Using Convolutional Neural Networks To Evaluate Long‐Term Groundwater Trends In Germany

Even water‐rich regions like Germany are more and more subject to climate stress. Severe droughts, followed by reduced
groundwater recharge and decreasing groundwater levels resulted in recent years in local shortages in drinking water
supply. We use a machine learning groundwater forecasting framework based on convolutional neural networks to
investigate the direct impact of climate change on Germany’s groundwater resources until 2100. Groundwater levels at 118
monitoring sites ‐ showing strong reactions to meteorological signals in the past and are spread across Germany ‐ were
predicted. The input data is taken from the RCP8.5 scenario with six climate models pre‐selected and pre‐processed by the
German meteorological service, covering a large part of the range of expected changes over the next 80 years. Our models
are using solely precipitation and temperature as input, are carefully evaluated using past observations and all achieve high
predictive skill scores. Only effects of natural climate change are considered, as reliable future input data on highly uncertain
human factors such as increased groundwater extraction or irrigation effects are basically unavailable. ... mehrFor most of the wells
considered in this study, we can show a significant (p<0.05) drop in the annual mean groundwater level, including longer
periods of low groundwater levels during the annual cyclealong with an increased number of extremely low groundwater
events. Spatially, results indicate stronger decreasing groundwater levels especially in northern and eastern Germany,
aggravating the already existing decreasing trend in these regions.