the difference between a tree and a house. A flat roof can look like a tree canopy from above.

The Paper's Trick: The computer learned to be a very strict bouncer. It used the 3D laser data to see that buildings are flat and smooth, while trees are bumpy and irregular.
The Result: The computer became very good at this, correctly identifying buildings 95% of the time and separating them from trees. This is crucial because if you mistake a roof for a tree, you might try to weigh a house!

4. Weighing the Trees

Once the computer found the trees and drew a circle around each one (delineation), it needed to guess how heavy they were.

The Paper's Trick: They couldn't climb every tree to measure it. Instead, they used a "recipe" (an allometric equation). They knew that bigger crowns and taller trees generally mean more wood. They calibrated this recipe using a list of 112,000 trees from Oakville that had already been measured by city workers.
The Catch: The computer is great at finding the trees, but it's not perfect at drawing the exact edge of every single crown. The paper found that the biggest source of error wasn't the "recipe" for weighing the trees, but rather the computer sometimes missing the exact edge of a tree's crown (especially for big, messy trees).

5. The Big Picture Results

After running this system over the 810 km² area for two different years (2018 and 2023), they found:

Total Weight: In 2018, the trees held about 1.73 million tons of biomass. By 2023, that grew to 1.81 million tons.
Carbon Gain: The trees absorbed an extra 39,000 tons of carbon over those five years.
Where the Trees Are: The heaviest trees were found along the Niagara Escarpment (a rocky cliff area) and in river valleys. The lighter, sparser trees were in residential neighborhoods and commercial areas.
Uncertainty Maps: The computer also created a "confidence map." It highlighted areas where it was less sure of its answers (mostly in agricultural fields or mixed areas), telling future researchers, "Check these spots first."

Summary

This paper built a smart, self-teaching computer system that uses winter laser scans and photos to find, outline, and weigh individual trees in a city without needing humans to draw every single one. It proved that we can get a very detailed, tree-by-tree map of urban biomass, showing us exactly where our city forests are growing and how much carbon they are storing, which helps cities plan better for the future.

Technical Summary: Self-Supervised Tree-level Biomass Estimation in Urban Environments

Problem Statement
Urban trees constitute a critical component of terrestrial carbon stocks, yet their biomass remains less spatially explicit and consistent than that of managed forests. Existing estimates often rely on sample inventories or coarse remote sensing products (e.g., 10–30 m resolution), which fail to resolve individual tree crowns or capture fine-scale heterogeneity in complex urban fabrics. Furthermore, scalable mapping is hindered by three methodological gaps: (1) the reliance on pixel-level manual annotation for supervised training, which limits coverage; (2) the unclear relative contributions of automated crown delineation errors versus allometric model errors to total uncertainty; and (3) the scarcity of multi-temporal, crown-scale biomass products accompanied by spatially explicit uncertainty layers.

Methodology
The study presents a self-supervised, multi-modal fusion framework applied to an 810 km² urban–rural landscape in southern Ontario, Canada, using bi-temporal (2018 and 2023) airborne LiDAR (8–10 pulses m⁻²) and near-infrared RGB orthophotography (0.16–0.20 m) acquired under leaf-off conditions.

Pseudo-Label Generation: To eliminate the need for manual annotation, the authors developed a rule-based framework to generate noisy pseudo-labels. This involved:
- Building Detection: Raster-based planarity analysis on the Canopy Height Model (CHM) combined with morphological filtering.
- Vegetation Classification: Differentiating needleleaf and deciduous trees using CHM height thresholds and spectral indices (NDVI and HSV color criteria) under leaf-off conditions, where needleleaf crowns retain high NDVI while deciduous crowns do not.
Self-Supervised Neural Network: A dual-stream cross-attention ConvNeXtV2 U-Net was trained on these pseudo-labels. The architecture processes optical (NIR, R, G, B) and structural (CHM) inputs through parallel encoders, utilizing cross-modal attention blocks to fuse features.
- Loss Function: The model employed a multitask objective combining land-cover segmentation, tree-type classification, and auxiliary input reconstruction. Crucially, it used an Active Negative Loss (ANL-CE) to mitigate overfitting to the noisy pseudo-labels.
- Uncertainty Quantification: A deep ensemble of 25 models (varying folds and hyperparameters) was used to generate per-pixel uncertainty maps, distinguishing aleatoric (data) and epistemic (model) uncertainty.
Crown Delineation: Individual tree crowns were delineated using multiscale watershed segmentation on the CHM within identified tree areas. Three scales were applied to address over-segmentation of small trees and under-segmentation of large, overlapping crowns.
Biomass Estimation: A power-law proxy model ( $AGB = e^\alpha \cdot (CA \cdot H)^\beta$ $A GB = e^{α} \cdot (C A \cdot H)^{β}$ ) was calibrated against species-specific allometric equations (Lambert et al., 2005) using 21,921 inventory trees from Oakville.
- Equation Selection: Crowns with low uncertainty were assigned functional-group-specific equations (deciduous vs. needleleaf); uncertain crowns were routed to a pooled "all-trees" equation to prevent unreliable extrapolation.
- Validation: Accuracy was assessed against a held-out test set of 90,726 trees, comparing predictions derived from inventory geometry (upper bound) versus operational watershed segmentation.

Key Results

Segmentation Performance: The ensemble model improved upon the initial rule-based pseudo-labels across all metrics. Against manually annotated tiles, the model achieved a mean precision of 0.86, recall of 0.83, and Dice coefficient of 0.84. Building detection was particularly robust (Dice 0.95), while needleleaf segmentation remained the most challenging class (Recall 0.66).
Biomass Estimation Accuracy:
- When using inventory-derived crown geometry (isolating allometric error), the model achieved $R^2 = 0.609$ .
- Under operational conditions using automated segmentation, $R^2$ decreased to 0.570.
- Needleleaf estimates were more accurate ( $R^2 = 0.646$ ) than deciduous estimates ( $R^2 = 0.553$ ), attributed to the clearer height peaks of needleleaf canopies in CHM data.
- The drop in performance from inventory to operational geometry identified crown delineation error as the dominant source of uncertainty, particularly for large trees which are prone to under-segmentation.
Spatial and Temporal Findings:
- Total Above-Ground Biomass (AGB) stocks increased from 1.73 Tg (811 Gg C) in 2018 to 1.81 Tg (850 Gg C) in 2023, representing a net carbon gain of 39 Gg C.
- High biomass densities (~140 Mg ha⁻¹) were concentrated along the Niagara Escarpment and riparian corridors, while residential and commercial areas showed lower densities.
- Deep-ensemble uncertainty maps successfully identified regions of high epistemic uncertainty (e.g., agricultural and transitional zones) where training data was underrepresented.

Significance and Claims
The authors claim this work provides a scalable, prototype workflow for urban biomass estimation that operates exclusively on standard provincial sensor data without manual annotation. Key contributions include:

Scalability: Demonstrating that rule-based pseudo-labeling combined with neural network refinement can achieve high accuracy in heterogeneous urban environments, overcoming the bottleneck of manual labeling.
Error Characterization: Explicitly quantifying that crown delineation errors, rather than allometric model errors, are the primary limitation in operational AGB estimation, with a specific focus on the under-segmentation of large trees.
Uncertainty-Aware Mapping: Providing a bi-temporal, crown-level database with an associated uncertainty layer, allowing users to distinguish between areas of high confidence and those requiring field validation.
Carbon Accounting: Offering a spatially explicit alternative to coarse-resolution or sample-based methods, yielding a publicly available database relevant to municipal management and regional carbon accounting for "trees outside forests."

The study maintains a modest tone regarding absolute biomass values, acknowledging that the use of forest-derived allometric equations for open-grown urban trees may introduce systematic bias. However, the authors assert that the relative spatial and temporal patterns remain robust, providing a foundational step toward more consistent urban carbon accounting.