Tracking how fast glaciers are shrinking is crucial for measuring the pace of climate change and projecting future sea level rises. This is normally a painstaking manual job, but a new approach that enables AI to analyze satellite images of glaciers anywhere in the world could help automate the monitoring process.

Glaciers that flow directly into the ocean play a crucial role in the earth’s climate, but global warming is making them retreat ever faster. This can have severe knock-on effects as ice that breaks away from “calving fronts”—the ends of glaciers where icebergs shear off into the water—dumps massive amounts of freshwater into the sea, which can alter ocean currents and cause sea levels to rise. Bright white glaciers also reflect a lot of sunlight. When they shrink, they expose dark seawater that absorbs heat from the sun.

All of this means that tracking glacier loss is critical for understanding how both local and global climate conditions will change over time. But the number of glaciers that need to be monitored around the world far outstrips the capacity of human analysts. There is hope that AI-based image analysis could help plug the gap, but previous models have performed poorly on regions not included in their training data. This severely limits the applicability of the approach, given how difficult it is to collect manually-labeled images.

Now, a paper accepted to the IEEE International Conference on Image Processing (ICIP) shows that a leading deep learning model for tracing glacier calving fronts can be adapted to new locations with minimal additional data. Researchers from the Friedrich-Alexander University of Erlangen–Nuremberg (FAU), in Germany, showed that the model’s error—the average distance between the modeled boundary and the real one—was cut from more than a kilometer to less than 70 meters by providing three pieces of information: one hand-labeled image per glacier, unlabeled summer reference images, and a map of the underlying rock.

In related research, some of the paper’s authors have already put the approach to work, using it to extract monthly calving front positions for all 145 glaciers in Norway’s Svalbard archipelago from 2015 to 2024. The team now hopes to extend the approach to another 1,500 glaciers in the Arctic.

“It’s about understanding glaciers better and how they react to changes in the climate,” says Nora Gourmelon, a Ph.D. student at FAU and co-lead author of the ICIP paper. “When you know about the past, then you will also hopefully be better able to understand how they will change in the future.”

Reducing the margin of error

Historically, delineating calving fronts has required students and researchers to pour over satellite radar images to manually trace the boundary between glaciers and the ocean, says Gourmelon. The process is time-consuming though, so numerous research groups have been experimenting with using computer vision models to automate the process.

In 2023, Gourmelon and her colleagues produced a dataset of 681 radar images of seven glaciers in Antarctica, Greenland, and Alaska, with manually annotated calving fronts to help train and benchmark new models. But when they took a state-of-the-art deep learning model trained on this dataset and applied it to previously unseen glaciers in Svalbard, they found it was off by an average of 1,131.6 m.

Gathering enough manually annotated data to retrain a model on every new glacier you want to analyze would clearly be infeasible, so the authors tried to find a more efficient way to boost performance. They produced one manually annotated calving front image for all 145 glaciers in Svalbard and combined this with several more raw satellite images of each glacier to create a new training set of 5,539 images. When they retrained the model on both this new data and the original benchmark data, the error dropped to 445.3 m.

They then developed two novel strategies to further improve accuracy. For both humans and AI, it can be tricky to distinguish the boundary of a glacier from ice melange—the mush of floating icebergs, sea ice, and snow that can accumulate at the calving front. So when the researchers uploaded a series of images of a glacier for the model to annotate, they included three images from the summer, when the melange is not present and the boundary of the glacier is clear. These acted as a reference point for the model and pushed the error down to 204.6 m.

As a final step, the researchers also provided the model with a static map of the rock underlying each glacier, derived from Open Street Map data that outlines the coast of Svalbard. This slashed the error to just 103.6 m. By running an ensemble of five different versions of their model and averaging their outputs, the researchers were able to get their final error down to just 68.7 m. While that may still sound fairly imprecise, Gourmelon says it’s comparable to manual annotation error rates. “Humans themselves are not really consistent in labeling, especially when there’s ice melange or when the resolution of the satellite image is not that good,” she says.

Automating glacier modeling

While the approach still requires some legwork, it can dramatically speed up the analysis of new regions. Most recent research of this kind has been done on annual or decadal timescales, says Dakota Pyles, a Ph.D. student at FAU, who led the second study that mapped nine years worth of glacier dynamics in Svalbard. In contrast to the lower frequency tracking, Pyles was able to generate monthly calving fronts for every glacier in that study—a total of more than 203,294 annotations—providing a much finer-grained view of how ice dynamics are changing on the archipelago.

“My project would not be possible at the scale that we’re going for if we didn’t have the model,” Pyles says. “So that is a great benefit for us and for advancing the field of glaciology [in general].”

In the long run, the approach could make it possible to partially automate the monitoring of glaciers around the world over extended periods of time, Gourmelon says. “We still need some labeled images from the specific region or satellite that you want to use for monitoring to train it on first, but then it can be used,” she says. “If how the image is captured and where you’re looking at stays consistent, then no recalibration would be necessary.”