Evaluation of a novel approach to estimate PM2.5 concentrations at high spatial resolution during smoke episodes by fusing low-cost sensor and reference monitor observations with chemical transport model forecasts

Near-real-time community-scale air quality information is essential to enable effective decision making that reduces exposure during pollution events such as those caused by smoke from fires. Air quality conditions during smoke episodes are complex, change rapidly, and can vary over small spatial scales. The accuracy and resolution of existing public information products that predict current air quality conditions during smoke events using simple interpolation of reference monitor observations can be improved by incorporating observations from low-cost sensors and chemical transport models (CTM). Based on recent methods reported by Schulte et al. (Environmental Research Letters, 2020), which focused on the Los Angeles basin in California, we demonstrate that high accuracy can be achieved across the U.S. by kriging the differences between gridded mean observations and CTM simulations. We use gridded-median observations of fine particulate matter (PM2.5) from U.S. Environmental Protection Agency AirNow reference monitors, PurpleAir low-cost sensors, and the National Oceanic and Atmospheric Administration (NOAA) NAQFC forecast model to predict hourly concentrations at a 1x1 km spatial resolution. We apply this method across diverse urban and rural geographies with variable topographies and spatial coverage of PurpleAir sensors and reference monitors. We evaluate the accuracy, bias, and precision of PM2.5 concentration predictions via 10-fold cross validation during wildfire events that occurred in 2021 and 2022. Our novel data fusion method offers a straightforward, computationally efficient, interpretable method that can improve the predictions of continuous air quality index (AQI) surfaces in near-real time during high pollution smoke episodes.