An artificial neural network as a quick tool to assess the effects of climate change and agricultural policies on groundwater resources

Groundwater is a strategic reserve that is often used to meet water demands in dry seasons and
during drought periods. However, the over-exploitation of this vital resources can jeopardize its
sustainability. Projected climate change is expected to further exacerbate the situation in many
regions of the world. Therefore, it is essential for decision makers to have simple tools to model
groundwater flow and to assist in aquifer management. These tools can reduce the computational
cost of complex physics-based models, without undermining the reliability of the results. The aim
of this work is to develop a surrogate model capable of simulating groundwater flow in the Konya
closed basin, a major agricultural region located in central Turkey. The model is used to analyze
different future water demand scenarios and evaluate the possible effects of climate change and
agricultural policies on groundwater. This aquifer is one of the pilot sites investigated within the
“Innovative and Sustainable Groundwater Management In the Mediterranean (InTheMed)” project,
which is part of the PRIMA programme. An Artificial Neural Network (ANN) was trained to provide
groundwater levels at 30 monitoring points for the period 2020-2039 accounting for different
climate and agricultural scenarios. The surrogate model replaces a full numerical surfacesubsurface
flow model implemented in MODFLOW and calibrated using field data recorded in the
period 2000-2019. To define the dataset that feeds the ANN, two multiplicative coefficients were
considered: one applied to the historical precipitation and the other to crop water demand. The
two coefficients and the current month were considered as input features of the ANN, while the
piezometric heads at the 30 monitoring points were the outputs. A dataset of 100 combinations of
precipitation and crop coefficients was generated using the Latin Hypercube Sampling method,
assuming an increase/decrease range in terms of precipitation equal to +/- 40% and water
demand equal to +/- 25%. For each combination of the coefficients, the full numerical model was
run starting from January 2020 to obtain piezometric heads at the 30 monitoring points with a
monthly time discretization. The final dataset was used to train (70%), validate (15%) and test (15%)
the network, highlighting a very good performance of the ANN for all three phases. The fully
trained network was used to predict groundwater levels considering three different precipitation
scenarios for the period 2020-2039: - 20% of the observed precipitation, no reduction of the
observed precipitation and + 20% of the observed precipitation. For each precipitation scenario,
the water demand was considered in the range -/+ 20%.
This work was developed under the scope of the InTheMED project. InTheMED is part of the PRIMA
programme supported by the European Union’s HORIZON 2020 research and innovation
programme under grant agreement No 1923.