Data challenges in large-scale transient groundwater modeling using machine learning

Large-scale transient groundwater monitoring using machine learning (ML) remains challenging, primarily due to limited observational data. Here we take the contiguous United States (CONUS) as a case study. Despite records from over one million water table depth (WTD) monitoring wells, groundwater observations remain spatially and temporally sparse. We systematically quantify multiple dimensions of these groundwater observation gaps. To evaluate the impact of data limitations on model performance, we train a random forest model for WTD seasonality estimation over the CONUS based on a balanced dataset that contains training and test sets of equal sizes across different land cover types. Overall, the model performs relatively well, with a median Spearman’s ρ of 0.53, but still fails to reproduce seasonal WTD variations in many locations, especially those impacted by human activities. Comparison with a random forest trained on physically-based simulation results at the same grid cells suggests the complexity in modeling human-impacted groundwater systems. Previous hydrometeorological conditions are identified as the most important input for all land cover types. A perturbation experiment further shows that, under the current groundwater monitoring network across the CONUS, the model predictive performance is more sensitive to additional observations in grasslands and farmlands than those in other land cover types, offering empirical guidance for future well installation. Our study provides insight into how data challenges affect ML model predictive performance and highlights the importance of optimizing and developing global groundwater monitoring networks.