Estimating ground-level PM2.5 concentrations in Beijing using a satellite-based geographically and temporally weighted regression model

摘要

Most time-sequenced ambient air pollution data in China is published through daily Air Quality Index (AQI). However, few studies have used the AQI data to calibrate satellite-based estimates of fine particulate matter (PM2.5, particles no greater than 2.5 mu m in aerodynamic diameter) concentrations, partly because the AQI-derived PM2.5 is not continuously obtained each day. Taking Beijing as an example, we developed a geographically and temporally weighted regression (GTWR) model that can account for spatial and temporal variability in the relationship between the non-continuous AQI-derived PM2.5 and satellite-derived aerosol optical depth (AOD). The GTWR model, which uses AOD values with a 3-km spatial resolution obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS), meteorological fields, and land-use variables as predictors, was fitted seasonally from April 2013 to March 2015. After being cross-validated against ground observations, the coefficient of determination (R-2) of PM2.5 ranged from 0.36 to 0.75, with a mean value of 0.58. The GTWR model outperforms several conventional models, such as the multiple linear regression (MLR) model, geographically weighted regression (GWR) model, temporally weighted regression (TWR) model, and linear mixed-effects (LME) model. Compared to a previous spatiotemporal model, the two-stage (LME + GWR) model, the GTWR model may be more feasible. When the number of daily records is >= 5, there is no obvious difference in prediction accuracy (cross-validated R-2 both valued at 0.68). However, when the number of daily records is <5, the GTWR model performs much better (cross-validated R-2 of 0.45 and 0.08). Our estimates indicate that the gridded annual mean PM2.5 values range from 62 to 110 mu g/m(3), denoting strong spatial variation. We find that when available, continuous daily PM2.5 observations can significantly improve model performance and therefore facilitate the estimation of surface PM2.5 concentrations at urban scales. The GTWR model may serve as a reference for studying regions where continuous air pollution data are limited. (C) 2017 Elsevier Inc. All rights reserved.