Cities are increasingly exploiting the use of the subsurface by placing buildings and infrastructure underground and by extracting resources. To predict and manage the effect of these physical developments, the transport of pollution in the groundwater, and the change of climate in urban areas, it is increasingly important to understand and be able to model the urban hydrogeology; the groundwaters interaction with piped networks, the local water level variations and flow paths.

This thesis is a synopsis of research conducted during the Ph.D. project Shallow Urban Groundwater Dynamics. The main objectives were to: (1) analyze the effect of urban geology in models of shallow groundwater levels and flow, (2) analyze the effect of spatial discretization when urban geology is represented in hydrological models, and (3) explore the application of machine learning for shallow groundwater level prediction in urban settings.

Two papers form the core of this work: Paper I “Impact of urban geology on model simulations of shallow groundwater levels and flow paths” (LaBianca et al., 2023), and Paper II “Machine learning for predicting shallow groundwater levels in urban areas”(LaBianca et al., n.d.). The thesis summarizes the results from the two research papers and discusses the outcome and future perspectives for research within urban hydrology.

The research included a case study, of a domain in the City of Odense, Denmark. It involved setting up a monitoring network of the shallow groundwater levels and stream discharge, as well as modeling experiments. Paper I focused on the first two main objectives. I tested the effect of different geological models in a process-based hydrological model. The hydrological model included integrated surface water and groundwater process, as well as overland and saturated zone drainage, and leaking sewer pipes. The geological models had different interpretations of the shallow urban geology within the domain. Furthermore, this paper tested the impact of the horizontal discretization. The study analyzed the effect of these two elements on the simulated groundwater levels and flow.

Paper II addresses the third objective and explored the application of machine learning for shallow groundwater level prediction for the same domain as in Paper I, in Odense City. In paper II the third object was addressed by training two machine learning models based on the same algorithm and target dataset, but with a variation of the feature variables within the train data. The models’ performances and predictions were compared with one of the process-based hydrological models developed in Paper I. The machine learning models were trained to predict an average yearly minimum water table depth, that is expected to occur every year within the domain.

Moreover, Paper II includes the development of a data-driven method for homogenizing a heterogeneous dataset of shallow groundwater level measurements. The homogenization was done to include data from wells with few and discontinuous water level measurements in the training of the machine learning model. Lastly, the SHapley Additive exPlanation (SHAP) method was applied to analyze the relations that the machine learning models established with the feature variables in the training dataset.

The thesis concludes that even though urban geology only takes up a small percentage of the geological model, a representation of urban geology in models of shallow urban groundwater affects the flow simulations. An effect of the urban geology can be detected even if the horizontal discretization is larger than the size of the urban geology elements. The results showed that the urban geology only affected the simulated shallow groundwater level locally. Yet, the results showed that a fine discretization close to the size of the urban geology elements’ resolution improves the hydrological model performance and suggests that the urban hydrogeological models need to be run in a fine horizontal discretization to simulate the local variability in shallow urban groundwater levels.

Moreover, the thesis concludes that machine learning for spatial predictions of the shallow water table depth shows great potential for application in urban areas. The machine learning models resulted in a spatially better fit to the water table depth observations than the processbased model. The detailed analysis of machine learning models' relationship with feature variables can potentially be used for improving the knowledge of where and how different features affect urban shallow groundwater. Lastly, the results from Paper II illustrate that datadriven methods can be utilized during data curation and that it has the potential for improving and supplementing the scarce urban monitoring data as model inputs.

The presented studies contribute to the understanding of urban hydrogeology by testing methods of modeling the urban shallow groundwater and the effect of urban geology and model discretization. The findings are relevant for future studies of shallow urban groundwater dynamics, data curation of urban geology and water level data, transport modeling, and studies of the effects of climate change and climate change adaptation measures in urban areas.