Because ambient ozone (O3) has fine spatial scale variability in addition to a large scale regional distribution, accurate exposure predictions for population health studies need to also capture fine spatial scale differences in exposure. To address these needs, we developed a 3-year average land use regression (LUR) and combined LUR and Bayesian maximum entropy (BME) by incorporating a national area variability LUR model for China from 2015 to 2017 along with data that take into account incompleteness of O3 monitoring data into a BME framework. Spatio-temporal kriging models that either included or did not include "soft" data were used for comparison. The final LUR model included five predictor variables: road length within a 1000 m buffer, temperature, wind speed, industrial land area within a 3000 m buffer and altitude. The 1-year predicted O3 concentrations based on the ratio method moderately agreed with the measured concentration, and the regression R2 values were 0.53, 0.57 and 0.59 in the year of 2015, 2016 and 2017, respectively. The LUR/BME model performed better (R2 = 0.80, root mean squared error [RMSE] = 23.5 μg/m3) than the ordinary spatio-temporal kriging model that either included "soft" data (R2 = 0.57, RMSE = 49.2 μg/m3) or did not include the "soft" data (R2 = 0.52, RMSE = 58.5 μg/m3). We have demonstrated that a hybrid LUR/BME model can provide accurate predictions of O3 concentrations with high spatio-temporal resolution at the national scale in mainland China.
Keywords: Air pollution; Bayesian maximum entropy; China; Land use regression; Ozone.
Copyright © 2020 Elsevier B.V. All rights reserved.