Spatiotemporal modeling with temporal-invariant variogram subgroups to estimate fine particulate matter PM_2.5 concentrations

详细信息	查看全文 | 推荐本文 |

• 作者：Chu-Chih Chen^a ; ^{ccchen@nhri.org.tw} ; Chang-Fu Wu^b ; ¹ ; ^{changfu@ntu.edu.tw} ; Hwa-Lung Yu^c ; Chang-Chuan Chan^b ; Tsun-Jen Cheng^b
• 关键词：Empirical variogram ; Geographic information system ; Kriging ; Land-use regression ; Temporal trend
• 刊名：Atmospheric Environment
• 出版年：2012
• 期刊代码：11_13522310
• 类别：env
• 出版时间：July, 2012
• 卷：54
• 期：Complete
• 页码：1-8
• 文件大小：556 K

摘要

Short-term exposure estimation of daily air pollution levels incorporating geographic information system (GIS) into spatiotemporal modeling remains a great challenge for assessing corresponding acute adverse health effects. Due to daily meteorological effects on the dispersion of pollutants, explanatory spatial covariables and their coefficients may not be the same as in classical land-use regression (LUR) modeling for long-term exposure. In this paper, we propose a two-stage spatiotemporal model for daily fine particulate matter (PM_2.5) concentration prediction: first, daily nonlinear temporal trends are estimated through a generalized additive model, and second, GIS covariates are used to predict spatial variation in the temporal trend-removed residuals. To account for spatial dependence on meteorological conditions, the dates of the study period are divided by the sill of the daily empirical variogram into approximately temporal-invariant subgroups. Within each subgroup, daily PM_2.5 estimations are obtained by combining the temporal and spatial parts of the estimations from the two stages. The proposed method is applied to the modeling of spatiotemporal PM_2.5 concentrations observed at 18 ambient air monitoring stations in Taipei metropolitan area during 2006-2008. The results showed that the PM_2.5 concentrations decreased whereas the relative humidity and wind speed increased with the sill subgroups, which may be due to the effects of daily meteorological conditions on the dispersions of the particles. Also, the covariates and their coefficients of the LUR models varied with subgroups and had in general higher adjusted R-squares and smaller root mean square errors in prediction than those of a single overall LUR model.