Multi -layer perceptron inversion of seafloor topography in the South China Sea using multi -source marine geodetic data

This paper uses a Multi-Layer Perceptron neural network (MLP) constructs a high-resolution (1' X1') bathymetry model for the South China Sea region (108 degrees E - 121 degrees E, 6 degrees N- 23 degrees N), known as MLP_Depth. This method integrates data from the Scripps Institution of Oceanography (SIO), including vertical deviation, gravity anomalies, and vertical gravity gradient data, alongside shipborne bathymetric data from the National Centers for Environmental Information (NCEI). Firstly, the positional information of 642716 shipborne bathymetric control points and the gravity information (including vertical deviation, gravity anomalies, and vertical gravity gradient) at nearby 4' X 4' grid points are used as input data, while the actual bathymetry at the shipborne bathymetric control points are used as the output data. Then, the MLP neural network model is trained with this dataset. At the end of the training, the coefficient of determination (R2) is 99% and the mean absolute error (MAE) is 39. 33 m. Then, by feeding the input data from the central points of the 1' X 1' grid cells within the study area into the MLP model, we can obtain the predicted bathymetry values at these central grid points. Finally, based on the predicted bathymetry values, we establish the MLP_Depth model within the study area with a resolution of 1' X 1'. The MLP_Depth model's predicted bathymetry are compared with the measured bathymetry at 160679 check points. The standard deviation (STD) of the difference is 75. 38 m, the mean absolute percentage error (MAPE) is 5. 89%, and the MAE is 42. 91 m. These metrics are superior to those of the GEBC0_2021, topo_23. 1, and ETOPO1 models, as well as the differences between the measured bathymetry at check points (STD: 108. 88 m, 113. 41 m, 229. 67 m; MAPE: 6. 11%, 6.94%, 18. 37%; MAE: 47.33 m, 52. 24 m, 130. 08 m), respectively. Additionally, to assess the accuracy of bathymetry models established using this approach in distinct regions, this study developed bathymetry models (MLP_Depth_A and MLP_Depth_B) within specific area A and area B of the study region. Validation results demonstrate that both MLP_Depth_A and MLP_ Depth_B offer superior precision compared to the MLP_Depth model, effectively capturing the nuances in bathymetry variations.