BackgroundThree decades of rapid economic development is causing severe and widespread PM2.5 (particulate matter ≤ 2.5 μm) pollution in China. However, research on the health impacts of PM2.5 exposure has been hindered by limited historical PM2.5 concentration data.ObjectivesWe estimated ambient PM2.5 concentrations from 2004 to 2013 in China at 0.1° resolution using the most recent satellite data and evaluated model performance with available ground observations.MethodsWe developed a two-stage spatial statistical model using the Moderate Resolution Imaging Spectroradiometer (MODIS) Collection 6 aerosol optical depth (AOD) and assimilated meteorology, land use data, and PM2.5 concentrations from China’s recently established ground monitoring network. An inverse variance weighting (IVW) approach was developed to combine MODIS Dark Target and Deep Blue AOD to optimize data coverage. We evaluated model-predicted PM2.5 concentrations from 2004 to early 2014 using ground observations.ResultsThe overall model cross-validation R2 and relative prediction error were 0.79 and 35.6%, respectively. Validation beyond the model year (2013) indicated that it accurately predicted PM2.5 concentrations with little bias at the monthly (R2 = 0.73, regression slope = 0.91) and seasonal (R2 = 0.79, regression slope = 0.92) levels. Seasonal variations revealed that winter was the most polluted season and that summer was the cleanest season. Analysis of predicted PM2.5 levels showed a mean annual increase of 1.97 μg/m3 between 2004 and 2007 and a decrease of 0.46 μg/m3 between 2008 and 2013.ConclusionsOur satellite-driven model can provide reliable historical PM2.5 estimates in China at a resolution comparable to those used in epidemiologic studies on the health effects of long-term PM2.5 exposure in North America. This data source can potentially advance research on PM2.5 health effects in China.CitationMa Z, Hu X, Sayer AM, Levy R, Zhang Q, Xue Y, Tong S, Bi J, Huang L, Liu Y. 2016. Satellite-based spatiotemporal trends in PM2.5 concentrations: China, 2004–2013. Environ Health Perspect 124:184–192; http://dx.doi.org/10.1289/ehp.1409481

Estimating ground-level PM2.5 from satellite-derived aerosol optical depth (AOD) using a spatial statistical model is a promising new method to evaluate the spatial and temporal characteristics of PM2.5 exposure in a large geographic region. However, studies outside North America have been limited due to the lack of ground PM2.5 measurements to calibrate the model. Taking advantage of the newly established national monitoring network, we developed a national-scale geographically weighted regression (GWR) model to estimate daily PM2.5 concentrations in China with fused satellite AOD as the primary predictor. The results showed that the meteorological and land use information can greatly improve model performance. The overall cross-validation (CV) R(2) is 0.64 and root mean squared prediction error (RMSE) is 32.98 μg/m(3). The mean prediction error (MPE) of the predicted annual PM2.5 is 8.28 μg/m(3). Our predicted annual PM2.5 concentrations indicated that over 96% of the Chinese population lives in areas that exceed the Chinese National Ambient Air Quality Standard (CNAAQS) Level 2 standard. Our results also confirmed satellite-derived AOD in conjunction with meteorological fields and land use information can be successfully applied to extend the ground PM2.5 monitoring network in China.

To estimate PM concentrations, many parametric regression models have been developed, while nonparametric machine learning algorithms are used less often and national-scale models are rare. In this paper, we develop a random forest model incorporating aerosol optical depth (AOD) data, meteorological fields, and land use variables to estimate daily 24 h averaged ground-level PM concentrations over the conterminous United States in 2011. Random forests are an ensemble learning method that provides predictions with high accuracy and interpretability. Our results achieve an overall cross-validation (CV) R value of 0.80. Mean prediction error (MPE) and root mean squared prediction error (RMSPE) for daily predictions are 1.78 and 2.83 μg/m, respectively, indicating a good agreement between CV predictions and observations. The prediction accuracy of our model is similar to those reported in previous studies using neural networks or regression models on both national and regional scales. In addition, the incorporation of convolutional layers for land use terms and nearby PM measurements increase CV R by ∼0.02 and ∼0.06, respectively, indicating their significant contributions to prediction accuracy. A pair of different variable importance measures both indicate that the convolutional layer for nearby PM measurements and AOD values are among the most-important predictor variables for the training process.

No abstract

No abstract