Assessing PM2.5 concentrations in Tehran, Iran, from space using MAIAC, deep blue, and dark target AOD and machine learning algorithms

This study aims to explore the spatial estimation of fine particulate matter (PM2.5) using 10-km merged dark target and deep blue (DB_DT) Aerosol Optical Depth (AOD) and 1-km Multi-Angle Implementation of Atmospheric Correction (MAIAC) AOD over Tehran. The ability of four Machine Learning Algorithms (MLAs) to predict PM2.5 concentrations is also investigated. Results show that the association of satellite AOD with surface PM significantly increases after considering the contribution of relative humidity in PM mass concentration and normalization of AOD to Planetary boundary layer height (PBLH). The examination of derived aerosol layer height (ALH) from 159 Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) profiles shows that PBLH could successfully represent the top of aerosol-laden layers. Surprisingly, the highest correlation was found between normalized 10-km DB_DT AOD and corrected PM2.5 measurements. Consequently, random forest (RF) fed by this AOD product has yielded the best performance (R-2 = 0.68, RMSE = 17.52 and MRE = 27.46%). Importance analysis of variables reveals that DB_DT and meteorological fields are of highest and least importance among selected variables, respectively. The RF performance is less satisfactory during summer which is assumed to be caused by the omission of unknown features representing the formation of secondary aerosols. The inferior accuracy of estimation in the north and east of Tehran is also linked to lacking features which could feed the transportation of PM(2.5 )from west to the east of the study area into MLAs.