doi:
This project aims to forecast groundwater levels under climate change scenarios following the previous work done by KIT and BGR-Berlin, Germany.
The current study applies a convolutional neural network (CNN) to understand future trends in groundwater levels (GWL) caused by climate change in 505 wells in Lower Saxony, northern Germany. Input features correspond exclusively to meteorological variables (precipitation (P), temperature (T), and relative humidity (RH)), projected over six climate models for the RCP8.5 scenario. The model is executed by including and excluding RH as an input variable. Model performance correlations with geospatial and time series features are computed to explore external influences. Results show a significant improvement in the model accuracy when RH is removed. Performance affectations are noticed with the distance to waterworks, land cover type, and time series anomalies. GWL projections present variations depending on the climate model but, in a significant proportion, estimate low annual declining changes up to 10% for 2020-2100 and local increasing changes mainly below 2.5%