Deep learning enables city-wide climate projections of street-level heat stress

Abstract

Urban areas are increasingly vulnerable to the impacts of climate change, especially heatwaves, due to their distinct characteristics. However, the influence of urban form and land cover on future outdoor thermal comfort remains inadequately quantified in climate models. This study addresses this issue by introducing the Unified Human Thermal Comfort Neural Network (UHTC-NN), a novel deep learning framework that efficiently and accurately maps pedestrian-level urban heat stress at a building-resolved scale of 1 m across entire cities. Using the city of Freiburg, Germany, as a case study, the model uses extensive spatial data to generate detailed Universal Thermal Climate Index (UTCI) maps by downscaling CMIP5 climate ensembles for the period 2070–2099. The model results show significant increases in heat stress hours under future climate scenarios (RCP4.5 and RCP8.5), with climate signals emerging as the dominant effect. Our model reveals that future heat stress hours will exhibit significant spatial variability, with contrasting day-night dynamics. While overall heat stress hours (UTCI ≥ 26 °C) increase more uniformly during the day, nighttime heat stress hours and daytime extremes (UTCI ≥ 38 °C) increase more heterogeneously. These patterns are significantly influenced by shading and radiation trapping at the meter scale, and by building density and land cover at the hectormeter scale. This work highlights the need for high-resolution models to accurately map urban heat exposure and inform adaptive urban planning. By facilitating comprehensive analyses, the model supports targeted interventions to mitigate urban heat, providing a local-specific tool for urban planning in a warming world.