Towards identifying scale-dependent impacts on groundwater level dynamics with Deep Learning

Detailed process-understanding of climatic and non-climatic drivers is generally required to estimate future groundwater availability under climate change. Groundwater level (GWL) dynamics are very sensitive to groundwater pumping, but information on their local effects and magnitude – especially in combination with natural fluctuations – is often missing or inaccurate. It has been shown by previous studies that complex hydrogeological processes can be learned from neural networks, whereby Deep Learning (DL) demonstrates its strengths particularly in combination with large data sets. However, there are limitations in the interpretability of the predictions and the transferability with such methods. Furthermore, most groundwater data are not yet ready for data-driven applications. This study aims at improving GWL predictions with DL by combining big data elements from a newly constructed global groundwater database with long-term short-term memory (LSTM) networks. Our underlying hypothesis is that scale-dependent processes can be learned for groundwater dynamics, similar to streamflow data. For our experiments we use continuous groundwater level observations from basins worldwide and basin attributes – spatially heterogeneous but temporally static catchment attributes (e.g. topography) and continuous observations of the meteorological forcing (precipitation and temperature). The initial results are consistent with previous studies in that GWL prediction performance is good with LSTM models trained with climate input on single wells. It is now being tested whether the LSTM model trained on many wells simultaneously is able to represent the climatic effects - but not the anthropogenic effects, e.g. with wells that are considered to be anthropogenically unaffected.