To Protect Rhinos, Researchers Turn to New AI Tech

Faked imagery has a complicated history: grifters selling forged Picassos; Soviets vanishing images of Leon Trotsky from their propaganda photos; recent deepfakes putting Taylor Swift in compromising positions.

But in an illustration that not all such manipulation is done with ill intent, University of Maryland researchers are making artificial intelligence (AI)-generated images of rhinoceroses in hopes of saving the real ones.

Rhino populations have been in peril for half a century or more due to poaching of their horns (an ingredient in folk medicines). To protect the massive mammals, scientists must tally them with accuracy—a surprisingly difficult pursuit. Many are inaccessible inside private reserves; drone-mounted cameras, meanwhile, can’t cover their sweeping habitats and sometimes duplicate counts as the beasts thunder about. Satellites cover wider expanses, but their grainy images often fail to distinguish rhinos from elephants—or trees or rocks, for that matter.

Now, researchers with UMD and Princeton University have developed a rhino-detection system pairing satellite data with AI to boost conservation efforts, outlined in a study published last month in Scientific Reports. By “showing” a computer numerous images of verified white rhinos in a South African reserve, the model learns enough about their telltale shapes and sizes to identify them in pixelated images captured from hundreds of miles above. 

Analyzing satellite imagery with AI is not a novel concept, but the new study goes much further. Traditional machine learning requires up to hundreds of thousands of sample images to be scrutinized by human “annotators”—often low-wage workers in the developing world toiling for large U.S. companies—who spend months drawing thick boxes around the objects of interest on millions of images. The resulting training data for machine learning systems can cost hundreds of thousands of dollars. 

The UMD-Princeton system eliminates time, cost and labor by feeding just a small sampling of annotated images into an AI model. Using gaming software, the researchers manipulated them into “fake” 3D representations—animated rhino prototypes—to teach the computer to pinpoint the 2-ton animals in their habitats.

“We train an AI model on the fake image, and it works on the real one,” explained Xiaomin Lin, an electrical engineering professor at the University of South Florida who co-led the study as a UMD Ph.D. student.

This “synthetic” data animation can identify rhinos in satellite images with 65% accuracy. There’s room for improvement, Lin said, “but the results are exciting because there was nothing there before.”

The breakthrough promises to be a boon to conservationists who note the endangered creatures are vital to ecosystems. International preservation groups are sending billions of dollars to countries like South Africa to address the issue. 

“If you’ve got a rhino population on the brink because you don’t have enough breeding females, but you don’t have accurate data to understand the population demographics, it’s going to be problematic,” said Isla Duporge, a Princeton postdoctoral researcher in ecology and evolutionary biology who co-led the study.

[AI-Generated Iceberg Models Could Help Sailors Navigate Arctic Waters]

Lin originally developed synthetic data techniques so aerial drones could track down water drones lost at sea during other experiments, then applied it to satellite images of whales splashing above the water. Duporge, who as a Ph.D. student used satellite images to detect elephants without the aid of synthetic data, conceived of the rhino project after reading a 2023 paper on his whale findings.

Duporge already had satellite images covering thousands of hectares of South Africa’s largest private rhino reserve and wondered if Lin could apply his ocean-based algorithms to land animals. 

Co-written by UMD computer science and math double-degree student Aadi Palnitkar ‘26 and physics and computer science double-degree student Arjun Suresh '26, both research assistants in the Maryland Robotics Center, the study is the first to pair satellite imagery with synthetic imagery tech to detect rhinos—and perhaps any land mammal—Lin and Duporge said.

The research builds off technology developed by UMD computer science Professor Yiannis Aloimonos, Lin’s Ph.D. mentor and coauthor of the study, who previously used synthetic animation to count screws on tables and oysters on sea bottoms.

In recent years, computer scientists have adopted “anti-annotation” approaches to curb the exploitative labor in impoverished countries, Aloimonos said.

“This is the dark side of AI,” he said. “With synthetic data, you don’t have to do that anymore. This is for the social good.”