Bridging gaps in sparse groundwater data: classical, ensemble, and deep learning approaches for single- and multi-well imputation

Groundwater-levels are essential for aquifer management and policy-making, yet national monitoring networks often contain substantial missing data. Imputing these gaps is especially challenging in systems with scarce and irregular measurements. This study evaluates groundwater-level imputation in the Grombalia shallow aquifer using five methods: Auto-Regressive Integrated Moving Average (ARIMA), Multivariate Imputation by Chained Equations (MICE), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory neural networks (LSTM). Both single-well and multi-well strategies are assessed within a feasibility framework integrating standard error metrics, wavelet-based multi-resolution analysis, and visual inspection to classify model performance from Excellent to Unacceptable and ensure physically realistic reconstructed trajectories. In the single-well case, 58% of wells meet feasibility criteria. XGBoost provides the most reliable performance, capturing full frequency dynamics. LSTM performs competitively but cannot reconstruct early-series values due to lag-window requirements. RF tends to oversmooth fluctuations, MICE preserves broad trends but misses higher-frequency dynamics, and ARIMA performs poorly across most wells. Multi-well modeling improves accuracy and enables reconstruction of early-period gaps, increasing the proportion of feasible wells to 67%. Feature selection based on Self-Organizing Map (SOM) clustering generally outperforms Pearson and Spearman correlations, although no single technique is optimal for all wells. While relying solely on groundwater monitoring networks offers practical advantages and no external data requirements more than 30% of wells remain infeasible. Further improvements requires integrating additional physical drivers, such as precipitation, evapotranspiration, or remote-sensing indicators, and exploring hybrid modeling strategies.