📄 earth science

Hot School Grounds, Mismatched Evidence: Facility-Scale Urban Heat in Barcelona and Madrid

This study introduces a facility-aligned screening framework using Landsat 8 data to assess urban heat across 434 public primary schools in Barcelona and Madrid, revealing significant inter-city temperature disparities and warming trends while demonstrating that Extra Trees models effectively predict school-ground surface temperatures when validated at the campus level rather than the pixel level.

Original authors: Jiarui Shen

Published 2026-09-08

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiarui Shen

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 getting hotter, a reality that has moved from a distant climate warning to a daily design challenge for urban planners. When the sun beats down on a city, the ground, buildings, and pavement absorb that energy and release it as heat, creating pockets of intense warmth that can be significantly hotter than the surrounding countryside. To understand this phenomenon, scientists often look at the "skin temperature" of the city—the heat radiating directly from the surface of the ground and roofs—which can be measured from space using satellites. They also use a system called Local Climate Zones, which acts like a standard vocabulary for describing different types of neighborhoods, from dense high-rises to open low-rise areas. While these tools are excellent for mapping heat across entire regions, a critical gap remains: city budgets and maintenance crews do not manage abstract regions or pixels on a map. They manage specific, bounded places like schoolyards, parks, and plazas. The question of how to translate broad, satellite-based heat maps into actionable plans for individual public assets has been a persistent bottleneck in urban planning.

A new study by Jiarui Shen, a researcher at the University of Chicago, tackles this mismatch by focusing on a specific, tangible asset: the public primary school. The researcher examined school grounds in two major Spanish cities, Barcelona and Madrid, to see how heat behaves at the facility scale. Instead of treating the school as just another patch of pixels in a larger city-wide image, the researcher treated each school campus as a distinct unit of management. They gathered data over a decade, from 2014 to 2024, looking at the surface temperatures of 158 schools in Barcelona and 276 schools in Madrid during the hottest months of July and August. By using satellite imagery that captures the ground temperature, they were able to track not just how hot a school was in a single year, but whether that heat was getting worse over time.

The findings reveal a clear difference between the two cities. The median summer surface temperature for school grounds in Madrid was 37.2 degrees Celsius, while in Barcelona it was 31.8 degrees Celsius. Beyond just the current heat levels, the study looked at the trajectory of warming. In Madrid, the heat on school grounds was rising faster, with an average increase of 0.18 degrees Celsius per year, compared to 0.09 degrees Celsius per year in Barcelona. Furthermore, a larger share of schools in Madrid showed a statistically significant, steady increase in heat over the decade, suggesting that the problem is accelerating in that city more than in the other. These results highlight that while both cities face heat challenges, the intensity and the speed of change are distinct, requiring different levels of urgency and attention.

To understand what drives these temperature differences, the researcher tested various factors that could influence the heat on a school campus. They looked at how much green vegetation was present, how much moisture was in the soil and plants, how much hard, paved surface existed, and how much bare, dry ground was exposed. They also measured how much sunlight the surfaces reflected. Using advanced computer models, they analyzed which of these factors were the strongest predictors of surface heat. The models showed that moisture and vegetation were the most powerful indicators. Specifically, the amount of moisture in the soil and plants, followed closely by the amount of green vegetation, were the top factors associated with cooler surfaces. Conversely, hard, paved surfaces and exposed dry ground were linked to higher temperatures.

A crucial part of this study was how the researcher tested their computer models. Often, studies train a model on some data and then test it on other data from the same location, which can give a falsely optimistic view of how well the model works. This study took a different approach: they trained the model on a set of schools and then tested it on entirely different schools that the model had never seen before. This method, known as school-grouped validation, is a stricter test because it mimics the real-world scenario of applying a solution to a new facility. When they used this stricter test, the model's performance dropped slightly compared to looser tests, but it provided a much more honest and reliable estimate of how well the findings could be applied to new schools. The best-performing model in this strict test was able to predict the heat levels of unseen schools with a high degree of accuracy, confirming that the relationship between moisture, vegetation, and heat is robust enough to be used for planning.

The ultimate goal of this research is not to prove that planting a tree will lower the temperature by a specific number of degrees, but to provide a reliable screening tool for city officials. The study concludes that the most effective way to use this evidence is to prioritize schools for further, on-the-ground inspection based on their moisture and vegetation levels. If a school has low moisture in its soil or very little green cover, it should be flagged as a high priority for a detailed audit. This audit would then involve field measurements and local experts to determine exactly what interventions, such as adding water features, planting trees, or changing paving materials, would be most effective. The study explicitly avoids claiming that the satellite data alone can solve the problem or that it measures the actual heat felt by children. Instead, it offers a practical, evidence-based filter to help cities decide which of their many public assets need immediate attention and resources.

By aligning the scientific measurement with the actual unit of management—the school campus—this research bridges the gap between high-level climate data and local decision-making. It shows that while the heat problem is complex, the path to addressing it can be simplified by focusing on the right indicators: moisture and green cover. For city managers in Barcelona, Madrid, and beyond, this means they can stop guessing which schools are the hottest and start using a clear, data-driven method to identify the facilities where intervention will make the most difference. The study does not offer a magic bullet, but it does provide a clear, reliable map for where to start the work of cooling our cities.

Technical Summary: Hot School Grounds, Mismatched Evidence: Facility-Scale Urban Heat in Barcelona and Madrid

Problem Statement
The paper addresses a "decision-unit mismatch" in urban heat research. While remote sensing and Local Climate Zone (LCZ) frameworks provide detailed thermal data at pixel or neighborhood scales, municipal adaptation decisions (funding, redesign, maintenance) are implemented at the scale of specific bounded public assets, such as school grounds. Existing models often report high predictive accuracy by splitting data randomly at the pixel level, which fails to account for the fact that pixels within the same facility share vegetation, materials, and acquisition conditions. Consequently, a model may appear accurate but fail to generalize to new, unseen facilities. The study argues that diagnostic scale (pixels) and implementation scale (managed assets) must be aligned to produce evidence relevant for municipal action.

Methodology
The study employs a longitudinal, comparative remote-sensing design focusing on public primary schools in Barcelona (158 schools) and Madrid (276 schools) from 2014 to 2024.

Data Sources: The analysis utilizes Landsat 8 Collection 2 Level-2 imagery (July–August annually) for Land Surface Temperature (LST) and spectral indices. LCZ data (100-m resolution) serves as contextual urban-form metadata, not as a causal predictor. School identities and campus polygons are derived from official municipal open-data inventories.
Variables: The dependent variable is the median summer LST across valid pixels within a school campus. Five predictors are used: NDVI (vegetation), NDMI (moisture), NDBI (built surfaces), NDBSI (bare/dry surfaces), and a shortwave-reflectance proxy (albedo).
Trend Analysis: Longitudinal trends are estimated using Theil-Sen slopes and standard Mann-Kendall tests with Benjamini-Hochberg false discovery rate (FDR) control within each city.
Machine Learning: Seven supervised regression algorithms (Decision Tree, Random Forest, Extra Trees, XGBoost, CatBoost, LightGBM, Gradient Boosting) are compared.
Validation Strategy: The core methodological innovation is the use of school-grouped cross-validation. Instead of random pixel splitting, the data is split by school_id (five-fold GroupKFold). This ensures that pixels from a specific school appear only in either the training or testing set, preventing data leakage and providing a realistic estimate of performance for held-out facilities. A random pixel split is retained only as a supplementary benchmark.
Interpretation: Feature importance is derived from the selected model but is explicitly interpreted as "predictive priority" for site audits, not as causal intervention effects.

Key Results

Thermal Levels and Trends: Madrid school grounds were significantly hotter than Barcelona's (median summer LST: 37.2°C vs. 31.8°C). Madrid also exhibited steeper median warming trends (0.18°C/year vs. 0.09°C/year) and a higher proportion of schools with statistically significant positive monotonic trends (40.6% vs. 23.4%).
Model Performance: Under school-grouped cross-validation, the Extra Trees regressor performed best, achieving a mean $R^2$ of 0.78 ± 0.05 and an RMSE of 2.41 ± 0.18°C. In contrast, a random pixel split yielded an inflated $R^2$ of 0.89. The 0.11 difference in performance between grouped and random splits was larger than the performance gap between the top two algorithms (Extra Trees and Random Forest), highlighting that validation structure is more critical than algorithmic choice.
Feature Importance: In the pooled Extra Trees model, NDMI (moisture) and NDVI (vegetation) were the top two predictors (importance scores of 0.29 and 0.27, respectively). NDBI, the albedo proxy, and NDBSI followed.
Site-Audit Priorities: Based on feature importance, the study identifies moisture management (NDMI) and vegetation/canopy (NDVI) as the primary domains for follow-up site audits. Hard surfaces (NDBI/NDBSI) and material reflectance are secondary domains.

Key Contributions

Decision-Unit Alignment: The paper proposes a framework where the unit of analysis (school campus), the unit of validation (held-out schools), and the unit of application (site audit) are aligned. This shifts the evidential basis from abstract pixel-level accuracy to facility-scale screening relevance.
Validation Discipline: The study demonstrates that standard pixel-level validation in urban heat studies often yields optimistic performance metrics that do not translate to new facilities. By enforcing school-grouped holdouts, the study provides a more conservative and operationally relevant performance estimate.
Bounded Interpretation: The paper explicitly decouples predictive attribution from causal claims. It argues that high feature importance identifies "priority for evaluation" (e.g., where to send auditors) rather than quantifying the temperature reduction of specific retrofits.

Significance and Claims
The paper claims that its primary contribution is decision-unit alignment, not algorithmic novelty or the discovery of universal thermal rules. It asserts that:

Facility-scale screening depends as much on how evidence is validated (matching the management unit) as on what is modeled.
Remote sensing can effectively identify facilities and audit domains for prioritization but cannot replace high-resolution field measurements for assessing human thermal comfort or retrofit efficacy.
The workflow serves as an intermediate evidence layer to reduce the search space for more expensive follow-up assessments (e.g., microclimate simulation or intervention studies).

The author is modest in their claims, noting that the study does not establish geographic external validity (as schools may be spatially proximate), does not claim autocorrelation-robust trend significance, and does not infer causal effects of interventions. The results are presented as a "screening layer" to guide municipal prioritization, with the understanding that local calibration and field verification are required for specific design interventions.

More like this