Machine learning-based prediction and classification of seawater intrusion in the hyper-arid coastal aquifer of Fujairah, UAE

Study region: This study focuses on an unconfined coastal aquifer located in the Emirate of Fujairah, United Arab Emirates—a hyper-arid region where the coastal aquifer is vulnerable to seawater intrusion (SWI), posing a threat to groundwater quality. Study focus: Fifteen machine learning (ML) algorithms were evaluated to predict and classify total dissolved solids (TDS) as an indicator of SWI. The models were trained using six hydrogeological parameters: rainfall, hydraulic head, distance from the coastline, aquifer saturated thickness, hydraulic conductivity, and specific yield. Model performance was assessed for both prediction and classification tasks. LightGBM yielded the highest prediction accuracy (R² = 0.9574), followed by CatBoost (R² = 0.9565) and XGBoost (R² = 0.9517). For classification, CatBoost and Gradient Boosting achieved the best performance with 97.6 % accuracy and AUC = 0.9986. The top-performing prediction models were utilized to derive empirical formulas for estimating groundwater TDS, informed by variable importance analysis of the six hydrogeological parameters. New hydrological insights for the region: Variable importance analysis highlighted hydraulic head and distance from the coast as key predictors of TDS, consistent with established SWI mechanisms. The study demonstrates that ML can be an effective alternative to traditional modeling approaches in data-limited environments, offering a robust SWI assessment. The developed empirical equations provide practical tools for local groundwater TDS estimation. Future work should address temporal dynamics and salinity vertical distribution patterns.