AI spots the ocean's ghost fleet

If you picture the open ocean as a vast, empty wilderness, it turns out our maps were just broken.

For decades, governments and conservationists tracked maritime traffic through the Automatic Identification System (AIS), a radio transponder vessels use to broadcast their location. The glaring flaw? If a vessel captain wants to fish illegally, they can simply flip the switch off or sail through patchy reception areas.

Now, a groundbreaking study published in Nature by Google-backed nonprofit Global Fishing Watch has exposed the sheer scale of that blind spot. Combining satellite radar imagery with deep learning, researchers found that a staggering 75% of the world’s industrial fishing vessels are not publicly tracked. Up to 30% of transport and offshore energy vessels are similarly operating in the dark.

Radar meets AI to crunch 2,000 terabytes of ocean

Standard optical satellite photos are useless when clouds roll in or night falls. To get around this, the team tapped radar data from the European Space Agency’s Sentinel-1 constellation, which pierces cloud cover and "sees" through pitch darkness.

Analyzing 2,000 terabytes of imagery manually would take lifetimes. Machine learning engineer Fernando Paolo and his team developed three deep-learning models to scan the data, spot offshore infrastructure, classify vessels, and estimate vessel sizes across 2017 to 2021.

Focusing on the 15% of the global ocean where roughly 75% of industrial activity happens, the AI revealed crowded maritime traffic in regions previously thought to have little to no activity, particularly around South and Southeast Asia, and off the coasts of Northern and Western Africa.

It also documented a massive offshore energy shift. By the end of 2020, offshore wind turbines outnumbered oil structures globally. By 2021, turbines made up 48% of all stationary ocean infrastructure compared to 38% for oil, concentrated almost entirely off Northern Europe and China. Yet oil platforms still generated five times more surrounding vessel traffic than wind turbines.

Why it matters

We think this work changes ocean conservation from a guessing game into an empirical science.

Nearly every nation recently committed to the Kunming-Montreal Global Biodiversity Framework to protect 30% of global waters by 2030. But until now, nobody could answer the most fundamental question: which 30% should we protect? You can’t negotiate conservation boundaries if you don’t even know where the factory trawlers and offshore oil clusters actually are.

What makes this approach smart is the use of open-source tooling. As researchers note, packaging these deep-learning tools for global imagery democratizes maritime surveillance, allowing low-income nations without expensive naval fleets to monitor their own territorial waters on a shoestring budget.

The obvious catch is enforcement. A computer vision model can flag an unlisted 60-meter steel hull drifting near a protected reef, but it can't launch an intercept boat, inspect holds for forced labor, or resolve international jurisdiction disputes. AI just turned the lights on across the global commons, but governments will still have to do the messy police work.