Interpretable and Transferable Machine Learning for Managed Aquifer Recharge Planning in Sandy Catchments

“Managed Aquifer Recharge (MAR) offers a promising strategy for enhancing groundwater resilience in drought-prone areas such as the eastern sandy region of the Netherlands, where freely draining soils and intensive drainage management increase vulnerability to desiccation. Machine learning (ML) models are increasingly used for MAR planning due to their ability to simulate multiple recharge scenarios efficiently. However, their limited interpretability and transferability across regions restricts their real-world application. To address this, we use two complementary ML models: U-Net and XGBoost, where U-Net is more detailed of the two while XGBoost provides interpretability. The groundwater response to MAR was decomposed into two physically meaningful components: the immediate MAR-induced storage rise and a decay coefficient representing drainage losses. Using SHAP values, we quantify the influence of surface water and subsurface properties on each component. Domain experts reported improved confidence in predictions due to this interpretability. Our results show that different inputs control the short-term storage; which are recharge rate, and specific yield dominate the MAR response, versus its long-term retention; aquifer properties, and river conductance. This distinction has direct implications for multi-criteria decision analysis (MCDA), offering a data-driven basis to prioritize criteria for both recharge efficiency and storage longevity. We also assess model generalizability across sandy catchments beyond the training region (Baakse Beek). While the performance in nearby catchments is lower than in the training region, including additional training sites improves accuracy. Notably, targeted sampling of underrepresented hydrogeological conditions improves performance more efficiently than random sampling. This study presents a practical, interpretable, and data-efficient ML approach for MAR assessment. It demonstrates how physically relevant model outputs and strategic training data design can support scalable, transparent decision-making in groundwater management.”

(Citation: Fernandes, V., P. De Louw, C. Ritsema and R. Bartholomeus (2025). “Interpretable and Transferable Machine Learning for Managed Aquifer Recharge Planning in Sandy Catchments.” AGU25 https://agu.confex.com/agu/agu25/meetingapp.cgi/Paper/1982631)