Assessing future changes in Baltic sea extreme wave heights using a machine learning approach

Abstract Extreme sea surface waves pose significant risks to coastal infrastructure, ecosystems, and human activities, not only in oceans but also in semi-enclosed basins like the Baltic Sea, where limited fetch and complex bathymetry can amplify local impacts. This study presents a novel machine learning framework to project significant wave height at local scales across the Baltic Sea from 1850 to 2100. Using Random Forest models trained on observed wave data and ERA5 atmospheric reanalysis, we simulate 3-hourly maximum Hs time series driven by four CMIP6 climate models under the SSP2–4.5 climate scenario. While SSP2-4.5 is the primary focus due to its alignment with current policy trajectories, the EC-Earth model was selected for detailed projection analysis as it demonstrated the highest skill in reproducing historical wave conditions and synoptic-scale atmospheric patterns. To explore scenario uncertainty, additional simulations using SSP1-2.6 and SSP3-7.0 were also included. Results indicate a basin-wide decline in the 95th percentile of Hs from the early 20th century until the middle of the 21st century, followed by a stagnation shift, with spatial variability across sub-basins. This change appears largely independent of the carbon emission scenario for the future, suggesting a dominant role of internal atmospheric variability and large-scale dynamics patterns, though the limited ensemble size and absence of formal trend significance testing warrant cautious interpretation. An independent analysis of Lamb Weather Types, based on ERA5-reconstructed wave time series, reveals that extreme wave events are most frequently associated with cyclonic and westerly synoptic patterns, with some site-specific differences. In contrast, calm meteorological conditions are negatively correlated with wave extremes. An inter-model comparison reveals substantial variability in projected extremes, underscoring the importance of ensemble approaches. The methodology is transferable to other regions, though its performance may vary with local atmospheric and oceanographic conditions. These findings provide valuable insights for coastal risk assessments and adaptation planning in a changing climate.