Bridging the Scale Gap: Machine Learning for Urban Heat Island Modeling and Mitigation

Urban Heat Islands (UHIs) pose a formidable challenge to cities worldwide, escalating surface and air temperatures beyond rural surroundings and intensifying heat risks amid global warming. Urbanization of the land surface drives this phenomenon, with cities featuring heat-absorbing materials, dense populations, and minimal green spaces, compared to rural areas. The intricate interplay of factors, including urban material properties, vegetation cover, population density, urban activities, land use, and local climate, govern UHI intensity and spatio-temporal variability. Despite recognition of these factors, uncertainties persist in quantifying their combined and interacting influence on surface temperatures and overall UHI characteristics. This research aims to bridge this knowledge gap by employing machine learning techniques to unravel the drivers of UHI variability at fine spatial scale, and to emulate future UHI characteristics. Our work leverages Landsat observations to extract land surface temperatures, and multi-headed deep learning models — based on Convolution Neural Networks, Attention, and Transformers — to relate meteorological information, land use and land cover data, and high-resolution satellite images to land surface temperature. By modeling Landsat observations from various UHI-related features, this approach offers insight into the intricate drivers of UHI dynamics, and can be applied to downscale future climate projections of urban land surface temperatures. The integration of attention mechanisms provides interpretability and better performance, essential for robust decision-making. Ultimately, the framework's ability to generate high-resolution projections of land surface temperature, alongside predictive analyses, opens paths for proactive UHI mitigation strategies.