A Machine Learning application towards a better representation of Madrid’s urban future climate

Cities are considered local “hotspots” of climate change. Urban areas concentrate a large fraction of global population, wealth, and emissions, exposing their inhabitants to climate change impacts. Therefore, the improvement of urban present climate description and future projections are paramount for designing adaptation and mitigation strategies. Global Climate Models are state-of-the-art tools for projecting future climate. However, most of the simulations have coarse resolutions and do not have physical urban parametrisations to adequately represent the physical properties and processes at the urban scale.

The advantage of applying a machine learning approach – Extreme Gradient Boosting (XGBoost) – is explored for better describing Madrid’s urban present and future climates, namely, the ability to reproduce the 2-m air temperature (Tmax, Tmin), land surface temperature (LST), urban heat island (UHI) and surface urban heat island (SUHI) effects. The XGBoost is evaluated at daily scales for local ground temperatures and, at both daily and hourly scales, to represent the spatial structure of LST w.r.t. remote sensing data. Firstly, for present climate, XGBoost is trained with a set of ERA5 predictors (at 0.25°), ground stations, and LST observations. Secondly, a number of sensitivity cases are performed to assess the results dependency to predictors and their resolution. Thirdly, the learned relationship between the set of predictors and predictands is applied to 4 Earth System Global Climate Models (ESGCM) predictors, providing historical and future climate projections for the 21^st century under four emission scenarios.

Overall, XGBoost results reveal a good performance and significant added value against ERA5 and the ESGCMs. XGBoost greatly improves the reproduction of the present climate Tmax, Tmin, LST, and more importantly, the UHI (-0.5°C and +3°C for Tmax and Tmin, respectively), and the SUHI (+1°C and +2°C for Tmax and Tmin, respectively). For future climate, XGBoost significantly corrects the ESGCM UHI misrepresentation but seems to underestimate the expected Madrid’s local warming (3.5°C anomaly under the SSP5-8.5 scenario).

Acknowledgments:

This work is supported by the Portuguese Fundação para a Ciência e Tecnologia, FCT, I.P./MCTES through national funds (PIDDAC):

UID/50019/2025 and LA/P/0068/2020 https://doi.org/10.54499/LA/P/0068/2020).

DCAL are supported by the Portuguese Foundation for Science and Technology (FCT) financed by national funds from the MCTES through grant https://doi.org/10.54499/2022.03183.CEECIND/CP1715/CT0004.