Trans-UNet Network for Predicting Bathymetry in South China Sea From Gravity and Geological Data

This paper proposes a method for constructing a seafloor topography model based on a Transformer-enhanced U-Net network (Trans-UNet) to improve the accuracy of bathymetry predictions. The method incorporates gravity data (gravity anomalies, vertical deflections, and vertical gravity gradients) and geological data (the sea surface to Moho, sediment thickness, mean dynamic topography, and seafloor age). The proposed method is tested in the South China Sea (112°E-119°E, 11°N-20°N), and a seafloor topography model (Trans-UNet model) with a resolution of 1′×1′ is constructed for the region. By comparison, a seafloor topography model (U-Net model) for this region was constructed using U-Net. Compared to the inspect points, the standard deviation of the difference between the Trans-UNet model and the measured bathymetry at those points is 124.11 m. Compared to the U-Net model, topo_25.1, DTU18BAT and GEBCO_2024 models, the accuracy of Trans-UNet model shows improvements of 7.64 m, 21.49 m, 76.91 m, and 4.94 m, respectively. These results effectively demonstrate the validity of the proposed method and verify the accuracy of the Trans-UNet model's performance. By comparing the accuracy of various seafloor topography models across different depth ranges, the Trans-UNet model shows superior performance in deeper water regions, with improvements of 7.37m and 7.28m compared to the GEBCO_2024 model in the depth ranges of 2000-4000m and greater than 4000m, respectively.