Spatial-temporal PM2.5 Prediction Using MODIS AOD Products

Abstract

In recently decade haze in China has severely hurt its economy and threatened the health of its population. There is often strong demand from the Ministry for the Environment for assessing, predicting, and trying to reduce the levels of PM2.5 around the country. In practice, PM2.5 data is difficult to measure. Monitor sites are not distributed uniformly, most of them built in urban area. Traditional air pollution epidemiology studies being conducted in large cities can be limited by the availability of monitoring. Satellite Aerosol Optical Depth (AOD) measurements offer the possibility of exposure estimates for the entire population. In this situation, the 10 km MODIS Aerosol Optical Depth (AOD) product can be used as predictor since recent studies has proved the statistical relationship between AOD and PM2.5. The traditional statistical study on AOD and PM2.5 are primarily Geographic Weighted Regression. Based on Gaussian process regression, this study developed a new regression approach to predict PM2.5 distribution in a Bayesian hierarchical setting from October 2016 to October 2017. The spatial non-stationarity was modeled by a Gaussian process with exponential covariance function. Parameters to explain factors like AOD, spatial random effects and non-spatial factors were estimated via a Bayesian hierarchical framework. The result illustrated that our model showed a good daily prediction on unknow sites by giving a 0.76 R^2 under 10 cross validation and a precise annual prediction with R^2 equal to 0.90. For daily model, we compared our result with GWR and a machine learning method support vector machine (0.68 and 0.75 respectively), which showed modeling spatial random effects via Gaussian process was able to improve the accuracy PM2.5 predicting using MODIS AOD data.