Supplementary material from "Micro-scale Modelling of the Urban Heat Island Hazard during heatwaves, a case study in Turin"

Ground-level air temperature maps at the agglomeration scale are vital for assessing hazards from the Urban Heat Island (UHI) effect during extreme heat events. Their prediction is nowadays challenging, requiring models that balance high spatial resolution with scalability. In this study, we develop Machine Learning (ML) algorithms, based on six high-resolution parameters describing topography, geometry, and land use of the urban environment. We evaluate two methods—Multiple Linear Regression (MLR) and Convolutional Neural Network (CNN)—for predicting the UHI effect (and related hazards) in Turin. Models are trained using temperature data from NetAtmo Citizen Weather Stations (CWS). We also assess the impact of adding a seventh predictor from a Numerical Weather Prediction (NWP) model. The CNN achieves a root mean square error (RMSE) below 1.19°C, slightly outperforming the MLR, which reaches an RMSE of up to 1.22°C. Notably, the CNN trained without NWP data performs similarly to the MLR model that includes it, demonstrating CNN’s robustness with limited input. Temperature maps and parameter analysis reveal the need to better understand spatial drivers of urban temperature variability and confirm the potential of ML tools in urban climate modelling. Leveraging on these insights, we discuss key factors to reduce uncertainties in data-driven temperature models.