A satellite-based model for estimating PM2.5 concentration in a sparsely populated environment using soft computing techniques

We applied three soft computing methods including adaptive neuro-fuzzy inference system (ANFIS), support vector machine (SVM) and back-propagation artificial neural network (BPANN) algorithms for estimating the ground-level PM2.5 concentration. These models were trained by comprehensive satellite-based, meteorological, and geographical data. A 10-fold cross-validation (CV) technique was used to identify the optimal predictive model. Results showed that ANFIS was the best-performing model for predicting the variations in PM2.5 concentration. Our findings demonstrated that the CV-R2 of the ANFIS (0.81) is greater than that of the SVM (0.67) and BPANN (0.54) model. The results suggested that soft computing methods like ANFIS, in combination with spatiotemporal data from satellites, meteorological data and geographical information improve the estimate of PM2.5 concentration in sparsely populated areas.