Deepfake whales could be a key conservation tool

By scrolling to social media, you may have killed on the coils of Leonardo DiCaprio Dancing or Tom Cruise Crooning, only to realize that they are created with artificial intelligence. Hyper realistic videos and images like these – also called Deepfakes – are notorious for celebrity pranks. But technology also has serious scientific applications. In the field of ecology, for example, Doppelgängers of the AI of rare species could improve efforts to understand, monitor and protect them.

More specifically, the Deepfake fauna could help form AI models to detect fauna in the images of satellites, planes and drones. Ecologists are counting more and more on such images as the crow flies to study the behavior of species and the trends of the population.

“We are really in the Big Data era regarding remote sensing in ecology and conservation,” explains Dave Johnston, director of Marine Robotics and the remote sensing laboratory of the Nicholas Environment School in Duke. “Over the past two decades, our ability to collect high -resolution remote control images has increased exponentially, largely due to the progress of drone technology and an increase in satellite capacities.”

Increase in data

Traditionally, the researchers had to use their own eyes to browse satellite and air images for target species. Now AI detection tools can speed up the process. The key lies in the data used to train IT models. The models must “see” a lot, many realistic examples of a species to know what to look for in images in the field.

For a common fauna, abundant images exist, so the assembly of training data is quite easy. But images are often limited for rare species, which mix in their environment or which live in inaccessible areas, such as war areas.

“One of the major challenges of ecology is the idea of data shortage,” explains Henry Sun, a graduate in 2025 of the Nicholas school which is double origin in biology and marine sciences and conservation, with a minor in computer science. “For a species where you only have several hundred individuals, you are just not going to have images that can be made different enough to be able to form a good AI detection model.”

What is an ecologist to do? A promising option is to strengthen rare training data with data generated by AI or synthetic, petrol, Deepfakes. This approach, called increased data, could allow new ecological information, according to a recent article in Nature.

Sun, a former Rachel Carson stock market from the Nicholas school and a undergraduate scholarship holder on the space of Caroline du Nord, recently studied the theme of the increase in data for his senior thesis, which he plans to publish. More specifically, he explored if AI could produce images sufficiently realistic to complete images of drones of the right whale from the North Atlantic in critical danger, whose population has decreased to less than 400 people. Theoretically, synthetic data could be used to help form other AI tools to detect the North Atlantic’s right whales in real aerial images.

Sun’s research has been inspired by greater collaboration between the Nicholas school and several Canadian organizations – including the Canadian Space Agency, Fisheries and Oceans Canada, the New Brunswick University and the Hatfield Consultants environmental group – to build a space detection system for the North Atlantic population.

“There are a lot of ocean, and despite the fact that the whales come back, there is still a very small number compared to the area you have to look for,” explains Johnston, who was thesis advisor to Sun. “And which means that we need very effective tools to find them. But it also means that we do not often have very good data archives to train these models to identify them.”

To create Deepfake whales, Sun and his team used broadcasting models, which generate images in response to prompts in the form of descriptive text, an exemplary image or both. Although other researchers have explored means of increasing the data to be used in whale detection, the team says that it is the first to use diffusion models for this purpose.

Researchers have used several dissemination models available in the trade that are pre-formed on Internet data trains. In other words, these basic models, as we have known, are ready to produce a variety of images in response to prompts.

Sun and his team experienced several methods of image generation, first using text prompts, then image prompts, and finally a method called fine-tuning, which included text and image prompts. Fine setting is a way to improve the performance of a basic model for a specific task – in this case, producing photos of high -resolution hyper -resolution whales – by leading it more on a smaller and more specific data set.

“Sometimes the diffusion model produces images of anatomically distorted whales, such as whales that are joint or whales with several fins, which shows that it has not exactly learned the most precise representation,” explains Sun. The fine adjustment can teach the model to avoid these errors.

Test credibility

All in all, the team has created hundreds of air images of whales to the right of the North Atlantic and, as a comparison, hundreds of aerial images of humpback whales. Because much more real images of the bump are available for the formation of the generative AI, the team has hypothesized that their models would produce more realistic synthetic hump images.

The last step was to test the veracity of their deep whales. Were they credible? To answer this question, researchers have nourished their photos in a Google tool called Reverse Image Search, which analyzes an entry image, Internet research similar images and produces results. In this case, the objective was to see if Google could recognize the whales represented in the synthetic data and to return the images of the same species.

In the false photos produced by text or image prompts, Google has confused many whales to the right of the North Atlantic for bumps. On the other hand, it correctly identified the two species of whales in almost all the images produced by a fine adjustment.

The team also found that the images of the right whales of the North Atlantic created by the fine adjustment were more precise than those generated with text or image prompts.

The next phase of research is to determine whether synthetic whale imagery can complete the training data for AI detection models. As a starting point, Sun recruited the first cycle of Duke Max Niu to start the basic tests.

“Max has formed deep learning models using both real images and some of the false images I made,” explains Sun. The idea is “to see if there is a proportion of false images that will benefit the model”.

Line

This fall, Sun will continue his studies as a doctorate. Student at Duke Marine Lab, working in the Juliet Wong laboratory. Although he plans to pay his attention from whales to sea urchins, he is committed to helping to demystify AI for researchers.

“Something that interests me extremely is capacity building for natural scientists in the field of artificial intelligence, because I am thinking more and more, they are skills that everyone needs,” says Sun. To this end, Sun hopes to plan awareness events linked to AI, such as the one hour session he organized during the Duke Oceans week last March on the use of AI in the Ocean Science.

While more and more ecologists are turning to AI, however, ethical considerations will become more urgent, explains Holly Houliston, a doctorate. Student at the British Antarctic Survey and the University of Cambridge, who helped supervise Sun’s work as a guest researcher at the Marine Lab. According to Houliston, the

“You must be really clear about the ecological question you are trying to answer. For example, if you look at calves – so baby whales – you may want to generate more images from these younger animals because you probably have only a few.” The use of dissemination and generative AI models in general has environmental impacts. Studies like this can help environmentalists understand how to use them in a responsible manner. “

As Johnston notes, “this intersection between IT and environmental sciences will only grow.”