Agricultural Managed Aquifer Recharge (Ag-MAR) is increasingly used to mitigate groundwater overdraft, yet interval-scale recharge estimates are rarely available in a form suitable for rapid field application. This study developed explainable machine-learning models to predict interval-scale groundwater recharge (Rₜ) during winter Ag-MAR using 76 recharge intervals collected from 2020 to 2022 in an alfalfa field at the University of California Kearney Agricultural Research and Extension Center (KARE). The predictor dataset was constructed from field observations and a calibrated HYDRUS-2D reference model, and included applied winter water (Iₜ), precipitation (Pₜ), actual evapotranspiration (ETₐ), soil-water storage change (ΔSₜ), and soil oxygen concentration (%O₂). Four input scenarios were evaluated to reflect realistic monitoring conditions: C1 (It, Pt, ETa), C2 (It, Pt, ETa, ΔSt), C3 (It, Pt, ETa, %O₂), and C4 (It, Pt, ETa, ΔSt, %O₂). Decision Tree, Random Forest, XGBoost, and CatBoost models were trained using Bayesian hyperparameter optimization with 5-fold cross-validation and then assessed using held-out testing and leave-one-year-out validation. The best held-out performance was achieved by XGBoost under C4, with R² = 0.93 and RMSE = 12.0 mm. When ΔSₜ was unavailable, C3 provided comparable performance (R² ≈ 0.92–0.93; RMSE ≈ 12.5–13.6 mm), indicating that %O₂ can serve as a practical proxy for recharge-relevant saturation dynamics. Leave-one-year-out validation showed more conservative but consistent results, with C4-XGBoost providing the strongest temporal transferability (overall R² = 0.53; RMSE = 29.6 mm). SHAP analysis identified Iₜ as the dominant predictor across all scenarios, while ΔSₜ and/or %O₂ provided the most influential secondary information. A desktop graphical user interface was also developed to support practical recharge estimation. Although the findings are currently site-specific, the results demonstrate a promising and operationally relevant pathway for low-sensor, explainable recharge prediction in Ag-MAR systems.

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

期刊论文
发表日期：

2026年6月3日

查看全文

Mohamed Galal Eltarabily,
Mohamed Kamel Elshaarawy,
Helen E. Dahlke,
Sultan Begna,
Dong Wang,
Khaled M. Bali

摘要

Agricultural Managed Aquifer Recharge (Ag-MAR) is increasingly used to mitigate groundwater overdraft, yet interval-scale recharge estimates are rarely available in a form suitable for rapid field application. This study developed explainable machine-learning models to predict interval-scale groundwater recharge (Rₜ) during winter Ag-MAR using 76 recharge intervals collected from 2020 to 2022 in an alfalfa field at the University of California Kearney Agricultural Research and Extension Center (KARE). The predictor dataset was constructed from field observations and a calibrated HYDRUS-2D reference model, and included applied winter water (Iₜ), precipitation (Pₜ), actual evapotranspiration (ETₐ), soil-water storage change (ΔSₜ), and soil oxygen concentration (%O₂). Four input scenarios were evaluated to reflect realistic monitoring conditions: C1 (I_t, P_t, ET_a), C2 (I_t, P_t, ET_a, ΔS_t), C3 (I_t, P_t, ET_a, %O₂), and C4 (I_t, P_t, ET_a, ΔS_t, %O₂). Decision Tree, Random Forest, XGBoost, and CatBoost models were trained using Bayesian hyperparameter optimization with 5-fold cross-validation and then assessed using held-out testing and leave-one-year-out validation. The best held-out performance was achieved by XGBoost under C4, with R² = 0.93 and RMSE = 12.0 mm. When ΔSₜ was unavailable, C3 provided comparable performance (R² ≈ 0.92–0.93; RMSE ≈ 12.5–13.6 mm), indicating that %O₂ can serve as a practical proxy for recharge-relevant saturation dynamics. Leave-one-year-out validation showed more conservative but consistent results, with C4-XGBoost providing the strongest temporal transferability (overall R² = 0.53; RMSE = 29.6 mm). SHAP analysis identified Iₜ as the dominant predictor across all scenarios, while ΔSₜ and/or %O₂ provided the most influential secondary information. A desktop graphical user interface was also developed to support practical recharge estimation. Although the findings are currently site-specific, the results demonstrate a promising and operationally relevant pathway for low-sensor, explainable recharge prediction in Ag-MAR systems.