Abstract
Most of the United States (US) population resides in cities, where they are subjected to the urban heat island effect. In this study, we develop a method to estimate hourly air temperatures at 0.01° × 0.01° resolution, improving exposure assessment of US population when compared to existing gridded products. We use an extensive network of personal weather stations to capture the intra-urban variability. The uncertainty associated with this crowdsourced data set is addressed through a spatiotemporal Bayesian model implemented with the Integrated Nested Laplace Approximation-Stochastic Partial Differential Equation approach. We evaluate the model on Philadelphia (PA), New York City (NY), Phoenix (AZ), and the Triangle area (NC). These case studies span different climatic zones and urban landscapes. They cover several meteorological events including a deadly heatwave in Phoenix and a snowstorm hitting part of the US in winter 2021. We obtain an overall root mean square error of 1.06°C , demonstrating the versatility of our model, and its applicability across various regions in the US. The high granularity of our model allows for the precise identification of hotspots that were previously undetected with daymet and gridMET products. Using the data generated by our method, we show that neighborhoods with high population concentration are more likely to experience elevated temperatures and prolonged hot nights, thus encouraging the use of our model for further epidemiological investigations on the impact of heat or cold stress on human health.