Shorelines as graphs: A spatio-temporal data-driven model for predicting shoreline dynamics

Accurate prediction of shoreline dynamics is essential for sustainable coastal management. Data-driven models have become increasingly important due to their efficiency and reduced reliance on detailed domain knowledge. However, most existing approaches treat spatial and temporal patterns separately, limiting their ability to capture spatio-temporal dependencies. This study introduces GraphShore, a novel model that represents shorelines as graphs and applies a spatio-temporal joint graph neural network to model both spatial structures and temporal sequences within the same framework. When evaluated against a benchmarking dataset on an embayed beach, the proposed model ranked fifth among more than 40 models. It also achieves exceptional accuracy (normalized root mean square error (NRMSE) < 1) for beaches influenced by inlets and engineering structures, where dynamics are more complex and non-stationary. Ablation experiments revealed that while incorporating spatial information offers limited benefit for embayed beaches such as Curl Curl, it substantially improves model stability and accuracy for non-stationary environments. Furthermore, the architecture of graph neural network enables direct extraction of learnt spatial and temporal correlations, which align closely with the variables used in physics-based cross-shore and longshore sediment transport models, respectively. This demonstrates that GraphShore can capture patterns that are consistent with physically meaningful proxies of shoreline dynamics. This study not only advances data-driven shoreline modelling but also offers a transferable methodology for representing and predicting geospatial vector dynamics in other geophysical systems. • A spatio-temporal data-driven model is proposed to predict shoreline change. • Model performs well across beaches with varying shoreline dynamics. • The spatio-temporal correlations learnt by the model are physically interpretable.