Spatiotemporal Regression Kriging For Estimating PM_2.5 Concentration Using Aerosol Optical Depth Remote Sensing Data

Particulate matter with aerodynamic diameters of less than 2.5 ?m is collectively referred to as fine particle matter,i.e.PM_2.5.PM_2.5 is one of the major air pollutants,which causes a seriously harmful effect on human health.With the rapid development of economy,air pollution in mainland China has been extremely serious,and East China with developed economy and high population density is one of the most heavily pollution areas.A large number of anthropogenic emissions of living and industrial pollutants bring high PM_2.5 concentration,resulting in frequent haze weather in East China.In order to effectively monitor PM_2.5 concentration and provide timely and accurate haze warnings,an air quality monitoring network throughout China has been gradually established from 2012.But the sparseness of air monitoring sites,the uneven distribution of sites between eastern and western areas,between urban and rural areas,and the lack of historical PM_2.5 observations,all greatly limit the epidemiological studies of PM_2.5 in Chinese population.Given the strong association between PM2 5 and aerosol optical depth?AOD?,it is an important way to overcome the above-mentioned issues in ground observation by using AOD satellite remote sensing data with long observation time series and wide coverage to indirectly estimate PM_2.5 concentration.MODIS C6 3 km AOD products are used to construct the PM_2.5 estimation model from March 1,2015 to February 29,2016 in East China.However,due to the contamination of cloud and snow,large-scale data gaps usually exist on AOD products.In order to improve the coverage of AOD data in East China,a two-step linear regression model is used to fill Terra AOD with Aqua AOD and fill DT AOD with DB AOD in the case that their daily correlation coefficients are not less than 0.8.Then the combined 3 km AOD data is decomposed into a deterministic component,a random component,dimension-reduction and observation errors according to a spatiotemporal mixed effects model.The EM algorithm is used to iteratively estimate the parameters until the numerical schema converges.Based on the optimal model of each season,the MODIS 3 km AOD is completely filled using the fixed rank smoothing method.The results statistically correlated with the AREONET AOD show that the R2 of the gap-filled MODIS 3 km AOD is 0.73 during the whole period.And the highest R2 is 0.84 in summer while the lowest is 0.57 in winter.Given the close relation between PM_2.5 and AOD and the spatiotemporal autocorrelation of PM_2.5,the gap-filled MODIS 3 km AOD data,the PM_2.5 ground observation data,and the assimilation data of relative humidity and atmospheric boundary layer height are used to construct a season-specific spatiotemporal regression kriging model for PM_2.5 estimation in East China.The linear and median-polished spatiotemporal trend is extracted from PM_2.5 concentration,and the interpolation on residuals is completed by spatiotemporal simple kriging method.The results of cross validation show that R2 is 0.88 and RMSE is 16.55?g/m3 in the whole period.The estimation accuracies in autumn and winter are close,that is,R2 are both 0.88,and RMSE are 14.51 and 22.50?g/m3 respectively.The estimation accuracy is worse in spring with R2 being 0.82 and RMSE 14.81?g/m3,and the worst in summer with R2 being 0.79 and RMSE 14.81 ?g/m3.Furthermore,the estimation accuracies on the annual and seasonal mean PM_2.5 concentrations are better,with R2 being above 0.95 and RMSE below 4?g/m3.The annual mean PM_2.5 concentration and AOD in East China exhibit a nearly consistent spatial pattern,i.e.higher in north and lower in south.Located in a flat terrain adjacent to heavily pollution areas,which facilitates the transmission of pollutants,the north part in East China becomes the most heavily polluted area,followed by the central part with relatively developed economy and high population density.The southern part located in a hilly area away from heavily polluted areas,has ideal weather conditions for pollutants diffusion,and therefore becomes the most slightly polluted area.The spatial pattern of seasonal mean PM_2.5 concentration is quite different from that of AOD.The PM_2.5 concentration keeps a stable spatial pattern across all four seasons,while the spatial pattern of AOD exhibits a significant seasonal change.In addition,the mean PM_2.5 concentration and AOD occur at inconsistent levels in different seasons.The overall PM_2.5 concentration is the highest in winter,the lowest in summer,and in the middle in both spring and autumn,while the overall AOD in both spring and summer is much higher than that in both autumn and winter.The spatiotemporal interpolation of PM_2.5 residuals is the most computation-intensive part during model execution.However,the interpolation task of each grid point is highly independent without any communication overhead,which can be regarded as a data-parallel problem.Specifically,GPU is well-suited to address this kind of problem,and therefore a GPU-accelerated spatiotemporal kriging interpolation parallel algorithm is implemented with the CUDA programming specification.The search of spatiotemporal nearest neighbors,the computation of weights based on a linear system and the weighted average of spatiotemporal observations are programmed into kernel functions running on the GPU.The spatiotemporal interpolation experiments on the grids of different spatial resolutions show that the GPU-accelerated parallel algorithm significantly reduces the computing time,and the speedup is much higher than that of the CPU quad-cores parallel algorithm,e.g.,when dealing with 3 km grid,the speedup can reach 106.27,which is 26.7 times faster compared to the CPU quad-cores parallel algorithm.In short,the large-scale data gaps in MODIS 3 km AOD remote sensing data are completely filled using the fixed rank smoothing method based on a spatiotemporal mixed effects model,and its coverage in East China is significantly improved.Multi-source datasets are applied to construction of a season-specific spatiotemporal regression kriging model,which is used to estimate spatially and temporally continuous 3 km PM_2.5 concentration grid data in East China.And the spatial patterns and pollution levels between PM_2.5 and AOD are compared and analyzed at both annual and seasonal scales.In addition,a GPU-accelerated spatiotemporal kriging interpolation parallel algorithm programmed with the CUD A specification is developed to overcome:the computational bottleneck.The efficient and accurate estimation of PM_2.5 concentration is of great significance to the environmental epidemiology research and atmospheric environment protection in China.