Using Deep Learning to generate future projections of temperature extremes and the urban heat island in Paris

Understanding and modelling the urban climate and the impacts of climate change on the urban environment is essential to underpin the development of adequate adaptation and mitigation measures and policies. City-scale climate projections require very high-resolution physically-based models which are commonly time-intensive and computationally expensive to run. To overcome this problem, cost and time effective alternatives, such as Deep Learning, are often sought.

Here, we present an application of two lightweight 3-layer Convolutional Neural Network (CNN) architectures to downscale 7 Global Climate Models (GCMs) from the Coupled Model Intercomparison Project Phase 6, for the 2015-2100 period using four socioeconomic pathways (SSP1-2.5, 2-4.5, 3-7.0, 5-8.5), for the city of Paris, France. The CNNs generate projections of 2-meter temperature (T2m) at point-level (using data from 7 in-situ observational stations) and Land Surface Temperature (LST) at a spatial resolution of ~5 km. The resulting dataset is used to analyse the CNNs representation of air temperatures, LST, the urban heat island (UHI) and temperature extremes, including heatwave frequency, in future climate. The CNN downscaled projections replicate the Parisian UHI effect, described in previous studies and in the observational data used to train the CNNs. The GCMs, on the other hand, due to their coarse resolution, are unable to capture the UHI effect. Moreover, the CNNs projections are consistent with the GCMs overall warming mean trend for maximum and minimum T2m and LST, 90^th percentile maximum T2m, and an increase in tropical nights throughout the 21^st century (for the warmest scenarios). However, CNNs underestimate the increase in heatwave frequency present in the GCMs under the warmest scenarios. Although further research is required to understand the shortcomings in heatwave DL projections, this work supports the potential of DL as a downscaling method for urban climate studies.

 

Acknowledgements: 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).

Frederico Johannsen was supported by FCT, I.P with the doctoral grant with the reference UI/BD/151498/2021 and DOI identifier 10.54499/UI/BD/151498/2021.