Well-Scale Groundwater-Level Forecasting and Gap Reconstruction in Qatar’s Arid Aquifer Using Interpretable XGBoost

Reliable use of machine learning in groundwater management requires a clear distinction between forecasting, reconstruction, and spatial extrapolation. This study evaluates eXtreme Gradient Boosting (XGBoost) as a data-driven methodological benchmark, not as a process-based groundwater-flow model or monitoring-network design tool. Using 25,133 observations from 879 wells in Qatar’s carbonate aquifer system (2011–2016), we assessed three operational tasks: sequential one-step-ahead forecasting at monitored wells, retrospective gap reconstruction, and strict spatial transfer without local groundwater-level history. For forecasting, the principal no-climate model achieved RMSE = 0.31 m, approximately 40% lower than persistence (0.51 m). The held-out-well design included 107 of 664 supervised wells (16.1%; 3143 observations); on the common subset of 2856 observations from 92 wells, XGBoost achieved RMSE = 0.17 m versus 0.40 m for persistence. Gap reconstruction yielded RMSE values ranging from 0.37 m to 0.46 m across random- and block-masking scenarios; XGBoost outperformed persistence but was not consistently more accurate than linear or PCHIP interpolation. Strict spatial transfer failed when local history was unavailable (mean R2 = −0.29 ± 1.52 across four K-means blocks). SHapley Additive exPlanations (SHAP) indicated that antecedent groundwater levels accounted for approximately 99% of predictive importance. Thus, the framework can complement monitoring through short-horizon forecasting and gap reconstruction where local observations exist but should not replace field monitoring or be used for extrapolation to unmonitored areas.