A high-resolution computationally-efficient spatiotemporal model for estimating daily PM2.5 concentrations in Beijing, China

Exposure to ambient air pollution is the largest environmental source of global disease burden. Accurate assessment of air pollutant exposure at finely-resolved spatial and temporal scales is critical for valid estimation of health effects. Limitations of exposure assessment approaches include computational burden and unreliable or complex input data. We aimed to develop an accurate, easy-to-use, high-resolution statistical model to estimate ambient PM2.5 concentrations in Beijing. We implemented a model that estimates long- and short-term trends in pollutant concentrations based on observations from regulatory monitors, geographic covariates, and spatial smoothing. It also allows for inclusion of spatiotemporal covariates. We used observations from 19 monitors around Beijing and 90 geographic covariates, including road density, meteorological covariates, population and building density, land use types, topography, and vegetation cover to produce predictions of daily PM2.5 concentrations for 2015–2017. The model yields predictions at any geographic point, and here we summarize results for a 500m × 500m scale. The daily and long-term average cross validated R2 were 0.96 and 0.93, with RMSE of 13.1 and 1.7, respectively. The addition of temperature as a spatiotemporal covariate did not change the results materially. Over the study period, annual and seasonal PM2.5 concentrations in Beijing declined substantially, although they remained high relative to recommended level by WHO. We developed an easy-to-use, less computationally-intensive statistical model utilizing readily available input data to provide highly accurate estimates of PM2.5 concentrations in Beijing at a fine spatiotemporal scale. Model predictions have already been used in a publicly available mobile application designed to provide exposure estimates for epidemiological analyses of the effects of PM2.5 among residents of Beijing.