Estimating the spatiotemporal variation of NO2 concentration using an adaptive neuro-fuzzy inference system

Abstract

Statistical modelling has been successfully used to estimate the variations of NO2 concentration, but employing new modelling techniques can make these estimations far more accurate. To do so, for the first time in application to spatiotemporal air pollution modelling, we employed a soft computing algorithm called adaptive neuro-fuzzy inference system (ANFIS) to estimate the NO2 variations. Comprehensive data sets were investigated to determine the most effective predictors for the modelling process, including land use, meteorological, satellite, and traffic variables. We have demonstrated that using selected satellite, traffic, meteorological, and land use predictors in modelling increased the R2 by 21%, and decreased the root mean square error (RMSE) by 47% compared with the model only trained by land use and meteorological predictors. The ANFIS model found to have better performance and higher accuracy than the multiple regression model. Our best model, captures 91% of the spatiotemporal variability of monthly mean NO2 concentrations at 1 km spatial resolution (RMSE 1.49 ppb) in a selected area of Australia.