PSO-Optimized Ensemble Learning with SHAP for Seasonal Groundwater Quality Prediction

Groundwater is the principal source of freshwater in hyper-arid regions such as the Algerian Sahara, making reliable groundwater quality assessment essential for sustainable water resource management. This study developed an interpretable hybrid modeling framework for predicting the groundwater WQI in the Ghardaia region by integrating Particle Swarm Optimization (PSO) with ensemble machine learning algorithms. A total of 540 groundwater samples were collected from 67 artesian boreholes during the dry/wet seasons over the period 2021 to 2025. To ensure robust model evaluation and minimize data leakage, a spatially independent validation strategy. Based on borehole locations were adopted. Three ensemble algorithms, namely Random Forest (RF), Adaptive Boosting (AdaBoost), and Extreme Gradient Boosting (XGBoost), were optimized using PSO and trained separately for the dry/wet-season datasets. Among the evaluated models, PSO-AdaBoost achieved the best predictive performance during the dry season, with R² = 0.976, whereas PSO-XGBoost performed best during the wet season, with R² = 0.985. Hydrochemical variables' contribution to WQI predictions was quantified by investigating model interpretability using SHapley Additive exPlanations (SHAP). SHAP analysis revealed that potassium (K⁺) was the most influential predictor during the dry season, reflecting strong evapoconcentration and water-rock interaction processes. pH and magnesium (Mg²⁺) became relatively more influential during wet-period sampling campaigns, suggesting temporal hydrochemical variability potentially associated with physicochemical equilibrium changes and indirect hydrodynamic influences. The proposed PSO-ensemble-SHAP framework provides an accurate and interpretable approach for groundwater quality prediction and supports efficient monitoring and management of groundwater resources in arid aquifer systems.