Monitoring and Predicting Saltwater Intrusion via Temporal Aquifer Vulnerability Maps and Surrogate Models

This article proposes a methodology to accurately monitor seawater intrusion (SWI) using time-varied GALDIT vulnerability maps. The properly produced samples are then used as input–output patterns for the approximate SWI simulation. As a novelty, the specific area of high susceptibility to SWI is proposed as the dynamic saltwater wedge position to suitably select the monitoring locations (MLs) from a narrowed area. It is observed that varied initial conditions over time periods have more influence than variable pumping rates on salinity at MLs far from the production wells. Support Vector Regression (SVR), Artificial Neural Network (ANN) and Gaussian Process Regression (GPR) models have been substituted for the numerical model of SWI. Input training patterns of the surrogate models are initial salinity concentrations at selected MLs plus transient pumping values via Latin hypercube sampling. The final salinity at MLs constitutes the output patterns. The paper applies this new methodology to a small study area subject to the SWI problem. The generalization ability of surrogate models for predicting new initial conditions-pumping datasets was evaluated using performance criteria considering the ML locations. All surrogates offered good results for predicting SWI at specified MLs. The SVR model had poor performance compared to ANN and GPR models in MLs near the pumping wells, due to their salinity fluctuations over time.