National PM2.S and NO2 Spatiotemporal Models Integrating Intensive Monitoring Data and Satellite-Derived Land Use Regression in a Universal Kriging Framework in the United States: 1999-2016

摘要

Background: Recent nationwide PM2.5 and NO2 exposure models increasingly incorporate land-use regression, satellite-derived observations and spatial smoothing to improve prediction accuracy for epidemiological studies. However, those studies exclusively relied on exposure data from routine monitoring sites that could be less useful for small spatial-scale predictions. Objectives: To develop regionalized national models in estimating PM2.5 and NO2 exposures with high spatiotemporal resolution from 1999 to 2016 in the United States. Methods: We collected monitoring data from numerous regulatory (number of monitors: 1495 [PM2.5], 754 [NO2]) and cohort-specific monitors (939 [PM2.5], 2573 [NO2]) that captured fine-scale residential and roadway exposure gradients. We established a novel modeling framework that incorporated all of the unbalanced monitoring data with land use regression and universal kriging using dimension-reduced predictor variables from satellite observations and a large geographic database. Results: Ten-fold cross-validations showed good model performances with total spatiotemporal R2s of 0.82 and 0.81 for PM2.5 and NO2 using the left out routine monitors. We observed larger within-city spatial variations (coefficient of variation [CV]) from the best models that accounted for cohort-specific monitors than the models that completely relied on routine monitoring sites (increased % of CV: 20-27% for PM2.5, 23-32% for NO2). Regional spatial R2 was highest in the southeast region (0.91) for PM2.5 and lowest in the northwest region (0.71), but was generally the same for NO2 across regions (0.84-0.87). Including satellite PM2.5 or NO2 data moderately improved predictions for points far from monitoring locations. Conclusion: Our models can make accurate point predictions of PM2.5 and NO2 concentrations at both short and long time scales when utilizing additional data from a large number of fine-scale monitors and satellite technology.