The 2018 Planetary Health Annual Meeting
Satellite aerosol optical depth (AOD) data have been used to assess population exposure to fine particulate matter (PM2.5), but are challenged by non-random missingness due to cloud/snow cover and high surface reflectance. Previous studies filled the data gap by spatially smoothing neighboring PM2.5 measurements or predictions; however, this strategy ignored the effect of cloud cover on aerosol loadings and did not perform well when monitoring stations are sparse or there is seasonal large-scale missingness. For example, in the Yangtze River Delta, the monsoon season (summer) leads to, on average, approximately 75% AOD missingness even after combining Aqua and Terra data. Here we present a Multiple Imputation (MI) method that fused the chemical transport model (CTM) simulations and the high-resolution satellite retrievals to fill missing AOD and provide PM2.5 predictions with fine resolution, complete coverage, and high accuracy. First, we fitted daily MI models that employed the spatiotemporal auto-correlation of AOD to impute missing AOD from available satellite data, CTM simulations, and meteorological parameters. Repeated imputations were conducted to account for random error. Then a two-stage hybrid model was fitted to estimate daily ground PM2.5 concentrations from complete-coverage MAIAC AOD, meteorology, and land use information. The daily MI models have an average R2 as 0.77, with an inter quartile range from 0.71 to 0.82. The overall model 10-fold cross validation R2 (relative prediction error) were 0.81 (34%) and 0.73 (29%) for year 2013 and 2014, respectively. Models fitted with only observational AOD or only imputation AOD performed similarly. Using models fitted by data of year 2013 and 2014 to predict monthly PM2.5 concentrations in 2015 gave an R2 as 0.70 and 0.71, respectively. This method provides reliable PM2.5 predictions with complete coverage that can reduce exposure error in air pollution health effect research.