Modeling spatial-temporal changes in PM2.5 concentration based on data imputation and the use of machine learning methods in different geographical contexts of the Tehran metropolis

Management of exposure and dealing with the consequences of the concentration of PM2.5 in urban environments requires accurate modeling of spatial-temporal changes of pollutant. Accurate modeling of spatial-temporal changes requires appropriate modeling methods and complete and accurate data. These data are measured by different sensors and with different accuracy, have different variability and due to unavoidable factors such as sensor damage. Missing data cause many problems such as loss of sample size and errors in data analysis; therefore, it is necessary to use solutions to estimate the missing data in modeling the concentration of PM2.5.  In this study, a method based on extra tree and decision tree models has been proposed to imputation the missing values of PM2.5 along with considering the relationships between variables while maintaining their variability and natural uncertainty. Meteorological variables and other main pollutants such as O3, Pm10, Co, So2, No2 were considered as effective variables in imputation the missing values of PM2.5. Meteorological variables including total precipitation, relative humidity, and temperature were extracted from the model of the European Center for medium-term weather forecasting. Using the ECMWF model, in addition to increasing the number of meteorological stations, provides the possibility of using hourly resolution with a very small number of missing data, as opposed to a limited number of three-hour resolutions with a large number of missing meteorological data. The results showed that the extra tree method has a higher accuracy than the decision tree method with an average of R2=0.813 due to the reduction of bias with an average of R2=0.653 in imputation of missing PM2.5 values. After managing the missing data using the extra tree method, the XGBoost method was used due to the non-linear evaluation of the importance of the effective variables with the aim of increasing the accuracy and reducing the computational cost for modeling the spatial-temporal changes of the PM2.5 pollutant in different geographical contexts.