Deep Learning-Guided Urban Climate Risk Mitigation Through Optimal Spatial Allocation of Green and Cool Roofs

Abstract

With cities facing increasing challenges due to climate change, we developed a deep learning-based surrogate modeling framework to optimize urban roofing strategies for climate risk mitigation. Applied to Seoul, South Korea, the framework utilized the Weather Research and Forecasting model coupled with an Urban Canopy Model (WRF-UCM) to generate objective indices for heat stress, flash floods, and wind circulation projected to the end of this century under four roof schemes: business-as-usual, 25% and 100% cool roofs (CR25 and CR100), and 100% green roofs (GR100). These indices were used to test four deep learning algorithms: UNet, UNet++, UNet3+, and Multi-ResUNet. Multi-ResUNet demonstrated superior performance, thus it was employed to develop the surrogate model, which was applied to 262,144 multi-type roofing scenarios. Two optimal roofing scenarios were identified using the Pareto method, balancing the three climate objectives and economic costs: the first with CR100 covering 95.9% of urban areas, reducing heat stress by over 50% in 34.3% of regions and wind circulation by 10% in 27.7% of regions, and the second with CR100 covering 60.2% of urban areas, achieving a similar heat stress reduction in 21.6% of regions but a stronger reduction in wind circulation. Both scenarios had minimal impact on flash flood mitigation. This study highlights the importance of spatial configuration in maximizing the benefits of urban roofing strategies due to the heterogeneous effects across urban areas. Furthermore, the considerably lower computational time increases the practical utility of the proposed surrogate modeling framework for use in a diverse range of urban contexts, advancing global efforts to mitigate urban climate risks.