Explainable sequence-aware deep learning based decadal shoreline modelling in the Southern Baltic until 2050

Understanding future shoreline evolution is fundamental to developing adaptive coastal management strategies under climate change scenarios. This study analyzes the Southern Baltic Sea’s ∼37-km stretch comprising Usedom and Wolin islands, where sandy coastlines face intensifying erosion threats under rising anthropogenic and climatic pressures. We introduce an explainable Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) framework designed to bridge the gap between Deep Learning (DL) performance and physical interpretability in decadal forecasting. The results show that the best model achieves a Root Mean Squared Error (RMSE) of 10.40 m, Mean Absolute Error (MAE) of 7.13 m, and R-squared (R 2 ) of 0.55. Deep SHapley Additive exPlanation (DeepSHAP) attribution reveals that erosion is driven by the compound interaction of sea-level rise (SLR), storm surges, and extreme waves. This transferable framework represents a significant methodological contribution, enhancing regional early-warning systems and providing a robust, "white-box" approach for Baltic Sea’s operational coastal management. • A sequence-aware neural network model forecasts decadal shoreline change until 2050. • The best model achieves an R-squared value of 0.55 for decadal shoreline change. • Explainable models show erosion is driven by sea levels and extreme waves. • Widespread erosion intensification is projected for the Southern Baltic by 2050.