Enhancing Urban Heat Island Mapping in Turin During a Heat Wave: A Machine Learning Approach with Citizen Science Data

Accurate prediction of ground-level air temperature is crucial for developing heat hazard maps that help anticipate the impacts of heat waves on vulnerable populations. Numerical weather prediction models typically operate at the mesoscale resolution (1-2 km) (Garbero et al., 2021), and their application at high spatial resolution within large urban areas faces significant challenges, such as the complexities of urban parameterization and the high computational costs. In this context, machine learning has emerged as a promising alternative or complementary approach to these traditional methods (Zumwald et al., 2021).

This study presents a machine learning-based model designed to reconstruct high-resolution temperature maps and predict hourly temperatures for the city of Turin during a heat wave in June 2022. The model leverages nine predictor variables related to urban morphology, including building density, building height, sky view factor, and vegetation density, combined with temperature data from citizen weather stations (CWS). The CWS data, sourced from the Netatmo meteorological network, support the model's potential for generalization to other cities. Furthermore, this study evaluates the impact of integrating the outputs of the COSMO meteorological model into the predictor set.

This study compares the performance of two modelling approaches, trained for each nighttime hour, to reconstruct temperature maps: (i) a baseline multi-linear regression (MLR) model and (ii) a convolutional neural network (CNN). The MLR model was trained at two spatial resolutions - 50 and 100 m. Results indicate that the 100 m resolution yields lower RMSE values, with a maximum error of 1.36°C (reduced to 1.23°C when COSMO outputs are included as additional predictors). This finding highlights the importance of averaging predictors over sufficiently large spatial areas around sensor locations. The CNN model outperforms the MLR, achieving a maximum RMSE of 1.21°C (further reduced to 1.17°C). Both models demonstrate substantial improvement over the COSMO model, which exhibits a notably higher RMSE exceeding 2.5°C when predicting Netatmo temperatures.

A sensitivity analysis highlights the slightly greater influence of specific predictors, such as the Sky View Factor or the altitude. However, the relatively low magnitude of sensitivity suggests an excessive number of predictors, leading to compensatory effects when individual predictors are excluded from the model.

This study demonstrates the effectiveness of machine learning techniques in reconstructing temperature maps in Turin. Future work should focus on reducing predictor redundancy, improving data cleaning processes to mitigate the impact of outliers, and assessing the generalizability of this methodology to other cities.

 

References :

Garbero, V., Milelli, M., Bucchignani, E., Mercogliano, P., Varentsov, M., Rozinkina, I., Rivin, G., Blinov, D., Wouters, H., Schulz, J.-P., Schättler, U., Bassani, F., Demuzere, M., Repola, F., 2021. Evaluating the Urban Canopy Scheme TERRA\_URB in the COSMO Model for Selected European Cities. Atmosphere 12, 237. https://doi.org/10.3390/atmos12020237

Zumwald, M., Knüsel, B., Bresch, D.N., Knutti, R., 2021. Mapping urban temperature using crowd-sensing data and machine learning. Urban Climate 35, 100739. https://doi.org/10.1016/j.uclim.2020.100739