Catherine McGrath

Photo Credit: Jade Arneson/Macaulay Library at the Cornell Lab

About a century ago, East Coast barred owls started moving west. Ecologists say these fierce predators were previously constrained by the Great Plains’ vast expanses, which largely prevented them from traveling through uninterrupted forest as they prefer. But by the 1950s, a few owls had found their way north into the vast forests of British Columbia. From there, they moved south into Idaho, Montana, Washington, and Oregon. An ecological success story perhaps, but only for barred owls.

Barred owls, larger and more aggressive than many other owl species, tend to force out their smaller relatives, including the three subspecies of spotted owls. In the Pacific Northwest they have decimated populations of the northern spotted owl, which is listed as threatened under the Endangered Species Act. And now they are moving south, endangering the California spotted owl, which is not federally protected. This owl lives exclusively in the old-growth forests of the Sierra Nevada mountains and is already at risk because of logging, drought, climate change, and wildfires. Like their cousins to the north, they are proving no match for the invaders: barred owls outcompete them for prey and territory. Between 1995 and 2017, California spotted owl populations declined between 65 and 85 percent, according to researchers.

But the barred owls may have met their match: a team of biologists armed with AI.

Connor Wood, a research associate at Cornell University’s K. Lisa Yang Center for Conservation Bioacoustics, first spent time in the Sierra Nevada in the summer of 2011. The mountain range lies between the Central Valley of California and the Great Basin. It runs 400 miles from north to south and 50 miles from east to west. Much of the Sierra Nevada is considered wilderness; it contains three national parks and ten national forests. It is also home to the highest point in the contiguous United States, Mt. Whitney. “It was very different from anywhere I’d been before,” said Wood, who grew up in Illinois. He immediately fell in love with the area’s scenic beauty.

A few years later, Wood became a research assistant for the University of Wisconsin and was tasked with designing region-wide monitoring plans for the California spotted owls and barred owls in the Sierra Nevada. It was becoming clear that the barred owls might drive the California spotted owl to extinction, and many observers knew that might mean an ecosystem-wide problem. Spotted owls are what is called an umbrella species, which means protecting them benefits other species in the same area: small rodents like squirrels and chipmunks and other birds like the western tanager and evening grosbeak.

Wood and his colleagues decided that the first step to helping the California spotted owl was to determine exactly how many barred owls there were in the region. They began to collect hours of audio, analyzing it for barred owl hoots— “who cooks for you”—and trying to identify individual birds within the recordings. This approach is part of the field of bioacoustics: the study of animal vocalizations and communication as a means of gathering information about wildlife behavior, population density, and ecosystem diversity. It was all new to Wood and he wasn’t sure it would work.

Wood and his three colleagues deployed two or three recorders for five to seven nights at a time in 167 sites throughout the northern part of the Sierra Nevada, where the owls entered the mountain range from the Pacific Northwest and began migrating south. They gathered 49,800 hours of audio in 2017, and 145,600 hours in 2018. Using Raven Pro, a software that allows ecologists to visualize data, the team teased out each barred owl hoot. They determined that the population of barred owls in the 167 sites increased by 163 percent during the two years of their study; to ecologists this suggested a low rate, and they concluded that the barred owl invasion was still in the beginning phases. They had caught it early.

Wood and his colleagues recommended that barred owls in the northern Sierra Nevada be shot. “You, ultimately, in the long run will have either spotted owls or barred owls. You will not have both,” Wood said. “And there’s probably hundreds of thousands of barred owls throughout eastern North America and there’s no spotted owls. The only place to find spotted owls is in their native range in the western U.S. and into the Baja Peninsula of Mexico.” Wood, generally an owl enthusiast, said that this was not an easy decision. Using a mid-sized gauge shotgun, two of Wood’s colleagues killed just under 100 barred owls between 2018 and 2020. The following year, barred owls returned to only six of the twenty-seven sites they had been removed from and the California spotted owls returned to 15 of those sites. At five of them, pairs of spotted owls were detected nesting.

It was a “rare conservation success story,” as Wood put it in a website post. But barred owls were still plentiful in other areas of the Sierra Nevada—in fact, Wood and his colleagues had only canvassed 65% of national forest land in just the northern part of the mountain range—in other words, just a tiny slice of the whole. The U.S. Forest Service then asked the team to take on the whole area, some 39,612 square miles. It was not do-able, not scalable, Wood thought. The technology he had used had produced numerous false positive detections that had slowed the analysis process significantly. “I had real concerns about our ability to do this on a big scale because there was going to be just so much analytical time during detector review,” Wood said.

For decades people have warned about the dangers of artificial intelligence, neural networks that mimic the human brain. But the recent advent of ChatGPT and Midjourney seem to have ushered in an era of unprecedented panic about AI. People fear it will replace actors, architects, lawyers, writers, musicians and on-screen news anchors. Even some baristas have been replaced by robots. It seems to many the most unnatural of tools.

But in the world of conservation, AI does not always have these dark connotations. It can be seen as a useful tool to detect patterns that emerge from a tsunami of data. AI is not replacing ecologists, but rather, many experts say, helping them become more effective, helping them track whales and wolves, monitor poachers in Africa and identify individual animals, like those in a population of brown bears. Wood thinks it now could be used to save the California spotted owl.

While Wood was spending hours analyzing invasive barred owl calls out west, a colleague back in Ithaca, New York, was developing technology that would transform the Sierra Nevada project.  BirdNET is one of the most recent AI tools to come out of bioacoustics. The field began in the early 1900s when the advent of audio equipment made it possible to record, analyze and store clips of sound. Despite the clunky nature of parabolic recorders, hydrophones and storage discs, ecologists began to collect vocalizations of birds, insects and marine mammals.

In the late 1900s, magnetic tape recording revolutionized the way scientists could record in the field. These new devices were portable, handheld systems that allowed researchers to review audio instantaneously. Vast numbers of recordings were stored in various archives, such as the Library of Natural Sounds at Cornell, later renamed the Macaulay Library, which houses two million hours of bird recordings, all of which can now be accessed through a public web portal called eBird.

The Macaulay Library also houses a museum of recording equipment, where a visitor can see the evolution of bioacoustic technology arrayed across one wall. The technology is arranged in chronological order. On one end, clunky recorders from the early 1990s resemble briefcases in size, shape and functionality. Moving left, the technology gradually becomes smaller, until the display ends with portable hand-held recorders the size of a large wallet.

At the start of the twenty-first century, acoustic recording units exponentially increased the speed of data collection. These recorders can be left in the field for extended periods due to their large data storage capacity and long battery life. Cornell developed its own version, called Swift One. “Where in the past for this field, researchers, including myself, we’re going out with our handheld recorders, and following birds around,” Leah Crenshaw, a graduate student at Cornell, said. With Swift One “we can take these microphone units and tie them to a tree and leave them out in the field recording for several weeks at a time.” Wood used Swift One recorders in the Sierra Nevada.

Bioacoustics has made numerous contributions to ecology. Analysis of whale songs has led to discoveries about whale communication and behavior and, recently, about how wind turbines and ships interfere with cetacean communication. Scientists have found that bats adjust their echolocation to environmental challenges such as the presence of other bats or urban acoustics, and that bat colonies have complex social interactions. Ecologists in the 1960s discovered that two types of cicadas emerge periodically and synchronize. In the same decade, scientists at the University of Michigan discovered that the sound of bees can cause nearby flowers to fill with nectar.

Cornell’s Lab of Ornithology has been a leader in avian bioacoustics. Arthur Allen, the founder of the lab, helped shape the field. Researchers have developed land manager guides, an app for birders to identify the sounds of birds and made discoveries about bird migration patterns, climate change and genetics. BirdNET is the lab’s most recent contribution to the field.

BirdNET was created by Stefan Kahl, a computer scientist from Germany. Kahl has a background in image processing and in 2016 developed software to monitor assisted-living facilities. His idea was that his algorithm could be trained to listen for residents who needed assistance. But he soon ran into a hurdle, one common to the field of AI.  Before any AI model can be used, it must go through a training phase in which it is fed hundreds or thousands of data points, what is called the training set. Kahl realized that there wasn’t enough training data on ambient living—very few recorded sounds of someone falling or calling for help existed—and so he couldn’t sufficiently train his model.

On a visit to Germany to visit a colleague, Holger Klinck, the director of the K. Lisa Yang Center for Conservation Bioacoustics, met Kahl and learned about his project. Klinck immediately saw a connection between Kahl’s algorithm and the work being done at the Yang Center. Klinck thought Kahl’s algorithm might be useful for bioacoustics, so he invited him to come to Cornell to see.

At first, Kahl tested the algorithm on a dataset of seventy-five birds. He knew nothing about birds. It quickly became clear that a shortage of data would not be a problem. “Within like one or two weeks, we noticed, oh, this works, this actually works, because birds do have a decent amount of training data and the training data is from the actual, real world,” Kahl said. Fascinated by the intricacies of avian acoustics, he decided to do his dissertation on birdsong recognition. “There’s finally like this acoustic domain that has a sufficient amount of training data for these data-hungry algorithms that we’ve been using,” he said.

In 2017, Kahl competed in the BirdCLEF competition, an annual challenge open to the public in which participants are tasked with developing a machine learning algorithm to identify all the bird calls in a recording specially made for the contest. That year, the competitors were tasked with identifying 1,500 bird species in a 6.5-hour sample. Kahl’s team won an award as one of the best-performing systems.

BirdNET works quickly and accurately. Each piece of audio is fed into software and transformed into a spectrogram, a visual representation of audio on a graph charting frequency against time. Within these spectrograms, each chirp, trill or song is displayed visually with distinct characteristics. For instance, the barred owl “who cooks for you” hoot looks like two or more isolated silos increasing over time and culminating in a crescendo and subsequent fade out. Since the algorithm has been trained on various bird calls—and on audio samples with no calls at all, negative samples that teach the algorithm what not to look for—it compares the newly uploaded spectrogram with ones it has “heard” before. It analyzes it and provides the user with probabilities and a final prediction.

Kahl explains it like this. Imagine there is an unreleased Beatles song discovered in a studio. It’s never been played and never been aired. You’ve never heard it before but once it begins to play, you know it’s the Beatles. There is something about the voices and the music that gives away that it’s a Beatles song. “And that’s how I like to explain this pattern: there’s somewhere hidden in a spectrogram, not necessarily human recognizable…What these embeddings are, they just are like a condensed version of all these features, like the pattern condensed into a sequence of numbers, which is not human interpretable anymore,” he said. Each sequence contains 1,024 numbers.

After BirdNET’s success at the competition, Kahl and his team began to increase the number of species in the dataset that the algorithm is trained on, from 1,000 to 2,000 to 6,000 species. Among those species are the barred owl, the California spotted owl, the northern spotted owl, and the Mexican spotted owl.

As well as non-avian species. “Now it’s not even just birds, now it’s everything,” Kahl said. Wood, Kahl and colleagues have used BirdNET to successfully detect wolves and coyotes in the Sierra Nevada. Other scientists are using BirdNET too: to identify gibbons in Indonesia, seals in Japan, and lions in India. A project in Norway uses BirdNET to monitor biodiversity across the entire nation in real-time. It is even helping to determine how snowmobile noise affects bird vocalizations during the winter in Yellowstone National Park. Cathleen Balantic, a biologist with the National Parks Service, said she uses BirdNET to identify the presence of birds and frogs. “BirdNET’s wood frog model is really strong,” said Balantic, adding that wood frogs signal the start of spring and that is interesting to track because climate change means spring arrives earlier. “I’ve never been happier for something to make previous aspects of my job obsolete,” she said.

But BirdNET doesn’t work for all species. Elephant vocalizations have presented a challenge, as very low-frequency sounds sometimes do. “It doesn’t work for elephant rumbles for whatever reason,” Kahl said. Marine mammals are hard to identify as well. 

It doesn’t even work for all bird species. Ingrid Molina, project coordinator for Cornell’s Our Coffee Our Birds program in Costa Rica and Colombia, studies the effects of coffee farm agrochemicals on birds. To do this, she deploys numerous acoustic recording units before and after the chemicals are applied. Her studies produce overwhelming amounts of audio, yet she chooses not to use BirdNET. She has noticed that it sometimes misidentifies the sounds of tropical birds. “Any piece of technology developed in the North…we have to prove it in the tropics,” she said. If a species is very common and has a distinct call, BirdNET is trustworthy. If a species is migratory or only found in Costa Rica, for example, the algorithm runs into trouble.

This limitation might be explained by the training data. “There’s this weird distribution of where we actually have bird data from,” said Tom Denton, a computer scientist at Google who works on bioacoustics. There is a ton of bird call data from North America and Europe, but very little data from Sub-Saharan Africa and South America. He also said that the different environments might contribute to decreased performance of the software: “How many birds are vocalizing at a time? That can have a big impact. And maybe if you’ve got more rain in the background. That might make it harder to detect things as well.”

Another scientist working with BirdNET has experienced similar limitations in North America. Katharine McGinn, a graduate student studying at the University of Wisconsin-Madison, found that BirdNET mistakes the hoot of a saw-whet owl with the beep of a truck backing up. The two sounds manifest on spectrograms as a simple, vertical line. “The important part of that modeling process is the ability to predict and understand how much error a machine learning algorithm like BirdNET has,” McGinn said.

BirdNET is not the only machine learning tool for ecologists. The Bear ID project led by Melanie Clapham, Ed Miller and Mary Nguyen monitors bears for conservation using photographs from camera traps and deep learning to identify the faces of individual bears in British Columbia. At the moment, it can identify 160 different bears. Justin Kitzes’s lab at the University of Pittsburgh is developing, training and experimenting with machine learning for bird detection. “You’ve got 100,000 hours of audio and there may only be 60 seconds of that sound in there. The question is how to find it. Deep learning broke this open,” Kitzes said. His lab studies how bird species are distributed across landscapes and how human activity impacts this distribution.

Birds all over the world are threatened by habitat loss, invasive species and collisions with glass and other structures. In North America, according to a 2019 study published in Science, three billion birds have been lost since 1970. The 2022 “State of the Birds” report released by the North American Bird Conservation Initiative identified 70 bird species that are at a “tipping point,” meaning the species’ population could decline 50% or more in the next 50 years, based on recent assessments. Eighty-nine North American bird species are listed as either threatened or endangered—including not only the northern spotted owl, but its southernmost relative, the Mexican spotted owl. Recently, the U.S. Fish and Wildlife Service proposed that the California spotted owl population living in the Sierra Nevada be listed in its own right as threatened and that those living in southern coastal areas of California be listed as endangered.

Wood recently used BirdNET to estimate the populations of California spotted owls and barred owls across the entire Sierra Nevada. From May to July of 2021, Wood and 17 others deployed 1,651 acoustic recording units at 851 sites throughout the mountain range. They collected 557,000 hours of audio, equivalent to 63.5 years of time. BirdNET identified between 2,218 and 2,328 California spotted owls, but few barred owls. This suggested to Wood that the eradication of barred owls in the north two years prior was successful in arresting their spread through the entire mountain range. Wood and his colleagues concluded that removing the pressure to outcompete barred owls equipped the California spotted owl with more resilience to face other environmental pressures like wildfires and excessive logging.

“Spotted owls are not…they haven’t died or really gone away. They’ve left these really preferred areas and they’re kind of just hanging out in adjacent, sort of low-quality habitat hoping to get back in,” Wood said. “BirdNET revolutionized how we can do bioacoustics. I don’t think that’s an overstatement.”