Harbor seals lounge on the seaweed-covered rocks of the Maine coast. Photo by Krista Ingram/Colgate University
BANGOR, Maine, Sept. 7 (UPI) -- Once scarce, harbor seals are now plentiful in the Gulf of Maine and along the Pine Tree State's rocky coast.
Despite their ubiquity, the life of the harbor seal isn't well understood. And while seal numbers have been dropping, the concern remains whether too many exist.
Now, new machine learning software designed to recognize the faces of individual seals could help scientists answer questions about the social behavior, site fidelity and movement of harbor seals.
Human facial recognition technology is frighteningly good, and scientists have had some success tweaking photo-reading algorithms to identify and recall the faces of primates.
"Seals are one step further away from the human face, so I wasn't sure if this would work," Krista Ingram, professor of biology at Colgate University, told UPI. "But I thought it was worth a try."
The try was successful, with much of the credit, according to Ingram, going to her research partner, Ahmet Ay, an associate professor of biology and mathematics at Colgate.
Training AI to recognize faces
To build their program, dubbed SealNet, Ay tweaked and refined the code of machine learning algorithms intended to recognize primates.
"We looked at a few other programs that were out there, and we took bits and pieces and combined them," Ingram said.
But Ingram said the result of Ay's algorithmic amalgamation is a machine learning program that's much more powerful than what came before.
"He doesn't just want to copy from someone else. He wants to improve it," Ingram said.
With the help of student researchers Zachary Birenbaum and Hieu Do, the research team trained the new algorithm -- using dozens of photographs collected from a haul-out in Maine -- to spot seal faces and differentiate between individual harbor seals.
The program works by identifying and recognizing facial patterns, like the shapes and arrangement of eyes and nose.
Previously, when scientists used a primate-spotting program called PrimeNet to identify seal faces, the software achieved 88% accuracy.
When researchers used 1,752 photos of 408 individual seals to test their SealNet software, the program proved 95% accurate.
Seal population changes
In the 19th and 20th centuries, seals were blamed for declines in commercial fish stocks. Maine and Massachusetts paid bounties for killed seals. Thousands were killed, and gray and harbor seal populations in the Gulf of Maine and around New England shrank dramatically.
Ingram says that when she first start visiting Maine several decades ago, seals were a rare sight.
"Now, they are becoming a major ecological factor here in Casco Bay," she said.
Harbor seals have experienced a strong rebound since the Marine Mammal Protection Act became law in 1972. Along Maine's cragged shores, the harbor seal has become omnipresent.
But scientists aren't sure whether they've reached full recovery, or perhaps, surpassed it.
"We don't have a good sense of how many seals there were in the Gulf of Maine historically," Kristina Cammen, assistant professor of marine mammal science at the University of Maine, told UPI.
"We know they were here in significant numbers, but we really don't know what that baseline before exploitation was."
Behaviors, impacts studied
While SealNet won't be able to reveal the size of historic seal populations, scientists hope the software can help them answer questions about seal behaviors, as well as the marine mammal's environmental and ecological impacts.
"I think [facial recognition software] will give us a better sense of seals as individuals, and that can offer us a better sense of their ecologies and behavioral patterns," Cammen said.
Scientists have a pretty good sense how many seals are living in Maine's waters. They also have a fair understanding of the marine mammal's life cycle and general movement patterns.
But less is known about how individual seals interact with one another at haul-out sites and when they hunt for fish.
"We know very little about whether they aggregate with the same seals or if it's totally random," Ingram said. "On top of that, that will inform more ecological-level population dynamic questions, like whether they are using the same resources year after year."
While seals do consume some of the same fish and shellfish targeted by fishers, they enjoy a rather diverse diet.
"Seals in the Gulf of Maine are fairly opportunistic. They tend to consume what's readily available and easy to catch," Cammen said.
Benefits of facial recognition
A better understanding of how exactly seals are using environmental resources could help alleviate concerns that local seal populations have grown too large.
Researchers suggest SealNet also could help scientists study links between the region's growing seal population and an increase in white shark sightings.
"In our photographs, we can see the bites of sharks on the bodies of seals," Ingram said. "So you can kind of track the increase in seal-shark interactions in the Gulf of Maine."
By looking more closely at where individual seals move, scientists may be able to discern where sharks are likely to migrate.
More precise movement data could also aid researchers like Michelle Berger, who uses seals to measure the prevalence of environmental toxins.
"Pretty much every chemical that we've looked for, we've found in the harbor seals," Berger, associate scientist at the Shaw Institute, told UPI. "They really integrate their exposure from everywhere they go throughout their lifetimes."
"And scientists still don't fully know where these individual seals move throughout the year," Berger said.
"If there was a way to track individual seals over an extended period of time using cameras and facial recognition technology, and to show where they were going and when, that's pretty exciting."