One of the best-documented environmental uses of artificial intelligence is finding ships that do not broadcast their position. A 2024 study in Nature combined Sentinel satellite images, ship position data, and three deep-learning models to map industrial activity at sea from 2017 to 2021, and found that 72 to 76 percent of industrial fishing vessels were not publicly tracked. The work shows what AI adds to environmental monitoring, which is a map of activity, and what it leaves undone, which is enforcement.
Key takeaways
- The study found that 72 to 76 percent of industrial fishing vessels were missing from public tracking systems, compared with 21 to 30 percent of transport and energy vessel activity.
- Three separate deep-learning models handled vessel detection, offshore infrastructure type, and fishing versus non-fishing activity.
- Small vessels and coastal strips are weak spots, so the authors say total fishing activity is likely underestimated.
- A map shows where vessels are. Deciding whether a vessel broke a rule remains a legal and human task.
The gap the study set out to measure
Most public maps of fishing rely on the Automatic Identification System (AIS), a transponder signal that ships broadcast so others can avoid collisions. Vessels without a transponder, or with it switched off, do not appear. The authors asked how much activity those maps miss. Their approach was to find ships directly in radar and optical satellite images, then match each detection against AIS positions. A ship seen from space but absent from AIS counts as untracked.
How the models were used
Radar imagery from the European Sentinel-1 satellites can see ships through cloud and at night, and optical imagery from Sentinel-2 adds detail in clear conditions. According to the paper, one model detects vessels and estimates their length, a second classifies offshore structures as oil, wind, or other, and a third separates fishing from non-fishing vessels. Reported accuracy on the test sets was above 97 percent for vessel detection, above 98 percent for structure type, and above 90 percent for the fishing label. The authors matched detections to billions of AIS messages with a probability-based score.
What the results said
| Finding | Reported result |
|---|---|
| Industrial fishing vessels not publicly tracked | 72 to 76 percent |
| Transport and energy vessel activity not publicly tracked | 21 to 30 percent |
| Change in fishing at the start of the COVID-19 pandemic (2020) | Down about 12 percent, not back to earlier levels by 2021 |
| Offshore structures by 2021 | Wind turbines outnumbered oil structures |
The study points out that much of the untracked fishing is in waters around South Asia, Southeast Asia, and Africa. The gap is a statement about public data, not a count of illegal fishing: a vessel can be legally fishing and still not carry or broadcast AIS.
Limits the authors state
The Sentinel-1 satellites do not sample most of the open ocean, objects very near the shore are not classified, and no imagery was available for Antarctic waters. Detection depends on size. Vessels under 15 meters are mostly missed, so the authors conclude that total fishing activity is likely underestimated. Sentinel-1B also stopped operating in December 2021, which bears on how far the method can be repeated with the same satellites. An illustrative reading: a 12-meter (about 39 feet) coastal boat fishing near shore would fall outside this method, while a 60-meter trawler offshore would very likely be seen.
Questions to ask of any environmental AI claim
- What quantity is measured, and is it observed directly or inferred?
- Which objects, places, or periods are outside the coverage?
- Were accuracy figures measured on data separate from the training data?
- Are the data and code open, so that others can reproduce the result?
- Who acts on the output, and under what authority?
FAQ
Does the study show illegal fishing?
No. It shows vessels that do not appear in public tracking data. Whether a given vessel violated a rule depends on licenses, location, and local law, which the images alone cannot establish.
Why use radar instead of ordinary photographs?
Radar images can be collected through clouds and in darkness, so ships can be detected on more dates. The optical Sentinel-2 images used in the study are limited by cloud cover.
Is the data public?
The authors state that the vessel and infrastructure data are available through the Global Fishing Watch portal, and that analysis code is open source, according to the paper's availability statement.
Can the same method find small boats?
Not well. The paper reports that vessels under 15 meters are mostly missed by the satellites used, which is why the total is described as an underestimate.
Related reading
- Effects of AI on sustainability, in both directions
- Alternatives to artificial intelligence
- How AlphaGo learned to play Go
- What GPT-based chatbots are
More in AI and Society.
Last verified: October 2026