����APP

On shore, the landing sounds straightforward enough. When we reach Green Rock on this blustery late-September morning, Jonathan Felis will nose the Zodiac up to the island, bump it lightly, and the person poised at the bow will jump onto terra firma. “It’s all about the timing,” he tells us. Bump. And jump.

After cutting four miles north through waves pounding the Northern California coast, we see the island’s formidable cliffs. Suddenly the bump-and-jump notion seems downright daunting. The inflatable boat heaves, plummeting six feet below the landing spot—a slippery, barnacle-covered surface supposedly ideal to launch oneself onto from this puny, lurching vessel. Emma Kelsey doesn’t look fazed; on an upswell she leaps with the powerful grace of a gymnast and sticks the dismount. I jump next, clumsily grasp some sharp protrusions, and haul myself to my feet. Photographer Jim McAuley, gear strapped to his back, goes last. He hits low on the rock and slowly slides down, unable to get purchase. He slips into the water.

While I fight the urge to laugh, U.S. Geological Survey scientists Felis and Kelsey respond professionally. This is, after all, a government operation, we’re all wearing flotation jackets, and the pair had prepared us for this situation—and worse (“If everyone goes overboard...”). As McAuley bobs, Kelsey shouts instructions. Seconds later Felis plucks him out of the Pacific Ocean, drenched but unharmed. “I think,” says McAuley, “I’ll take photos of this island from the boat.”

Man down, I follow Kelsey as she scrambles to an alcove above our landing spot, passing empty cormorant nests and stepping over the desiccated remains of Common Murre chicks. After poking around, Kelsey calls, “Got it!” She’s crouched before a sandwich-size, guano-covered box containing an acoustic sensor, a device that has recorded thousands of hours of calling seabirds, barking sea lions, and crashing waves since she placed it here in March.

Each night during her six-month absence from Green Rock, the gadget has recorded a minute of sound for every five. Amid the cacophony it has stored, the biologists are interested in only one sound in particular: the call of Ashy Storm-Petrels, nocturnal, pint-size seabirds from central California to northern Mexico.

Ashies are hardy birds, but they’re vulnerable to many threats, including predation by rats and gulls on their island breeding grounds and oil spills from the offshore rigs along the state’s southern coast. And in coming years, ����APP change could take a big toll; warmer, more acidic oceans , and rising seas threaten to inundate nest sites. Closer monitoring of the birds’ health would not only assist in their conservation; because their lives are so tied to the sea, it would also help scientists track how this swath of the Pacific Ocean itself is changing. The only problem: Ashies are among the world’s most difficult seabirds to study.

That’s where the recorders come in. Kelsey and Felis are on a multi-week trip to collect the devices they’d deployed on 22 of the 20,000 outcroppings of the California Coastal National Monument, protected waters along the state’s 1,100-mile shore.

Buried in the sound files is information about where Ashy Storm-Petrels are located and, perhaps, how many are at certain sites. If the scientists can filter through the noise and pinpoint Ashies, over time they could more readily detect changes, such as shifts in when they use breeding sites or even a colony collapse. The challenge is identifying the birds’ calls amid thousands of hours of clamor, and then translating the data into estimates of their abundance. It’s the sort of gargantuan undertaking a grad student might spend years on. They’re hoping a computer can do it, in a fraction of the time.

Ashy storm-petrels spend most of their lives on California’s cool, open waters consuming small organisms that well up to the surface. Their terrestrial nests—in crevices on islets and a few caves on sheer cliffs—are, as we experienced, difficult to access. And they only come out at night. So visual counts of breeding colonies by air or boat, the gold standard for seabird surveys, are useless. To tally Ashies, biologists have to get on the rocks to look for them. Nobody knows if the birds are breeding on Green Rock, for instance. It’s unlikely because it’s north of their known range, but the recorders could reveal if they’ve moved up the coast. Prior to our arrival, the island hadn’t been surveyed in half a decade.

In their quest to monitor the birds, Kelsey and Felis are following in the sure-footed steps of , a biologist who scoured island after islet for three decades in search of breeding colonies. Carter died in 2017 but left behind detailed records of every site with confirmed or suspected storm-petrel nests. (And, helpfully, landing notes.)

That painstaking approach helped reveal the bird’s range and, combined with catch-and-release surveys at two of the largest breeding colonies, provided a general sketch of the species’ movement and population. Yet despite all the effort, scientists still don’t have a more efficient way to decipher whether the Ashy population—loosely estimated at around 10,000 individuals—is changing or what its true abundance is.

“You could argue, ‘Well, let’s just keep funding the Harry approach. Guys going out in little rubber boats jumping on slippery rocks,’” says Matthew McKown, an ecologist and co-founder of Conservation Metrics, a company that provides automated technologies for wildlife monitoring. “I’d say, if you’re going to be there, drop this gadget off so you can understand what’s happening the whole season—not just the brief window you’re there.”

Acoustic monitoring has long been used to determine whether specific bat or bird species are present in an area. Each one has unique vocalizations, and scientists have traditionally either picked out calls by ear from recordings or used software that transforms noise into spectrograms—pictures of the sound frequencies—that they manually scan for their target species’ hallmark signatures.

With the storm-petrel project, the scientists want to take acoustic analysis to a new level by employing an increasingly powerful tool: artificial intelligence (AI). Pushed forward by advances in processing power and machine learning, computers are now crunching enormous amounts of data to automate tasks scientists have long done manually or struggled to do at all. Using the same AI tools that power Amazon’s Alexa, Facebook’s facial-recognition technology, and Google’s self-driving cars, scientists are speeding and scaling up animal population studies and improving wildlife protections. Computers, for example, can now recognize and count individual animals in photos and safeguard endangered species by predicting where poachers may strike. 

For Ashies, combining acoustic monitoring with AI means that the daring island excursions scientists have long undertaken won’t just provide a blip of data from the moment they visit. Over entire breeding seasons they could collect unprecedented insights into how the birds—and, by extension, the larger marine ecosystem—are faring.

To accomplish this, however, McKown must first prove to the Bureau of Land Management, which oversees the monument, USGS, and other agencies that manage Ashy Storm-Petrels and their habitat, that his company’s technology can reliably identify the bird’s calls and not confuse the sound of an Ashy with, say, the five other storm-petrel species found regularly in California’s waters. It’s no small challenge. Today the most advanced AI still has its limits; even Amazon’s Alexa and Apple’s Siri don’t always get your song request right.

F

elis and Kelsey delivered the first season’s worth of guano-caked acoustic sensor boxes to in late 2017. “They don’t wash them off,” McKown tells me the following summer when I visit the company’s offices, a few trailers at the coastal science campus of the University of California, Santa Cruz. There, McKown and his small crew of T-shirt- and flannel-clad analysts walk me through how they’re transforming recordings into maps of where Ashies are present.

First they compress the audio files and convert them into a half-million spectrogram windows, pictures that each represent a two-second-or-so snippet of sound. Then their algorithm—a set of steps used to train a computer to identify the bird calls—goes to work. They could have built a simple algorithm to identify Ashy calls alone, but to improve their accuracy, they’ve developed a more complex one that aims to tease out the calls of 14 species commonly found in the region, including the squeaky sound of an Ashy, the more melodious twitter of a Leach’s Storm-Petrel, and the screech of a Western Gull. As the model sifts through the spectrograms, it learns the signatures unique to each species and makes more precise deductions. “Instead of, is it an Ashy Storm-Petrel or no,” says McKown, “it can decide, this is not actually an Ashy Storm-Petrel, it’s a Leach’s.”

Analyzing the 2017 acoustic data—8,796 hours, more than a year’s worth—took their “fancy data center” a couple of days, says McKown. (He’s referring to a bathroom closet. Earlier, when I saw a warning sign to keep the closet doors locked at all times, I’d wondered about their high-security cleaning supplies.) Analyst Kerry Dunleavy then spent three weeks meticulously validating the results, looking at the spectrograms identified as likely Ashies to see if the model was right.

When she finished, the team was thrilled to see that the model had accurately detected Ashies and other storm-petrels at the same sites Carter did. At the handful of sites north of the Ashy Storm-Petrel’s known range, including Green Rock, it detected Leach’s Storm-Petrels, but no Ashies. What’s more, at a few sites where Carter and colleagues found evidence of some type of storm-petrel, such as egg shells or feathers, but no breeding birds, the model named specific species.

Nobody was surprised that the model didn’t reveal a long-lost Ashy Storm-Petrel breeding colony. “Their range is really well described,” says Josh Adams, a USGS biologist who has studied the species for 20-plus years. “There’s no mysterious floating island filled with Ashies out there.” Besides, locating a new colony wasn’t the aim; it was to determine if AI-powered acoustic monitoring could identify the species. And it can.

But for the approach to be really useful, says Adams, it must help scientists monitor trends. McKown’s team has proven that the recorders can reveal if and when Ashies are present. The next challenge is using the data to identify how many individuals are at each site. That could shed light not only on overall population trends, but also on whether specific colonies are declining—a possible indication of food shortages, increased predation, or other emerging threats. AI can’t make that leap alone; it requires data gathered by old-fashioned fieldwork.

T

he best time to capture ashies is during the moon’s dark phase, when the fine mesh of the mist nets in which biologists snare them is least visible. The best places to go are the two largest breeding colonies: the South Farallones, off the coast of San Francisco, and the Channel Islands, about 30 miles from Santa Barbara. These nocturnal quests underpin existing species estimates.

In June McAuley and I once again join Kelsey in the field. We’re 600 miles south of Green Rock, accompanied by Adams, her boss, for three nights on Prince, an islet amid the Channel Islands where researchers have been mist-netting Ashies since the 1970s.

The bump-and-jump onto a flat-topped rock is the easy part of navigating this 35-acre landscape. The one small level area serves as a kitchen during the day and Kelsey’s bed at night. Adams, who spent five summers mist-netting Ashies on Prince in the 2000s, curls around a favorite rock to sleep. McAuley and I hang hammocks from yet more rocks.

The island is more welcoming to birds than people. They’re everywhere, and the sharp ammoniac odor of guano fills the air. Cormorants dominate one end; Cassin’s Auklets have an apartment complex of burrows on another. Black Oystercatcher parents squeak frantically as we wash dishes at the landing spot, their intrepid chicks toddling among the rocks. Western Gulls pose a more serious threat. Several are nesting near our designated bathroom area, forcing us to duck and cover while doing our business, or be swiped by dive-bombing adults.

Work begins when the sun goes down. Armed with fresh coffee, at 9 p.m. Adams and Kelsey unfurl the net, set up near the acoustic device, and start playing a recording of the Ashy Storm-Petrel’s flight call. Within 20 minutes birds attracted by the shrieks become tangled in the mesh. Adams and Kelsey expertly remove and band them, take beak and wing measurements, and release them. At 2 a.m. they stop the recording. In three nights of this they catch 78 Ashies.

To determine the birds’ relative abundance at these sites, Adams explains, biologists use a metric called “catch per unit effort”—in this case, the number of Ashies caught in the five-hour stretch after sunset. CPUE has long been the only technique for tracking changes in their abundance on islands.

Now McKown’s team is trying to deduce the number of birds on Prince based on the rates of the calls recorded by the gadgets. Then they’ll compare the algorithm’s results with the number generated by the CPUE formula to see if the two measurements sync up. “If there is a correlation—if catch per unit effort is related to call rates—then potentially we could use call rates as a monitoring tool,” says Adams. 

In other words, if they’re successful, they’ll have unlocked the ability to track changes in the number of birds anywhere there’s an acoustic recorder, not just at the long-term mist-net sites. If they only had to drop off and pick up sensors at the beginning and end of breeding seasons, they could expand monitoring and improve early warning of changes to the species’ overall health. “Getting a better grasp on the population size would be invaluable,” says Channel Islands National Park wildlife biologist David Mazurkiewicz.

On Prince, everyone is exhausted when we close the nets the first night. We all conk out in minutes. Except McAuley. When he lowers himself into his hammock, the bottom scrapes a rock and the material splits, dumping him on the ground.

C

onservation Metrics’ server crunches data for more than just this project. Analysts are also identifying the vocalizations, and subsequently the range, of Night Parrots, an Australian bird written off as extinct until one was rediscovered in 2006. With camera-trap photos they’re detecting cats and rats preying on nesting seabirds. They’re working with Cornell University to map the movement of elephants in the Congo with an eye toward understanding how the endangered mammals react to forestry practices.

The company isn’t alone in doing this work; countless academic groups, start-ups, and nonprofits deploy AI in conservation. Rice University researchers, for instance, are using AI to predict extreme-weather patterns, which could help governments better prepare for hurricanes and heat waves. There’s , an animal-world counterpart to Facebook that distinguishes individual zebras by their stripes and whale sharks by their skin patterns. And ����APP is exploring whether AI could improve the accuracy and scope of bird surveys and identify critical restoration needs after natural disasters.

“AI is now such a buzzword,” says Fei Fang, a Carnegie Mellon University computer scientist who works on the , an effort to predict locations poachers will target and optimize enforcement patrol routes. It’s technology, she notes, that isn’t always used for good and doesn’t live up to its hype in all cases. But in the field of conservation, where human resources are limited and geographies far-flung, she believes the buzz is real: “It can help address the most significant challenges we face.”

In all this work, processing huge data streams has become one of the biggest hurdles. Conservation Metrics has been working up against that limit lately. “When we started in 2012, a project had a hundred gigabytes or less,” says McKown. “Now we have projects that are bringing in 100 terabytes of data a year. We’re rapidly bursting at the seams.”

Take their Congo elephant project, which has recordings from 52 sensors that run 24 hours a day. The company’s single server requires three weeks to analyze three months of data. “It’s a bottleneck,” he says.

For many researchers working with AI, closer collaborations with tech companies could dramatically ease the strain. Microsoft’s five-year, $50 million , started in 2017, recently awarded McKown’s company a grant; Google launched its own $25 million program last year. “We’re trying to move what happens in the bathroom closet into the cloud,” says McKown. His team began pilot testing this approach with Microsoft’s cloud-computing platform in January. The effort slashed elephant data analysis to 15 hours from 22 days.

Already AI for Earth has funded more than 200 projects in four areas: biodiversity, ����APP change, agriculture, and water. Its goal is to make it easier for time- and cash-strapped researchers to store, manage, and analyze enormous amounts of ecological and environmental data, thus informing better conservation decisions. The vision is to put top-notch models created by grantees—and their data sets—in the cloud and provide open access for anyone to use. So a group on the other side of the world could, for instance, use the basic code underlying the storm-petrel models to identify calls of other species.

“Not every lab or conservation group is going to be spinning up and creating its own AI models,” says AI for Earth project manager Bonnie Lei, a former penguin researcher. “This is all toward a vision of truly democratized access, truly easy to use,” she says. “And, fingers crossed, it will truly move the conservation needle.”

A

rtificial intelligence may be on its way to becoming an indispensable conservation tool, but it isn’t going to supplant humans. “We’re trying to increase the footprint and the observation power of the Harry Carters of the world,” says McKown. “This is not going to replace those people.”

We still need rangers on the ground in Uganda to deter and catch poachers, and community scientists who upload photos, essential to Wildbook and other crowdsourced monitoring projects. We also need intrepid researchers like Kelsey and Felis to visit inhospitable environs, knowing the exact rock to jump on to access breeding sites and deploy acoustic devices, and to continue mist-netting to gather age and health data that sound recordings can’t capture.

The Ashy study ended last year, but Kelsey may soon be back in the Zodiac, bumping and jumping to deploy even more recorders along the West Coast. This winter McKown shared his team’s success at reliably detecting storm-petrels with a multi-agency group currently crafting the first-ever, range-wide Ashy Storm-Petrel management plan.

The exact protocols, including how many and what kinds of acoustic sensors would be deployed and where, haven’t been hammered out yet. If these tools are adopted, they could create a network that gathers data about Ashies and other storm-petrels, auklets, and murrelets that share their breeding grounds.

With it, scientists would be in position to glean invaluable insights into how some of the world’s most cryptic birds are faring, without birds realizing they were there. And if one day changing conditions push Ashies north to Green Rock, they’ll know.

This story originally ran in the Spring 2019 issue “Cryptic Seabirds Can Hide From This Scientist—But Not the Spy Network She Leaves Behind." To receive our print magazine, become a member by .