Xiaoli Deng (School of Engineering, University of Newcastle, Newcastle, Australia, Australia); Ole Baltazar Andersen (DTU Space, Technical University of Denmark, Lyngby, Denmark., Denmark); Richard Coleman (Institute for Marine and Antarctic Studies (IMAS), University of Tasmania, Tasmania, Australia, Australia)

We introduce BathDNN25, a deep neural network model purpose-built to enhance global bathymetric prediction using satellite-derived gravity data. The model is trained on shipborne bathymetry and leverages geophysical inputs derived from the Surface Water and Ocean Topography (SWOT) mission’s wide-swath altimetry, including gravity anomalies, deflections of the vertical (DoV), vertical gravity gradients (VGG), and their band-pass filtered variants.

BathDNN25 is specifically designed to overcome limitations posed by sparse in-situ data and complex geological variability. A key architectural innovation is the integration of both raw and band-pass filtered VGG features, enabling the model to resolve subtle and multi-scale seafloor features such as ridges, escarpments, trenches, and seamounts. Through adaptive feature extraction and multi-scale learning, the model generalizes effectively across geologically diverse marine regions.

Extensive validation shows BathDNN25 significantly outperforms existing models, achieving residual standard deviations of 97 m for shipborne data and 205 m for seamounts. These results represent accuracy improvements exceeding 55% and 74%, respectively, over state-of-the-art methods (Harper and Sandwell, 2024).

BathDNN25 demonstrates the potential of deep learning to transform bathymetric mapping by fusing satellite altimetry with carefully engineered geophysical features. Its scalability, precision, and robustness offer a valuable tool for advancing global ocean modeling, geophysical interpretation, and marine resource exploration.