Predicting Local Climate Zones using Urban Morphometrics and Satellite Imagery

This study evaluates the efficacy of using urban morphometrics, alone and fused with satellite imagery, to predict Local Climate Zones across five sites, finding that while morphometrics capture a broader range of urban form properties, the predictive relationship is site-dependent and inconsistent, suggesting the LCZ framework should be used with caution in morphological studies.

The Gist

Imagine you are trying to describe a city to a friend who has never seen it. You could take a photo from a helicopter (satellite imagery) and say, "Look, that area looks like a dense forest of tall buildings, while that one looks like a sparse park with small houses." This is how scientists usually map Local Climate Zones (LCZs)—they use satellite photos to guess what kind of weather a neighborhood might have based on how it looks from above.

But what if you didn't have a camera? What if you only had a detailed map of the streets and the outlines of every single building? Could you still guess the climate zone just by looking at the shapes and arrangement of the city?

This is exactly what Hugo Majer and Martin Fleischmann set out to test in their new study. They asked: Can we predict a city's climate zone just by measuring its "skeleton" (the buildings and streets), or do we absolutely need the "skin" (the satellite photos)?

Here is a simple breakdown of their adventure:

1. The Two Tools: The "Skeleton" vs. The "Skin"

• The Skin (Satellite Imagery): This is the standard method. Scientists use AI to look at satellite pictures (like Google Earth) and learn to recognize patterns. "Oh, that patch of green and gray means it's a park. That dense grid of gray means it's a downtown."
• The Skeleton (Urban Morphometrics): This is the new approach. Instead of looking at colors, they used math to measure the city's structure. They counted 321 different things, like:
  • How long are the building walls?
  • How far apart are the houses?
  • Are the streets straight or curvy?
  • Are the buildings square or weirdly shaped?
  • Think of this as measuring the city's DNA rather than taking a selfie.

2. The Experiment: Five Cities, Four Guessing Games

They tested their ideas in five very different cities: Berlin, Hong Kong, Paris, Rome, and São Paulo. They set up four different "guessing games" (classification schemes):

1. The Skeleton Only: Guessing the climate zone using only the math measurements of buildings and streets.
2. The Skin Only: The standard method using only satellite photos (the baseline).
3. The Hybrid (Stacking): Taking the math measurements and literally stacking them on top of the satellite photo like extra layers of a sandwich, then letting the AI guess.
4. The Hybrid (Fusion): A more complex mix where the AI looks at the photo, extracts its "brain" (features), and then combines that with the math measurements before guessing.

3. The Results: A Mixed Bag

Here is what they found, using some analogies:

• The "Skeleton Only" approach was a hit-or-miss.

  • The Good: In some cities, just knowing the building shapes was enough to tell the difference between a "dense city center" and a "spread-out suburb." It was like recognizing a person just by their silhouette.
  • The Bad: In other cities, the math failed. It couldn't tell the difference between a "tall, open area" and a "short, open area," or between an "industrial zone" and a "residential zone." It was like trying to guess someone's job just by their shoe size; sometimes it works, but often it's a wild guess.
  • The Verdict: The math is useful, but it's not a magic crystal ball. It depends heavily on the specific city.

• The "Hybrid" approach (Skin + Skeleton) was... okay, but not a miracle.

  • In two cities (Hong Kong and Rome), combining the photo with the math measurements made the AI significantly smarter. It was like giving a detective both a fingerprint and a witness description.
  • However, in the other three cities, adding the math measurements didn't help much. In fact, sometimes it made the AI more confused! It's like trying to solve a puzzle by adding extra pieces that don't quite fit; sometimes they help, but often they just clutter the picture.

4. The Big Takeaway

The researchers concluded that Local Climate Zones are a bit of a tricky concept.

They found that the relationship between a city's physical shape (the math) and its climate zone label is "tenuous" (weak and shaky). Just because two neighborhoods look different on a map doesn't always mean they have different climate zones, and vice versa.

The Lesson for the Future:
If you are a city planner or a scientist studying how cities affect the weather, don't rely solely on the LCZ labels. They are a bit like a "one-size-fits-all" t-shirt; they might fit some cities perfectly, but for others, they are too loose or too tight.

The study suggests that while we can use math to describe cities, we need to be careful not to oversimplify. The LCZ framework is a useful tool, but it shouldn't replace a deep, detailed look at how a city is actually built.

In a Nutshell

• Can we map cities using just math? Yes, but it's inconsistent.
• Does adding math to satellite photos help? Sometimes, but not always.
• What's the main warning? Don't trust the "Local Climate Zone" labels too blindly. They are a simplification of a very complex reality, and using them as a substitute for real urban analysis can lead to misunderstandings.

The authors are essentially saying: "The map is not the territory." Just because we can measure the city's bones doesn't mean we fully understand its climate soul.

Technical Summary

1. Problem Statement

The Local Climate Zone (LCZ) framework is a standard typology for representing urban form in climate and morphological analyses. However, current LCZ mapping relies predominantly on satellite imagery (Remote Sensing, RS), which captures visible surface properties but may overlook the underlying geometric and spatial relationships of urban elements. While Urban Morphometrics (quantitative measures of urban form) offers a rigorous way to describe urban structure, its potential to predict LCZs independently or in fusion with imagery remains under-explored.

Existing studies utilizing urban form descriptors for LCZ prediction have been limited by:

• Using small, isolated sets of metrics rather than comprehensive morphometric descriptions.
• Lack of robust evaluation across diverse geographic sites.
• Inconsistent findings regarding whether adding morphometric data improves upon standard RS-based classification.

The study aims to evaluate the efficacy of 2D urban morphometrics in predicting LCZs and determine if fusing morphometric data with satellite imagery enhances classification accuracy compared to RS-only baselines.

2. Methodology

The study employs a four-step framework evaluated across five diverse cities (Berlin, Hong Kong, Paris, Rome, São Paulo) using reference data from the 2017 IEEE GRSS Data Fusion Contest.

A. Data Preparation

• Morphometric Data: Derived from Overture Maps (2025) building footprints and street networks. Using the momepy Python library, the authors calculated 321 contextual 2D morphometric attributes per Enclosed Tessellation Cell (ETC).
  • Input: 107 primary attributes covering dimension, shape, spatial distribution, intensity, and connectivity across multiple scales (element, neighbors, neighborhood).
  • Contextualization: Percentiles (25th, 50th, 75th) were computed across topological neighbors to capture spatial patterns.
• Satellite Imagery: Cloud-free Sentinel-2 Level-2A scenes (2017/2018) with 10 spectral bands (resampled to 10m).
• Reference Data: Ground-truth LCZ polygons (LCZs 1–10 for urban, A–G for natural).

B. Classification Schemes

Four distinct schemes were designed and compared:

1. S1 (Morphometric-Based): Classifies ETCs into urban LCZs (1–10) using Random Forest (RF) based solely on the 321 morphometric attributes.
2. S2 (RS-Based Baseline): A scene-based Convolutional Neural Network (MSMLA-50) classifying 320×320m image patches. This serves as the state-of-the-art baseline.
3. S3 (Fusion - Stacked): Rasterizes the top 20 morphometric attributes (identified by S1) and stacks them as additional bands with Sentinel-2 imagery. Classified via RF.
4. S4 (Fusion - Feature Concatenation): Extracts deep features from the MSMLA-50 CNN (pre-classification layer) and concatenates them with aggregated morphometric descriptors (mean, min, max, std, median). The fused vector is classified via RF.

C. Evaluation

• Metrics: Weighted F1-score (F1), Overall Accuracy (OA), and class-specific scores for Urban (F1U) and Natural (F1N) types.
• Validation: Stratified 5-fold cross-validation per site to account for spatial heterogeneity and class imbalance.

3. Key Contributions

• Comprehensive Morphometric Application: This is the most exhaustive application of urban morphometrics for LCZ mapping to date, utilizing 321 attributes rather than a handful of isolated metrics.
• Dual Fusion Strategies: The study compares two distinct fusion techniques (band stacking vs. deep feature concatenation) to determine the optimal method for integrating geometric and spectral data.
• Robust Multi-Site Evaluation: Unlike previous studies limited to one or two cities, this research evaluates performance across five geographically and culturally distinct cities, providing a more rigorous test of generalizability.
• Critical Re-evaluation of LCZ Framework: The study challenges the assumption that LCZs are purely morphological, highlighting the tenuous link between measurable 2D form and LCZ typology.

4. Results

Morphometric-Only Prediction (S1)

• Performance: Highly variable and site-dependent. F1 scores ranged from 64.2% (Rome) to 92.5% (Paris).
• Capabilities: Successfully distinguished compact (LCZ 1–3) vs. open (LCZ 4–6) arrangements and large low-rise (LCZ 8) vs. open low-rise (LCZ 6).
• Limitations: Poor at distinguishing building heights within the same density class, failed to identify sparse built (LCZ 9) vs. open low-rise (LCZ 6), and struggled significantly with industrial areas (LCZ 10).
• Stability: The model showed high instability, with performance fluctuating wildly between cross-validation folds, indicating sensitivity to training data composition.

Fusion vs. RS Baseline (S2 vs. S3/S4)

• Overall Impact: Fusion yielded modest and inconsistent improvements.
  • Improvements: Notable gains in Hong Kong (+4% F1U) and Rome (+6.4% F1U).
  • Negligible/Decline: Gains in Paris and Berlin were minimal (~1%), while São Paulo saw a decline in performance.
• Method Comparison:
  • S3 (Stacked) performed better for Urban LCZs (F1U) in 3/5 sites.
  • S4 (Feature Concatenation) performed better for Natural LCZs (F1N) in 3/5 sites.
• Specific Class Behavior:
  • Heavy Industry (LCZ 10): S3 significantly improved detection across all sites compared to the baseline, whereas S4 generally degraded performance.
  • Sparse Built (LCZ 9): Fusion did not resolve the systematic misclassification of LCZ 9 as LCZ 6; in some cases, it worsened the confusion compared to the RS baseline.
  • Natural Types: Fusion often increased confusion between natural classes (e.g., scattered trees vs. dense trees) and did not consistently resolve urban-natural confusion.

5. Significance and Conclusion

The study concludes that the relationship between 2D urban morphometrics and LCZ typology is selective and inconsistent.

• Implications for Morphometrics: While a broad range of morphometric properties (shape, connectivity, intensity) are relevant for distinguishing LCZs, 2D metrics alone are insufficient for robust, universal LCZ mapping. They work best in sites dominated by clear density contrasts (compact vs. open) but fail in complex scenarios involving height variability or industrial zones.
• Implications for Data Fusion: Combining morphometrics with satellite imagery offers only marginal benefits over advanced RS-based CNNs. The gains are not universal and depend heavily on the specific urban context and the fusion technique used.
• Conceptual Warning: The authors argue that the LCZ framework should be used with caution in morphological studies. The framework's reliance on a simplified typology may not fully capture the complexity of urban form, and using LCZs as a proxy for morphological analysis risks oversimplification and reduced comparability across different cities.

Future Directions: The authors suggest exploring 3D building metrics (height, volume) to complement 2D morphometrics and developing transferable subsets of morphometric attributes to improve model stability.