Summary Traditional bathymetry inversion methods often fail to capture the complex nonlinear relationship between gravity data and bathymetry and lack the capability to quantify prediction uncertainty. To address these limitations, we investigated deep learning and Bayesian methods that enhance prediction accuracy and provide estimates of prediction uncertainty. Three methods—the Fully Connected Neural Network (FCNN), normalizing Flow model (Flow) and FCNN-Markov Chain Monte Carlo (FCNN-MCMC) —were developed to construct high-resolution (1′×1′) bathymetry models (FCNN, Flow, and FCNN-MCMC models) of the South China Sea (113°E-119°E, 12°N-19°N). The input data included positional, topographic, and gravity information from 4′×4′ grid points surrounding each training, validation and prediction point, while the output data corresponded to the measured bathymetry at training and validation points. At the check points, the standard deviation (STD) of the FCNN, Flow, and FCNN-MCMC models decreased by 7. 13 m, 14. 24 m, and 15. 51 m, respectively, compared with topo₂7. 1, and by 18. 19 m, 25. 30 m, and 26. 57 m, respectively, compared with ETOPO2022. The distribution of prediction uncertainties (STD) showed that over 90 per cent of the area exhibited an STD below 130 m. The prediction uncertainties exhibited a spatial distribution similar to the predicted results, with higher uncertainties mainly concentrated in shallow waters and steep areas.
Zhou et al. (Fri,) studied this question.