A full-coverage estimation of PM concentrations using a hybrid XGBoost-WD model and WRF-simulated meteorological fields in the Yangtze River Delta Urban Agglomeration, China.

In spite of the state-of-the-art performances of machine learning in the PM estimation, the high-value PM underestimation and non-random aerosol optical depth (AOD) missing are still huge obstacles. By incorporating wavelet decomposition (WD) into the extreme gradient boosting (XGBoost), a hybrid XGBoost-WD model was established to obtain the full-coverage PM estimation at 3-km spatial resolution in the Yangtze River Delta Urban Agglomeration (YRDUA). In this study, 3-km-resolution meteorological fields simulated by WRF along with AOD derived from Moderate Resolution Imaging Spectroradiometer (MODIS) were served as explanatory variables. Model MW and Model NW were developed using XGBoost-WD for the areas with and without AOD respectively to obtain a full-coverage PM mapping in the YRDUA. The XGBoost-WD model showed good performances in estimating PM with R of 0.80 in the Model MW and 0.87 in the Model NW. Moreover, the K-value of Model MW increased from 0.77 to 0.79 and that of Model NM increased from 0.81 to 0.86 compared with the model without the step of WD, indicating an improvement on the problem of PM underestimation. Due to a better ability of capturing abrupt changes in the PM concentrations, the spatial evolution of PM during a typical pollution event could be mapped more accurately. Finally, the analysis of variable importance showed that the three most important variables in the estimation of the low-frequency coefficients of PM (PM_A4) were temperature at 2 m (T2), day of year (DOY) and longitude (LON), while that in the high-frequency coefficients of PM (PM_D) were CO, AOD and NO. This study not only provided an effective solution to the PM underestimation and AOD missing problems in the PM estimation, but also proposed a new method to further refine the sophisticated correlations between PM and some spatiotemporal variables.