Into Vanishing Ice

Having more ship traffic (and people) in the Arctic raises the chances that one or many could be caught unprepared during the months that ice pack forms and melts. Mary Keller is keenly aware of that impending risk.

“In many ways, sea ice is a brave new world,” she said. “The ice is thinning, and precipitation is changing from snow to rain. Both will accelerate the seasonal spring/summer sea ice melt, which begins, on average, about a week earlier every decade. Currents and winds will set that ice in motion, so areas that appear clear in satellite imagery can be ice-blocked just a day or two later, spelling trouble for ships that thought they were on a safe course through a strait or bay.”

“In many ways, sea ice is a brave new world...areas that appear clear in satellite imagery can be ice-blocked just a day or two later, spelling trouble for ships...”

Mary Keller

Remote Sensing Scientist

Ships will need near-term sea ice forecasts to ensure they can safely pass through the region, but according to Keller, what is currently available is not up to the challenge.

“It’s as inaccurate as weather forecasting had been in, say, the 1970s,” she said.

Ice-tracking methods at the National Ice Center today rely on an intensive process involving analysts manually clicking tens of thousands of points along the ice edge in satellite images. They then incorporate the longitude and latitude coordinates of the points into a geographical information system to create maps they can deliver to interested parties.

The Ice Center is producing daily 48-hour forecast maps for certain regions of the Arctic, but the resolutions are still too low to route ships.

Keller and her colleague Christine Piatko, a computer scientist at APL, had the nascent idea to somehow leverage the capabilities of machine learning and artificial intelligence to improve the forecasting capabilities, but exactly how was at first hazy.

Much in the same way companies like Microsoft have leveraged photos and artificial intelligence (AI) to analyze animal movements and help combat mass extinction, Piatko’s idea was to leverage tracking data of animals that depend on sea ice, such as polar bears and walruses, to predict where the ice will be. And there was plenty of data to work with.

For nearly 30 years, scientists have equipped various animals with trackers to show how their movements and migrations are each individually shifting with the changing climate.

Piatko and her team found tracking data for walruses, caribou, ringed seals, Arctic hares, Arctic foxes, numerous sea birds and even Dall sheep. By far the most complete data, obtained from nearly 300 polar bears, was collected from 1985 to 2016 in the Beaufort Sea. It included the U.S. Geological Survey’s (USGS) GPS-enabled video camera collar data, which offered not only a glimpse into the bears’ movements but a chance to virtually hunt, sleep and swim with the world’s largest terrestrial carnivores.

“Using PINNs for a remote sensing problem like this is kind of new...because people don't typically use them to make predictions over large areas and long time spans like we're doing.”

Heather Hunter

Senior Research Scientist (APL)

Analyzing the bears’ paths and laying them over sea ice data, the team was able to clearly see that before 2000, polar bears hunting in the Beaufort Sea concentrated in areas that, today, are ice-free and no longer accessible to them. The animals subsequently moved eastward into regions that still retained thick ice in the years after.

“It was interesting from a biological perspective and probably quite useful for understanding migration trends,” Piatko said. “But for our purposes, it was just too patchy and coarse.” Many of the animals’ paths were localized and difficult to analyze. And given that the data also spanned years, the resolution wouldn’t be much better than what ice prediction modeling already does. They needed to be able to predict in the order of days or, at most, weeks.

Setting the polar bear effort aside, Piatko and Keller started instead working to create a machine-learning model that could do exactly what they needed: generate high-resolution daily sea ice forecasts out to seven days.

They trained the AI using data and forecasts from various atmospheric and oceanographic models of the Beaufort Sea, including ocean currents, sea surface temperature, winds and sea ice extent during the freeze-up period from September to November in 2016 and 2018-2020.

The system proved strikingly accurate. Compared with real-world sea ice data, the team’s model predicted ice extent with 97% and 93% accuracy at one and seven days out, respectively.

In one highlighted example from Oct. 24 to Oct. 26, 2019, the model predicted the closure of sea ice above Canada’s Northwest Territories within one day of when it occurred. Because the model showed the potential for ice formation on the correct date, Keller and Piatko’s team said it would have provided sufficient information for any ship to pass through safely.

The model’s initial success spurred a follow-on project — from ideas born out of Walrus — to ensure the model is grounded in real-world physics, which Piatko and Keller said was not guaranteed in the original design. “There currently is no built-in knowledge to say what it’s doing is physically sensible,” Piatko said.

Heather Hunter, a senior research scientist at APL with expertise in applications of computer vision and machine learning to remote sensing, spearheads the project. She and her team are using a physics-informed neural network, or PINN, a relatively new (and theoretically faster) type of deep learning framework. It incorporates the quite complex physics equations believed to control the interrelationships of a system’s various parts.

Hunter said that such models are notoriously difficult to train and unstable because they’re essentially creating a solution from a bunch of noise. “That’s difficult, even giving it support data to help it come to a solution,” she said. “That would be difficult to even ask a person to do.”

Given the complexity of the system, Hunter’s group is starting small. They’re trying to forecast just sea ice movement in the Greenland Sea, where the East Greenland Current strongly forces ice drift. And even that has proven challenging.

“Using PINNs for a remote sensing problem like this is kind of new, and I think a lot of the reason for that is because people don’t typically use them to make predictions over large areas and long time spans like we’re doing,’” Hunter said. “A lot of the PINNs work I see is at finer scales — on the order of seconds to hours, and from the nanoscale up to a few meters.”

Even so, the team has already gotten the model to start working. In its latest phase, the network will accurately spit out a one-day forecast of which direction sea ice in the Greenland Sea will move, marking a huge feat in its development already. The trouble has been getting it to accurately predict how far the ice will move and how complex that movement will be. That’s something Hunter suspects has to do with simplifications they had to make to the underlying physics equations. “PINNs seem to be sensitive to how complex — or simple — those are,” she said.

“Satellites deliver great images and data of what's happening to ice patterns...but we want to deliver something that can be body-worn...”

Jackie Sharp

Project Lead (APL)

Now in its second year of model development, the group is trying to eliminate those simplifications while also pushing the model one step further: making it so it predicts not only sea ice movement but also sea ice deformation, a characteristic essential to knowing where openings in the ice will form.

“These openings, called leads, are usually big enough to allow ships to pass through, though they can quickly close overnight and trap those same ships,” Hunter said. “They also absorb a lot of heat from the Sun, which accelerates ice melt, so knowing where these leads form is important from both a climate science and operational perspective.”

Beyond projecting large-scale changes to sea ice for ships, there’s another need for localized resolution that can help personnel navigate the melting environment in vehicles and on foot.

“Fact of the matter is, the ice is melting faster than it’s being created, and it’s cracking unpredictably,” said Jackie Sharp, who was co-principal investigator with Kelles Gordge on a project at APL called Detecting Ice Variations Over Time (DIVOT). “Satellites deliver great images and data of what’s happening to ice patterns at large-scale effects, but we want to deliver something that can be body-worn and let an operator know in real time where they should be safely driving or walking, sort of like Waze for ice.”

Sharp and Gordge quickly learned there wasn’t much relevant data available in terms of localized ice fracture imagery, so their team took matters into its own hands. The group started freezing massive blocks of ice around APL and creating synthetic ice fractures by applying heat and, well, brute force. “We were literally stabbing it with ice picks,” Sharp said.

Using spectrometers, infrared cameras, and hyperspectral imaging to capture the jags and features of ice cracks, the team developed an algorithm that could autonomously recognize the ice fissures. The initial progress on DIVOT proved promising, having confirmed that it’s possible to detect ice defects using body-worn cameras and that large, surface cracks were easy to spot. But Sharp and Gordge would like to take the project a step further and deliver a more holistic view.

“We’d love to develop some sort of augmented reality [AR] deliverable, maybe it’s glasses and the operator can see where hot spots are on the ground,” Sharp said. They’ve made attempts to bring the pieces together to make AR an actual reality.

Team members froze large ice blocks at APL and then pressed their fingers onto the surface to warm a few spots. Using open-source AR tools connected to a computer camera, they could clearly see the warm red dots where their fingers had touched the ice and transferred a bit of heat.

It was a start, but Sharp notes there’s still a ways to go before seeing a tool like this deployed in the field.