Accurate prediction of groundwater salinity in arid and semi-arid regions remains a major challenge due to complex, nonlinear interactions among climatic, geochemical, and anthropogenic factors. This study presents a novel hybrid machine learning framework for forecasting monthly groundwater electrical conductivity (EC) in the Shamil–Ashkara aquifer, located in northeastern Hormozgan Province, Iran, a region characterized by high evaporation, intensive agriculture, and complex hydrogeological conditions. Utilizing 26 years of monthly hydrochemical data from 34 monitoring wells, the workflow integrates phase space reconstruction (PSR) to capture latent temporal patterns, Boruta feature selection algorithm (BR) to identify key hydrochemical variables, and systematic hyperparameter tuning using grid search (GR). Both support vector machine (SVM) and radial basis function neural network (RBFNN) models were developed, with uncertainty quantified through quantile regression and bootstrap resampling. Model interpretability was achieved via Shapley additive explanations (SHAP) analysis, revealing the dominant roles of total dissolved solids (TDS), sodium (Na + ), and chloride (Cl - ) in EC variability. The optimized hybrid model (BR-PSR-RBFNN-GR) achieved superior performance (testing R 2 = 0.993, RMSE = 0.067) with the lowest predictive uncertainty. These findings demonstrate that integrated, interpretable machine learning approaches offer robust and reliable solutions for groundwater quality assessment in complex arid basin environments.
Zamani et al. (Mon,) studied this question.