Publications

Wu, JS; Yao, F; Li, WF; Si, ML (2016). VIIRS-based remote sensing estimation of ground-level PM2.5 concentrations in Beijing-Tianjin-Hebei: A spatiotemporal statistical model. REMOTE SENSING OF ENVIRONMENT, 184, 316-328.

Abstract
Satellite-based remote sensing data have been widely used in estimating ground-level PM2.5 concentrations as it can provide spatially detailed information. Most modern satellite-based PM2.5 estimates use statistical models that demand dense PM2.5 monitoring networks. As the national PM2.5 monitoring networks in China were not finished until the end of 2012, the research related to PM2.5 is relatively unsubstantial. To further improve the accuracy and application of remote sensing based estimation models for PM2.5 and take advantage of the newly established networks, we employed a time fixed effects regression model and geographically weighted regression model to develop a spatiotemporal statistical model that estimated ground-level PM2.5 concentrations in Beijing-Tianjin-Hebei. The time fixed effects regression model used the aerosol optical depth (AOD) data from the VIIRS (Visible Infrared Imaging Radiometer Suite) instrument as the major predictive variable along with several other dependent variables, including some factors uncommonly discussed in previous literature, i.e., the satellite-derived NO2 concentrations of the previous day (NO2_Lag) and four directional wind vectors, and estimated day-by-day ground-level PM2.5 surfaces. The geographically weighted regression model used the residuals from the time fixed effects regression model as the dependent variable and the AOD value as the independent variable. Through adding the estimated residuals back to previous surfaces, we obtained the final prediction maps of ground-level PM2.5 concentrations in Beijing-Tianjin-Hebei with a spatial resolution of 6 km x 6 km. The results were as follows. i). The spatiotemporal statistical model performed satisfactorily in that it successfully captured both the temporal and spatial variations in the PM2.5-AOD relationships. The coefficient of determination (R-2), mean prediction error (MPE), and root-mean-square error (RMSE) were 0.88301, 8.1331 mu g/m(3), and 13.0574 mu g/m(3), respectively, during model fitting and 0.71889, 12.2712 mu g/m(3), and 19.2927 mu g/m(3), respectively, during model validation. ii). Incorporating the NO2_Lag in the time fixed effects regression model significantly improved the model's performance and it played a positive role in ground-level PM2.5 concentrations. Replacing the simple wind speed with four directional wind vectors was helpful for the model's performance. iii). Meteorological factors and land use characteristics significantly affected the PM2.5-AOD relationships. The temperature and surface relative humidity (SRH) played a positive role, whereas the rainfall, planet boundary layer height (PBLH), average relative humidity in the PBLH (RH_PBLH), four directional wind vectors, and normalized difference vegetation index (NDVI) played a negative role. iv). The prediction maps revealed that fine particle pollution in Beijing-Tianjin-Hebei is severe and the pollution pattern presents relatively strong seasonal heterogeneity and southeast-northwest spatial heterogeneity. (C) 2016 Elsevier Inc. All rights reserved.

DOI:
10.1016/j.rse.2016.07.015

ISSN:
0034-4257