A machine learning framework based on multibeam bathymetry for predicting submarine dune migration rates

Predicting submarine dune migration rates is important for understanding seabed evolution and supporting marine engineering safety. This study develops a machine learning framework using two-period multibeam bathymetric data. A dataset of 111 samples was constructed using seven input features: dune wavelength, height, water depth, gentle and steep slope lengths, and gentle and steep slope angles. Eight models were evaluated: EO-Stacking, Stacking, XGBoost, LightGBM, MLP, BPNN, SVR, and RF. EO-Stacking achieved the best test performance, with R² = 0.97, RMSE = 2.29 m/yr, MAE = 1.59 m/yr, and MAPE = 8.43%. Training and testing correlation coefficients were 0.989 and 0.985, respectively. SHAP analysis identified water depth and steep slope length as the most important predictive features. The framework may complement field monitoring and support seabed evolution studies and marine engineering risk assessment.