Machine learning methods applied to sea level predictions in the upper part of a tidal estuary

Sea levels variations in the upper part of estuary are traditionally approached by relying on refined numerical simulations with high computational cost. As an alternative efficient and rapid solution, we assessed here the performances of two types of machine learning algorithms: (i) multiple regression methods based on linear and polynomial regression functions, and (ii) an artificial neural network, the multilayer perceptron. These algorithms were applied to three-year observations of sea levels maxima during high tides in the city of Landerneau, in the upper part of the Elorn estuary (western Brittany, France). Four input variables were considered in relation to tidal and coastal surge effects on sea level: the French tidal coefficient, the atmospheric pressure, the wind velocity and the river discharge. Whereas a part of these input variables derived from large-scale models with coarse spatial resolutions, the different algorithms showed good performances in this local environment, thus being able to capture sea level temporal variations at semi-diurnal and spring-neap time scales. Predictions improved furthermore the assessment of inundation events based so far on the exploitation of observations or numerical simulations in the downstream part of the estuary. Results obtained exhibited finally the weak influences of wind and river discharges on inundation events. (c) 2021 Institute of Oceanology of the Polish Academy of Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ).

Automated summary

The study assessed the performance of machine learning algorithms in addressing sea level variations in the upper part of estuary, showing that different algorithms can capture and predict sea level temporal variations, and in some aspects, they are more effective than traditional numerical simulations.