Using statistical and machine learning approaches to describe estuarine tidal dynamics

Abstract

Estuaries are ecologically valuable regions where tidal forces move large volumes of water. To understand the ongoing physical processes in such dynamic systems, a series of estuarine monitoring stations is required. Based on the measurements, estuarine dynamics can be described by key values, so-called tidal characteristics. The reconstruction and prediction of tidal characteristics by suitable approaches is essential to discover natural or anthropogenic changes. Therefore, it is of interest to inter- and extrapolate measured values in time and to investigate the spatial relationship between different stations. Normally, such system analyses are performed by deterministic numerical models. However, to facilitate long-term investigations also, statistical and machine learning approaches are good options. For a Weser estuary case study, we implemented three approaches (linear, non-linear, and artificial neural network regression) with the same database to enable the prediction of tidal extrema. Thereby we achieve an accuracy of 0.4–2.5% derivation (based on the RMSEs) while approximating measured values over 19 years. This proves that the approaches can be used for hindcast studies as well as for future analysis of system changes. Our work can be understood as a proof of concept for the practical potential of neural networks in estuarine system analysis. HIGHLIGHTS Estuarine water-level characteristics (hydrograph) can be successfully approximated based on weighted machine-learning regression models.; Comparison of different regression models with the same data set shows the potential of optimized neural networks.; Initial feature testing and selection of ML variables provides information on physical processes driving estuarine water-level dynamics.;

Keywords

• feed-forward neural network
• hindcast of tidal extrema (minimum and maximum)
• multiple regression modelling
• weser estuary
• xai