PM_2.5 Concentration Estimation And Influencing Factors Analysis Based On Multi-source Dat

High-resolution PM_2.5 spatial distribution data is of great significance for dynamic monitoring and control of PM_2.5 pollution.Himawari-8 AOD data,ERA5 meteorological reanalysis data,DEM,land use data,and night light remote sensing data were selected as Estimating variables,using an improved resampling method for data matching,and an improved linear mixed model combined with Geo-intelligent random forest to estimate PM_2.5concentration.Based on monthly average gridded PM_2.5 concentration data in Sichuan Province created by the i LME+Geoi-RF model,we use stacked machine learning to analyze the influencing factors of PM_2.5 concentration and their interaction effects in Sichuan Province on a monthly time scale and with a 5 km×5 km grid as the evaluation unit.The results reveal that（i）The prediction accuracy of the i LME+Geoi-RF model has been greatly improved compared to other models.Model fitted R²,RMSE,and MAE were 0.94,5.72μg·m^-3,and3.92μg·m^-3 and cross-validated R²,RMSE,and MAE were 0.82,10.20μg·m^-3,and 6.44μg·m^-3,respectively.The model can provide a more reasonable scientific reference for regional air quality assessment,human exposure risk assessment and environmental pollution control.（ii）There was a significant seasonal difference in PM_2.5 concentration in Sichuan Province.In terms of spatial distribution,the PM_2.5 concentration of Sichuan Province in the east area was generally higher than the west,and the local pollution level was relatively high.The high-valued areas were mainly distributed in the eastern region where the cities have been developing rapidly and the population was densely distributed,the low-valued areas were mainly distributed in the western region where economic development is backward and sparsely populated.（iii）Although the overall distribution of PM_2.5concentration estimated by different models is basically the same,the i LME+Geoi-RF model can more accurately and effectively estimate the spatial distribution of pollution in this study area.（iv）Compared with a convolutional neural network,a deep belief network,adaptive boosting,extreme gradient boosting and random forest,the stacked machine learning model has a better ability to simulate the relationship between PM_2.5 concentrations and influencing factors,with higher accuracy and a lower error rate.（v）There was significant spatial autocorrelation and spatial heterogeneity of the PM_2.5concentration in Sichuan Province from September 2015 to August 2020,and the regional and clustering characteristics of air pollution are obvious,with PM_2.5 concentrations mainly showing high-high clustering and low-low clustering.（vi）Among the influencing factors selected in this paper,PS,TEMP,DEM and CVL are the key factors influencing PM_2.5 concentrations in Sichuan Province,but the key influencing factors of PM_2.5 concentrations differ in by month.Therefore,identifying the key influencing factors of PM_2.5 concentrations in Sichuan Province in different months is crucial for the accurate management of air quality.（vii）Among the influencing factors selected in this paper,PS∩TEMP,PS∩NL,TEMP∩NL,BLH∩PS,RH∩PS,DEM∩NL,PS∩RAIN,NDVI∩PS,RH∩PS,WS∩TEMP,TEMP∩RAIN and BLH∩TEMP are the key interactions that influenced PM_2.5concentrations in Sichuan Province,but the influence of interacting factors on PM_2.5concentrations varied by month.Therefore,the identification of key interaction factors influencing PM_2.5 concentrations in Sichuan Province on a monthly time scale can provide some scientific evidence to support PM_2.5 pollution prevention and control.