A wavelet-ANN-based framework for estimating air pollutant concentrations using remotely sensed data in Tehran metropolitan area

In developing countries, most major cities are increasingly exposed to air pollution as a factor affecting the quality of life and public health of the community. High population density in Tehran causes this metropolitan area to be one of the most important region in Iran. Polluting industry and the use of polluting transportation are two of the main sources of air pollutant in Tehran and have turned this city to the most polluted metropolitan area in Iran. Consequently, the need for the air pollution reduction is too necessary in this area. The air pollutant concentration predictions can improve decision making for appropriate solutions to reduce air pollution. Since more precise methods are required to predict air pollutants for better management of this problem, using hybrid methods can be an important step in modeling different pollutants. This study examined the performance of the random forest feature selection and wavelet transformation methods when they combine with the multiple-linear regression and multilayer perceptron artificial neural network to achieve an efficient model to estimate several pollutants including carbon monoxide, nitrogen dioxide, sulfur dioxide, and PM2.5 in Tehran metropolitan area. For these purpose four groups of remotely sensed-derived and spatial data including spatial data, meteorological data, traffic information, and the air pollutant concentrations in the days before the prediction day were applied as the input data of the models. Results showed that the modeling of all pollutants by the multilayer perceptron neural network along with the wavelet transform method provides higher accuracy than the other models. Furthermore, the estimation accuracy of the carbon monoxide pollutant (with error of estimation=19.8% ) was lower than the other pollutants while PM2.5 (with error of estimation=17.0%) was estimated with higher accuracy compared to that derived for other pollutants. Moreover, it was shown that the pollutant concentrations for the days before the day for that the estimation is implemented are the most important attributes, according to the random forest feature selection method.