On the deep learning approach for improving the representation of urban climate: The Paris urban heat island and temperature extremes
Frederico Johannsen, Pedro M. M. Soares, Gaby S. Langendijk
Abstract
As cities encompass most of the global population, it is crucial to understand the effects of climate change in an urban context to develop tailored adaptation and mitigation strategies. Physically-based numerical models are computationally demanding and present scale limitations that complicate the representation of cities and their effects on the regional-to-local climate and vice-versa. Therefore, often, methodologies (e.g. statistical downscaling) are sought to complement the output of numerical models for urban climate studies. Deep Learning (DL) is a growing technology that has become a universal presence in society and the scientific community, geosciences included, showing promising results all around. In this study, we applied DL models, namely Convolutional Neural Networks (CNNs), to downscale land surface temperature (LST) and predict 2-m maximum and minimum temperatures (T2mmax and T2mmin, respectively) over Paris between 2004 and 2022, and compared the results with ERA5, the most recent atmospheric reanalysis of the European Centre for Medium Range Weather Forecasts. Several experiments featuring different sets of ERA5 predictors were used as input data to the DL models. Afterwards, the quality of the DL models in representing the urban heat island (UHI) over Paris was assessed. Our results showed substantial improvements in LST, T2m and UHI downscaling with DL (using a small number of predictors) in comparison to ERA5. Particularly, the best-performing DL experiments presented nighttime UHI and daytime SUHI biases (RMSE) below 0.80 °C and 0.50 °C (2.8 °C and 2.3 °C), respectively. This study supports the potential of using DL as a downscaling technique to help improve the representation of temperature extremes in an urban context in the historical period.