Monitoring an array of 225 cameras throughout Tanzania’s Serengeti National Park presented wildlife researchers with an enviable dilemma. They had captured millions of images of life on the African savanna over 18 months, including lions taking down zebras and bat-eared foxes chasing aardwolves. The time had come to classify the curious creatures in every photograph.
University of Oxford postdoc Alexandra Swanson and her colleagues found the solution in citizen science, tapping into a network of more than 28,000 registered volunteers who identified the animals in the photos with 96 percent accuracy. The resulting set of 322,653 wildlife photographs is now publicly available, thanks to the Snapshot Serengeti camera survey, a research project that has aggregated visual data on animal movements and spatial relationships across a survey area of 1,125 square kilometers.
Every image is catalogued according to the species it captures, along with information on location, time of day and observable behavior. Shot between 2010 and 2013, collectively they represent the first continuous systematic data on large predator and prey species in the Serengeti ecosystem across several years.
This study began when principal investigator Swanson—at the time a PhD candidate at the University of Minnesota—set up a network of camera traps to monitor how lions shared space with other Serengeti predators. “Based on the photographic evidence, we could see that hyenas and lions tended to follow each other in their search for food whereas species like the cheetah would almost always enter an area after the lions had left,” says Swanson, whose project is now part of Zooniverse, a suite of projects produced, maintained and developed by the Citizen Science Alliance and headquartered at Oxford's physics department. The researchers found these relationships were consistent across their survey area.
Swanson’s cameras operated continuously, 24/7, and when her team began checking memory cards they were inundated with images not just of predators but also of game herds, large birds and, occasionally, nothing. To manage the flood Swanson turned to her fellow doctoral candidate, Margaret Kosmala, to see whether she could automate image classification through a computer program. “I told her computer-vision research wasn’t there yet,” says Kosmala, whose undergraduate degree is in computer science. “But I was blown away by how attractive the images were. They were fun to look at, so I thought we could easily get 100 people to help us. That’s when we started thinking about how to develop a citizen science project.”
Swanson and Kosmala soon got help from Zooniverse to build snapshotserengeti.org, a digital interface for volunteers to help process camera-trap data. Within the first three days volunteers had gotten through all the photographs, and made over 1.3 million classifications. The site reached 10.8 million classifications and is still active today. Earlier this month the researchers published their results thus far on the open-access journal Scientific Data’s Web site. “This proves that citizen science is more than a tool for outreach and education, it’s a method for legitimate research,” Swanson says. “Our volunteers parsed through data that would have taken us years to unpack in a matter of daysand created a resource for further developments in both ecological and computer-vision research.” [Scientific American and Scientific Data are both part of Nature Publishing Group.]
Programmers need a large baseline data set with accurate information to establish rules for classification before they can automate computer recognition of visual information. This includes photos with correct identifications as well as properly classified images of moving foliage. By publishing their data under a creative commons license, the Snapshot Serengeti team hopes to provide that baseline for new programs and pave the way for citizen scientists to play bigger roles where the machines fall short. “We want to save human effort for the really hard problems,” Swanson says. “Ecologists are collecting more data than ever before and a lot of it goes unused, but if we can meet this need for tools to process data, it expands the range of questions we can ask [about an ecosystem].”
Sites such as Snapshot Serengeti can be expensive to build but Zooniverse plans to release an open-access platform soon that enables people to build their own citizen science Web sites based on the original source code. The platform is currently in beta testing, according to Swanson, and will be announced on Zooniverse.org when it goes live.
Meanwhile, the cameras are still rolling.