Towards an improved representation of the urban heat island effect : A multi-scale application of XGBoost for madrid

Cities are considered local "hotspots" of climate change, therefore, the improvement of urban present climate description as well as future projections is paramount for designing adaptation and mitigation strategies. Physically-based numerical models often have coarse resolutions and do not present parametrisations to adequately represent physical processes at the urban scale. This article presents an innovative application of XGBoost (a machine learning approach) as an alternative to explore and improve the description of the climate of Madrid. XGBoost's ability to reproduce the present climate of 2-m air temperature and land surface temperature (LST), and the urban heat island (UHI) effect, was assessed. XGBoost was trained with a set of ERA5 predictors (0.25°) and calibrated with observations from ground stations (2000−2022) and remote sensing data (2004–2022). Several sensitivity cases were performed to assess the results dependency to predictors and their resolution. XGBoost was evaluated at daily scale for the 2-m maximum and minimum temperatures (Tmax and Tmin, respectively) and LST, and at hourly scale for LST. Overall, XGBoost reveals a good performance and significant added value against ERA5 for all variables considered and for both UHI and surface UHI. This study presents XGBoost as a promising technology to describe urban climate.