It is a fact of life for birders that it is difficult to differentiate certain species — notably, sparrow and drab songbirds are called “little brown work”. It is almost impossible to distinguish individuals. Now, photographs and a computer program to analyze the video has accomplished that feat. It is promised to reveal new information on bird behavior in advance.
“We spend a lot of time with binoculars, staring at birds and their feet,” says Ikon Levis, a behavioral expert from Kenya. Reason: Over the years, researchers have identified birds by placing colored bands on their feet. They use those bands to identify birds in the forest – and in photographs and videos in the laboratory. Work can often be laborious, Levine says.
Outfit tags in particular can make the work easier by incorporating GPS and proximity sensors, which interact with animals. Passive integrated transponder (PIT) tags are also used to ping the antennas within a few centimeters when a bird is attached to the ground and to identify pets. Behavioral ecologist Claire Dottlerant of the French national research agency CNRS and her colleagues have associated these small tags with the toes of sociable birds (Philatelic society) Since 2017.
In southern Africa, sociable weavers often work to build large nests in acacia trees. The nest can weigh up to 1 ton and can house up to 200 birds in individual chambers. Their cooperative practices also include chicken farming and defense against snakes and hawks. To study these behaviors, researchers identify and track hundreds of individual birds.
The antennas on the feeders keep track of which birds are living in the colony. But more granular information – such as birds contributing the most to communal activities – is not possible to achieve that path. And Doutrelant and his colleagues did not place antennas in the entire nest: birds are wary of them, and their chambers are too close to each other for reliable data collection.
Hence team member Andre Ferreira, a Ph.D. Students at the University of Montpellier, decided to try a kind of artificial intelligence. The device, called a convoluted neural network, is used to search through thousands of images to determine which visual characteristics can be used to classify a given image; It then uses that information to classify new images. Conversational neural networks have already been used to identify various plant and animal species in the wild, including 48 types of African animals. He has also achieved a more complex task for elephants and some primates: to distinguish between individuals of the same species.
Ferreira fed the neural network several thousand photographs of 30 sociable weavers that had already been tagged. “No one came up with an efficient method to collect these training data sets,” he says. To take photos, they placed cameras near bird feeders equipped with radio-frequency antennas. As the birds landed, a small computer recorded their identities using their PIT tag, and a camera photographed their backs every 2 seconds. (The rear view is the part of the most frequently seen bird while they are nesting or netted.)
After only 2 weeks, Ferreira had enough photographs to train the neural network. “We weren’t sure it would work,” Doutrelant recalls. “We’ve seen these birds a lot, and we’ve never managed to recognize them without rings of color.” But when the photos given were not previously seen, the neural network correctly identified individual birds 90% of the time, they report in this week Methods in Ecology and Development. Doutrelant says that it is about the same accuracy as humans trying to color rings with binoculars.
Ferreira then tried the approach on two other bird species studied by Damian Farren, a behavioral ecologist at the Max Planck Institute of Animal Behavior. The tool was accurate in identifying the zebra finch in captivity and the great breast in the wild. Both species are widely studied by ecologists.
But Gail Petricli, a behavioral ecologist at Davis’ University of California, sees some limitations to the approach. For example, with species that are difficult to capture and tag, it can be difficult to obtain the thousands of identifiable photographs needed to train neural networks. She studies more and more sage grouse, a species, in the fall, and she tries to avoid handling them because it emphasizes birds. Another possible limitation: When birds molt, the neural network cannot detect them and will need to retreat. Ferraria is collecting photos of other symptoms such as the head aspect to improve the tool.
The biggest limitation with current neural networks, Ferreira says, is that it tries to identify every bird that it already knows, so it cannot recognize a new person. Ferreira is now working with Farine to try a different kind of neural network that could do this – it would need to be trained on many more bird paintings. If the data set was large enough, the tool could also be used by researchers who have not tagged their birds. “I think it will be a complete game changer,” Farin says.
Despite those limitations, Patricy called the new work “exciting”, and it opens up possibilities for the study of many other bird species and behaviors. “The fact that this algorithm was able to tell them apart – when they look similar to the naked eye – is certainly striking.”