Research On Spatial Reconstruction Of Aerosol Optical Depth Data

In the context of rapid urbanization and industrialization,airborne particulate matter pollution has become an inevitable environmental problem.AOD is an important characterization of aerosol content and atmospheric particulate matter pollution,so it is essential to reconstruct AOD spatially to further understand air environment quality,better provide data support for air compound pollution research,and provide long-term basic information for epidemiology and health and hygiene effect research in China.AOD data are mainly obtained by satellite remote sensing,which can provide long time and large scale data sets,but due to the different design principles of sensors,calibration process and inversion algorithms,the inversion AOD products have certain overestimation and underestimation;and due to other factors such as cloud cover and orbit interval,the AOD data are missing.In this thesis,the research direction is to improve the accuracy and coverage of AOD data,and conduct spatial reconstruction of AOD data.Taking MODIS AOD data as the research object,combined with machine learning technology,we explore more efficient AOD spatial reconstruction algorithms to improve the accuracy and coverage of AOD data.The key research contents are as follows:1.A machine learning based AOD bias correction model is proposed to solve the problem of overestimation and underestimation of AOD data.MODIS AOD data are used as the research object,and MODIS AOD data at different scales are used with EAR5 meteorological data and elevation data for spatio-temporal configuration,characteristic selection,and normalization to construct the data set,and machine learning methods(ANN,SVR,XGBoost)are applied to the constructed data model for bias correction and comparison analysis,and the ANN model is verified as the best AOD bias correction model.2.The random forest(GA-RFR)AOD data regression prediction model based on genetic algorithm optimization is proposed to solve the missing AOD data problem.Using the bias corrected data,the AOD data is filled with the random forest regression(RFR)model,in the process of parameter optimization using grid search cross-validation method in the RFR model,the constant search step length may cause the underfitting and overfitting problems.The GARFR regression model can effectively avoid the overfitting and underfitting problems by using the excellent global search capability of genetic algorithm,and the GA-RFR model has higher accuracy of data with guaranteed coverage by comparative analysis.3.A prototype system for spatial reconstruction of aerosol optical depth data was designed and implemented.Based on the above algorithms research,the platform architecture,database,and system functions were designed,and the “AODSRSystem”,which integrates data preprocessing,data management,aerosol optical depth deviation correction,filling,and thematic mapping,was realized.“AODSRSystem” can support the study of aerosol optical depth data spatial reconstruction method,and has important application value.