Multivariate Statistical Methods and Neural Networks for the Identification of Pollution Sources and Classification of the Apulia Region Ground Waters

Abstract

Multivariate statistical techniques such as Discriminant Function Analysis (DFA), Cluster Analysis (CA), Principal Component Analysis (PCA), Absolute Principal Component Score (APCS) and Neural Networks (NN) have been applied to a data set, of Apulian ground waters, formed by 1009 samples and 15 parameters: pH, Electrical Conductivity, Total Dissolved Solids, Dissolved Oxygen, Chemical Oxygen Demand, Na+, Ca2+, Mg2+, K+, Cl-, NO3-, SO42- and HCO3-, vital organism at 22 C and 36 C. Principal Component Analysis and Absolute Principal Component Scores allowed to identify, for each province, as well the sites diverging from the mean cluster, as the pollution sources (due to fertilizer applications, marine water intrusion, etc) pressurizing the sampling sites investigated. Discriminant Function Analysis allowed on the hand to identify variables with bigger discriminatory power, on the other to obtain good results in discriminating among the considered provinces and in forecasting. The application of Radial Basis Function Neural Networks gives results with bigger accuracy than DFA and confirms the electrical conductivity has the bigger relative importance.