BioEd Online

Lucy, the famous Australopithecus afarensis skeleton, was found by accident when palaeoanthropologist Donald Johanson took a detour back to his Land Rover in Ethiopia in 1974. Such luck will always have a place in fossil hunting, but artificial intelligence now promises to assist, after a team trained a computer neural network to recognize fossil sites in satellite images.

The network, described in a paper in Evolutionary Anthropology¹, independently identified several places from which palaeontologists had unearthed mammal fossils, and researchers are now set to use its predictions to explore further sites in the Great Divide Basin in Wyoming.

"The plan is to ground-truth this in July 2012," says the study's lead author, Bob Anemone, a palaeontologist at Western Michigan University in Kalamazoo. Since the 1990s, Anemone has been scouring the Great Divide Basin for fossils of mammals from the early Eocene epoch, about 50 million years ago. "We're going to go to some areas we've never been to, that we wouldn't have been aware of, and see what we find," he says.

Dated methods

Modern palaeontologists tend to employ much the same fossil-hunting strategy as their nineteenth-century forebears: read the literature to see where other people have found specimens, scour geologic and topographic maps for exposed rocks of a particular age and then meander around these places, eyes fixed on the ground.

"The role of luck in vertebrate palaeontology is legendary," says Anemone. "People tell you, 'I was out taking a piss one day and found a fossil.' Everybody recognizes that it's kind of a crapshoot."

Some fossil hunters have turned to satellite-imaging tools such as Google Earth to focus their searches. Beginning in the late 1980s, Tim White, a palaeoanthropologist from the University of California, Berkeley, and his team used images captured by the space shuttle to identify parts of Ethiopia worth exploring on foot; one of the sites produced Australopithecine teeth nearly 4 million years old².

Another Berkeley group, Leslea Hlusko and her team, has used high-resolution satellite images to find 28 sites containing bones or archaeological artefacts in Tanzania³.

But these approaches still involve hunting through reams of images by eye, relying on gut instinct to locate promising search areas.

In search of a less haphazard means of exploring the Great Divide Basin, which covers thousands of square kilometres, Anemone's team turned to software.

Patterns of pixels

Neural networks learn to spot patterns in known data sets, and can use these patterns to make predictions about other data. They are used in applications including image-recognition software, robotics and e-mail spam filtering.

To train his network to hunt for fossils, Anemone took satellite images of the Great Divide Basin and assigned pixels in six bands of light wavelengths, including infrared, to different kinds of terrain. He also marked whether the pixel represented a fossil site or not.

By comparing the attributes of 'fossil' and 'non-fossil' pixels, the network learned to accurately distinguish fossil sites — typically covering hundreds of square metres and found around eroded sandstone — from other kinds of terrain, such as forest, scrubland and wetland. The researchers then set the network loose on satellite images from the same area that it hadn't seen before.

In the unfamiliar images, the model correctly identified 79% of the pixels that were known to represent fossil sites. Of the pixels that it flagged, 99% held fossils.

Once they had trained the network on the Great Divide Basin, the researchers gave it images from a different location: the nearby Bison Basin, which is made up of older rocks. They compared the results against fossil-location data provided by Christopher Beard, a palaeontologist at the Carnegie Museum of Natural History in Pittsburgh, Pennsylvania. The computer correctly identified four fossil sites, including one that Beard did not tell Anemone's team about until after the study was complete.

A targeted search

Such a tool could be invaluable for palaeontologists heading to previously unexplored areas, says Anemone. In theory, it could be used anywhere, as long as it was first trained using satellite images from a geologically similar place, he adds.

Glenn Conroy, a palaeoanthropologist at Washington University in St Louis, Missouri, and a co-author of the neural-network paper, is currently using the approach to look for caves that might contain ancient human fossils in the Cradle of Humankind, near Johannesburg, South Africa.

As researchers identify more regions to explore, "these sorts of approaches will become more and more important, because they will allow us to target our searches better" without wasting grant money, says Peter Ungar, a palaeoanthropologist at the University of Arkansas in Fayetteville.

But he doubts that scientists will hand over all the responsibility for site location to a computerized black box, suggesting instead that they will use these sophisticated approaches to guide their own hunches. "You're never going to lose the gut feeling," he says.