Satellite-derived high resolution PM2.5 concentrations in Yangtze River Delta Region of China using improved linear mixed effects model

Satellite remotely sensed aerosol optical depth (AOD) provides an effective way to fill the spatial and temporal gaps left by ground PM2.5 monitoring network. Previous studies have established robust advanced statistical models to estimate PM2.5 using AOD data in China. However, their coarse resolutions (∼10 km or greater) of PM2.5 estimations are not enough to support the health effect studies at urban scales. In this study, 3 km AOD data from Moderate Resolution Imaging Spectroradiometer (MODIS) collection 6 products were used to estimate the high resolution PM2.5 concentrations in Yangtze Delta Region of China. We proposed a nested linear mixed effects (LME) model including nested month-, week-, and day-specific random effects of PM2.5-AOD relationships. Validation results show that the LME model only with day-specific random effects (non-nested model) used in previous studies has poor performance in the days without PM2.5-AOD matchups (the R2 of day-of-year-based cross validation (DOY-based CV) is 0.148). The results also show that our nested model cannot improve the performance of non-nested model in the days with PM2.5-AOD matchups (sample-based CV R2 = 0.671 for nested model vs. 0.661 for non-nested model), but can greatly improve the model performance beyond those days (DOY-based CV R2 = 0.339 for nested model vs. 0.148 for non-nested model). To further improve the model performance, we applied the “buffer models” (i.e., models fitted from datasets which ground PM2.5 were matched with the average AOD values within certain radius buffer zones of gridded PM2.5 data) on the 3 km AOD data since the “buffer models” has more days with PM2.5-AOD matchups and can provide more day-specific relationships. The results of this study show that 3 km MODIS C6 AOD data can be used to estimate PM2.5 concentrations and can provide more detailed spatial information for urban scale studies. The application of our nested LME model can greatly improve the accuracy of 3 km PM2.5 predictions.