Study On PM_2.5 Inversion In The Beijing-tianjin-hebei Region Using Satellite Remote Sensing

Recently,the severe haze pollution over eastern China has received extensive attention.PM2.5 consists mainly of the ash haze and sharply influences on the environment and air quality as well as public health.Detailed surface-based PM2.5 concentrations of spatial-temporal data are needed to support of the assessment of PM2.5 exposure to health risks,carrying out the simulation and prediction research,and tracking of the emission sources.Routine Stationary ground measurement of PM2.5 with high sample frequency and accuracy,but restrained to its excessively high construction costs,sporadic and uneven spatial coverage,and limited monitoring range.Besides,different inversion algorithms of aerosol optical depth(AOD)products are inadequate in the utilization and coordination of data accuracy and data size;In addition,some simple statistical inversion methods of low precision and has the limitation of modeling data type and data quantum.All of these limit the research and application of PM2.5.To solve the above mentioned problems,In this study,we comprehensively utilize of groundbased PM2.5 observations and GEOS meteorological parameters together with MODIS Terra and Aqua satellites collection 6 of fusion of Dark Target(DT)and deep blue algorithm(DB)aerosol products in the year of 2015 over the beijing-tianjin-hebei region of china,settig up PM2.5 estimation statistical models based on the widely used multiple linear regression and machine learning algorithms(including regression tree,random forests,support vector machine)in two satellite data separately.The same training and testing datasets as well as methods are conformed for both Terra and Aqua satellite data in our study area separately to make all these models entirely comparable.The 10 repeat 5 fold cross-validation(CV)methods are used to compare with the rationality and estimation ability of our models.Results showed that the performance of AQUA corresponded data sets on the four models are better than TERRA data set,Random forests(RF)algorithm we firstly applied in this field has the best performance on both validation sets and testing sets in four kinds of models(TERRA: CV R = 0.767,RMSE = 43.512;AQUA: CV R = 0.843,RMSE = 33.904),Combining with some valuable functions of the algorism itself,RF algorithm was implied to have a great deal of potentials to improve the ground truth PM2.5 estimation.Applying this optimal machine learning algorithm--RF,we evaluate the daily PM2.5 concentration.Randomly selected three monitoring sites are used to plot the time series and validating a good model performance.At the same time,PM2.5 site kriging interpolation maps are plotted in the time of MODIS passing territory.We compare with the estimation difference of these two datasets,daily and seasonal averaged maps are demonstrated to explore the spatial distribution and variation characteristics.Research shows that the RF model can improve the low-value overestimate and high value underestimate phenomenon in the process of PM2.5 inversion and effectively improve the retrieval accuracy;Inversion image has a continues and fine space coverage,can basicly reflect the spatial distribution of PM2.5 in the study area accurately,while for the lack of AOD,the retrieval maps can't reflect the seasonal variation characteristics accurately.Anyway,this study provides a new way for PM2.5 inversion,and research data sets can provide a reference for related applications.