Applying linear and nonlinear models for the estimation of particulate matter variability

In this study, data collected from an urban air quality monitoring network are being used for the purpose of evaluating various methodologies used for spatial interpolation in the context of proposing an effective yet simple to apply scheme for PM spatial point estimations. The examined methods are the Inverse Distance Weighting, two linear regression models, the Multiple Linear Regression and the Linear Mixed Model, along with a Feed Forward Neural Network (FFNN) model. These schemes utilize daily PM₁₀ and PM₂.₅ concentrations collected from five and three air quality monitoring sites respectively. In order to obtain the resulted estimations, the leave-one-out cross-validation methodology is used for all methods. The evaluation of their predictive ability is performed by using a combination of difference and correlation statistical measures, scatter plots and statistical tests. The results indicate the usefulness of FFNNs as they are found to be statistically significantly superior for modelling the particulate matter spatial variability. The model performance statistics show that in most cases the error values are considerably lower for the FFNN model. Additionally, the rank and Wilcoxon rank tests reveal that the null hypothesis for equal predictive accuracy is rejected for the majority of monitoring sites and schemes (values lower than the critical t-value). According to the comparison results, the FFNN model is selected for forecasting air quality limit exceedances set by the European Union and World Health Organization air quality standards. For two monitoring sites in which the largest number of exceedances occurred, the probability of detection is high while the probability of false detection is very low, further establishing the neural networks’ predictive ability.