Machine learning-based downscaling of modelled climate change impacts on groundwater table depth

Abstract. There is an urgent demand for assessments of climate change impact on the hydrological cycle at high spatial
resolution. In particular, the impacts on shallow groundwater levels, which can lead to both flooding and drought, have
major implications for agriculture, adaptation and urban planning. Predicting such hydrological impacts is typically
performed using physically-based hydrological models. However, such models are computationally expensive, especially at
high spatial resolutions.

This study is based on the Danish national groundwater model, setup as a distributed, integrated surface-subsurface model at
500 m horizontal resolution. Recently, a version at a higher resolution of 100 m was created; amongst others, to better
represent the uppermost groundwater table and to meet end user demands for water management and climate adaptation. The
increase in resolution of the hydrological model, however, also increases computational bottleneck. To evaluate climate
change impact, a large ensemble of climate models was run with the 500 m hydrological model, while performing the same
ensemble run with the 100 m resolution nation-wide model was deemed infeasible. The desired outputs at 100 m resolution
were produced by developing a novel, hybrid downscaling method based on machine learning.

Hydrological models for five subcatchments, covering around 9 % of Denmark and selected to represent a range of
hydrogeological settings, were run at 100 m resolution with forcings from a reduced ensemble of climate models. Random
Forest algorithms were established using the simulated climate change impacts (future – present) on water table depth at 100
m resolution from those submodels as training data.

The trained downscaling algorithms then were applied to create nation-wide maps of climate change-induced impacts on the
shallow groundwater table at 100 m resolution. These downscaled maps were successfully validated against results from a
validation submodel at 100 m resolution excluded from training the algorithms, and compared to the impact signals from the
500 m hydrological model across Denmark.

The suggested downscaling algorithm also opens for the spatial downscaling of other model outputs. It has the potential for
further applications where, for example, computational limitations inhibit running distributed hydrological models at fine
resolutions.