Application of the Transformer model to Sentinel-2 data for accurate water depth estimation in a coastal environment

Accurate bathymetric mapping is vital for coastal management and navigation, yet traditional sonar remains costly and limited in shallow areas. This study evaluates a Transformer model using Sentinel-2 satellite imagery to estimate water depth in the turbid coastal waters of Nanshan Port, China. We compared the Transformer against Stumpf, log-linear, and neural network models using approximately 4,000 field-measured control points.Outperforming all other methods, the Transformer achieved a coefficient of determination (R2) of 0.90, mean absolute error (MAE) of 0.55 m, mean relative error (MRE) of 0.11 m, and root mean square error (RMSE) of 0.81 m. Although error rates increased across all models at depths exceeding 9 meters, the Transformer maintained robust performance. These findings demonstrate that integrating advanced machine learning with satellite remote sensing provides a cost-effective, accurate solution for bathymetric mapping in challenging turbid environments.