Multi-Year Predictability of Sandy Shoreline Change from Remote-Sensing Reconstruction and a Spatiotemporal Transformer

Most studies of sandy shoreline forecasting address relatively short time scales. Under limited annual observations and strong shoreline persistence, the added value of a Transformer over simple baselines and the influence of remotely sensed shoreline definitions remain insufficiently tested. Using Xichong Beach, Shenzhen, we constructed 40-year shoreline series for 80 transects from 284 quality-controlled Landsat waterlines acquired during 1986–2025. We compared a quality-controlled annual landward-envelope composite waterline with annual median waterlines and examined the effects of transect spacing and positional error on long-term change rates. We then developed a residual spatiotemporal Transformer that uses 10 years of shoreline states and historical wind–wave exposure to directly predict five future horizons, and compared it with persistence, rolling linear trend, and random forest models. The annual landward-envelope composite waterline was systematically landward of the annual median waterline (positional RMSE, 12.78 m), but their alongshore LRR patterns were strongly correlated (r = 0.982) and identified consistent major erosion–accretion zones. After Monte Carlo error propagation, the beach-mean LRR was −0.250 m yr−1 (95% interval, −0.299 to −0.203 m yr−1), whereas the direction of change remained uncertain at 39 local transects. Across 400 year–transect locations in the independent 2021–2025 evaluation period, the Transformer produced the lowest RMSE, MAE, and Dynamic RMSE (9.403, 7.399, and 12.042 m, respectively), with an RMSE skill of 27.0% relative to persistence. Environmental features yielded a small gain during rolling validation but no stable improvement in the independent evaluation period. SHAP attribution identified recent shoreline state as the dominant predictive information, followed by wind–wave exposure. Direct forecasts for 2026–2030 gave a beach-mean displacement of −8.527 m in 2030 (95% conditional residual bootstrap interval, −12.514 to −4.950 m), although every local-transect interval crossed zero. Multi-year predictability is therefore scale dependent: beach-mean trends are more resolvable, whereas local change directions remain constrained by observation error and model residuals.