Estimation of PM2.5 concentrations at a high spatiotemporal resolution using constrained mixed-effect bagging models with MAIAC aerosol optical depth

Exposure estimation of fine particulate matter with diameter < 2.5 mu m (PM2.5) at high spatiotemporal resolution is crucial to epidemiological studies that examine acute or sub-chronic health outcomes of PM2.5. However, exposure assessment of PM2.5 has been negatively affected by sparsely distributed monitoring stations. In addition, several limitations exist among the existing methods for high spatiotemporal resolution PM2.5 estimation, including ignorance or limited use of spatial autocorrelation, single-model methods, and use of aerosol optical depth data with non-random missingness. These limitations probably introduce bias or high uncertainty in model estimation. In this paper, we proposed an approach of constrained mixed-effect bagging models to leverage advanced algorithm of the high-resolution AOD retrieved by Multi-Angle Implementation of Atmospheric Correction (MAIAC), with other spatiotemporal predictors and spatial autocorrelation to reliably estimate PM2.5 at a high spatiotemporal resolution. Our base model was a daily mixed-effect spatial model that accounted for spatial autocorrelation using embedded structured and unstructured spatial random effects. Point estimates from the base models were then averaged based on the bootstrap aggregating (bagging) to reduce variance in prediction. Then, constrained optimization was developed to minimize the impact of missing AOD and to capture a full time-series of PM2.5 concentration. Our daily-level bagging allowed AOD-PM2.5 association and spatial autocorrelation to vary daily, which substantially improved the model performance. As a case study of daily PM2.5 predictions in 2014 in Shandong Province, China, our approach achieved R-2 of 0.87 (RMSE: 18.6 mu g/m(3)) in cross validation, and R-2 of 0.75 (RMSE: 20.6 mu g/m(3)) in an independent test, similar to or better than most existing methods. We further extended the 2014 models to simulate 2014-2016 full time-series of biweekly average PM2.5 concentrations with no use of covariates in 2015-2016 but constrained optimization over 2014 daily point estimates; the results showed well-captured temporal trend with a total correlation of 0.81 between the simulated and observed values from 2015 to 2016. Our approach can be applied for other regions for exposure estimation of PM2.5 when measurements alone are not able to capture the desirable spatial and temporal resolutions.