View in EDS

A hybrid approach for groundwater level prediction: integrating water balance model state variables and machine learning algorithms.

Saved in:
Bibliographic Details
Title: A hybrid approach for groundwater level prediction: integrating water balance model state variables and machine learning algorithms.
Authors: EL Bilali, Ali1 (AUTHOR) ali1gpee@gmail.com, El Khalki, El Mahdi1 (AUTHOR), Ait Naceur, Khaoula1 (AUTHOR), Jaffar, Oumar1 (AUTHOR), El Ouafi, Said1 (AUTHOR), Hadri, Abdessamad1 (AUTHOR)
Source: Environmental Earth Sciences. Jan2026, Vol. 85 Issue 1, p1-17. 17p.
Subject Terms: *Soil-Water Balance Model, *Machine learning, *Computer simulation, *Support vector machines, *Boosting algorithms, *Hydrologic cycle, *Water table, *Sensitivity analysis
Abstract: Effective water resources planning and management require a robust understanding of groundwater level (GWL) dynamics, which in turn depends on the reliability of simulation models. Modeling GWL in anisotropic settings remains a significant challenge, particularly in poorly monitored basins. This study investigates a hybrid modeling framework that integrates state variables from the Water Partition and Balance (WAPABA) model into Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), and Shapley Additive Explanations (SHAP) algorithms to improve GWL prediction in the Bouregerg Catchment in Morocco. The WAPABA model demonstrated strong performance in simulating runoff, achieving Nash–Sutcliffe Efficiency (NSE) values of 0.88 and 0.77 during calibration and validation, respectively. Sobol sensitivity analysis identified key parameters, including the catchment consumption curve and groundwater yield proportion. Incorporating WAPABA-derived state variables into SVR and XGBoost models substantially enhanced GWL prediction, yielding NSE values between 0.53 and 0.96 and Kling–Gupta Efficiency (KGE) values between 0.779 and 0.962. SVR exhibited a slight performance advantage over XGBoost, and the hybrid models consistently outperformed standalone machine learning approaches. SHAP-based interpretability analysis highlighted the dominant influence of hydrological state variables such as potential evapotranspiration, soil water storage, and groundwater fraction on GWL dynamics, with their relative importance varying according to geological conditions. Overall, the proposed hybrid framework offers a powerful and process-consistent approach for modeling GWL in anisotropic environments, supporting improved decision-making in water resources management. [ABSTRACT FROM AUTHOR]
Database: Energy & Power Source
Description
Abstract:Effective water resources planning and management require a robust understanding of groundwater level (GWL) dynamics, which in turn depends on the reliability of simulation models. Modeling GWL in anisotropic settings remains a significant challenge, particularly in poorly monitored basins. This study investigates a hybrid modeling framework that integrates state variables from the Water Partition and Balance (WAPABA) model into Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), and Shapley Additive Explanations (SHAP) algorithms to improve GWL prediction in the Bouregerg Catchment in Morocco. The WAPABA model demonstrated strong performance in simulating runoff, achieving Nash–Sutcliffe Efficiency (NSE) values of 0.88 and 0.77 during calibration and validation, respectively. Sobol sensitivity analysis identified key parameters, including the catchment consumption curve and groundwater yield proportion. Incorporating WAPABA-derived state variables into SVR and XGBoost models substantially enhanced GWL prediction, yielding NSE values between 0.53 and 0.96 and Kling–Gupta Efficiency (KGE) values between 0.779 and 0.962. SVR exhibited a slight performance advantage over XGBoost, and the hybrid models consistently outperformed standalone machine learning approaches. SHAP-based interpretability analysis highlighted the dominant influence of hydrological state variables such as potential evapotranspiration, soil water storage, and groundwater fraction on GWL dynamics, with their relative importance varying according to geological conditions. Overall, the proposed hybrid framework offers a powerful and process-consistent approach for modeling GWL in anisotropic environments, supporting improved decision-making in water resources management. [ABSTRACT FROM AUTHOR]
ISSN:18666280
DOI:10.1007/s12665-025-12738-8