Field-based explainable machine learning for interval-scale groundwater dynamics in agricultural managed aquifer recharge (Ag-MAR)

Authors: ELTARABILY, MOHAMED - University Of California, Davis; ELSHAARAWY, MOHAMED - Horus University-Egypt; DAHLKE, HELEN - University Of California, Davis; Begna, Sultan; Wang, Dong; BALI, KHALED - Kearney Agricultural Center

Submitted to: Stochastic Environmental Research and Risk Assessment (SERRA)
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/14/2026
Publication Date: 6/3/2026
Citation: Eltarabily, M., Elshaarawy, M., Dahlke, H., Begna, S.H., Wang, D., Bali, K. 2026. Field-based explainable machine learning for interval-scale groundwater dynamics in agricultural managed aquifer recharge (Ag-MAR). Stochastic Environmental Research and Risk Assessment (SERRA). 40,145. https://doi.org/10.1007/s00477-026-03282-3.
DOI: https://doi.org/10.1007/s00477-026-03282-3

Interpretive Summary: Sustainability of agriculture is threatened by depletion of Aquifer. Agricultural Managed Aquifer Recharge (Ag-MAR) could be a potential, cost-effective strategy to minimize groundwater depletion by infiltrating excess surface water from storm flows using agricultural fields. However, quantification of ground water recharge amount is lacking for permitting and water accounting purposes with local and state agencies in California. Four Machine Learning (ML) models (Decision Tree, Random Forest, Extreme Gradient Boosting, and CatBoost Gradient Boosting) were examined for predicting ground water recharge amount under Ag-MAR. Variables consist of applied winter water, precipitation, evapotranspiration, soil water storage change, and soil oxygen concentration obtained from an alfalfa-field experiment were used for training, testing and evaluation of model’s fitness. Findings revealed that Extreme Gradient Boosting model consistently outperformed the other models with R2 value of 0.93 and mean square error value of 12.0 mm with all variables included in the model. Optimal model was deployed through a desktop graphical user interface, providing a user-friendly platform for making recharge amount predictions. This study highlights the potential of ML techniques in Ag-MAR applications, offering valuable insights to inform water resource management for enhancing water and agriculture sustainability, particularly in regions facing water scarcity.

Technical Abstract: Agricultural Managed Aquifer Recharge (Ag-MAR) has emerged as one of the most cost-effective strategies to minimize aquifer depletion by infiltrating excess surface water from storm flows using agricultural fields. Yet, accurate quantification of the groundwater recharge amount (Rt) is critical for permitting and water accounting purposes with regulatory agencies. This study employed four machine learning (ML) models: Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGB), and CatBoost Gradient Boosting (CGB) to predict groundwater recharge amount, Rt under Ag-MAR. The measured variables such as applied winter water (It), precipitation (Pt), evapotranspiration (ETa), soil water storage change ('St), and soil oxygen concentration (%O2), were obtained from an alfalfa-field experiment conducted at University of California- Kearney Agricultural Research and Extension Center in 2020-2022. Four distinct variable combinations were evaluated: C1 (It, Pt, ETa), C2 (It, Pt, ETa, 'St), C3 (It, Pt, ETa, %O2), and C4 (It, Pt, ETa, 'St, %O2), to assess their fit with impact on model performance. A total of 76 records were used for model training and testing, representing recharge intervals from three winter flooding seasons. Results indicated that the XGB model consistently outperformed the other models, followed by CGB, RF, and DT exhibiting higher R2 of 0.93 with root mean square error (RMSE) of 12.0 mm during the testing stage for the combination of all variables; C4 (It, Pt, ETa, 'St, %O2). The SHAP analysis showed that when measurements of soil water storage change ('St) and soil oxygen concentration (%O2) are not included or unavailable, evapotranspiration (ETa) emerges as the second most influential variable, following winter-applied water (It), affecting groundwater recharge estimates. Whenever soil water storage change ('St) or soil oxygen concentration (%O2) is measured and incorporated in the model, each emerges as the second most influential factor, following applied winter water (It), in predicting groundwater recharge estimates. The optimal model was deployed through a desktop graphical user interface (GUI), providing a user-friendly platform for making Rt predictions. This study highlights the potential of ML techniques in Ag-MAR applications, offering valuable insights to inform water resource management for enhancing water sustainability, particularly in regions facing water scarcity.