New AI Approach Reveals Ocean Currents in Unprecedented Detail

Story by:

• Alex Fox- [email protected]

Media contact:

• Lauren Fimbres Wood- [email protected]

Share This:

• Share this story on Linkedin
• Share this story on Facebook
• Share this story on Threads
• Share this story on Twitter
• Share this story via email

Article Content

The Gulf Stream is full of interleaving temperature gradient structures that hint at complex underlying current dynamics. GOES-East satellite observations and machine learning have, for the first time, connected this observed structure to the much more difficult problem of observing ocean currents. Video credit: Luc
Lenain/Scripps Institution of Oceanography.

Scientists have developed a new method to measure ocean surface currents over large areas in greater detail than ever before. Called GOFLOW (Geostationary Ocean Flow), the approach applies deep learning to thermal images from weather satellites already in orbit, requiring no new hardware to achieve what the researchers describe as a major advancement in ocean observation.

The study, co-led by Luc Lenain, an oceanographer at UC San Diego's Scripps Institution of Oceanography, and Kaushik Srinivasan, a Scripps alumnus now at UCLA, was published today in the journal Nature Geoscience. The study's two other co-authors, Roy Barkan of Tel Aviv University and Nick Pizzo of the University of Rhode Island, are also Scripps alumni. The project was supported by grants from the Office of Naval Research, NASA and the European Research Council.

Ocean currents play a huge role in shaping Earth's weather and climate, transporting heat around the globe, moving carbon between the atmosphere and ocean interior, and redistributing nutrients that sustain life in the sea. Understanding ocean currents are also vital for search-and-rescue operations and tracking the movement of oil spills. Yet measuring currents over large areas of the ocean has remained extremely challenging. Some satellites estimate currents indirectly by measuring variations in sea-surface height, but they typically image the same location only every 10 days or so - too infrequently to track currents that can appear and disappear within hours. Ship-based measurements and coastal radar systems can capture rapid changes but only for limited areas.

This has left a persistent gap in observations at the scales where most of the ocean's vertical mixing occurs - when shallower waters are mixed deeper or vice versa. The phenomena that drive vertical mixing can be less than 10 kilometers (six miles) wide and transform in hours. Understanding vertical mixing matters because it powers key processes such as bringing nutrients up from depth to sustain marine ecosystems and pumping carbon dioxide from the surface to deeper waters where it can be stored long-term.

In 2023, Lenain was examining thermal imagery of the North Atlantic Ocean from the geostationary satellite GOES-East, which is primarily used for observing weather. The images, produced as frequently as every five minutes, showed passing clouds and swirls of warm and cool water evolving on the sea surface. As he looked, Lenain could see the imprint of major currents such as the Gulf Stream in the temperature patterns and began exploring how to convert what his eye could see in those images into a new way to measure ocean currents.

To accomplish this, the team trained a neural network to recognize how ocean surface temperature patterns shift and deform when pushed by underlying currents. The network learned from a high-resolution computer simulation of ocean circulation, which provided examples of temperature patterns and the water velocities that created them. By tracking how complex temperature patterns moved across consecutive images taken by the GOES-East satellite, the trained network could infer the currents responsible for those changes.

"Weather satellites have been observing the ocean surface for years," said Lenain. "The breakthrough was learning how to turn that time-lapse into hourly maps of currents by tracking how temperature patterns bend, stretch and move from one hour to the next."

The researchers tested GOFLOW's accuracy by comparing its output to velocities recorded by shipboard instruments in the Gulf Stream region in 2023, as well as standard satellite methods using ocean topography. GOFLOW's measurements agreed with the data collected with ships and traditional satellite techniques, and revealed much greater detail for smaller, faster-moving eddies and boundary layers where existing methods showed only blurred averages. This newfound detail allowed the researchers to measure for the first time key statistical signatures of small, intense currents that drive vertical mixing in the ocean that previously had been documented only in computer simulations.

"This opens a range of exciting possibilities in physical oceanography that, until now, were largely accessible only through simulations," said Lenain. "Using GOFLOW, we can now measure key signatures of these small, intense currents using real observations rather than relying almost entirely on simulations. This opens the door to testing long-standing ideas about how the ocean takes up heat and carbon."

Because the method works with existing geostationary satellites it does not require new instruments to be launched into space. Over time, GOFLOW could be incorporated directly into weather forecasts and climate models, and may ultimately help improve forecasts by resolving rapidly evolving currents that influence air-sea exchange, marine debris transport and ocean ecosystems.

The researchers note that cloud cover remains a limitation, since clouds block the thermal imagery GOFLOW relies on. Future work will incorporate other types of satellite data to fill in the gaps when clouds block satellites' views and achieve continuous coverage. The team is currently working to extend the method globally. The study's data products and computer code are being made publicly available to support further research and applications.

Learn more about research and education at UC San Diego in: Artificial Intelligence