A physically interpretable transfer learning framework for improved generalization of data-driven groundwater level...

This study develops an interpretable transfer learning (TL) framework for groundwater level (GL) prediction that addresses two challenges: accurate prediction under data-scarce conditions and hydrogeological interpretation of data-driven model parameters. The framework integrates Gated Recurrent Unit (GRU) networks with one-dimensional (1D) convolution and TL techniques. A pretrained backbone model jointly trained across 62 monitoring wells spanning diverse hydrogeological settings in South Korea learned generalizable precipitation-GL response patterns. The framework was rigorously evaluated through leave-one-well-out cross-validation (LOWOCV) under one-, two-, and three-year target-well data-scarcity scenarios. The TL framework consistently outperformed conventional GRU models across all scenarios. Notably, the TL framework trained using only one year of target-well data exceeded the predictive performance of the baseline GRU model trained using three years of data, demonstrating that knowledge transfer effectively compensates for limited records. Quantitative extrapolation analysis further showed that the TL framework maintained stable predictive performance beyond the historical training range, indicating robust generalization under non-stationary hydrological conditions. In addition, convolutional kernel weights showed strong agreement with independently estimated residence times for short-term precipitation-groundwater responses, while TL similarity weights corresponded closely with hydrogeological properties including residence time and vadose-zone permeability characteristics. These results demonstrate that the framework enables reliable prediction and hydrogeologically meaningful interpretation using only short-term GL records, providing a practical tool for early-stage assessment of newly established monitoring wells.