Local refinement of a national-scale groundwater model

Julian Koch , Jun Liu , Lars Troldborg

This method article presents a local refinement framework for a national-scale, machine learning-based groundwater model that predicts typical summer and winter water table depth at 10 × 10 m resolution at national scale of Denmark. While the existing baseline model provides high-resolution national coverage and is suitable for screening purposes, its accuracy remains insufficient for local groundwater management applications. The proposed method integrates new groundwater observations into the existing training dataset and retrains the baseline model to enhance local fidelity while preserving large-scale consistency. The approach is demonstrated for a 212 km² case study area using 100 synthetic groundwater observations. Cross-validation results show consistent improvements in mean error, mean absolute error, and root mean squared error compared to the baseline model, particularly when increased weights are assigned to new observations. The refined models produce improved precision in areas with new observations while maintaining baseline behaviour elsewhere, demonstrating its suitability for combining national datasets with local monitoring networks for improved decision support in climate adaptation, nature conservation and other applications.