📄 earth science

Novel Use of Local Climate Zones: Urban Form-based Method for Carbon Sink Assessment using Convolutional Neural Network

This study proposes a novel urban form-based method for estimating city carbon sinks by integrating Local Climate Zone (LCZ) maps with a Convolutional Neural Network (CNN) model, demonstrating that lower-intensity built environments (LCZ 5, 6, and 9) offer greater carbon sequestration potential while achieving high estimation accuracy despite sensitivity to LCZ map precision.

Original authors: Tsz-Kin Lau, Li-Tzu Yu, Tzu-Ping Lin

Published 2026-09-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tsz-Kin Lau, Li-Tzu Yu, Tzu-Ping Lin

Original paper licensed under CC BY 4.0 (https://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

Cities are constantly changing, growing taller and denser as more people move in. This expansion often pushes out the green spaces that naturally clean our air, creating a pressing need to understand how much carbon these urban environments can still absorb. To tackle this, scientists have developed a way to categorize the physical shape of a city into distinct neighborhoods based on how buildings and land are arranged. They call these categories "Local Climate Zones," a system that groups areas by their height, density, and surface materials, ranging from compact high-rise districts to open low-rise areas and dense forests. Alongside this, the concept of a "carbon sink" refers to the natural ability of vegetation, like trees and grass, to pull carbon dioxide out of the atmosphere and store it. As cities strive to meet global goals for reducing greenhouse gases, figuring out exactly how much carbon their remaining greenery can capture has become a vital task for urban planners and environmentalists.

A team of researchers recently set out to find a faster, simpler way to measure this carbon-capturing potential within complex cityscapes. Instead of relying on slow, detailed surveys of every single tree, they combined the city zoning maps with a type of computer program known as a convolutional neural network. Think of this program as a highly trained eye that can look at a map of city zones and instantly recognize patterns to predict how much carbon is being stored. The team tested this approach in two very different Taiwanese cities: Taipei, a bustling metropolis with a high density of tall buildings, and Chiayi, a smaller city with a mix of low and mid-rise structures. They first created detailed maps showing where trees and green spaces existed in both cities, calculating the exact amount of carbon each area could absorb. Then, they fed these maps into their computer model, teaching it to link the visual patterns of the city zones to the carbon data.

The results revealed a clear and somewhat surprising connection between the shape of a city and its ability to store carbon. The study found that the most effective carbon sinks were not always the densest forests, but also included specific types of built-up areas. In particular, neighborhoods with lower-rise buildings and more open space, such as those with scattered low-rise structures or sparsely built areas, showed a high potential for carbon storage. This suggests that when cities are less intensely developed, there is simply more room for vegetation to grow and thrive, allowing trees and shrubs to capture more carbon. The computer model proved remarkably successful at learning these relationships in Taipei, where it could predict carbon levels with high accuracy, matching the real-world data closely. However, the model struggled when applied to Chiayi, largely because the maps used to describe the city's zones contained errors. In Chiayi, some natural areas were mistakenly labeled as built-up zones on the map, which confused the computer and led to inaccurate predictions.

This mismatch highlighted a crucial lesson: the success of this new method depends entirely on the accuracy of the underlying city maps. If the map says a forest is a parking lot, the computer will make the wrong guess about the carbon storage. Despite this limitation, the study demonstrates that using these zoning maps combined with smart computer algorithms offers a powerful, cost-effective tool for estimating carbon sinks across entire cities. The researchers concluded that while the method works best when the city data is precise, it provides a much simpler alternative to traditional, labor-intensive measurement techniques. By understanding which urban forms support the most green growth, city planners can make better decisions to preserve or create spaces that help fight climate change, ensuring that even as cities grow, they continue to breathe for the planet.

Technical Summary: Novel Use of Local Climate Zones for Urban Carbon Sink Assessment

Problem Statement
Global warming necessitates strategies for carbon emission reduction, with carbon sinks (specifically vegetation-based carbon sinks, or $CS_v$ ) playing a critical role. While oceans and forests are primary sinks, urbanization often destroys natural landscapes, reducing a city's capacity to sequester carbon. Current methods for estimating urban $CS_v$ often require complex calculations, detailed vegetation inventories, or high-resolution digital models, making them difficult to apply broadly. There is a need for a simplified, cost-effective method to quantify the relationship between urban form and carbon sequestration to aid urban planning and net-zero emission targets.

Methodology
This study proposes a novel framework combining Local Climate Zones (LCZ) and Convolutional Neural Networks (CNN) to estimate $CS_v$ without requiring detailed vegetation data.

Study Areas and Data: The research focused on two Taiwanese cities with varying urbanization intensities: Taipei City (high density) and Chiayi City (lower density).
- LCZ Mapping: Global LCZ maps (100m resolution) were obtained for both cities to quantify urban form. These maps classify environments into 17 classes (10 built, 7 natural).
- Carbon Sink Calculation ( $CS_v$ ): A 3D vegetation model was generated using SPOT 6 imagery, K-means clustering, and the ETH Global Sentinel-2 Canopy Height map. Based on Taiwan's Green Building Guidelines, $CS_v$ was calculated by categorizing vegetation into big trees, small trees, shrubs, and green fields, assigning specific sequestration rates (kgCO2e/m²/yr) to each.
- Data Preparation: The calculated $CS_v$ maps were resampled to a 500m grid. The LCZ maps were resampled to 10m and divided into 50m × 50m grids (2,500 m²) to serve as input images for the CNN.
Model Architecture: A standard CNN model was constructed with 4 convolutional layers, 4 pooling layers, and 1 flattened layer.
- Training: The model was trained on LCZ grid inputs to predict the corresponding $CS_v$ values. Training was conducted with sampling percentages ranging from 50% to 80%.
- Evaluation: Performance was assessed using the Coefficient of Determination ( $R^2$ ), Mean Absolute Error (MAE), Cosine Similarity, and Pearson Correlation. The models were tested on both "seen" data (same city) and "strange" data (cross-city application).

Key Results

Urban Form and Carbon Sink: Statistical analysis revealed that built environments with lower urbanization intensity—specifically LCZ 5 (Open mid-rise), LCZ 6 (Open low-rise), and LCZ 9 (Sparsely built)—have a high potential for contributing to $CS_v$ . This is attributed to the availability of space for vegetation growth. Natural landscapes (LCZ 11, 12, 14) also showed high contribution, consistent with expectations.
CNN Performance (Taipei): The CNN model trained on Taipei City data performed strongly, achieving an $R^2$ score over 0.75 and a low MAE (approximately 29,000 kgCO2e/yr) with 50% sampling. The model demonstrated a high correlation (Pearson ~0.999) between observed and predicted values within the same city.
CNN Performance (Chiayi): The model trained on Chiayi City data performed poorly, with negative or low $R^2$ scores and high MAE. The authors attribute this to misclassifications in the LCZ map for Chiayi (e.g., natural landscapes incorrectly classified as built zones), which led the model to learn unreliable relationships between urban form and carbon sink.
Cross-City Applicability: When models trained on one city were applied to the other ("strange datasets"), performance dropped significantly (low $R^2$ , high MAE). Correlation analysis confirmed that Taipei and Chiayi have distinct spatial patterns of $CS_v$ and LCZ, indicating that the models learn city-specific relationships rather than generalized global patterns.

Key Contributions

Novel Application of LCZ: The study moves beyond the traditional use of LCZ for urban climate and heat island studies, proposing its use as a proxy for quantifying urban form to estimate carbon sinks.
Simplified Estimation Method: The proposed CNN-LCZ method offers a simpler alternative to traditional $CS_v$ assessment, as it does not require detailed, ground-level vegetation inventories or complex 3D modeling for the prediction phase, relying instead on LCZ maps and remote sensing.
Quantified Urban Form Impact: The research empirically demonstrates that lower-density built environments (LCZ 5, 6, 9) can significantly contribute to urban carbon sinks, providing actionable insights for urban planning.

Significance and Claims
The authors claim that this study presents a novel, cost-effective, and efficient method for estimating urban carbon sinks. By leveraging the relationship between LCZ patterns and $CS_v$ , the method can assist in urban planning to ensure development maintains a high level of carbon sequestration.

However, the paper maintains a modest tone regarding its limitations. It explicitly states that the method's reliability is heavily dependent on the accuracy of the LCZ map. In the case of Chiayi, where the LCZ map contained classification errors, the estimation results were unreliable. Therefore, the authors conclude that while the method is feasible, it requires high-quality, locally validated LCZ maps and city-specific model training to be effective. The study does not claim the method is universally applicable without local calibration but rather highlights its potential as a tool for cities with accurate environmental quantification.

Technical Summary: Novel Use of Local Climate Zones for Urban Carbon Sink Assessment

More like this