Work

Short‐Wavelength Seafloor Topography From SWOT Marine‐Gravity Products Using a Transformer U‐Net With Hybrid Loss

Year: 2026

Type: article

Abstract: Classical satellite altimetry has limited sensitivity to seafloor topography at wavelengths shorter than about 150 km, leaving short‐wavelength structure poorly constrained away from ship tracks. The Surface Water and Ocean Topography (SWOT) mission improves gravity recovery at these scales. We evaluate U‐Net‐family encoder‐decoder models and use a Transformer U‐Net to map four SWOT‐derived marine‐gravity fields on a 1 arcminute grid—the east–west and north–south components of the deflection of the vertical, gravity anomaly, and vertical gravity gradient—band‐pass filtered to 8–150 km, to bathymetry residuals on a 15 arcsecond grid matching GEBCO 2024. A hybrid loss combining mean squared error with gradient and structural‐similarity terms, together with pixel weights from the GEBCO Type Identifier, limits oversmoothing of steep slopes. Agreement with the band‐passed GEBCO 2024 target is reported as product consistency, which improves with multi‐basin training and modest transfer learning across ridge systems. Independent accuracy is assessed against spatially independent, quality‐controlled 1 arcminute ship soundings. On this benchmark, the model improves on a conventional Nettleton inversion of the same SWOT gravity data (SWOT_NET), indicating that the framework can extract useful short‐wavelength structure from SWOT marine‐gravity products. Although it does not yet exceed ship‐constrained global grids, which already incorporate the soundings, the result supports the method as a reproducible gravity‐only framework and a candidate tool for short‐wavelength mapping in under‐surveyed regions.

Source: Journal of Geophysical Research Machine Learning and Computation

DOI: 10.1029/2026jh001392