Explainable deep learning based decadal shoreline forecasting in the Southern Baltic

The coastal environments of the Southern Baltic Sea are of high ecological and socio-economic importance. Understanding future changes along its extensive and complex shorelines can help us comprehend the climatic and natural pressures arising from extreme weather events of compound and cascading nature, providing valuable insights for effective coastal management and the prevention of future adverse erosional changes. Current shoreline forecasting methods have limited capabilities to capture nonlinear forcings, have limited temporal forecasting and lack explainability. We present sequence-aware LSTM-RNN framework with optimized lookback functionality designed for end-to-end recursive shoreline forecasting. The model integrates 15 environmental factors spanning climatic, hydrometeorological and geomorphological indicators to enhance spatiotemporal representation, capture compound characteristics and maintain physical consistency. Trained with ERA5 reanalysis products, Landsat satellite observations, and CMIP6 SLR projections, our LSTM-RNN model achieves high forecasting skill of over 25 years, yielding an RMSE of 10.40, MAE of 7.13, and R2 of 0.55. The model was then allowed to make predictions for three proposed sectors, revealing consistent increase in erosional tendencies from 2030 to 2050 across nearly whole study region. Explainable AI method, DeepSHAP reveals that the increasing erosion in these sectors is governed by rising sea levels under high emission scenario when combined with storm surges and maximum significant wave height which far outweigh the accretion caused by wind-wave variables. The progression aligns closely with the established theories of shoreline evolution under the influence of rising sea levels and storm surges, underscoring the model’s ability to identify physically meaningful drivers. The framework demonstrates strong potential for advancing explainable AI in Earth observation, combining predictive accuracy with physical explainability for operational shoreline monitoring and climate change mitigation applications.