AI fishing vessels research has revealed a surprising gap between the activity taking place at sea and the ships visible on public tracking maps. By combining satellite images with machine learning, scientists found that a large share of industrial fishing vessels in their global analysis were not matched to public location broadcasts. The result does not mean most fishing is illegal. It means the public picture of where industrial fleets operate was much less complete than many readers might assume.
The researchers published their findings in the journal Nature in January 2024 after analysing imagery and vessel records covering 2017 to 2021. Their original study estimated that 72 to 76 per cent of industrial fishing vessels were not publicly tracked in the observed data. It is a historical research result about a particular period and mapping method, not a live count of the world’s boats in October 2026.
Why AI fishing vessels research had to look from space
Public ship-tracking services are useful because many vessels transmit their positions through systems designed to improve navigation and awareness of other traffic. But public signals do not cover everything that moves on the water. Some vessels are not required to broadcast, equipment can fail, coverage can differ and a ship can operate without producing a position that appears on a public map. The absence of a signal is therefore a gap in information, not by itself evidence of wrongdoing.
Satellite imagery provides a different kind of observation. Rather than relying on a vessel to announce its location, an orbiting sensor can detect objects on the sea surface. Radar imagery is particularly important because it can identify many vessels when cloud and darkness make ordinary photographs less useful. Optical satellite images and existing location records can then help analysts understand which objects are likely to be ships and which may already have a matching public position.
The challenge is scale. Looking manually through enormous collections of imagery would be slow, expensive and inconsistent. Machine-learning models can help identify objects, estimate their size and distinguish likely fishing activity from other uses of the sea. That classification is not infallible, and the study includes uncertainty ranges and limitations. Its value comes from combining many observations into a broader picture rather than pretending every satellite image is a perfect eyewitness.
An important reason to combine different sources is that each has blind spots. A position broadcast can provide regular information about a cooperating vessel but is no help when a location is unavailable. A radar image can reveal an object without telling an analyst who owns it or what licence it holds. Optical pictures may offer visual detail when conditions permit, yet clouds interfere. Bringing the records together reduces some uncertainty, while the remaining uncertainty has to be recorded honestly rather than hidden behind a single coloured symbol on a map.
The researchers also needed to distinguish a vessel from a fixed structure. An offshore platform, turbine or other installation may appear in images even though it is not travelling. The broader map includes shipping and energy activity as well as fishing, helping reveal how different human uses of coastal waters overlap. The study illustrates why the phrase ocean monitoring encompasses far more than finding a few suspicious boats.
What the hidden fleet figure actually means
The headline result is striking because it concerns public tracking visibility, not simply the quantity of fish caught. The team estimated that between 72 and 76 per cent of industrial fishing vessels in its analysis were absent from public vessel-tracking records. It also found missing public tracking information for a meaningful share of other vessel activity. The percentages depend on the study’s methods, coverage and vessel classifications, and should not be casually applied to every fishing boat anywhere on Earth.
In its explanation of the mapping project, Global Fishing Watch describes combining satellite data, location signals and machine-learning predictions to find vessels that did not broadcast positions publicly. Much of the mapped industrial fishing activity was concentrated around Asia, while Africa and other regions also had substantial gaps in public visibility. The point is not that a certain country must be behaving badly. It is that the most visible public tracking network can give a distorted picture of which regions are busiest.
An untracked vessel could be operating lawfully in an area where broadcasting is not required. Conversely, some illegal activity may occur even on vessels that transmit their positions. Enforcement therefore needs licensing information, local law, timing, gear, protected-area rules and human investigation. A map can indicate where to look, but it cannot deliver a courtroom finding or label every operator a criminal.
This distinction matters for journalism as much as science. A dramatic claim that satellites caught the world’s illegal fishing fleet would go far beyond the evidence. The more important and defensible story is that governments, researchers and the public can misunderstand ocean activity if they rely solely on the vessels that announce their locations.
Why the information matters beyond fishing fleets
The ocean supplies food, supports coastal livelihoods and contains habitats that are difficult to monitor from land. Authorities attempting to manage fish stocks need to understand how much industrial activity occurs and where it is concentrated. When a public map omits a large amount of work at sea, even well-intended decisions can start from incomplete assumptions.
The same evidence may support conversations about marine protected areas. A designation on a map is useful only if organisations can understand which activities take place inside and around it. Satellite tools can improve the ability to examine patterns over time, while vessel operators and regulators still need fair procedures to establish what an observation means. Better visibility should support accountability rather than replace it.
There is also a direct link to everyday consumers. Many people cannot personally verify the route by which seafood reaches a shop or restaurant. Improved monitoring may help organisations examine supply chains, but it does not automatically certify a product as sustainable. Certification and enforcement remain separate processes with their own evidence, incentives and limitations.
For coastal communities, the stakes are also economic. Local fishers may feel the effects of industrial activity but have little capacity to observe offshore operations themselves. More transparent mapping can inform public debate about access to resources and the enforcement of existing rules. It cannot resolve conflicts over fishing rights by itself, and it must avoid treating small boats as though they all face the same reporting obligations as industrial fleets. Meaningful transparency makes differences visible instead of flattening them.
Similar uses of environmental AI are appearing in other areas. LiveAIWire has covered tools that identify orcas in the sea, satellite analysis of small farms and efforts to detect wildlife trafficking. They illustrate a common principle: the model is most valuable when it reveals evidence people can check and act upon, not when its output is treated as an unquestionable verdict.
What satellite AI still struggles to see
Satellites do not observe every place continuously. Sensor resolution, orbital schedules, coastal weather, sea conditions and the size of vessels all affect what may be detected. A small boat close to land or mixed with other objects can present different challenges from a large industrial vessel on open water. For those reasons, the study’s estimates should not be recast as a perfect census of all maritime activity.
The research also involved classifications, including judgments about whether an object was likely to be engaged in fishing. Those models depend on training data and assumptions that deserve independent scrutiny. Improving a classifier or adding different imagery can change the resulting map. That is normal in developing research, provided the limitations are available for others to examine.
Another complication is interpretation. An apparent change in activity might reflect real behaviour, better observation or different data coverage. Analysts need to distinguish those explanations before declaring that an enforcement measure has succeeded or failed. This is particularly important when using a historical dataset to comment on conditions several years later.
The encouraging result is not that artificial intelligence has solved illegal fishing. It is that a broad blind spot can be investigated with evidence that was previously difficult to combine. Fishing fleets need not disappear from the map simply because they are missing from the public broadcasts. The harder next step is turning that visibility into fair, effective and transparent decisions about an ocean shared by everyone.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.