Tracking wildlife migration has been historically difficult in the rugged terrain of Alaska. Researchers primarily rely on either surveys or GPS tracking to understand bird migration patterns. Both methods are expensive, either in terms of time or money. And the trackers are often too large or heavy.
One way to sidestep these common issues is to record audio from frequently used nesting grounds. Using birdsong allows researchers to unobtrusively study the animals, although there's a downside. Each day produces a flood of audio recordings from multiple microphones placed around nesting grounds. It takes trained listeners endless hours to search the noisy soundscape for birdsong.
In a recently published paper in the journal Science Advances, U.S. researchers explain how they got around these tracking troubles. Columbia University ecologist Ruth Oliver and her fellow collaborators replaced the human ears with machine learning algorithms to listen to birdsong.
Oliver told VOA News, "Arrival times of migratory song birds is really important for their reproductive success. And obviously sending people to the Arctic to do field work is very expensive and takes a lot of time" — hence, the scientists' interest in creating an automated method for tracking bird species.
Oliver and her colleagues focused on migratory songbirds who fly to northern Alaska during their mating season. These birds tend to chirp more frequently as soon as they reach the breeding grounds to attract a mate. Spring is short in Alaska and the birds must breed and hatch their clutch before winter.
The team of researchers recorded the springtime soundscape of northern Alaska for five sequential years. They placed microphones at four sites in the foothills of the Brooks Range, which recorded 1,200 audio hours.
However, Oliver admitted the recordings weren't always perfect. "There's a lot of other noise in these recordings" Oliver said. "Even in May in northern Alaska there's lots of wind, lots of rain, and all of that is confounding when you're listening to birds."
The scientists fed hours of audio into two types of machine learning algorithms — one that used human expertise to help train it and one that relied solely on the collected audio. Both algorithms were based on the same model that's used by applications like Siri and Alexa.
Oliver told VOA that in creating the human-supervised algorithm, she "wrote a little program to randomly sample about 1 percent of the data set" and then listened to 4-second clips. She scored these clips as either containing or not containing songbird vocalizations and then fed this information into the program.
Both algorithms were fairly accurate at estimating when the avian commuters arrived in the foothills. The models showed the importance of snowmelt for the arrival of the traveling birds. The human-trained model was slightly better at recognizing the relationship between weather conditions and bird calls, although neither model specifically tracked individual species.
This technique has great potential according to Emily Jo Williams, vice president of migratory birds and habitat at the American Bird Conservancy, "This kind of technique that allows you to survey populations in those remote areas is really exciting and could allow us to even discover new places where protection and conservation efforts are needed," she said.
This study looked at nesting grounds near the Alaskan Arctic Refuge, which is a summer home for birds from nearly every continent. For example, the Northern Wheatear travels approximately 21,000 kilometers (13,000 miles) from Africa to summer in the refuge.
Williams told VOA, "We know from some research that some birds' ranges have actually changed, and they've moved in response to what we think is a warming climate." She went on to explain that "the timing of that migration has evolved over eons, and in large part it's relative to what food sources are available over a particular time, what weather patterns are or aren't favorable. So you could end up with bird migration out of sync with insect hatches or the phenology of plants that birds have a relationship to."
Tools like the algorithm created in this study could be used to track how migratory patterns of many species may shift in response to climate change. Using machine learning is a new way to follow these shifting patterns in birds, insects and other animals.