Recovering Bathymetry Using BP Neural Network Combined with Modified Gravity–Geologic Method: A Case Study in the South China Sea

Abstract

The gravity–geologic method (GGM) is widely used for bathymetric predictions. However, the conventional GGM cannot be applied in regions without actual bathymetric data. The modified gravity–geologic method (MGGM) enhances the accuracy of bathymetric models by supplementing short-wavelength gravity anomalies with an a priori bathymetric model, but it overlooks the significance of actual bathymetric data in the prediction process. In this study, we used the BP neural network (BPNN), incorporating shipborne depth soundings and coastline data as zero-depth estimates combined with the MGGM to produce a bathymetric model (BPGGM_BAT) for the South China Sea (105°E–122°E, 0°N–26°N). The results indicate that the BPGGM_BAT model decreases the root-mean-square (RMS) of bathymetry differences from 154.33 m to approximately 140.43 m relative to multibeam depth data. Additionally, the RMS differences between the BPGGM_BAT model and multibeam depth data show further improvements of 19.63%, 20.10%, and 19.54% when compared with the recently released SRTM15_V2.6, GEBCO_2022, and topo_V27.1 models, respectively. The precision of the BPGGM_BAT model is comparable to that of the SDUST2023BCO model, as verified using multibeam depth data in open sea regions. The BPGGM_BAT model outperforms existing models with RMS differences of 8.54% to 32.66%, as verified using Electronic Navigational Chart (ENC) bathymetric data in the regions around the Zhongsha and Nansha Islands. A power density analysis suggests that the BPGGM_BAT model is superior to the MGGM_BAT model for predicting seafloor topography within wavelengths shorter than 15 km, and its performance is closely consistent with that of the topo_V27.1 and SDUST2023BCO models. Overall, this integrated method demonstrates significant potential for improving the accuracy of bathymetric predictions.