Permafrost covers more than 10 percent of the planet and stores vast amounts of carbon in frozen soil, making it a crucial—and dangerous—carbon sink. Understanding how (and why) permafrost is changing with the climate is critical to estimating (and mitigating) carbon emissions, but these dynamics remain largely opaque to researchers. A trio of researchers are working to change that, applying supercomputer-powered deep learning to better assess permafrost dynamics.
Satellite-based remote sensing has changed the game for permafrost analysis, but in some ways, it just introduced a new roadblock: combing through the massive amounts of data produced by those sensors.
The research team includes Chandi Witharana from the University of Connecticut, Kenton McHenry of the National Center for Supercomputing Applications (NCSA) and Anna Liljedahl of the Woodwell Climate Research Center. The trio, supported by the NSF’s “Navigating the New Arctic” program, gained access to more than a million satellite images of the Arctic.
The goal: to identify, and characterize, the telling ice wedges in permafrost that researchers can use to detect how that area of permafrost has been changing (or not) by noting their size, shape and characteristics.
“I was on Facebook some years ago and noted that they were starting to use facial recognition software on photos,” Liljedahl said in an interview with Aaron Dubrow of the Texas Advanced Computing Center (TACC). “I wondered whether this could be applied to ice wedge polygons in the Arctic.”
“That’s where we brought in AI-based deep learning methods to process and analyze this large amount of data,” Witharana said.
The trio annotated 50,000 ice wedge polygons by hand, marking them as either low-centered or high-centered relative to their ridges, with high-centered wedges indicating melting. This annotated data was used to train a neural network, which initially struggled when it was first tested a few years ago.
Those intervening years of fine-tuning, however, have yielded strong results, and with accuracy now between 80 and 90 percent, the team sought to scale up by employing supercomputing power. For that, they used Longhorn, a 2.3-Linpack petaflops system at TACC, and the Bridges-2 system at the Pittsburgh Supercomputing Center (PSC)—both allocated to the researchers through the NSF’s Extreme Science and Engineering Discovery Environment (XSEDE).
Using their model across those one million Arctic images on these powerful systems, the researchers had (by the end of 2021) mapped 1.2 billion ice wedges. “Permafrost isn’t characterized at these spatial scales in climate models,” Liljedahl said. “This study will help us derive a baseline and also see how changes are occurring over time.”
“Every year, we get an almost near real-time pulse meter on the Arctic in the form of sea ice extent,” she added. “We want to do the same with permafrost. There are so many rapid changes. We need to be able to really understand, and communicate, what’s happening in the permafrost.”
To learn more about this research, read the reporting from TACC’s Aaron Dubrow here.