Research On Remote Sensing Retrieval Of Multi-type Observation Data Of Near-ground PM_2.5 Concentration Based On Deep Neural Network

In recent years,PM_2.5 pollution in China has greatly affected the daily life of Chinese people.How to effectively use satellite remote sensing data to invert PM_2.5 is of great significance to explore the spatiotemporal evolution law of PM_2.5.At present,most Of the satellite remote sensing methods for atmospheric PM_2.5 inversion are based on the Aerosol Optical Depth（AOD）products,which are usually retrieved from Top-Of-Atmosphere Reflectance（TOA）.The mapping model between TOA products and the PM_2.5 concentration monitored by ground stations can effectively reduce the error transfer caused by AOD inversion,but TOA actually couples the surface reflectance and atmospheric reflectance at the same time at present.Based on this,this paper uses the Himawari-8（H8）satellite data and carries out atmospheric correction by 6S model to statistically obtain the surface reflectance relationship between the first 6 bands of H8.Then,the surface reflectance of the first 5 bands is estimated by using the apparent reflectance of the sixth band of the satellite which is close to the surface reflectance.After deducting the surface reflectance.And finally,Atmospheric Reflectance（ATM）is obtained to complete the ground-air decoupling of TOA.In addition,based on the deep neural network,a near-surface PM_2.5 inversion model was constructed with atmospheric reflectance,satellite brightness temperature data,observation Angle,time,meteorological data,DEM and other characteristic parameters.The main research contents and conclusions are as follows:1.In this paper,taking part of the Yangtze River Delta as an example,three machine learning algorithm models were built respectively based on AOD inversion method(called AOD-PM_2.5 method),TOA inversion method(called TOA-PM_2.5method)and ATM inversion method(called ATM-PM_2.5 method).The comparison of the inversion results of these models shows that the proposed ATM-PM_2.5 method has the highest accuracy,compared with AOD-PM_2.5 method and TOA-PM_2.5 method.On the verification site,the R² and RMSE values of the proposed ATM-PM_2.5 method are0.88 and 14.70 ug/m³,which are 8.64%and 1.14%higher than that of AOD-PM_2.5method and TOA-PM_2.5 method respectively.RMSE value decreased by 3.98 ug/m³and 0.76 ug/m³,respectively.2.In order to further explore the feasibility of nighttime PM_2.5 inversion,combined with the special type of Himawari-8 satellite with high time resolution（10min/time）,and considering the time sequence factor,the daytime PM_2.5 value of the current date and the daytime PM_2.5 value of the next day observed by the satellite were taken as input.The PM_2.5 ground monitoring value of the current night（17:00 on the same day-07:00 on the next day）was used as the Y value for model training.In addition,PM_2.5concentration was estimated（hourly mean）for part of the night period from December17 to 18,2018（20:00 that day to 4:00 am the next day）by the model,and the monitoring values of the station were compared with the estimated values of the model.The accuracy of the inversion model in the night period was decreased compared with that in the day period.The highest R² is 0.66,the lowest R² is 0.5,the lowest RMSE is 18.61ug/m³,and the highest is 25.84 ug/m³.3.In order to verify the accuracy of the deep learning network in this article on different sensor data sets,a PM_2.5 inversion model was constructed on the MODIS data set.Taking Anhui Province as an example,the accuracy of the test set and the accuracy of the new validation set not participating in the training R² are both above 0.6,and the RMSE is below 29 ug/m³.However,due to the amount of data and the frequency of observations are less than that of H8,its accuracy is not as good as H8.4.Aiming at the problem that the current inversion results need to be quantified and standardized,this paper designed an automated and intelligent PM_2.5 inversion software based on the previous H8 inversion method model to realize the batch automation of PM_2.5 inversion.Based on the developed software,PM_2.5 high-frequency monitoring was carried out in typical areas.The results show that the proposed method has the potential to provide data support for the study of PM_2.5 dynamic distribution and real-time monitoring.