Detection and Simulation of Urban Heat Islands Using a Fine-Tuned Geospatial Foundation Model for Microclimate Impact Prediction
Jannis Fleckenstein, David Kreismann, Tamara Rosemary Govindasamy, Thomas Brunschwiler, Etienne Vos, Mattia Rigotti · 2025
Read Detection and Simulation of Urban Heat Islands Using a Fine-Tuned Geospatial Foundation Model for Microclimate Impact Prediction by Jannis Fleckenstein, David Kreismann, Tamara Rosemary Govindasamy, Thomas Brunschwiler, Etienne Vos, Mattia Rigotti (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-18773.txt.
Introduction
Jannis Fleckenstein and his five co-authors apply a geospatial foundation model to the urban heat island problem, where cities run hotter than their surroundings. Planning shade and green space requires detailed temperature maps, but conventional models trained on scarce local data often predict poorly, especially in underserved neighborhoods. Foundation models, pretrained on global data, promise strong general performance with only a little local fine-tuning.
The team first built an empirical picture of real heat patterns by measuring how much green spaces cool their surroundings, and checked the model's predictions against it. They then fine-tuned the model to predict land surface temperatures under future climate scenarios. A simulated inpainting exercise, where a changed landscape is painted into the data, showed how planners could test mitigation ideas before spending on them. The authors argue such models are a powerful tool for data-scarce regions aiming at climate-resilient cities.
Catalog record
| CREATOR | Jannis Fleckenstein, David Kreismann, Tamara Rosemary Govindasamy, Thomas Brunschwiler, Etienne Vos, Mattia Rigotti |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-18773
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Detection and Simulation of Urban Heat Islands Using a Fine-Tuned Geospatial Foundation Model for Microclimate Impact Prediction Jannis Fleckenstein, David Kreismann, Tamara Rosemary Govindasamy, Thomas Brunschwiler, Etienne Vos, Mattia Rigotti Abstract As urbanization and climate change progress, urban heat island effects are becoming more frequent and severe. To formulate effective mitigation plans, cities require detailed air temperature data, yet conventional machine learning models with limited data often produce inaccurate predictions, particularly in underserved areas. Geospatial foundation models trained on global unstructured data offer a promising alternative by demonstrating strong generalization and requiring only minimal fine-tuning. In this study, an empirical ground truth of urban heat patterns is established by quantifying cooling effects from green spaces and benchmarking them against model predictions to evaluate the model’s accuracy. The foundation model is subsequently fine-tuned to predict land surface temperatures under future climate scenarios, and its practical value is demonstrated through a simulated in-painting that highlights its role for mitigation support. The results indicate that foundation models offer a powerful way for evaluating urban heat island mitigation strategies in data-scarce regions to support more climate-resilient cities. 1 Introduction As cities grow, increasing building and population densities drive land use changes that impact the local energy balance and shape the microclimate conditions of urban areas [ 1 ] . With over 70% of the global population projected to live in cities by 2050, the urban heat island (UHI) effect has become a major concern [ 2 , 3 ] . Factors such as heat-retaining materials, vegetation loss and dense structures can cause cities to be over 5 °C warmer than their surrounding areas [ 4 ] . These temperature increases lead to greater energy consumption, more heat-related health issues, and declining air quality [ 5 , 6 ] . One major challenge in addressing UHI impacts is the lack of high-resolution, timely air temperature data, which is essential for early warnings and effective heat risk mitigation plans [ 7 , 8 ] . Current forecasting approaches often require large amounts of input data, specialized expertise, and significant computational resources, making real-time forecasting at fine scales largely inaccessible [ 9 , 10 ] . Data limitations from sparse stations and satellite revisit intervals further constrain accuracy [ 11 , 12 ] . While artificial intelligence (AI)-based models offer increased flexibility, they are often limited by their dependence on large labeled datasets and poor generalization across domains [ 13 ] . In response to these limitations, recent developments in geospatial foundation models (GFMs), show strong generalization across spatial resolutions and regions with minimal fine-tuning, making them a promising alternative for urban climate analysis [ 14 ] . Building on this potential, this study explores the use of a fine-tuned GFM by establishing an empirical ground truth of how urban green spaces affect temperature through internal and spillover cooling in UHI areas [ 15 , 16 ] . The model is validated against these cooling effects to assess its ability to replicate observed processes. It is then used to forecast urban temperatures under future climate scenarios and to simulate targeted in-painting as a strategy to mitigate UHI impacts. The result is a unified workflow that integrates data analysis, forecasting, and simulation to inform climate-resilient urban planning. 2 Methodology GFMs have demonstrated strong performance in geospatial tasks, with promising generalization across space and time, making them suitable for UHI applications in the context of increasing urbanization and climate-induced temperature extremes [ 17 ] . However, existing studies rarely conduct in-depth evaluations of physical realism for GFMs, such as examining how urban green spaces influence modelled UHI dynamics through simulations like inpainting or future-scenario forecasting, which are essential for informed urban heat mitigation and planning [ 18 ] . To address this gap, this study follows a three-phase experimental workflow. First, an empirical baseline of how green spaces influence temperature is established by overlaying land-use/land-cover (LULC) maps with high-resolution land surface temperature (LST) imagery to extract park areas and analyse how urban temperature gradients evolve with distance from these green spaces, thereby quantifying internal cooling within parks and spillover cooling into surrounding built-up areas, with all results aggregated over summer daytime data from 2017 to 2025 across all cities, and assessing how well the GFM is able to replicate these effects for mitigation planning. Second, the model’s extrapolation ability under future climate conditions is assessed through a focused experiment in Bra s , ov, Romania, a city intentionally selected because it was not part of the original training set. Although GFMs are designed to generalize across regions, isolating a single unseen city allows for a clearer evaluation of performance and highlights the feasibility of future climate projections. In this setting, the data are ordered by temperature and the model is fine-tuned only on the cooler 90% of observations, while the hottest 10% are withheld to test generalization to extreme heat in Bra s , ov. The difference between predicted and observed values in this upper percentile serves as an estimate of how far the model can reliably extrapolate beyond its training range. Under the assumption that this 10% threshold represents its extrapolation limit, the model is then used to project future UHI conditions by replacing ERA5 with EURO-CORDEX climate inputs under different future Representative Concentration Pathway (RCP) scenarios. Third, a hypothetical urban greening intervention is simulated by guided inpainting of the satellite image, replacing built-up pixels with those representing urban green areas. The input spectral indices are therefore adjusted accordingly, and the model is run to predict the resulting thermal impact to support simulated UHI mitigation planning. 2.1 Training procedure and geospatial foundation model Building on the work of Bhamjee et al. 2024 , the Granite-Geospatial-Land-Surface-Temperature (Granite-GFM) model [ 20 ] is used as a GFM for predicting LST at high spatial and temporal resolution. The model is evaluated in two configurations, V1 which is fine-tuned using Harmonised Landsat Sentinel-2 (HLS L30 [ 21 ] ) imagery and ERA5-Land [ 22 ] 2 m air temperature statistics from 2013–2023 across 28 cities, and V2 which is fine-tuned on an extended dataset covering 52 cities across a wider range of hydro-climatic zones. Granite-GFM incorporates a Shifted Window (SWIN) Transformer architecture and builds on the Prithvi-SWIN-L Earth Observation foundation model [ 23 , 24 , 17 ] . While state of the art approaches for LST prediction focus on single regions of interest, the Granite-GFM model enables the estimation of LST at a 30m spatial and hourly temporal scale, for any city of interest using fewer input samples. 2.2 Study regions and data Our analysis covers thirteen European cities that were not part of the training data, selected to test the model’s out-of-sample generalization across diverse urban environments. A regional map of the selected cities is provided in Appendix 5 . Twelve of these cities, marked in red, are used in the first workflow step to assess internal and spillover cooling effects across varying hydro-climatic settings. For the second step, Bra s , ov (marked in blue) is used as an unseen evaluation city. In the third intervention step, Prague is selected for in-painting. While these cities are highlighted, the workflow can in principle be applied to any urban area, which