Global bathymetry prediction from satellite gravity and vertical gravity gradient using a deep neural network

Abstract

Global bathymetry is essential for marine sciences, yet over 73% of the seafloor remains unmapped. While satellite derived free-air gravity anomalies are widely used to predict depth, these models often fail to resolve high-frequency  structures in rugged terrains like trenches and seamounts. To address this, we developed a deep neural network (DNN) that integrates vertical gravity gradient (VGG) alongside gravity anomalies to generate improved global bathymetric grid. By training on data grouped into 18-arc-minute bins, we evaluated a VGG-integrated model against a gravity-only baseline. Because VGG mathematically amplifies  short-wavelength signals and filters out regional isostatic noise, our multi-feature architecture better captures complex geomorphological features. An independent evaluation along the steep topography of the Philippine Trench demonstrated the significant advantage of VGG integration which improved RMSE by approximately 22%. These results establish that integrating VGG into deep learning inversions provides a robust pathway for mapping uncharted regions of the seafloor with better detail, particularly across complex bathymetry.