Graph Fourier Deep Learning for Spatiotemporal and Hydrogeological Interpretation of Groundwater Levels in the Yellow River Basin

Groundwater systems exhibit strong spatial heterogeneity and nonlinear dynamics, complicating the parameterization of physically based hydrological models and limiting adaptability at large scales. Deep learning (DL) offers a promising data‐driven alternative but faces challenges in prediction accuracy and interpretability. This study proposes a novel Graph Fourier Network (GFN) model that integrates Fourier domain modeling, dynamic graph structure generation, and multi‐head graph attention. Hydrogeological units are embedded as prior information, leveraging data from 473 groundwater monitoring wells across the Yellow River Basin (YRB), China. We show that the proposed GFN model significantly outperforms baseline models, including Long Short‐Term Memory (LSTM), Extreme Gradient Boosting (XGBoost) and Spatio‐Temporal LSTM (ST‐LSTM), achieving an R 2 of up to 0.90. By incorporating hydrogeological prior knowledge, the GFN model enhances the extrapolation capability of groundwater level predictions, particularly for extended lead times up to 25 days. The predicted groundwater level dynamics show lagged and nonlinear responses to climatic and ecological drivers, with their spatial heterogeneity shaped by hydrogeological conditions. These response patterns include the prolonged influence of precipitation, and spatially varying effects of evapotranspiration and vegetation across hydrogeological units. These findings highlight the GFN model informed by hydrogeological information effectively captures characteristics of regional groundwater system. This study provides a fast and reliable approach for short‐ and long‐term regional groundwater prediction under varying hydrogeological conditions.