Coupling Satellite-Derived Coastal Change Metrics with Data-Driven Modelling for Near-Future Predictions

Abstract Modelling short-term coastal change and ongoing resilience pre- and post-storm events can help coastal managers make better-informed decisions. Hazards exacerbated by climate change often affect coasts differently, so approaches to predicting physical responses need to be broadly applicable. When compared to traditional process-based approaches, data-driven models require less a priori information, and can be simpler and more efficient in construction. However, these models rely on regular measurements of coastal metrics to form a statistical understanding of past trends and make robust predictions. Gathering observations via ground-based techniques is costly and logistically difficult, leading to spatiotemporal record gaps. Public satellites offer a solution with regular, scalable, moderate–high resolution data giving greater insights into coastal change. Seizing upon this opportunity for integration, we present a framework for satellite-data-driven, short-term coastal change predictions in near-real-time. A variety of satellite-derived observations of coastal change are obtained automatically for a site in northeast Scotland, using the opensource toolkits CoastSat and VedgeSat. These include the waterline and vegetation edge derived from Sentinel-2 imagery, and wave-related metrics from Copernicus Marine Service. A deep learning model is trained with these regular observations (stored as timeseries on cross-shore transects) over the 10-year catalogue of Earth Observation data. The trained model is then forced with near-future wave conditions from Copernicus, and predicts 10-day cross-shore waterline and vegetation edge positions. With validation, neural networks trained on satellite-derived observations could offer a revolution in automated, scalable, efficient solutions for coastal management in a changing climate.