TL;DR
Researchers used Random Forest regression to model the Urban Heat Island (UHI) effect in Brussels and Liege, Belgium, bypassing computationally expensive climate simulations. The study demonstrates that converting flat rooftops to "intensive" green roofs could lower average land surface temperatures by up to 0.68°C across an entire city.
What Problem Does This Paper Address?
As cities grow denser, they turn into "heat islands"—urban areas that are significantly warmer than their rural surroundings due to heat-absorbing concrete and a lack of vegetation. This leads to higher energy consumption for cooling and increased health risks.
While "green roofs" (rooftops covered in vegetation) are a known solution, predicting their impact at a city-wide scale is difficult. Traditional micro-scale climate models (like ENVImet) are too computationally heavy for entire cities, while mesoscale models (like WRF) often lack the fine-grained resolution needed to see individual buildings. The authors address this gap by using a machine learning approach that combines high-resolution satellite data with 3D urban morphology to predict cooling outcomes quickly and accurately.
Methods
The team employed Random Forest (RF) Regression, a supervised learning algorithm that builds an ensemble of decision trees to predict a target variable—in this case, Land Surface Temperature (LST).
What sets this study apart is the inclusion of both 2D and 3D parameters. Beyond just looking at "how much green" is present, the model considers:
- Morphological Parameters: Building density, average height, and height variation.
- Sky View Factor (SVF): How much of the sky is visible from the ground (crucial for heat radiation).
- Frontal Area Index (FAI): A measure of how buildings obstruct wind flow.
- Distance to Water: Proximity to rivers like the Meuse or Senne.
To simulate green roofs, the researchers identified all flat roofs larger than 100 m². They then "digitally planted" vegetation by adjusting the NDVI (greenness index) and NDBI (built-up index) values for those specific pixels in their dataset and re-ran the model to see how the predicted temperature changed.
Datasets
The study utilized a blend of local GIS data and global satellite imagery:
- Building Footprints & Heights: Sourced from the PICC dataset (Liege) and UrbIS (Brussels).
- Satellite Imagery:
- Landsat-8: Used to derive the actual Land Surface Temperature (LST) at 30m resolution.
- Sentinel-2: Used to calculate NDVI and NDBI at 10m resolution.
- Training Points: 11,500 randomly generated points across both cities were used to train and test the RF model.
Key Insights and Ideas
- The Dominance of Indices: The most important predictors for urban temperature weren't building heights or wind indices, but rather NDBI (Built-up Index) and NDVI (Vegetation Index). This confirms that the material of the surface (concrete vs. plants) is the primary driver of the surface heat island.
- Wind Matters: The inclusion of the Frontal Area Index (FAI) provided a significant boost to the model's logic, as it accounts for how urban geometry traps or moves air.
- The "Intensive" Advantage: The study distinguished between "extensive" green roofs (thin soil, moss/grass) and "intensive" green roofs (thick soil, shrubs/gardens). The model showed that intensive roofs provide roughly double the cooling benefit of extensive ones.
- Non-Linear Accuracy: The model was highly accurate for temperatures between 21°C and 33°C but struggled with extreme outliers. This suggests that while RF is great for general urban planning, it might underpredict the severity of extreme heatwaves.
Key Results
- Model Performance: The RF model achieved an R-squared of 0.76 for Brussels and 0.63 for Liege, with a Root Mean Square Error (RMSE) of approximately 1.55°C to 1.97°C.
- Cooling Potential:
- In Liege, intensive green roofs (NDVI 0.8) reduced the average city-wide LST by 0.46°C.
- In Brussels, the same scenario reduced average LST by 0.48°C.
- When looking specifically at the building areas (rather than the whole city), the cooling effect was even more pronounced, reaching up to 0.68°C.
- Pixel Distribution: The researchers observed a significant shift in the temperature histograms; after adding green roofs, the number of "hot pixels" (32–35°C) dropped sharply.
Limitations and Future Work
The authors identify several areas for improvement:
- Extreme Temperatures: The model tends to underpredict temperatures at the very high and low ends of the spectrum. Adding data from more cities could help the algorithm learn these extremes.
- Multi-collinearity: Some variables (like building height and volume) are naturally correlated, which can muddy the "feature importance" rankings. Future work could use Principal Component Analysis (PCA) to clean this up.
- Neighborhood Effects: The current model changes the values of a roof but doesn't explicitly account for how that roof cools the air for the building next door. Incorporating a "neighborhood effect" parameter is a suggested next step.
- Flat Roof Constraint: The study only simulated greening on flat roofs > 100 m². Future iterations could explore the impact of greening slanted roofs or smaller residential structures.
Summary updated