Seafloor Topography Prediction from Altimetry-Derived Gravity Data Using a Wavelet-Assisted and High-Frequency Enhancement Neural Network

Seafloor topography (ST) has important significance for earth science research, marine resource exploration and underwater navigation. The conventional ST inversion methods are limited by linear approximation and poor small-scale topographic feature prediction. This study proposes a novel Wavelet-Assisted and High-Frequency Enhancement Neural Network (WAHFENN), an architecture integrating discrete wavelet transform (DWT), low-frequency retainment module (LFRM) and high-frequency enhancement module (HFEM) to enhance bathymetry prediction accuracy and capture small-scale topographic features. We apply the WAHFENN to predict the ST in a local area of the South China Sea (SCS). The results demonstrate that the WAHFENN model achieves a standard deviation (STD) of 50.66 m against shipborne single-beam check points, outperforming the topo_27.1 and SDUST2023BCO models by 31.46% and 28.49%, and surpassing the conventional Smith and Sandwell (SAS) method, gravity-geological method (GGM), and convolutional neural network (CNN) method by 78.23 m, 65.18 m, and 4.4 m, respectively. The WAHFENN model achieves a STD of 103.80 m against shipborne multibeam bathymetry data, representing improvements of 38.75%, 25.16%, and 15.58% over the SAS, GGM, and CNN models, respectively. The topographic detail comparisons and power spectral density analysis demonstrate that the WAHFENN model has the potential to outperform conventional methods in identifying small-scale topographic features.