Machine learning for predicting shallow groundwater levels in urban areas

In this study, the potential of machine learning (ML) for shallow groundwater level predictions in urban areas is explored. It focuses on curating a training dataset that represents the spatial variability of the water table depth, tests the effect of using different feature variables in ML modeling, and finally, compares two ML models with a physically-based (PB) urban hydrological model. To curate a consistent training dataset, a method of transferring low-frequency groundwater level measurements to a minimum water table depth (MWTD) was developed. Two ML models, one with national maps as feature variables and the other including local high-resolution urban feature variables, were trained against the same 280 groundwater level data points and applied to predict the MWTD at a 10 m spatial resolution for the city of Odense, Denmark. The ML models reached a similar fit to the observations, with an RMSE of 1.1 m and 1.3 m, respectively, and outperformed the urban PB model. In densely urbanized areas, the ML models and the PB model showed up to a 1.5 m difference in predictions of MWTD. The results suggest that ML modeling has the potential to provide spatially high-resolution predictions of the shallow groundwater table in urban areas, which represents a challenge for PB models because of their model structure and the lack of hydrological knowledge hindering meaningful parameterization schemes. Furthermore, a SHapley Additive exPlanation (SHAP) analysis of the feature variables illustrates that ML models can be utilized to explore the hydrological relations in urban domains, by analyzing the feature variables' relations.