Advancing the prediction accuracy of satellite-based PM2.5 concentration mapping: A perspective of data mining through in situ PM2.5 measurements.

Ground-measured PM2.5 concentration data are oftentimes used as a response variable in various satellite-based PM2.5 mapping practices, yet few studies have attempted to incorporate ground-measured PM2.5 data collected from nearby stations or previous days as a priori information to improve the accuracy of gridded PM2.5 mapping. In this study, Gaussian kernel-based interpolators were developed to estimate prior PM2.5 information at each grid using neighboring PM2.5 observations in space and time. The estimated prior PM2.5 information and other factors such as aerosol optical depth (AOD) and meteorological conditions were incorporated into random forest regression models as essential predictor variables for more accurate PM2.5 mapping. The results of our case study in eastern China indicate that the inclusion of ground-based PM2.5 neighborhood information can significantly improve PM2.5 concentration mapping accuracy, yielding an increase of out-of-sample cross validation R2 by 0.23 (from 0.63 to 0.86) and a reduction of RMSE by 7.72 (from 19.63 to 11.91) μg/m3. In terms of the estimated relative importance of predictors, the PM2.5 neighborhood information played a more critical role than AOD in PM2.5 predictions. Compared with the temporal PM2.5 neighborhood term, the spatially neighboring PM2.5 term has an even larger potential to improve the final PM2.5 prediction accuracy. Additionally, a more robust and straightforward PM2.5 predictive framework was established by screening and removing the least important predictor stepwise from each modeling trial toward the final optimization. Overall, our results fully confirmed the positive effects of ground-based PM2.5 information over spatiotemporally neighboring space on the holistic PM2.5 mapping accuracy.