ISSN 2330-717X

Monitoring Arctic Permafrost With Satellites, Supercomputers, And Deep Learning


Permafrost — ground that has been permanently frozen for two or more years — makes up a large part of the Earth, around 15% of the Northern Hemisphere.


Permafrost is important for our climate, containing large amounts of biomass stored as methane and carbon dioxide, making tundra soil a carbon sink. However, permafrost’s innate characteristics and changing nature are not broadly understood.

As global warming heats the Earth and causes soil thawing, the permafrost carbon cycle is expected to accelerate and release soil-contained greenhouse gases into the atmosphere, creating a feedback loop that will exacerbate climate change.

Remote sensing is one way of getting a handle on the breadth, dynamics, and changes to permafrost. “It’s like a virtual passport to see this remote and difficult to reach part of the world,” says Chandi Witharana, assistant professor of Natural Resources & the Environment at the University of Connecticut. “Satellite imaging helps us monitor remote landscape in a detailed manner that we never had before.”

Over the past two decades, much of the Arctic has been mapped with extreme precision by commercial satellites. These maps are a treasure trove of data about this largely underexplored region. But the data is so large and unwieldy, it makes scholarship difficult, Witharana says.

With funding and support from the U.S. National Science Foundation (NSF) as part of the “Navigating the New Arctic” program, Witharana, as well as Kenton McHenry from the National Center for Supercomputing Applications, and Arctic researcher Anna Liljedahl of the Woodwell Climate Research Center, are making data about Arctic permafrost much more accessible.


The team was given free access to archives of over 1 million image scenes taken in the Arctic. That’s a lot of data — so much that traditional analysis and features extraction methods failed. “That’s where we brought in AI-based deep learning methods to process and analyze this large amount of data,” Witharana said.

One of the most distinctive, and telling, features of permafrost are ice wedges, which produce recognizable polygons in satellite imagery.

“The ice wedges form from the freezing and melting of soil in the tundra,” said Liljedahl. “Some of them are tens of thousands of years old.”

The shape and dimensions of ice wedge polygons can provide important information about the status and pace of change in the region. But they short-circuit conventional analysis.

“I was on Facebook some years ago and noted that they were starting to use facial recognition software on photos,” recalled Liljedahl. “I wondered whether this could be applied to ice wedge polygons in the Arctic.”

She contacted Witharana and McHenry, whom she had met at a panel review in Washington, D.C., and invited them to join her project idea. They each offered complimentary skills in domain expertise, code development, and big data management.

Starting in 2018, Witharana began using neural networks to detect not friends’ faces, but polygons from thousands of Arctic satellite images. To do so, Witharana and his team first had to annotate 50,000 individual polygons, hand-drawing their outlines and classifying them as either low-centered or high-centered.

Low-centered ice wedge polygons form a pool in the middle of the ridged outer part. High-centered ice wedges look more like muffins, Liljedahl said, and are evidence of ice wedge melting. The two types have different structural hydrological characteristics, which are important to understand in terms of their role in climate change, and to plan future infrastructure in Arctic communities.

“Permafrost isn’t characterized at these spatial scales in climate models,” said Liljedahl. “This study will help us derive a baseline and also see how changes are occurring over time.”

Training the model with the annotated images, they fed the satellite imagery into a neural network and tested it on un-annotated data. There were initial challenges — for instance, images trained for Canada were less effective in Russia, where the ice wedges are older and differently shaped. However, three years later, the team is seeing accuracy rates between 80 and 90%.

Leave a Reply

Your email address will not be published.