A jaguar disappearing into a rainforest can be almost impossible to follow. If a camera captures it again several weeks later, how can a scientist be sure that the animal is the same one? Researchers have developed an AI animal identification method that compares distinctive coat patterns, offering a faster way to recognise individual jaguars, zebras and other patterned animals without catching or physically marking them. It is a deceptively simple idea: let the animal’s own markings act like an identity card.
The research, published in Methods in Ecology and Evolution in June 2026, describes RAPID, a system designed to recognise an animal on relatively modest hardware. That matters because wildlife monitoring often takes place a long way from fast internet connections or powerful computers. Scientists may have photographs from drones and camera traps but still face the time-consuming job of deciding which individual appears in each image.
Why AI animal identification is harder than recognising a species
Knowing that a photograph shows a jaguar is only the beginning of many conservation questions. Researchers may need to know whether a particular jaguar has returned to a territory, whether the same female appears with cubs, or whether an individual has stopped appearing in an area. Without individual recognition, photographs of several animals can blur into a rough count that misses movement and changes over time.
Stripes and spots provide an unusually useful clue. Two zebras can look almost identical at a glance, yet their stripe arrangements differ. The same is true of the rosettes and markings on jaguars. A computer can describe those visual details mathematically and compare them with records from known animals. The problem is doing that accurately when the photograph is taken from another angle, under different light, or with part of the animal hidden by vegetation.
A zoo animal standing still beside a familiar wall is comparatively easy to photograph. A moving jaguar crossing a rainforest camera’s view offers a much less controlled scene. A drone may see only part of a zebra’s flank, while another picture may be blurred or show an animal turning. In a useful field system, speed is important, but a confidently incorrect match can be worse than an honest uncertain result because it may distort the history of an individual.
The research team’s University of Stuttgart explanation describes how RAPID begins with a reference collection of animals whose identities are already known. A new observation is compared with that collection. The output is not the machine inventing an identity from nowhere; it is an attempt to recognise a known individual from the pattern visible in the image. This requirement is easy to overlook in demonstrations of image recognition.
What the wild tests actually found
The researchers evaluated their method on several collections of images, including video of zebras in Kenya and camera-trap footage of jaguars in Ecuador. The university reports identification accuracy of 80 per cent on its new zebra dataset and 93 per cent on the new jaguar dataset. Across four public datasets, the reported range was 89 to 99 per cent. Those figures describe performance on the tested data and should not be interpreted as a guarantee that every wild animal can be identified in any photograph.
The comparison also involved speed. The team reported processing dozens of cropped animal images per second on a standard computer and fewer, but still useful, images per second on a more constrained device. A smaller system that can operate close to where images are collected may be more practical than a demanding model that must send every photograph to a distant server. The trade-off is familiar from everyday devices: computing power, battery life and available connections can be just as important as theoretical accuracy.
It is important to understand what was being processed. RAPID is built around images cropped to show a particular animal, not an unrestricted claim that a drone can film a vast savanna and instantly produce flawless identities for every creature. Finding an animal in a busy frame, selecting the right view and linking the result to an observation record can involve other parts of a monitoring system. The published method is a significant component, not a complete conservation programme.
What an animal’s markings can tell a conservation team
Suppose researchers repeatedly record a jaguar near the same forest path. Recognising the individual can help them build a timeline without touching the animal. If the jaguar later appears at a different monitoring site, the two observations can be connected. That could offer evidence about movement and territory use. It may also show whether an animal that appears injured in one image is seen moving normally at a later date. Each of those conclusions still requires careful interpretation of the photographs and the local ecology.
There are limits to any passport made of fur. A pattern has to be visible enough to compare with a trusted reference. A partly hidden animal or poor-quality image may not provide enough information. Animals without distinctive patterned coats may be unsuitable for this particular method. The researchers explicitly acknowledge that RAPID’s current approach does not provide the same solution for elephants. The tool’s name and examples should not be mistaken for a universal identification system for wildlife.
A growing reference collection is useful but also introduces questions about keeping records accurate. If one animal is mistakenly assigned another animal’s identity, later matches can compound the problem. Conservation projects therefore need processes for reviewing uncertain cases, recording where pictures were taken and correcting mistakes. A useful AI system should help experts organise evidence, not quietly convert guesses into permanent facts.
Faster counting is not the same as saving a species
Conservation decisions are shaped by information about habitat, movement, breeding, hunting and changing conditions. Better identification can make that information more timely and less expensive to gather. Yet software cannot protect a threatened animal simply by recognising it. The people managing protected areas still need resources, local knowledge, policies and the ability to respond when monitoring reveals a problem.
This is where the development fits with broader attempts to use AI in environmental observation. LiveAIWire has examined satellite mapping of small farms, a different example of turning large numbers of images into practical information about a changing landscape. Both applications benefit when the systems answer a specific field question rather than merely generating an impressive picture or score.
A field researcher may also want to avoid disturbing the animal. Camera traps can record visitors without requiring a human observer to approach each one, although camera placement and field practices still need to minimise disruption. Individual recognition from existing photographs could reduce some forms of repeated handling. It does not eliminate the ethical and practical questions surrounding wildlife monitoring, particularly when surveillance technology is deployed in sensitive environments.
Why open tools matter in the field
The team released RAPID as open-source software, making it possible for other groups to examine and adapt the method. That does not mean installing it is effortless. Researchers still need suitable reference imagery, equipment, storage, testing and people able to interpret the results. An open implementation does, however, give conservation organisations more opportunity to assess whether a tool works for their species, camera setup and budget.
The research also illustrates why smaller, specialised systems deserve attention alongside general-purpose AI. A model designed to recognise the markings of an animal can be valuable even if it cannot chat, write essays or perform unrelated tasks. In conservation, an accurately repeated identification may be much more useful than a sweeping promise of artificial intelligence. Technical restraint can be a strength when the job is narrow and the evidence can be checked.
There are other uses for patterns in nature. Scientists examining an animal over time may study changes in behaviour, interactions and survival, not just count photographs. The identity assigned to each observation is the thread connecting these events. Incorrect identity breaks that thread. A fast and testable matcher can improve the quality of a long-running dataset, while a poorly validated one could make the picture less reliable.
LiveAIWire previously covered AI-enabled orchard robotics, where perception is only one step before a machine interacts physically with the environment. RAPID tackles a different part of the problem: making sense of what has been seen without necessarily acting upon it. That distinction matters because environmental monitoring can improve decisions long before autonomous machines are trusted to intervene.
The next test happens outside the demonstration
The authors want to improve robustness when animals are partly obscured and extend recognition to species whose coats are less useful as identifiers. Whether they can do so while retaining speed on modest hardware remains an empirical question. A result that works in one park may need separate testing on another population or with a different camera. Serious deployment should report those local limitations rather than transferring laboratory figures unchanged.
A useful comparison is LiveAIWire’s coverage of AI methane detection from space. In both cases, images become evidence only when a system can identify something meaningful and people can check the result. The value is in a better record of the real world, not the spectacle of the algorithm itself.
The appealing image is still a simple one: a jaguar walks past a hidden camera, and a researcher can recognise it from the markings nature already supplied. The technology is not a substitute for conservation expertise. It is a tool that might make repeated encounters easier to connect, helping scientists understand which animals they are actually watching and what happens to them after they leave the frame.
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.