Shoreline Forecasting and Uncertainty Assessment with Gaussian Process Regression Along the Coastline of Hainan Island, China

This study develops a probabilistic framework for forecasting coastline change on Hainan Island, China, by integrating 30 years of Landsat-derived shoreline data (1990–2019), the Digital Shoreline Analysis System (DSAS), and Gaussian Process Regression with Automatic Relevance Determination (GPR–ARD). ERA5 climatic and hydrodynamic variables were assigned to individual coastal transects as annual summary indices. Strict temporal validation demonstrated reliable shoreline-distance prediction performance (R2 = 0.781, RMSE = 11.81 m, MAE = 7.70 m, and bias = −0.66 m), while providing spatially explicit confidence intervals. Feature-importance analyses using ARD and permutation importance identified longitude, latitude, and year as the principal predictors. In contrast, annual wave, wind, temperature, and precipitation summaries showed limited independent predictive contribution, likely because annual aggregation obscures seasonal and event-scale coastal processes. Projections for 2030 and 2050 indicate contrasting shoreline responses, with substantial retreat in southern areas west of Sanya (up to approximately −70 m) and shoreline advance in the northwestern Qiongzhou Strait region (up to approximately +55 m). Projection uncertainty increases over time, reaching approximately ±32–40 m. This framework provides a business-as-usual baseline for coastal management and highlights the value of sub-annual observations and probabilistic machine-learning approaches for improving future shoreline forecasts.