MODFLOW-informed machine learning for interpretable groundwater level prediction under sparse monitoring conditions in the Eastern Mitidja Aquifer, Northern Algeria

Predicting groundwater levels in semi-arid regions is challenging because groundwater systems are controlled by complex processes and monitoring data are often scarce. This study compares six machine learning models—Random Forest (RF), XGBoost, LightGBM, RBF-SVR, LSTM, and CNN—for groundwater level prediction in the El Hamiz sub-watershed, northern Algeria. A dataset of 3944 records from 136 monitoring wells (2005–2022) was combined with hydroclimatic variables, satellite-derived indicators, and groundwater abstraction data. Two modelling strategies were evaluated: one trained with observed groundwater levels and another using groundwater heads simulated by a calibrated MODFLOW 6 model during a period with limited observations. The latter was intended to examine whether physically based simulations could support machine learning when groundwater observations were unavailable. RF achieved the best performance (R² = 0.883, NSE = 0.873, RMSE = 2.096 m, KGE = 0.850), followed by XGBoost and LightGBM. The MODFLOW-informed RF also performed well (R² = 0.830; RMSE = 2.320 m), although it did not surpass the observation-based model. A Leave-One-Well-Out analysis using MODFLOW heads as an additional predictor produced a pooled R² of 0.869 and an RMSE of 1.744 m. SHAP analysis identified spatial location, evapotranspiration, groundwater abstraction, and precipitation as the most influential predictors. Prediction intervals showed slight undercoverage, indicating that further uncertainty calibration is needed for groundwater management.