Balancing accuracy, interpretability, and cost in groundwater quality analysis and prediction using advanced hybrid models

Groundwater quality degradation driven by climate change and anthropogenic pressures represents a major challenge for sustainable irrigation in arid and semiarid regions. Although artificial intelligence (AI) methods have been widely applied for Irrigation Water Quality Index (IWQI) prediction, many existing models lack interpretability and do not consider monitoring costs. Therefore, this study develops an integrated framework that combines predictive performance, explainability, and cost-effectiveness for groundwater quality assessment. A dataset of 621 groundwater samples from Béchar, Algeria, including 10 physicochemical parameters, was analyzed using two modeling approaches: (i) a stacked ensemble integrating Isolation Forest, Random Committee, Random Forest, and CatBoost with logistic regression, and (ii) a CNN encoder–decoder model. SHAP analysis was used to interpret model predictions, while feature-based optimization was applied to evaluate cost-effectiveness. The results identified electrical conductivity (EC), sodium (Na + ), and chloride (Cl − ) as the dominant IWQI drivers, with a strong negative correlation between IWQI and Na + (r ≈ − 0.94). The CNN encoder–decoder achieved the best overall performance, with Model M4 producing micro-F1 and macro-F1 scores of 0.957 and 0.947, respectively. SHAP analysis confirmed Na + and EC as the most influential variables. Cost analysis showed that Model M9, based on only four salinity-related parameters and a total cost of USD 39, maintained an accuracy above 92% and achieved the highest cost-effectiveness score of 0.02358. These findings demonstrate that interpretable and cost-optimized deep-learning models can support reliable groundwater quality assessment and provide practical decision-support tools for sustainable irrigation water management in data-limited regions.