Genetic Programming and Gaussian Process Regression Models for Groundwater Salinity Prediction: Machine Learning for Sustainable Water Resources Management

Abstract

Degradation of the quality of groundwater due to saltwater intrusion is considered as a major constraint limiting the use of water resources in coastal areas. Groundwater salinity prediction models can be used as surrogate models in a linked simulation-optimization methodology needed for developing and solving computationally feasible sustainable coastal aquifer management models. The present study utilizes two machine learning algorithms, namely, Genetic Programming (GP) and Gaussian Process Regression (GPR) models to approximate density dependent saltwater intrusion processes and predict salinity concentrations in an illustrative coastal aquifer system. Specifically, the GP and GPR models are trained and validated using pumping and resulting salinity concentration datasets obtained by solving a numerical 3D transient density dependent finite element based coastal aquifer flow and transport model. Prediction capabilities of the developed GP and GPR models are quantified using standard statistical parameters such as root mean squared error, coefficient of correlation and the Nash-Sutcliffe coefficient calculations. The results suggest that once trained and tested, both the GP and GPR models can be used to predict salinity concentration at selected monitoring locations in the modelled aquifer under variable groundwater pumping conditions. The performance evaluation results for the illustrative aquifer study area also show that the predictive capability of the GPR models are superior to those of the GP models. Therefore, GPR prediction models can be used as a substitute for the complex numerical simulation model in a linked simulation-optimization approach requiring numerous solutions of the simulation model to develop computationally feasible regional scale sustainable coastal aquifer management strategies.