This study aims to assess performances of eleven Machine Learning (ML) methods in predicting the Groundwater Quality Index (GWQI) for Yazd, an arid province in Iran. The ML models encompass Multiple Linear Regression (MLR), Support Vector Regression (SVR), K-Nearest Neighbors, Decision Tree Regression, Adaptive Boosting or AdaBoost, Random Forest Regression, Gradient Boosting Regression (GBR), XGBoost Regression (XGBR), Gaussian Process (GP), Artificial Neural Network (ANN), and Multi-Gene Genetic Programming (MGGP). After optimizing ML hyperparameters, ML-based estimation models were developed for three scenarios depending on which water quality parameters were used as input data: (1) K+ and pH; (2) K+, pH, Na+, Ca2+, SO4 2-, HCO3 - and Mg2+; and (3) K+, pH, Na+, Ca2+, SO4 2-, HCO3 -, Mg2+, Cl-, EC, TH, and TDS. For each scenario, ML-based models were assessed further by conducting (i) reliability analysis, (ii) ranking analysis, and (iii) confidence limits check. The results of the first scenario (with two input data) demonstrated the superiority of ANN, MGGP and GP, whereas ANN, MGGP and GBR were the most robust for the second scenario (with seven input data). Furthermore, the ranking analysis indicated that MLR, GP and ANN achieved the first highest ranks when eleven water quality parameters (third scenario) were used. The reliability analysis revealed that GP, MGGP, MLR, ANN, GBR, and XGBR achieved the highest reliability percentages across different scenarios, with ANN consistently ranking among the top models. Finally, the comprehensive comparative analysis of ML performances in this study reveals their potential for predicting GWQI.

Keywords