Predicting Seiche-Impacted Estuarine Water Levels with Machine Learning Methods

In estuarine environments, machine learning (ML) methods have been widely applied to predict water-level variations prone to flooding. However, most studies have focused on low-frequency components driven by tides and surges, neglecting high-frequency oscillations such as seiches. This study addresses this gap by assessing the ability of ML methods to predict seiche-influenced water levels. The application was conducted in the upper Elorn estuary (France), where seiches exceeded 0.6 m in height, with first-mode periods of 45–70 min. The ML procedure relied on a series of recurrent neural networks (RNNs, LSTM, and GRUs) and was implemented in a two-step framework to separately predict (i) low-frequency water-level variations and (ii) high-frequency seiche oscillations. The model accurately reproduced low-frequency dynamics (with a coefficient of determination of 0.98) and captured a substantial portion of seiches-related variability during major events. The integration of seiches improved peak total water-level predictions, reducing the mean absolute error by 30% during tidal cycles characterized by strong seiches (amplitude exceeding 0.1 m). Furthermore, the inclusion of seiches enhanced the estimation of the highest 10% peak water levels while reducing the tendency to underestimate measurements. These findings emphasize the importance of integrating seiche-generating physical processes into ML-based forecasting frameworks.