In a context where anticipating future trends and long-term variations in water resources is crucial, improving our knowledge about most types of aquifer responses to climate variability and change is necessary. Aquifers with variability dominated by seasonal (marked annual cycle) or low-frequency variations (interannual to decadal variations driven by large-scale climate dynamics) may encounter different sensitivities to climate change. We investigated this hypothesis by generating groundwater level projections using deep learning models for annual, inertial (low-frequency dominated) or mixed annual/low-frequency aquifer types in northern France from 16 CMIP6 climate model inputs in an ensemble approach. Generated projections were then analysed for trends and changes in variability. Generally, groundwater levels tended to decrease for all types and scenarios across the 2030-2100. The variability of projections showed slightly increasing variability for annual types for all scenarios but decreasing variability for mixed and inertial types. As the severity of the scenario increased, more mixed and inertial-type stations appeared to be affected by decreasing variability. Focusing on low-frequency confirmed this observation: while a significant amount of stations showed increasing variability for the less severe SSP 2-4.5 scenario, low-frequency variability eventually showed slight yet statistically significant decreasing trends as the severity of the scenario increased. For the most severe scenario, almost all stations were affected by decreasing low-frequency variability. Finally, groundwater levels seemed, in most instances, slightly higher in the future than in the historical period, without any significant differences between emission scenarios.

The development of groundwater levels (GWL) simulations, based on deep learning (DL) models, is gaining traction due to their success in a wide range of hydrological applications. GWL Simulations allow generating reconstructions to be used for exploring past temporal variability of groundwater resources or provide means to generate projections under climate change on decadal scales. Owing to the diversity of large-scale and local scale forcing factors involved in explaining GWL variability, machine learning or even deep learning approaches reveal relevant tools to simulate GWL. In addition, such methods do not require too much-extended knowledge of physical variables in the links between climate variables and GWL.In this paper, we investigated the capacities of three deep learning models (Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and Bidirectional LSTM (BLSTM)) to reproduce GWL variations over time. Among the three deep learning models, GRU performed relatively better in most cases. Another aspect was to evaluate the input data’s impact and usefulness of wavelet pre-processing considering its limitations and best practices. Two different input datasets are compared to each other, one considering Effective Precipitation only, the other considering Precipitation and Temperature.Maximum Overlap Discrete Wavelet Transform (MODWT) preprocessing was used to decompose the input variables to explore the impact of wavelet transform in improving the simulations on several types of GWL time series by unravelling “hidden” though useful information in input data. Results show that the preprocessing (MODWT) helps the models generate better simulations. This improvement is higher with raw climate data (precipitation & temperature) as compared to when effective precipitation was used as input. Finally, the Shapley Additive exPlanations (SHAP) approach was used to interpret the impact of input variables on the model simulations. Analysis of SHAP values indicated that the sources of the information content preferentially learned by the models to achieve best simulations. For instance, it was clear that simulation of inertial and mixed GWL required the models to learn from low-frequency variability presented in the input data.