ABSTRACT Robustly modeling salinity in tidal estuaries under changing conditions remains a scientific challenge. This study focuses on the Scheldt Estuary, which faces the dual threats of saltwater intrusion and climate change, and compares the performance of three machine learning models in simulating daily salinity, including artificial neural network (ANN), random forest (RF), and support vector regression (SVR). The ANN is selected and combined with a detailed full hydrodynamic model implemented in MIKE11 via three ensemble methods, including generalized linear model (GLM), cubist regression (CUBIST), and boosted smoothing spline model (BSTSM). Results underline the effectiveness of machine learning models, showcasing the superiority of ANN, particularly in the middle and lower regions. Ensemble methods prove vital in addressing local simulation biases seen in the base models, particularly in predicting high values, with the ensemble of ANN and MIKE11 via GLM performing better downstream. This approach significantly improved predictions, reducing the root mean squared error (RMSE) by up to 20% compared to the standalone models at the downstream. In addition, the impact analysis of the external boundary conditions emphasizes the challenges with the extrapolative robustness of machine learning. The ensemble model demonstrates the potential to reduce simulation uncertainties by amalgamating predictions from base models.
Zhu et al. (Thu,) studied this question.