Around 2018, Anna Liljedahl was browsing Facebook when she came across something that made her stop. The platform began employing facial recognition software to identify individuals in images; it was able to tag faces in the background of birthday photos and extract names with startling accuracy from crowd shots. When Liljedahl, an Arctic researcher, saw it, ice came to mind.
She specifically considered the ice wedge polygons—geometric shapes pressed into the frozen ground by thousands of years of freezing and thawing cycles—that stretched across the tundra in patterns visible from satellite imagery. She pondered whether an algorithm could learn to identify those if it could identify a human face. After meeting two researchers at a panel in Washington, D.C., she contacted them. What came next was one of the most significant recent uses of machine learning in climate science.
| Category | Details |
|---|---|
| Subject | AI & Machine Learning Applied to Arctic Permafrost and Iceberg Mapping |
| Lead Institution | University of Texas at Austin / Texas Advanced Computing Center (TACC) |
| Key Researcher (Permafrost) | Chandi Witharana, University of Connecticut; Anna Liljedahl, Woodwell Climate Research Center |
| Key Researcher (Icebergs) | Dr. Anne Braakmann-Folgmann, Centre for Polar Observation and Modelling, Northumbria University |
| AI Algorithm Used (Permafrost) | Deep learning neural network — trained on 50,000 annotated ice wedge polygons |
| AI Algorithm Used (Icebergs) | U-net neural network — trained on Sentinel-1 ESA satellite imagery |
| Scale of Permafrost Mapping | 1.2 billion ice wedge polygons identified across Arctic satellite data |
| Iceberg Mapping Speed | One-hundredth of a second per iceberg vs. several minutes manually |
| Supercomputer Used | Longhorn (TACC) + Bridges-2 (Pittsburgh Supercomputing Center) |
| Funding Source | U.S. National Science Foundation — “Navigating the New Arctic” program |
| Permafrost Coverage | ~15% of Northern Hemisphere land surface |
| Accuracy Rate | 80–90% for ice wedge polygon detection |
| Output Platform | Permafrost Discovery Gateway — open-access Arctic data archive |
| Reference Website | tacc.utexas.edu |
About 15% of the land area in the Northern Hemisphere is covered by permafrost. In its most basic form, it is simply ground that has been frozen for two years or longer, often tens of thousands of years. As global temperatures rise and the permafrost thaws, the massive amount of carbon stored inside that frozen ground—locked away as carbon dioxide and methane—begins to release.
One of the more sobering dynamics in climate science is the feedback loop that results from this: warming speeds up thawing, thawing releases carbon, and carbon speeds up warming. To comprehend this feedback loop, one must be somewhat accurate about the permafrost’s actual appearance and rate of change. That kind of information was almost impossible to collect on a large scale until recently. No one had the time or processing power to thoroughly analyze the satellite imagery of the vast and remote Arctic.
Chandi Witharana and his associates were able to help with that. In order to identify ice wedge polygons from thousands of Arctic satellite photos, Witharana started training a neural network in 2018. This technique is similar to the pattern recognition used in facial recognition, but it is applied to the geometric signatures of freezing ground. 50,000 individual polygons had to be manually annotated, their outlines drawn, and two different types of polygons classified as part of the training process.
High-centered ice wedge polygons, which resemble muffins and show that melting has already started, and low-centered ones, which contain water in a central pool. The difference is more important than it might seem. Because the two types behave differently hydrologically, they have different effects on how quickly a landscape changes and what those changes mean for the infrastructure and communities that are situated on top of it.
The Longhorn supercomputer at the Texas Advanced Computing Center and the Bridges-2 system in Pittsburgh produced results that were startlingly large. The team estimated they were only halfway through the entire dataset, but by the end of 2021, they had located and mapped 1.2 billion ice wedge polygons throughout the Arctic satellite archive.
Preprocessing, detection, and post-processing for each individual image analysis can be finished in less than an hour. The entire process is now possible in a way that was not previously possible by running them in parallel across a sizable supercomputer. The team’s accuracy rate is between 80 and 90 percent, which is not perfect—a trained human’s manual interpretation is still more accurate—but at this scale, perfection is not an option. Perfect accuracy applied to a thousand data points is not the same as eighty percent accuracy applied to a billion.
When Dr. Anne Braakmann-Folgmann and her colleagues at the Centre for Polar Observation and Modelling tackled the iceberg problem independently, they encountered similar limitations and discovered a similar solution. Icebergs are important for maritime safety because they can drift into shipping lanes with consequences that don’t need to be explained.
Monitoring icebergs manually using satellite imagery is time-consuming. Some icebergs are the size of small countries. A skilled human analyst can accurately outline a single iceberg in a matter of minutes. It can be completed in a hundredth of a second by an AI system that has been trained on images from the European Space Agency’s Sentinel-1 satellites using the U-net algorithm. There is more to the difference between those two timelines than just convenience. It’s the distinction between a monitoring system that can adapt to a shifting polar environment and one that is constantly catching up to past events.
A detail in the research on iceberg mapping merits more consideration than it usually gets. Due to their inability to differentiate the iceberg from the surrounding sea ice, the older automated algorithms, k-means and Otsu, were producing estimates of iceberg area that were between 150 and 170 percent too large. In actuality, they were measuring the incorrect thing with remarkable consistency. That error was lowered to an average of roughly 5% by the U-net algorithm.
Such an increase in measurement precision is more than a technical detail. A 150 percent overestimation of area leads to conclusions that are significantly incorrect when scientists are monitoring how much ice is calving from glaciers, how quickly icebergs are melting, and what freshwater and nutrients are being released into polar seas as a result. Accurate measurement is crucial for all subsequent processes.
The infrastructure this work is creating, such as the Permafrost Discovery Gateway, the open-access archive of mapped ice wedge data, and the workflow designs optimized for high-speed analysis on supercomputing systems, may be more valuable than any one discovery.
According to Liljedahl, it wants a near-real-time pulse meter on the Arctic, similar to the yearly data already provided by sea ice extent. Research facilities in Texas, Connecticut, and Northumbria are still putting that vision together, dataset by dataset, algorithm by algorithm. The urgency of comprehending a landscape that is changing more quickly than science has historically been able to follow is the goal of this slow work made fast by machines.
