Bias-correcting the high-resolution urban meteorology for impacts dataset (HUMID) temperature data using the personal weather stations (PWS) network and a spatial temporal Bayesian regression coupled with random forest

Heat exposure is associated with mortality and several adverse health outcomes, including heat strokes and heart diseases. The availability of accurate and high spatial temporal resolution of meteorological information is crucial to evaluate these associations. In this study, we use data from personal weather stations (PWS) to bias correct the recent 1 km high-resolution urban meteorology for impacts dataset (HUMID) daily temperature in the Atlanta metropolitan area for three summer periods from 2016 to 2018. We first developed a 4-step quality control scheme to remove potential erroneous data from PWS. Then, we assessed several bias correction models, a geostatistical Bayesian downscaler modeling approach, a random forest model, and a hybrid approach that uses residuals from the random forest model as the outcome for the Bayesian regression to account for spatial-temporal dependency not captured by random forest. Our results show that the Bayesian downscaler combined with random forest was the best predictive model with root mean square errors (RMSE) of 1.577 °C for maximum temperature and 0.696 °C for minimum temperature. In addition, our models resulted in bias-corrected temperatures that are, on average, lower than HUMID, suggesting that HUMID tends to overestimate temperature in most areas.