Ringed seal, Karlee Zammit / WCS Canada
Ringed seal © Karlee Zammit / WCS Canada

Can AI successfully detect Ringed Seal barks in an ocean of acoustic data?



A young scientist uses artificial intelligence to help protect the Arctic

How do you find a career that blends AI and wildlife conservation? How do you score a job that involves a key (and oh-so-adorable) arctic species, the ringed seal?

Karlee Zammit began by studying remote stars and galaxies. She enrolled in astrophysics at the University of Victoria.

“My path hasn’t been linear,” she confesses. She shifted mid-degree to pure physics, started analyzing sounds, and explored machine learning. After graduation, she worked as a bear research tech. She also found jobs at tourist lodges.

She worked in underwater acoustics for several years. Her passion for underwater life kept growing. Then one of her former professors sent her a job posting. "This job might interest you," he said. "It involves underwater acoustics!"

Karlee applied, kept her fingers crossed, and her application succeeded! And that’s how she started working as a Master's student with Dr. William Halliday.


Can you detect a ringed seal’s bark better than AI can?

Winners will be entered in a draw for WCS Canada swag.

  • Karlee Zammit photographing ringed seals in the Arctic.

    Karlee Zammit photographing ringed seals in the Arctic.
    © WCS Canada

  • Bill Halliday and Karlee Zammit in the field.

    Bill Halliday and Karlee Zammit in the field.
    © Debbie Sharpe / University of Victoria

Passive acoustic monitoring – how to do sound research in dark, icy waters

Dr. Halliday leads the Arctic Acoustics Initiative for WCS Canada's Western Arctic Program.

In other words, he eavesdrops on aquatic mammals living in the Arctic Ocean.

Nothing beats “Passive Acoustic Monitoring” (known as PAM to its close friends) when it comes to finding out what’s going on under the Arctic ice. After all, it’s dark under there. And cold! Dropping a camera under the ice can tell us only so much. An “ears first” approach works better – especially since sound travels further underwater than in air and arctic mammals tend to be a chatty bunch.

Dr. Halliday’s program has 13 acoustic recorders in the water, recording sounds in different locations in the western Canadian Arctic. That’s no small task. Think remote locations accessible only by sea planes. Think drilling through two metres of ice. Think extreme, battery-draining cold conditions. Think corrosive saltwater.

Each recorder collects ocean sounds for at least a year – up to two if needed. To save its battery, it turns on for only five minutes every half hour.

If one recorder takes 10 minutes of recordings every hour for one year, how long does it take to listen to it all? And remember, there are 13 recorders out there!

“If we’re getting back 13 full-year data sets, I’d need a crew of 10 people, full time, to go through that,” says Dr. Halliday. “Each recorder holds 500 gigabytes of data after a year.”

That’s where Karlee’s work will come in handy. Karlee’s AI detector finds the needles (ringed seal barks) in a haystack (years of recorded data). With a detector powered by AI, a job that might take five years shrinks to a mere three months.

How to build a ringed seal bark detector, step one

To build her detector, Karlee needed two things. First, she needed annotated data: recordings with the ringed seal barks already tagged.

Thanks to Dr. Halliday’s work, she started with 196 hours of annotated ringed seal sounds. Six people worked for two years to tag each ringed seal bark in hours of recorded data. Karlee spent another two months making the annotations compatible with deep learning.

“I had to make sure that a distinct start and stop time had been marked for each bark,” she explains. “The bark cannot be loosely defined. To train AI, you need to show a precise start and stop time.”

Luckily, Karlee can use her eyes to tag sounds, using a spectrogram. A spectrogram shows sounds as lines and vertical bars. On a spectrogram, a seal’s bark rises shows up as a short vertical bar and is often accompanied by yelps – short sounds that show up as short descending or ascending lines on the spectrogram.  Louder barks look brighter, stronger.

The sound signature of a Ringed Seal bark, WCS Canada
The sound signature of a Ringed Seal bark © WCS Canada

How to build a ringed seal bark detector, step two

Once Karlee had a good supply of annotated seal barks, she had to connect with the right kind of artificial intelligence. She needed an interface, a place to develop and train an AI detector. For that, she relied on the work of the MERIDIAN Data Analytics Team at the Institute for Big Data Analytics at Dalhousie University.

The MERIDIAN team developed an interface called Ketos. (It’s named after the ancient Greek word for whale.) MERIDIAN made Ketos specifically for researchers working with underwater acoustics.

A researcher like Karlee can enter 196 hours of seal barks in Ketos. Ketos becomes a base on which she can build and train a detector. And Ketos is free to use!

Ketos provides some of the software needed to handle acoustic data. Karlee still needed to “write the script” for her detector. (For those interested in coding, she used Python as the programming language.)

Once the detector had been trained on annotated data, Karlee set it loose on “raw” acoustic data, to see how well it did.

“It still picked up false positives,” she says. “Any short sound that looks like a vertical bar on the spectrogram can be mistaken for a ringed seal bark. If the rope fastening the hydrophone hits the side of the hydrophone, it can register as a false positive. The detector can get confused by bubbles and boats, too.”

Karlee went through the new data manually. She corrected the detector’s mistakes, then she retrained it on the corrected data. “With more data, the detector will become more accurate,” says Karlee. “The more it is trained, the more accurate and helpful it becomes.”

Karlee is now making an easy-to-use interface for the completed detector. She and Dr. Halliday will offer it, free of charge, to any researcher with acoustic data from the Arctic Ocean.

“Other researchers use PAM to record data for other species,” explains Dr. Halliday. “With a detector, they can afford to check their recordings for ringed seals.”

The ringed seal: a species on thin ice

“Ringed seals are understudied,” says Dr. Halliday. “They’re a species of special concern because they rely on the Arctic ice and that ice is changing rapidly.”

“We now have years of underwater acoustic recordings of the Arctic Ocean. Biologists can learn more about ringed seals by observing changes in where and when ringed seal sounds appear on the recordings.”

Just how does a ringed seal bark underwater without running out of breath?

“The seal vibrates air up and down its windpipe,” explains Dr. Halliday. The air stays in the seal’s body, so the seal has enough breath to stay under water – typically for 5-10 minutes for a normal dive. Then it comes up through the ice to breathe. Ringed seals use the claws on their fore-flippers to keep air holes open – in 2-metre-thick Arctic ice!

Ringed seals play a central role in the Arctic food chain. They dive to catch fish, shrimp, and small crustaceans. In turn, polar bears eat ringed seals.

“They are the favourite food of polar bears,” says Dr. Halliday. “If ringed seals disappear, it has implications for all the wildlife in the Arctic ecosystem. Ringed seals are also culturally important, part of the Inuit cultural identity.” Inuit living in the region rely on ringed seals for food and clothing.

But because ringed seals live, hunt, and raise their pups on and under the Arctic ice, they are facing a huge threat because of climate change. The Arctic is warming up to four times as fast as the rest of the world. Climate change means more than vanishing ice. Ship traffic in the Arctic increases as the ice retreats. WCS Canada’s research is trying to shed light on how ship noise can negatively affect mammals such as ringed seals.

What management practices can help ringed seals survive and thrive? What will reduce stress on this species?  What choices lead to a world where ringed seals continue to delight and nourish us for generations to come?

We don’t know today. But by using AI to speed up the learning process, Karlee, Dr. Halliday, and the Arctic Acoustics Initiative team have brought us closer to the answers.

Assessing seal diet and health

Assessing seal diet and health

We are working with Inuit in the Inuvialuit Settlement Region to assess seal diets and contaminants as indicators of a changing Arctic ecosystem.

Research fellowships

Research fellowships

Weston Family Boreal Research Fellowship

Arctic

Arctic

Climate change is happening two to four times faster in the Arctic than the rest of the world.