Prediction of beach profile changes using spatiotemporal graph neural networks for beach morphological evolution modeling

초록

In this study, a deep learning framework based on spatiotemporal graph neural networks (GNNs) is proposed to improve the prediction of beach morphological changes by effectively capturing nonlinear relationships and spatiotemporal dependencies among multiple input variables. The model is applied to the Wonpyeong (WP) region on the eastern coast of Korea, comprising four sandy beaches. It incorporates beach characteristics and oceanic forcing data collected through field surveys and ocean observations from 2010 and 2024, along with satellite optical imagery as supplemental remote sensing input. The WP region is subject to frequent anthropogenic interventions, such as submerged breakwater installation and beach nourishment, implemented to mitigate erosion resulting from coastal development. Accordingly, continuous monitoring and accurate morphological change prediction are critical for effective coastal management. An extensive set of ablation experiments was conducted to identify the optimal input variables, which include not only oceanic forcing factors, such as wave conditions and tidal levels, but also beach characteristics such as beach-face slope, median grain size, shoreline position, and cross-sectional profiles indicative of sediment transport and morphological change. To accommodate diverse temporal resolutions, variable-length data embeddings were developed for each input, along with a dedicated subnetwork for extracting shoreline features from satellite imagery. Despite abrupt anthropogenic alterations in beach morphology, the model exhibited robust predictive performance. Although it tended to underestimate bed level changes induced by artificial interventions, the predicted patterns of erosion, deposition, and shoreline movement aligned well with observed trends. Furthermore, the proposed model outperformed recently developed deep learning approaches in terms of prediction accuracy. © 2025 Elsevier Ltd

키워드