Performance and Limitations of Machine Learning Models for Groundwater Level Prediction Using Hydro-Climatic Variables

Refined simulation of groundwater level dynamics based on a machine learning stacking framework

Accurate and rapid prediction of groundwater levels (GWL) is essential for effective groundwater management. Machine learning models are efficient tools for GWL prediction, but individual models often suffer from limited generalization due to inherent randomness. This study proposed a stacking-based GWL prediction framework suitable for arid regions in Northwest China. Feature variables affecting GWL were selected using variable importance in projection (VIP). Then, three machine learning models—artificial neural networks (ANN), random forests (RF), and Light Gradient Boosting Machine (LightGBM)—were developed, and their outputs were integrated using a support vector regression (SVR)-based stacking method to enhance the accuracy of GWL prediction. The results show that the factors of influencing GWL changes vary significantly across different regions, and selecting the most contributive feature variables is beneficial for model construction. Among the individual models, the RF model demonstrated higher accuracy and more stable performance, outperforming the ANN and LightGBM models. However, individual models exhibited poor generalization during validation. In contrast, the stacking model maintained high performance, demonstrating superior generalization. Compared to the best-performing individual model (RF) in validation period, the Nash–Sutcliffe efficiency ( NSE ) and Kling–Gupta efficiency ( KGE ) of stacking model improved by 0.11–0.66 and 0.05–0.41, the correlation coefficient ( R 2 ) increased by 0.05–0.3, and root mean square error ( RMSE ) reduced by 0.01–0.1 m. In the stacking simulation, RF had the highest average contribution (80.2%), followed by ANN (13.9%) and LightGBM (5.9%). This study provides a stacking simulation framework based on machine learning methods for precise groundwater level simulation, which can serve as a reference for groundwater level simulation in other regions.

Interpretable machine learning for groundwater-level prediction incorporating pumping-well power consumption

Groundwater-level forecasting is essential for adaptive groundwater management in regions where pumping pressure and climatic variability interact, yet direct abstraction records are often unavailable. This study developed an interpretable extreme gradient boosting (XGB) framework, selected for its suitability for nonlinear tabular prediction under limited-data conditions, to predict monthly groundwater levels at nine monitoring wells in the mid-fan region of the Choushui River alluvial fan, Taiwan, using observations from January 2013 to June 2023. Rather than converting electricity use into pumping volume through a fixed and potentially uncertain coefficient, monthly pumping-well power consumption was used directly as an operational predictor of withdrawal-related stress. Three input configurations were evaluated to distinguish the predictive roles of hydro-meteorological forcing, local pumping-related stress, and cross-well groundwater-memory information. Out-of-sample testing revealed substantial between-well variability under the baseline setting (RMSE = 0.341–3.254 m; MAE = 0.247–2.783 m), with markedly larger errors in several pumping-sensitive wells. Adding pumping-well power consumption and cross-well lagged groundwater-level information produced directional reductions in average testing MAE, but these improvements were substantial in magnitude rather than statistically conclusive across wells under the present sample size. Specifically, average testing MAE decreased by 24.04% from Case A to Case B and by a further 9.27% from Case B to Case C, although one-sided Wilcoxon signed-rank tests on paired well-wise testing MAE values indicated that the differences between Cases A and B and between Cases B and C were not statistically significant at the across-well level. In contrast, relative to a one-step persistence benchmark, the corresponding skill scores were 6.23% for Case A, 28.79% for Case B, and 35.39% for Case C, and the final model significantly outperformed persistence. SHAP analysis showed that lag-1 groundwater level was the dominant predictor, while pumping-related power-consumption variables provided substantial additional explanatory information in pumping-sensitive wells. These results indicate that operational power-consumption data can be integrated into an interpretable machine-learning framework to improve monthly groundwater-level prediction beyond simple persistence in data-scarce aquifers, while also providing a practical basis for month-ahead groundwater-risk screening and future research on human-induced groundwater stress.

Predictive modeling of groundwater quality for irrigation using integrated indices and machine learning: case study in Tunisia

Understanding Long-Term Groundwater Storage Variability Using GRACE Data and Explainable Machine Learning

The decline in groundwater storage (GWS) poses a critical threat to water security in semi-arid regions where increasing agricultural water demand and climate variability are increasing pressure on aquifers. This study presents a novel hybrid modeling framework integrating multi-source satellite and climate data (GRACE, GLDAS, TerraClimate, and MODIS) with machine learning and explanatory artificial intelligence techniques for the long-term assessment and interpretation of GWS anomalies in the data-poor Iğdır Basin. Three different modeling approaches were developed: XGBoost, Long Short-Term Memory (LSTM) networks, and their combined model, and interpreted using the Shapley Additive Explanations (SHAP) method. The results showed a significant long-term decreasing trend in groundwater storage anomalies at a rate of −0.87 mm per month during the 2002–2016 period, indicating continuous depletion. The LSTM model demonstrated the best performance with R2 of 0.59, RMSE of 19.5 mm, and MAE of 15.1 mm, revealing the dominant role of temporal dependencies in groundwater systems. SHAP analysis identified lagged groundwater anomalies (especially GWS_lag3) as the most effective predictors; this may reflect the memory effect and lagged response specific to semi-arid aquifer systems, but this interpretation needs to be validated in different study areas. Snow water equivalent and total water storage anomalies also emerged as significant determinants, while the direct effect of instantaneous precipitation was found to be limited. This study addresses significant gaps in the literature by combining sequence-based modeling with model interpretability in a semi-arid closed basin. The findings highlight the necessity of using system memory and explainable artificial intelligence together for reliable groundwater prediction. While the proposed hybrid approach has the potential for application in other semi-arid regions, its broader usability needs to be supported by independent validation studies under different hydrogeological and climatic conditions.

Modelling Shallow Groundwater Level Fluctuations in Very Flat Landscapes Based on Satellite Data and Machine Learning

Groundwater level fluctuations play a critical role in shaping hydrological extremes in flat sedimentary landscapes, where shallow water table depth (WTD) and strong surface-subsurface connectivity modulate the impacts of floods and droughts. The Western Pampean Plain (Argentina) exemplifies these dynamics; however, accurate modeling is often hindered by the lack of continuous in situ monitoring. In this context, manual WTD measurements collected by local farmers represent an underexploited source of information for modeling. In this study, we developed a Random Forest framework integrating farmer-operated observations with climatic and satellite-derived data and evaluated its ability to reconstruct and predict WTD. We tested seven modeling strategies, integrating: (i) climatic variables (including effects up to 15 months); (ii) high-resolution satellite-derived Surface Water Cover Index (SWCI) from Landsat; and (iii) coarse-resolution Terrestrial Water Storage Anomalies (TWSA) from GRACE. The best-performing model integrated climatic variables and SWCI, yielding strong reconstruction (R2 = 0.861, RMSE = 0.266 m) and robust prediction (R2 = 0.752, RMSE = 0.344 m) performances under cross-validation and rolling-origin validation, respectively. Model interpretation revealed SWCI as the dominant predictor, reflecting the strong surface-subsurface connectivity that characterizes this environment. This study provides a practical framework that integrates farmer-operated groundwater monitoring with freely available satellite observations to support agricultural decision-making in flood and drought risk management across flat sedimentary landscapes.

A Remote Sensing-Based Groundwater Level Monitoring System Using Machine Learning

Groundwater is an essential natural resource for human societies and ecosystems. Traditional groundwater monitoring relies on in situ wells, which are susceptible to discontinuity, influencing water resource management. To overcome this deficiency, this study proposes a remote sensing-based groundwater level (GWL) monitoring system that uses machine learning (ML) algorithms and remotely sensed hydrological parameters to reconstruct well-specific GWL time series. Four machine learning algorithms, including K-Nearest Neighbor (KNN), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and a weight-mean Ensemble strategy, were adopted to construct the models at each well individually for monitoring GWL. Specifically, the GWL data for ~770 wells across the conterminous United States (CONUS) were modeled using remotely sensed precipitation (P), evapotranspiration (ET), terrestrial water storage anomaly (TWSA), and soil moisture (SM) datasets during the period from 2004 to 2019. Afterwards, the performances of models were evaluated during an independent period from 2020 to 2023. The results show that the Ensemble model outperforms the individual baseline models evaluated in this study (i.e., KNN, RF, and XGBoost), achieving a mean coefficient of determination (R2) of 0.81, root mean square error (RMSE) of 0.34 m, normalized RMSE (NRMSE) of 11.8%, and Nash–Sutcliffe efficiency (NSE) of 0.78. The results demonstrate that the proposed system can effectively reconstruct GWL dynamics for most wells. This can be a compensation for missing records for hydrologically significant wells, which are those with historical groundwater observations.