Forecasting and Hindcasting Coastal Scarp Dynamics Using Forcing Composites and Machine-Learning Models in Saaremaa, Estonia

Due to climate change and global sea-level rise (GSLR), the Estonian coasts, like many other previously depositional coasts around the world, are becoming erosional. The objective of this study is to statistically analyse and model (both forecast and hindcast) changes in the coastal scarp at the Järve coast in southwestern Saaremaa, Estonia, taking into account recently documented and projected shifts in regional climatological forcing and sedimentological background. Due to postglacial rebound with a local uplift of 2.2 mm/a over the Middle and Late Holocene, the area emerged from the sea about 4000 years ago. However, the growth of the uplifted coastal barrier, which has fully merged with the Saaremaa mainland, has essentially ceased, and the coast has largely become erosional. GSLR has recently outweighed the local postglacial sea-level lowering; storminess patterns have changed; the number of ice days has nearly halved in Estonia over the past ~100 years, and these tendencies are expected to continue.To statistically analyse past changes and project future coastal developments, matrices of annual forcing data, including 13 selected and presumably most influential parameters (such as wind speed components, storminess indicators, average and maximum sea levels, air temperatures, and sea-ice statistics from neighbouring meteorological-hydrological stations), are juxtaposed with parameters describing the geomorphic outcomes observed at five selected cross-shore profiles on the currently erosional Järve coast, as well as at the downdrift, accretional Mändjala site. Variations in volumetric (erosion-accretion) changes in the scarp above the mean sea-level elevation and changes in scarp position are examined for the period 1990-2025.Although the forcing parameters were chosen to minimize mutual duplication, the multivariate statistical analysis (correlation matrices and principal component analysis) yielded three major forcing composites. These were related, first, to gently varying, mainly NAO-related fluctuations in winds, storms and relative sea level; second, to the continuous warming trend, characterized by increasing air temperature and decreasing numbers of ice days; and third, to occasional (catastrophic and NAO-independent) storm-surge events. It was found that extreme storms can cause significant geomorphic changes not only during the event but also over longer periods. Such events produce immediate scarp erosion of several meters and deliver large volumes of fine-grained sediment to the nearshore zone, where it can be readily redistributed along the coast in subsequent years, leading to enhanced accretion in depositional areas even under relatively low forcing conditions. Furthermore, multivariate statistical models (including Principal Component Regression and forecasting), as well as machine-learning techniques (e.g., Random Forest, Boosted Trees, Support Vector Machines), relating the suite of forcing conditions to geomorphic outcomes for each profile, were tested. The forcing sets were manipulated to predict future developments along the Järve-Mändjala erosional-accretional coast. In addition, using older (observed and partly reconstructed) meteorological-oceanographic data on relative sea level, air temperature, ice conditions, and storm surges, the coastal changes were back-traced to pinpoint the mode change from areal increase and barrier growth to coastal erosion, which is hypothesized to have occurred approximately 50–100 years ago.