Abstract's details
Toward an Ocean Foundation Model Integrating SWOT and Conventional Altimetry
Event: 2025 SWOT Science Team Meeting
Session: Oceanography: Inversion/Assimilation
Presentation type: Poster
The Surface Water and Ocean Topography (SWOT) satellite provides daily global sea-surface-height measurements at kilometre-scale resolution, generating petabytes of data that surpass the processing capabilities of traditional oceanographic workflows. To overcome this data-volume challenge, deep learning, particularly unsupervised architectures, offers promising complementary tools to conventional numerical models and data assimilation frameworks.
We introduce an Ocean Foundation Model primarily trained on SWOT’s high-resolution surface-height fields, demonstrating its ability to effectively integrate additional datasets such as global ocean circulation simulations and traditional nadir altimeter measurements. By embedding these heterogeneous sources into a unified latent space, the model enables efficient similarity searches and rapid generation of ensemble outputs for spatiotemporal gap filling. Furthermore, this generative AI framework provides probabilistic uncertainty estimates comparable to those from traditional data-assimilation methods.
The unified latent representation significantly accelerates the extraction of key oceanic dynamics, such as eddy interactions and SST-SSH coupling, from SWOT’s extensive archives providing a powerful method for obtaining global statics. Our results illustrate how deep-learning foundation models can leverage SWOT observations to enhance oceanographic research and help advance towards an ocean digital twin.
Back to the list of abstractWe introduce an Ocean Foundation Model primarily trained on SWOT’s high-resolution surface-height fields, demonstrating its ability to effectively integrate additional datasets such as global ocean circulation simulations and traditional nadir altimeter measurements. By embedding these heterogeneous sources into a unified latent space, the model enables efficient similarity searches and rapid generation of ensemble outputs for spatiotemporal gap filling. Furthermore, this generative AI framework provides probabilistic uncertainty estimates comparable to those from traditional data-assimilation methods.
The unified latent representation significantly accelerates the extraction of key oceanic dynamics, such as eddy interactions and SST-SSH coupling, from SWOT’s extensive archives providing a powerful method for obtaining global statics. Our results illustrate how deep-learning foundation models can leverage SWOT observations to enhance oceanographic research and help advance towards an ocean digital twin.