SWOT Global Bathymetry Modeling Using Deep Neural Networks Trained on Multiple Geophysical Features

This paper presents BathDNN25, a global bathymetry model developed using gravity data derived from wide‐swath altimetry collected by the Surface Water and Ocean Topography (SWOT) mission, with shipborne bathymetry serving as training data in a deep neural network (DNN) framework. BathDNN25 integrates multiple geophysical inputs, including gravity anomalies , vertical gravity gradients , their band‐pass filtered forms , the north and east components derived from the deflection of the vertical (, ), their band‐pass versions (, ), low‐pass filtered bathymetry , and both low‐pass and band‐pass filtered gravity (, ), to capture both large‐scale trends and fine‐scale bathymetric features. A key innovation lies in its use of multi‐scale geophysical features, enabling enhanced sensitivity to morphological complexity such as ridges, escarpments, and seamounts, while adapting well to varying geological conditions and data sparsity. Model performance was assessed using residual statistics against independent data sets, including global shipborne soundings and seamount summits, with BathDNN25 achieving residual standard deviations of 99 and 167 m, respectively. Compared to existing methods (Harper & Sandwell, 2024, https://doi.org/10.1029/2023ea003199 ), this represents reductions in residual error of over 51% and 113%. SHAP analysis across 14 regions and ablation tests using four model variants further confirmed the complementary value of SWOT‐derived gravity features. Overall, BathDNN25 demonstrates accuracy, robustness, and scalability, underscoring the importance of high‐quality geophysical inputs and the potential of SWOT‐derived data and artificial intelligence in advancing global bathymetric modeling.