Deep Learning-Driven Downscaling for the Urban Climate of the Megacity of São Paulo at High Resolution

Urban areas concentrate a significant (and growing) fraction of the global population and economic activity, making them particularly vulnerable to the evolving climate change risks. While Earth System Global Climate Models (ESGCMs) provide essential long-term climate projections, their coarse spatial resolution and lack of urban parameterizations often fail to capture the complex physical processes within the cities. Therefore, downscaling urban climate characterization under future climate projections is essential for providing high-resolution data, necessary for developing effective mitigation and adaptation strategies, which are ultimately paramount for safeguarding societal well-being and urban resilience.

Within this framework, Artificial Intelligence methodologies – an evolving branch of statistical downscaling methods – offer an alternative approach to traditional ESGCMs for characterizing climate at the urban scale. In this study, Deep Learning (DL) is applied to generate high-resolution (0.05º) present and future urban climate projections for the megacity of São Paulo (SP), Brazil. Firstly, sensitivity cases were performed to evaluate the performance of convolutional neural network (CNN) architectures in predicting near-surface air temperature (daily maximum and minimum, T2max and T2min, respectively) and Land Surface Temperature (LST), using observational datasets as local predictands and ERA5 reanalysis as large-scale predictors. Secondly, a multi-model ensemble of CNN-based downscaled projections was developed to project T2max, T2min, and LST, along with their associated Urban Heat Island phenomena (UHI and SUHI, respectively), throughout the 21st century. These projections were developed for four Shared Socioeconomic Pathway (SSP) scenarios at  a daily scale. The resulting DL-downscaled projections demonstrate overall agreement with the CMIP6 ESGCM ensemble in the magnitude for the projected temperatures

This work is supported by FCT, I.P./MCTES through national funds (PIDDAC): LA/P/0068/2020 - https://doi.org/10.54499/LA/P/0068/2020, UID/50019/2025, https://doi.org /10.54499/UID/PRR/50019/2025, UID/PRR2/50019/2025. The authors would like also to acknowledge the project “Elaboração do Plano Municipal de Ação Climática de Barcelos (PMACB). A.B. also acknowledge individual funding from FCT, I.P./MCTES grant UI/BD/01324/2024