UrbanScale: Physics Informed Deep Learning Framework to Generate High-Resolution Urban Temperature Data

Elevated temperatures observed in urban areas have significant implications for human health, energy consumption and infrastructure reliability, and their negative effects disproportionately impact disadvantaged populations. Increasing scientific and practical understanding of urban heat islands is an important step in identifying those regions at elevated risk, and effective mitigation strategies. The urban heat island effect is most readily quantified through the measurement of land surface temperature (LST) in urban regions. Despite the importance of LST to informed decision making and preparing for extreme heat events, there is a current lack of widely available tools that can determine and predict LST at the high-resolution appropriate for the geometric complexity and high degree of heterogeneity present in urban environments. Physics-informed deep learning approaches combining data-driven approaches and physical modeling have demonstrated significant promise in a number of earth-science applications. We propose to evaluate several architectures of physics-informed neural networks to the problem of downscaling or generating high-resolution predictions of LST. Identification of the most promising candidates during Phase I will enable inclusion of these algorithms in a direct user-facing software application during Phase II.