Hourly PM_2.5 Estimation Using Apparent Reflectance And AOD Data From Himawari-8 Satellite

With the development of industrialization and urbanization,air pollution situation is becoming more and more serious in China.Due to its small particle size,large area,strong activity,easy to attach toxic and harmful substances,long residence time in the atmosphere and long transportation distance,PM_2.5 is very harmful to human health and atmospheric environment quality.At the end of 2012,China built a ground monitoring network of PM_2.5 in major cities,however,the spatial distribution of the stations is uneven and mainly concentrated in the central and eastern urban areas.Estimation of ground PM_2.5 based on Aerosol Optical Depth（AOD）products of satellite remote sensing is a research hotspot,which can compensate for the shortage of ground monitoring stations,with the characteristics of large area and high density coverage.At present,most of AOD products are provided by polar-orbiting satellites with low frequency of observations,and the hourly trend of PM_2.5 cannot be obtained.Moreover,the strict surface assumptions in the AOD inversion process make it impossible to obtain AOD in certain areas or periods.Therefore,this paper used the Apparent Reflectance（AR）and AOD data of geostationary satellite“Himawari-8”,meteorological data,population data,DEM,LUCC and ground PM_2.5 concentration observations data to estimate ground PM_2.5 concentration in China in January 2016and the Yangtze River Delta region in 2016 based on machine learning algorithm.Besides,the vertical distribution characteristics of aerosol obtained by ground-based LiDAR were used to correct AOD,and the AOD was used for PM_2.5 estimation.The main findings are as follows:（1）In order to reveal the advantages and disadvantages of the AOD-PM_2.5 and AR-PM_2.5 methods based on geostationary satellites in China,the PM_2.5concentrations estimated by the two methods in January 2016 were compared and analyzed.The results showed that the performance of the Deep Neural Network（DNN）model was better than that of the Gradient Boosting Regression Tree（GBRT）and the Random Forest（RF）models.The coefficient of determination（R²）was 0.86and the Root Mean Square Error（RMSE）was 26.81μg/m³.The AOD-PM_2.5 method performed better in China,while the AR-PM_2.5 method had less training samples in the sparse areas of the observation sites,which was not enough to represent the overall estimated sample;（2）In order to test the estimation ability of the AR-PM_2.5 method in the high-density site area,the Yangtze River Delta region was taken as an example to compare and analyze the PM_2.5 concentration estimated by the two methods in 2016.The results showed that the average,monthly and hourly mean values of the two estimation methods were basically same;However,due to the strict surface assumptions in the AOD inversion process,the PM_2.5 samples used in the AOD-PM_2.5method were less than one-half of the AR-PM_2.5 method,and most of them were in the case of pollution(PM_2.5 greater than 75μg/m³ was 26.10%),so the estimation result was obviously overestimated.There were many missing values in spatial distribution map of the single-day estimation results of PM_2.5.The AR-PM_2.5 method was closer to the actual situation than the AOD-PM_2.5 method,showing the advantages of this method in high-density sites;（3）Since the vertical distribution of aerosol is an important factor affecting the remote sensing of PM_2.5 estimation,the long-term（2013-2015）ground-based LiDAR data in Wuxi was used to reveal the vertical distribution characteristics of the aerosol extinction coefficient on both monthly and seasonal scales.The correlation of AOD-PM_2.5 was improved based on this,and the accuracy of remote sensing of PM_2.5estimation in heavily polluted areas was improved.