Comparing different methods for statistical modeling of particulate matter in Tehran, Iran

Particulate matter has major impacts on human health in urban regions, and Tehran is one of the most polluted metropolitan cities in the world, struggling to control this pollutant more than any other contaminant. PM2.5 concentrations were predicted by three statistical modeling methods: (i) decision tree (DT), (ii) Bayesian network (BN), and (iii) support vector machine (SVM). Collected data for three consecutive years (January 2013 to January 2016) were used to develop the models. Data from the initial 2 years were employed as the training data, and measurements from the last year were used for testing the models. Twelve parameters, covering meteorological variables and concentrations of several chemical species, were explored as potential predictors of PM2.5. According to the sensitivity analysis of PM2.5 by SVM and derived explicit equations from BN and DT, PM10, NO2, SO2, and O3 are the most important predictors. Furthermore, the impacts of the predictors on the PM2.5 were assessed which the chemical precursors’ influences indicated more in comparison with meteorological parameters. Capabilities of the models were compared to each other and the support vector machine was found to be the best performing, based on evaluation criteria. Nonetheless, the decision tree and Bayesian network methods also provided acceptable results. We suggest more studies using the SVM and other methods as hybrids would lead to improved models.