High Resolution Water Table Modeling of the Shallow Groundwater Using a Knowledge-Guided Gradient Boosting Decision Tree Model

Summary

This study demonstrates a knowledge-guided machine learning framework for high-resolution shallow groundwater modelling across Denmark, integrating physically-based hydrological simulation with gradient boosting decision trees. By augmenting limited observational data with simulated water table seasonality and proxy observations from hydrological features, the authors achieved improved model generalisation and spatial accuracy beyond conventional numerical models. The approach addresses a significant gap in applying modern machine learning to groundwater systems, where data scarcity has historically limited adoption of such techniques.

UK applicability

The knowledge-guided framework and CatBoost methodology are transferable to UK conditions and could support water resource management, flood risk assessment, and climate adaptation planning at high spatial resolution. However, model performance would depend on data availability; the UK's denser network of groundwater monitoring wells may provide advantages, though geological complexity (particularly in regions with fractured bedrock) may present additional challenges.

Key measures

Water table depth (cm); mean absolute error; confidence intervals via quantile regression; spatial resolution (10 m); coverage area (43,000 km²)

Outcomes reported

The study developed a machine learning model (CatBoost) to map shallow groundwater water table depth at 10 m spatial resolution across Denmark under summer and winter conditions. Model performance achieved mean absolute error of ~115 cm using well observations alone, and <50 cm when proxy observations were included.

Topic tags