National scale spatiotemporal land-use regression model for PM2.5, PM10 and NO2 concentration in China

Abstract

BackgroundAir pollution epidemiological studies increasingly rely on high-resolution exposure prediction models. However, to date, few models of this type exist for use in China. ObjectivesWe produced a national land-use regression model (LUR) to estimate monthly average PM2.5, PM10 and NO2 from 2014 to 2016 in China. MethodsWe developed a spatiotemporal semi-parametric model using generalized additive mixed models. A variety of predictor variables were included in model: time varying meteorological data, high resolution land cover data from Globaland30, satellite measures of aerosol optical depth, and Geographic Information System (GIS)-derived predictors. We assessed model performance with two cross-validation (CV) approaches, including hold-out CV, and 10-fold CV. ResultsOver 22,000 monthly observations at 1382 monitoring locations were included to estimate the air pollution exposure. The time-varying spatial terms explained 87%, 71%, and 69% of variability with a hold-out cross-validated R2 of 0.85, 0.62, and 0.62 for PM2.5, PM10 and NO2 models, respectively. Models show that meteorological variables, population density, elevation, distance to road, and land cover types were important predictors for air pollution exposure. Conclusionswe have developed a new nationwide model to estimate residence-level air pollution exposures, which can be used in studies of the chronic adverse effects of air pollution.