Pranciškus Brazdžiūnas, Darius Jarmalavičius, Gintautas Žilinskas, Donatas Pupienis

Machine learning-based prediction of cross-shore profile evolution in the southeastern Baltic Sea, Lithuania

Abstract This study explores the use of neural network-based machine learning techniques to predict yearly changes of the foredune height, beach height, and beach width. The research focuses on the evolution of the Curonian Spit coast based on cross-shore profile long-term field measurements (1993 to 2018) supported by the analysis of empirical and modelled annual wind, wave, and sea level data. The performance of two types of recurrent neural network models – LSTM and GRU – was assessed by monitoring training progress and validating on unseen data. Both demonstrated the ability to predict general trends; however, the LSTM model exhibited a superior performance in accurately discerning the direction of cross-shore profile parameters development. However, the models showed limited accuracy in predicting more stable morphometric parameters, such as foredune height, likely due to constrained variability in the available data. The signs of overfitting observed during training further highlight the insufficiency in both the variability and duration of the training data set. The findings demonstrate a high potential of machine learning methods to support coastal change forecasting, although the effectiveness of these models remains highly dependent on the quality, temporal span, and spatial coverage of the input data. Incorporating a longer time series with additional factors, such as nearshore seabed morphology and sediment type, may further enhance model performance.

Doi https://doi.org/10.5200/baltica.2025.2.7

Keywords beach; foredune; coastal morphology; field surveys; algorithm