Remote Sensing Retrieval Of Aerosol Optical Depth Based On Deep Learning And Analysis Of Its Spatiotemporal Patterns And Driving Forces

Atmospheric aerosol has far-reaching effects on global climate change,cloud properties,land-atmosphere radiation balance,air quality and human health.Aerosol optical depth(AOD)retrieved by remote sensing images is one of the most basic and critical optical parameter,its quantitative high precision retrieval,spatial and temporal distribution characteristics and pollution origin is indispensable for investigations concerned with climate change and the formulation of air pollution control measures.As an efficient detection method,space-based optical satellite remote sensing has the technical advantages of high spatial coverage,high timeliness and high-precision monitoring,which effectively makes up for the shortcomings of traditional ground-based remote sensing sites,such as discontinuous spatial distribution and high cost for aerosol monitoring.Compared with traditional,complex and inefficient physical models,machine learning,as an empirical statistical model,has become a new research method in the field of aerosol remote sensing.With the continuous development of aerosol remote sensing observation platforms and retrieval methods,reducing the uncertainty of satellite sensor data,optimizing models to improve the retrieval efficiency and accuracy of AOD,and fully exploring the driving factors of regional aerosol pollution have always been difficulties and challenges in this field.Based on the radiation transmission model theory,this study at first focused on cloud masking,band extraction,resampling and other pre-processing operations on Sentinel-2 high-resolution image data;constructing three aerosol retrieval algorithms based on machine learning,and selects the optimal model to achieve high-precision retrieval of AOD in Henan Province by combining ground monitoring data,AOD product data and high-resolution remote sensing data.Thirdly,the distribution characteristics of the AOD at different spatio-temporal scales in the study area were discussed in detail,and the driving force of the AOD was explained by meteorological factors,human factors and topographic factors.This effectively explored the application issues of high-resolution satellite data for a regional monitoring of atmospheric parameters by remote sensing techniques,and also provided a theoretical basis for a subsequent regional environment change monitoring.The main results and conclusions of this study are as follows:(1)Combining with the Sentinel-2 MSI satellite sensor’s multi-band minimum apparent reflectance data and MODIS AOD data,the surface reflectance estimation method is optimized based on different strategies in dark targets and bright surface areas,and the surface reflectance of high-precision blue light band(B2),red light band(B4)and short-wave infrared band(B12)is obtained,and a high-resolution surface reflectance library in Henan is constructed.Compared with the existing Sentinel-2 MSI sensor surface reflectance estimation method,the reliability and availability of data are improved,and the accuracy and applicability of AOD remote sensing inversion model are improved.(2)Construction of an optimal retrieval model for measuring AOD in the study area based on the sample data:The aerosol retrieval model based on a back propagation(BP)neural network has the best fitting effect,with a correlation coefficient of 0.907,and an RMSE of 0.128.The long short-term memory(LSTM)retrieval model can handle the long-term dependence relationship in the time series data,whereby the correlation coefficient results in 0.843,and the RMSE in 0.142.The correlation coefficient of a convolutional neural network retrieval model is 0.823 and the RMSE is 0.124.The accuracy of LSTM model is better than that of convolutional neural network(CNN)model.Overall,the BPNN model has the best iteration effect and robustness to outliers.The results indicate that this model greatly improved the accuracy and efficiency of the quantitative aerosol retrieval,and thus,can effectively reflect the regional variation of urban aerosols.(3)The Sentinel-2 AOD product and the LGHAP AOD were used to analyze the spatio-temporal distribution characteristics of aerosols at different scales over the Henan province.The analysis showed that the annual mean value of the AOD in the Henan Province showed a decreasing trend from 2018 to 2020,and obvious directional and seasonal differences in spatial distribution.The air quality in winter and spring was poor and strongly directional,while the air quality in summer was the best and a spatial distribution was not obvious.The spatial pattern of the AOD was high in the east and low in the west,and there was a strong correlation between different regions.Differences in landscape properties had a great influence on the AOD,and the positive contribution of the AOD at impervious areas was the largest.Woodland,grassland,shrub and water bodies contributed to the mitigation of the aerosol pollution.These findings effectively reveal the spatial and temporal distribution patterns of aerosol pollution in the Henan Province.(4)The driving factors of atmospheric aerosol optical depth and their interactions on the distribution characteristics of the AOD were quantified and explained through geographic detector and principal component analysis.In the principal component analysis,the load of wind speed was 0.915,which was much higher than other factors.The meteorological factor had the most obvious effect on the AOD in the study area.The results of the geographic detector show that the AOD increases correspondingly with the increase of wind speed,temperature or precipitation,and the interaction between meteorological factors strengthens the driving force of the AOD.During the epidemic period,the dust transport caused by adverse meteorological conditions such as the monsoon was the main contribution of the aerosol pollution in the Henan Province.This conclusion can provide a theoretical guidance for the government’s air quality management in the Henan Province.Summarizing it can be stated that the derived results clearly reveal the necessity to intensify the study of the mechanisms of regional aerosol pollution and their driving forces and to further improve the AOD retrieval algorithms,especially the estimation accuracy by investigating the reflectance characteristics of various complex surface formations.