Predictive modelling and seasonal analysis of water quality indicators: three different basins of Şanlıurfa, Turkey

The objective of this paper is twofold. First; we demonstrate the application of data mining techniques to predict quality indicators (TDS, Hardness, Na, Cl, SO $$_{4}$$ 4 ) of groundwater data measured in three different basins of Şanlıurfa. The determination of the potable water classes was predicted using six well-known classifiers with respect to the concentrations of five groundwater quality indicators collected from a total of 1240 sampling points in these basins. Among six data mining algorithms (KNN, J48, Naive Bayes, ANN, JRip and Random Forest), JRip and J48 performed best with 99% accuracy to correctly predict the water quality class at all stations. Second; we studied the effect of seasonal variation on the level of contamination and the physico-chemical properties. According to the seasonal average F-measure values of the classifiers, only the three worst water quality classes (C4, C5, C6) were observed in the Harran Plain. There was a similar seasonal class distribution with C4 and C5 classes in Sarım–Karataş and C5 and C6 classes in Ceylanpınar Plain. The highest contamination was detected in the summer period. When compared in terms of chemical quality indicators, the groundwater situation in Ceylanpınar Plain is better than Harran Plain. The groundwater quality conditions in Sarım–Karataş are substantially similar to Ceylanpınar Plain. According to the results of Kappa statistics, Random Forest, J48 and JRip value results were “Nearly Perfect”, while for Naive Bayes, means indicate “Moderate” to “Significant” level.