Author(s): Franziska Lauer; Frank Kosters
Linked Author(s):
Keywords: Comparative analysis; Deep regression; Multi response machine learning; Recurrent LSTM networks
Abstract: In coastal hydraulic engineering applications, machine learning (ML) approaches — particularly those using (deep) neural networks — offer an interesting alternative to the standard 3D hydro-numerical modelling. However, these ML-methods are often only applied to isolated aspects rather than integrating multiple responses of a system. As an example, we investigate the implementation of a multi-response ML in the context of tidal characteristic salinity. For our case study we have chosen salinity measurements from multiple monitoring stations located in the northern German Elbe estuary. Through a comparative analysis, we explore how different multi-response setups capture spatial variations in the estuarine salinity. The results demonstrate that the precision of predictions made by simple, deep and recurrent neural networks depends strongly on the optimization of their architectures. In all configurations, however, the multi-output option effectively approximates salinity across the entire study area, highlighting its potential for an application in coastal engineering.
Year: 2026