The shifting patterns of ocean currents shape our climate and weather. Even today, understanding how ocean currents behave is challenging. But researchers have developed an AI tool that can map currents across large swaths of the ocean with a level of detail previously unachievable. 

The team from the University of California, San Diego, published their work in the journal Nature Geoscience on April 13. They trained an AI network using thermal imagery from weather satellites in an approach they call GOFlow (Geostationary Ocean Flow). 

"We can now observe small, fast-changing ocean currents from space with much greater detail and frequency than before," Luc Lenain, an oceanographer at UC San Diego's Scripps Institution of Oceanography and the study's coauthor, told CNET. "Those currents are important because they help control how heat, carbon, nutrients and pollutants move through the ocean."

Ebbs and flows

A few years ago, while Lenain was looking through thermal satellite images of the North Atlantic Ocean, he noticed visual patterns in temperature changes from major currents, such as the Gulf Stream. He had an idea for a new way to measure ocean currents by taking what his eye could see and putting it into an AI tool.

The researchers trained the GOFlow neural network on simulated ocean currents, then used it on real images from a weather satellite. The AI tool used satellite images to track surface temperatures, which shift due to underlying ocean currents. 

By tracking temperature changes in the images, GOFlow inferred which current caused them. 

The team checked their work against data that was gathered by ships in the Gulf Stream region. They also tested GOFlow's results against more traditional satellite methods that rely on tracking height changes in the ocean surface. 

The researchers found that their outputs aligned with other ship and satellite results, but they say that GOFlow also provided greater detail on ocean currents that had previously been documented only in computer models. 

"These kinds of [AI] driven approaches are not replacing physics," Lenain said. "Instead, AI is helping us extract physical information that is already present in satellite observations, but has been difficult to recover with traditional methods until now." 

The view from above

Despite GOFlow's achievements, the researchers note that the tool has some limitations, such as cloud cover: Cloudy days can block a satellite's view of the ocean. They say future work will incorporate additional satellite data to fill these gaps. 

The computer code developed by the researchers will be made publicly available to help further work. 

"We wanted to make this work transparent, reproducible and useful to the broader community," Lenain said. "We see GOFlow as a stepping stone toward more routine use of large remote-sensing datasets combined with machine learning."

Gathering satellite images to learn about ocean currents is an example of Earth observation. Governments and militaries, as well as farmers and insurance companies, rely on this data for decision-making.

"The coolest thing about this work is that it repurposes existing satellite data, meant primarily for weather monitoring and forecasting, to extract ocean-critical information that people hadn't considered before for this purpose," Kaushik Srinivasan, a researcher at the UCLA Atmospheric and Oceanic Sciences dept. and the study's coauthor, told CNET. 

"I feel like this idea, combining existing infrastructure with AI/Machine Learning techniques, is a fertile ground for future science," Srinivasan said.

The GOFlow project is part of a larger AI trend, as AI tools can speed up and improve the accuracy of data analysis. NASA, the European Space Agency and private space firms have begun building and testing AI tools that can analyze such data.