Neural Networks for Ocean Monitoring: How AI Tracks Marine Life

The ocean covers about 71 percent of Earth's surface, roughly 361 million square kilometers, and most of it has nobody watching. Neural networks, software that learns patterns from labeled examples, are starting to change that. They find ships in radar images, pick whale calls out of years of underwater audio, and sift through satellite records for signs of strain on reefs. This article explains what ocean monitoring with neural networks actually delivers, how accurate it is, and where the popular claims run ahead of the evidence.

Why the ocean is so hard to watch

Start with scale. The world ocean covers about 361 million square kilometers, or 70.8 percent of Earth's surface, and its average depth is 3,688 meters, according to Wikipedia's ocean article. The Seabed 2030 project, which aims to produce a full map of the seafloor, reported that 27.3 percent of the seafloor had been mapped at high resolution as of June 2025. Our explainer on why the ocean is still mostly unexplored covers why that number is so low, and the one on how sonar maps the seafloor covers how the mapping is done.

Imagine policing every road in a country with a handful of speed cameras. You would catch some drivers, learn little about where the rest go, and never know how many you missed. Ocean observation works like that. In-situ sensors are sparse: the Argo program, which has run since 2000, uses about 4,000 drifting floats that dive to roughly 2,000 meters every 10 days and measure temperature and salinity. Argo's own site says the network was collecting 13,000 data profiles each month in 2024. That is a remarkable instrument, but it samples the full depth of the global ocean with only a few thousand moving points.

So the ocean is watched mostly from above and from below, with four data streams that complement each other.

Data stream	What it is	Strength	Limit
Optical satellites	Sentinel-2 has 13 spectral bands at 10 to 60 meter resolution, a 290 km swath, and about a 5-day combined revisit	Rich color detail; free data	Blocked by clouds, haze, and polar winter darkness
Radar satellites (SAR)	Sentinel-1 carries a C-band synthetic-aperture radar; its main mode gives 5 × 20 m resolution over a 250 km swath	Works day and night, through cloud and rain	Each satellite repeats its orbit every 12 days; resolution too coarse for small boats
Ship transponders (AIS)	VHF radio messages broadcast by vessels; about 200,000 ships appear on Global Fishing Watch at any moment	Identity, position, and speed	Only required on larger vessels; can be switched off or spoofed
Hydrophones	Underwater microphones recording sound continuously	Detects animals that cannot be seen from the surface	Point sensors; huge audio volumes

The satellite facts come from Wikipedia's Sentinel-1 article and its Sentinel-2 article. Sentinel-1B lost power on December 23, 2021, and Sentinel-1C launched on December 5, 2024, so radar coverage was thinner in between. The AIS facts come from Wikipedia's AIS article: the International Maritime Organization's SOLAS convention requires AIS on international-voyage ships of 300 gross tons or more and on all passenger ships, and the system "offers essentially no security," so a vessel can turn it off or broadcast false positions.

How neural networks find ships that do not want to be found

The case for automation begins with money and catch. A 2009 study in PLoS ONE by David Agnew and colleagues estimated that illegal and unreported fishing costs between $10 billion and $23.5 billion a year, roughly 11 to 26 million tonnes. Across the 54 exclusive economic zones and 15 high-seas regions they analyzed, the losses ran 13 to 31 percent of catch, with a mean of 18 percent. That study was published in 2009, so its numbers should be read as an order-of-magnitude range, not as a current measurement.

The enforcement problem is that illegal fishing has no reason to announce itself. Fleets that break rules often switch AIS off, a behavior that monitoring groups call "going dark." So the question for a neural network is: can you detect a ship from an image alone, whether or not it is broadcasting?

The most thorough answer so far is a 2024 study in Nature by Fernando Paolo and colleagues, titled "Satellite mapping reveals extensive industrial activity at sea." The team combined about 2 petabytes of Sentinel-1 radar and Sentinel-2 optical imagery from 2017 to 2021 with 53 billion AIS messages, using deep learning to detect and classify vessels and offshore structures, then checking each detection against AIS to see whether it was publicly tracked.

The finding that frames the problem: 72 to 76 percent of industrial fishing vessels detected in the study were not publicly tracked, compared with 21 to 30 percent of transport and energy vessels.

That does not mean three quarters of fishing is illegal. Many vessels lack any requirement to carry AIS, and some broadcast but are not publicly visible. What it means is that public tracking data alone gives a badly incomplete picture, and that images fill the gap. The authors write that vessels engaged in illicit activity often switch AIS off or manipulate their positions, and that the largest recent illegal fishing and forced-labor cases involved fleets that mostly did not use AIS.

A detector and three classifiers, step by step

The pipeline is a good example of how neural networks are actually used: as parts of a larger system, not as one magic model.

Stage	Method	Input	Output	Reported performance
1. Find bright targets	CFAR algorithm (not a neural network)	Radar image	Candidate detections	Flags bright targets against sea background
2. Vessel or noise, and length	ResNet-based convolutional network	80 × 80 pixel dual-band radar tiles	Presence and length	F1 0.97 on test data; R² 0.84 for length
3. Infrastructure type	ConvNeXt-based, two-branch network	100 × 100 pixel radar tiles plus Sentinel-2 optical tiles	Wind, oil, other, or noise	98.9% accuracy
4. Fishing or not	ConvNeXt-based network	Estimated length plus an 11-layer environmental stack (depth, distance from port, chlorophyll, temperature)	Fishing or non-fishing	90.5% test accuracy

CFAR stands for constant false alarm rate. It flags a pixel as a candidate if it is much brighter than its surroundings, with the threshold set relative to local sea clutter so that choppy water does not trigger a flood of alerts. It works like spotting a lighthouse on a dark coast: you do not need to know what the object is, only that it stands out from the dark around it. The neural networks then do the part that is hard to write as rules: deciding whether a bright blob is a ship, how long it is, and whether it is probably fishing.

The training data matters as much as the architecture. The presence and length model used about 12,000 samples drawn from high-confidence matches between radar detections and AIS, labeled by domain experts. The fishing classifier used about 120,000 tiles, with labels taken from AIS vessel identities. That choice has a consequence the authors acknowledge: a classifier trained on vessels that broadcast AIS may be less accurate in places where fishing is largely untracked.

What the detector misses

Detection depends on the size of the target and the resolution of the sensor. The paper reports recall above 95 percent for vessels longer than 50 meters, around 80 percent for 25 to 50 meters, and a steep drop below 25 meters. The authors estimate the false-positive rate at around 2 percent at most. Sentinel-1 does not sample most of the open ocean, sea-ice areas are not classified, and objects within 1 kilometer of shore are skipped. The study covers more than 15 percent of the ocean, which holds more than 75 percent of industrial activity, concentrated in coastal waters, and the authors note that some Chinese and Asian fishing activity is cut off at the edge of the satellite footprint and probably underestimated.

Those headline accuracy figures also need context. Suppose a scan produces 1,000,000 candidate patches containing no vessel at all. At a 2 percent false-positive rate, that is 20,000 false alerts. Whether 20,000 false alerts is acceptable depends on how many true detections there are and what it costs a coast guard to check each one. Our explainer on why 99 percent accuracy can be useless walks through that arithmetic, which is the reason ocean-monitoring papers report recall and precision by vessel size rather than a single accuracy number.

Open datasets and public tools

Progress in this field depends on shared data. The xView3-SAR dataset, introduced at NeurIPS 2022, consists of nearly 1,000 analysis-ready Sentinel-1 radar images, each averaging about 29,400 by 24,400 pixels, and is described by its authors as the largest labeled dataset for training machine-learning models to detect vessels and ocean structures in radar imagery. Each image comes with co-located bathymetry and wind rasters, and the data and code were released publicly. The accompanying xView3 challenge invited teams worldwide to build "dark vessel" detectors.

On the tool side, Global Fishing Watch launched its website in September 2016, created by Google with Oceana and SkyTruth, and became a nonprofit in 2017, according to Wikipedia's summary. The Allen Institute for AI's Skylight platform provides near-real-time maritime monitoring for conservation at no cost, draws on satellite imagery and about 100 million daily vessel position points, and lists NASA, the European Space Agency, Maxar, and the International Union for Conservation of Nature as partners. Skylight's page does not publish independent performance metrics, so claims about its detection rates are the developer's own.

How algorithms listen to whales

Sound travels farther underwater than light does, which is why whales rely on it, and why researchers deploy hydrophones. Baleen whale sounds range from about 10 hertz to 31 kilohertz, according to Wikipedia's article on whale vocalization. Blue whale moans have source levels of 155 to 188 decibels relative to 1 micropascal at 1 meter, and humpback song units run from about 20 hertz to above 24 kilohertz.

Sound source	Feature	Value
Baleen whales overall	Frequency range of sounds	About 10 Hz to 31 kHz
Blue whale moans	Source level	155 to 188 dB re 1 µPa at 1 m
Humpback whale song	Frequency of song units	About 20 Hz to above 24 kHz
Toothed whales (odontocetes)	Clicks and buzzes	0.2 to 150 kHz

A year of continuous recording produces far more audio than people can listen to. The standard solution is to convert the audio into a spectrogram, a picture with time on one axis, frequency on the other, and loudness as brightness. A whale call then becomes a recognizable shape: a rising sweep, a repeated pulse train, a stack of harmonics. Think of sheet music for a concert hall full of noise. A trained eye, or a trained network, can pick out one melody among the ship engines and snapping shrimp. For the biology behind the music, see our piece on why whales sing underwater.

A worked example from Hawaii

The best-documented example is a collaboration between Google and the NOAA Pacific Islands Fisheries Science Center. Google's October 29, 2018 write-up describes a ResNet-50 convolutional network trained on spectrogram images drawn from about 15 years of underwater recordings, which came to 9.2 terabytes after being downsampled from 200 kilohertz to 10 kilohertz. The model used per-channel energy normalization, a step that suppresses steady narrow-band noise such as boat engines. On 75-second test clips it reached more than 90 percent precision and 90 percent recall. The authors cautioned that the test data came from similar equipment and conditions as the training data, which means performance at a new site with different hydrophones is not guaranteed.

That caution is the standard limit of this approach. Networks trained at one site learn the site's noise as well as the whales' calls, so each new deployment needs checking, and often retraining.

What acoustic monitoring changes for right whales

The North Atlantic right whale shows why detection matters. NOAA's most recent estimate is about 380 whales, of which only about 70 are reproductively active females. Entanglement in fishing gear and vessel strikes are the leading causes of death and serious injury, NOAA estimates that more than 85 percent of right whales have been entangled at least once, and an Unusual Mortality Event has been ongoing since 2017. Research suggests only about a third of deaths are detected, so the true toll is likely larger than the recorded one.

NOAA's vessel rules combine fixed and flexible tools. In Seasonal Management Areas along the East Coast, most vessels 65 feet or longer must slow to 10 knots or less at certain times of year. Voluntary Right Whale Slow Zones are triggered by underwater acoustic detections, and NOAA is deploying passive acoustic devices called SoundTraps to track whale presence in the Southeast. NOAA's page does not say which detection algorithms feed the slow zones, so it would be wrong to credit neural networks for any particular alert. What the evidence supports is narrower and still important: continuous acoustic listening lets managers react to whales that no one has seen.

What satellites can and cannot tell us about coral reefs

Coral bleaching is the ocean problem most often linked to AI in headlines, so the details matter. Corals live in partnership with symbiotic algae, called zooxanthellae, that supply up to 90 percent of the coral's energy and give it its color. When water gets too warm, the coral expels the algae and turns white. A bleached coral can recover if the algae return, but it is more vulnerable to starvation and disease.

Heat stress is monitored with a physical rule, not a neural network. NOAA's Coral Reef Watch flags "hot spots" where sea surface temperature runs at least 1 degree Celsius above the long-term monthly average, then accumulates that stress over time in a measure called degree heating weeks. The program sends alerts for about 190 reef sites, according to Wikipedia's summary of coral bleaching. The method is simple by design, and because it rests on satellite temperature, it lets scientists spot rising stress earlier than diving surveys can.

The scale of recent events is large. The fourth global bleaching event began in February 2023 and was confirmed by NOAA in April 2024. From January 1, 2023, to September 30, 2025, bleaching-level heat stress affected about 84.4 percent of the world's coral reef area, with bleaching documented in at least 83 countries and territories. On the Great Barrier Reef, a study of 462 coral colonies at One Tree Island found that by July 2024 only 92 were unaffected, 193 had died, and 113 were still bleached. For the biology of what is at stake, see our guide to how coral reefs work, and for a related long-term stressor, our explainer on ocean acidification.

What role do neural networks play here? It is smaller than the headlines suggest. A degree-heating-week forecast is already an accumulation rule applied to temperature. Learned models might improve local forecasts by adding factors such as depth, currents, past stress, and reef composition, but whether they beat the simple rule has to be shown against observed bleaching, and bleaching events are rare and unevenly surveyed, which makes training and testing hard. The honest summary is that coral heat-stress monitoring is a success of satellites and physics, and machine learning is a candidate for refinement rather than the engine.

Where the data runs out

The strongest limits on ocean AI are not computational. They are about what data exists.

Coverage. Radar satellites sample a fraction of the ocean. The Nature study covered more than 15 percent of it, and its sample excluded sea-ice regions and the nearshore strip.

Bias in the labels. Models learn from labeled examples, and labeled examples come from vessels that broadcast AIS, regions with research budgets, and instruments whose recordings were annotated. A model that excels in the Gulf of Maine or off Hawaii may stumble elsewhere, and the evidence for that risk is in the papers' own caveats about similar equipment and untracked fishing.

Small boats and small things. Detection falls steeply below 25 meters, which is exactly the size range of many artisanal and coastal fishing boats. Ghost gear, the abandoned or lost fishing nets that keep catching animals, is harder still. A Wikipedia review of ghost nets cites estimates of 48,000 to 100,000 tonnes of lost gear generated each year, and a separate figure of about 640,000 tonnes, or 10 percent of marine litter, without reconciling the two. It describes spotting drifting nets as a needle-in-a-haystack problem, and the newest approaches it lists, such as a 2024 net-harvesting buoy and GPS tracking of gear, do not rely on image recognition. Our piece on AI and ocean plastic cleanup covers the machines that try to collect what the detectors find.

Shifting conditions. Ocean conditions, vessel types, and fishing methods change, so a model trained on 2017 to 2021 data can drift. Continuous retraining and fresh labeled data are the standard remedy, and both cost money.

Access and equity. Sentinel data is free and open, and some tools such as Skylight are offered at no cost to conservation users, which narrows the gap. Gaps remain in the capacity to act on the results, which is the subject of the next section.

Detection is not enforcement

An alert is only the first link in a chain that runs through verification, a legal authority, a vessel that can reach the spot, and a court that will act. The hardest links are legal, not technical.

The main international instrument is the FAO's Port State Measures Agreement, brokered in 2009 and in force since 2016. Wikipedia's article on illegal, unreported, and unregulated fishing describes it as closing ports to vessels suspected of illegal fishing, so that permission to dock can be denied, but it does not compel countries to apply the measures to their own flagged ships. The same article notes that false positives and manipulated data reduce the effectiveness of satellite AIS and vessel monitoring systems.

Falsified signals are a real feature of the data. The Wikipedia AIS entry cites more than 500 cases of ships manipulating their systems to hide their locations, linked to illegal fishing, tanker and container movements, and reportedly smuggling. That is one reason imagery-based detection matters: it does not depend on the vessel's cooperation. It is also why accusations built on a single detection are risky, since a wrong identification can become a diplomatic dispute. For the geography of who can enforce what, see our ranking of the largest exclusive economic zones.

Detection data can also support conservation outside of enforcement. Mapped fishing effort and offshore infrastructure feed planning for protected areas, a link explored in our article on AI and biodiversity monitoring by satellite. The Nature study found that about 86 percent of fishing activity occurs in waters shallower than 200 meters, which are about 7 percent of the ocean, so protection planning can concentrate on a small share of the sea. It also found that wind turbines more than doubled in number since 2017 and that about 28,000 offshore structures existed by the end of 2021, a reminder that the ocean is not only a fishery.

What people get wrong about AI ocean monitoring

"Untracked means illegal." A vessel missing from public AIS data may be legally exempt, too small to carry a transponder, or broadcasting to authorities only. The 72 to 76 percent figure is about visibility, not about law-breaking.

"AI watches the whole ocean in real time." No dataset does. Radar coverage is partial, revisit intervals are measured in days, and the study with the broadest scope still covered about 15 percent of the sea, concentrated where activity is densest.

"A model's accuracy carries over to any region." Both the Hawaii whale detector and the global vessel classifier come with warnings that performance depends on similar conditions and on how labels were created.

"AI predicts coral bleaching weeks ahead." Heat-stress forecasting relies mainly on satellite temperature and the degree-heating-week rule. Learned models are a research direction, not the established tool.

"Hearing a whale protects it." Detection helps only if someone acts on it, as with slow zones for ships. Entanglement in fishing gear remains a leading cause of death for right whales, and no listening system removes a net.

The takeaway

The most useful thing neural networks do for the ocean is turn raw data nobody can read into counts and maps people can use: how many large vessels are out there, which of them are invisible to public tracking, and where a particular whale species is calling. They are best understood as a way to widen the area covered by a small number of expert eyes. They do not replace those eyes, they rarely see the smallest boats, and their accuracy figures hold only under the conditions where they were measured.

The harder questions are not about algorithms. They are about who owns the data, who pays to check an alert, and which authority is willing to act when the alert points to a ship in someone else's waters. Better detectors make those questions harder to avoid. The technology narrows what can hide in the ocean, and leaves what to do about it to people.

FAQ

Q: How accurate is vessel detection in satellite images?

A: In the 2024 Nature study, recall exceeded 95 percent for vessels longer than 50 meters and was about 80 percent for those between 25 and 50 meters, falling steeply below 25 meters. The authors estimated a false-positive rate of around 2 percent at most. A separate classifier labeled fishing versus non-fishing vessels with 90.5 percent test accuracy.

Q: Does an untracked vessel mean illegal fishing?

A: No. The study found that 72 to 76 percent of industrial fishing vessels were not publicly tracked, but many are legally exempt from carrying AIS or broadcast privately. Missing from public tracking data signals a gap in visibility, and the authors note that vessels engaged in illicit activity often switch AIS off.

Q: How do neural networks recognize whale calls?

A: They are trained on spectrograms, pictures of sound with time and frequency as axes, labeled by experts. Google and NOAA trained a ResNet-50 network on about 15 years of Pacific recordings and reached more than 90 percent precision and 90 percent recall on 75-second clips, with the caveat that test data resembled training data.

Q: Can AI predict coral bleaching?

A: Heat stress is already forecast from satellite temperatures using NOAA Coral Reef Watch's degree heating weeks, which flag stress when temperatures run at least 1 degree Celsius above the long-term monthly average. Machine learning could refine local forecasts, but it has to be tested against observed bleaching, and such events are rare.

Q: Is the data free to use?

A: Much of it is. The European Sentinel-1 and Sentinel-2 satellites have free and open data policies, Global Fishing Watch offers a public map, and the Allen Institute's Skylight is provided at no cost to conservation users. Commercial imagery is typically sold by subscription.

Q: Can AI find ghost nets?

A: Not reliably. Drifting nets are small, often submerged, and hard to see from space. Reviews describe detection as a needle-in-a-haystack problem, and recent efforts such as net-harvesting buoys and gear tracking do not depend on image recognition.

Q: Who acts on the alerts?

A: Flag states, coastal states, and port authorities, within limits set by law. The Port State Measures Agreement lets ports deny access to vessels suspected of illegal fishing, but it does not compel countries to police their own flagged ships.

Sources

• Paolo, F. S., et al. (2024). Satellite mapping reveals extensive industrial activity at sea. Nature.
• Agnew, D. J., et al. (2009). Estimating the worldwide extent of illegal fishing. PLoS ONE.
• Paolo, F., et al. (2022). xView3-SAR: Detecting dark fishing activity using synthetic aperture radar imagery. NeurIPS.
• Google Research (2018). Acoustic detection of humpback whales using a convolutional neural network.
• NOAA Fisheries. North Atlantic right whale.
• Argo Program. About Argo.
• Allen Institute for AI. Skylight.

---

Read next:

• AI-Powered Carbon Capture Technology — Can Machine Learning Scale Carbon Removal?
• Machine Learning for Climate Prediction — How AI Forecasts Climate Change
• Smart Cities and AI — How Machine Learning Reduces Urban Energy Waste