Estimation of PM2.5 as a Harmful Environmental Hazard in Tehran by Fusion of MODIS Aerosol Products through a Machine Learning Approach

Air pollution is one of the most harmful natural hazards in Tehran metropolitan. Particles with a diameter of less than 2.5 micrometers (PM2.5) as one of the most harmful pollutants have endangered the health of people living in Tehran. One of the PM2.5 estimation techniques is the use of Aerosol Optical Depth (AOD) products derived from satellite observations. Various AOD products are retrieved with different algorithms that do not have the same accuracy and spatial resolution. Due to the differences in many assumptions and approximations adopted in the AOD retrieval process, the generated AOD products involve uncertainties. This issue reduces the accuracy of PM2.5 concentration estimation. The purpose of this study was to investigate the possibility of fusing AOD products obtained from MODIS sensor observations (retrieved by Deep Blue and Dark Target algorithms) to estimate PM2.5 more accurately. First, the performance of different machine learning algorithms in estimating PM2.5 from AOD data was evaluated. As a result, the XGBoost algorithm with the highest performance was selected as the base model for PM2.5 estimation. Then, the AOD products were fused using a weighted averaging based on the retrieval quality of the primary products. Finally, the fused AOD product along with meteorological data were employed to estimate PM2.5 using XGboost. The results demonstrated that the accuracy of PM2.5 estimation from the fused AOD product is better than when the AOD products are used individually (= 0.77, MAE = 7.00 , RMSE = 9.59 ). Thus, the retrieval quality of AOD products will lead to more accurate estimation of PM2.5 in the end.