���Ҵ�ý�ƽ������

Ute Herzfeld (PI), Harald Sandal (pilot), Gustav Svanstroem (helicopter technician) and Matthew Lawson (research assistant) during the��Negribreen Glacier System Airborne Geophysical campaign (Photo Credit: Thomas Trantow).��
��

Glaciers are constantly changing and reshaping the Earth’s surface.��

���Ҵ�ý�ƽ������ researchers have developed a new machine learning tool to better understand how Arctic glaciers suddenly accelerate or “surge”. �� ��

The team, led by��Ute Herzfeld, a research professor in the Department of Electrical, Computer and Energy Engineering,��created an open-source cyberinfrastructure called GEOCLASS-image, designed to decode the physical processes behind glacier motion using high-resolution satellite imagery and machine learning.��

Glacier surges are sudden bursts of movement in otherwise slow-flowing ice.��

Normally, glaciers move at a steady pace, but during a rare “surge”, that rate can accelerate up to 200 times faster than usual. The ice fractures into deep crevasses and pushes large volumes of ice toward the ocean. These dramatic events provide scientists with new insight into the unpredictable drivers of sea-level rise. ��

“Most deep machine learning systems don’t know what to look for in images,” said Herzfeld, who is also the director of the Geomathematics, Remote Sensing and Cryospheric Sciences Laboratory. “We have built a system that understands the physics of ice deformation, so the classifications actually mean something.”

Understanding how a glacier surges

Unlike traditional artificial intelligence systems that often struggle to interpret complex natural phenomena, the team created a new neural network approach—VarioCNN—to better understand glacial acceleration.

“Surging glaciers are one of the deep uncertainties in sea-level rise projections,” Herzfeld said. “They can move much faster than normal and current earth system models do not yet have the ability to account for them.”

To tackle this problem, Herzfeld and her team merged two powerful approaches: a deep convolutional neural network (CNN), common in the field of computer science and remote sensing and a physics-informed neural network model that captures how crevasses in the ice form, widen and intersect during motion.��

“Think of neural networks as Lego blocks,” Herzfeld said. “We’ve taken some from physically informed models, some from deep learning and built a new kind of AI that’s meaningful.”

Putting AI to the test��

The team tested their approach on a real-world event: the unexpected 2016 surge of Negribreen, a glacier located in the Arctic archipelago of Svalbard a 1,000 km south of the North Pole.��

Using Maxar WorldView satellite imagery collected in 2016-2018, the researchers tracked subtle changes across the glacier’s surface with remarkable detail.

They discovered that crevasse patterns, which change dramatically during a surge, hold information about surge dynamics that can be retrieved using their neural network approach.����

One-dimensional crevasses appeared at the leading edge of the surge, while deeper within the surge area, complex patterns tell the story of the transformation and deformation of the ice, which can be of use in numerical modeling of the glacial acceleration.��

Shear, a type of deformation that plays a key role in glacial acceleration, is easily misclassified in deep learning, but correctly identified using VarioCNN.

With their new VarioCNN model, they classified different types of crevasses from satellite images and used those patterns to interpret how the glacier moved and changed.

Results of the classification were then used to understand how the surge expanded and affected the entire Negribreen glacier system. Ultimately, ice mass equivalent to 1% of global annual sea-level rise transferred to the ocean.

Published in��, their results demonstrated how integrating physical knowledge into a neural network model, carried out at the computational level, can advance machine learning and glaciological understanding of glacier surges. The paper was selected as the cover story of Remote Sensing receiving record downloads during the first two weeks after publication.

Student Connor Meyers setting up a GPS station at the edge of Negribreen (Photo Credit: Ute Herzfeld).��

“The problem of task-oriented machine learning is especially intriguing to me,” said Silas Twickler (Phys’25) who was a research assistant on the project. “While simply applying pre-existing neural networks may be sufficient for certain applications, the augmentation of these networks can allow for a drastic improvement in machine learning.”

AI for the geosciences��

A major hurdle in applying machine learning to studying glaciers is the limited amount of labeled data.��To overcome this, Herzfeld’s team developed a way that allows scientists to gradually refine the model using a relatively small number of hand-labeled satellite images.��

VarioCNN was trained on just a few thousand of examples, far fewer than the 100,000 images than typical deep learning models require. Due to its modular design, the GEOCLASS cyberinfrastructure can be adapted to study other geophysical processes and potentially surfaces of other planets.

“Our tool is not just for glaciologists, but for anyone working with remote sensing and physical systems,” Herzfeld said. “Ultimately, we hope to give scientists better tools to understand how the Earth is changing.”��

This research was funded by the National Science Foundation Office of Advanced Cyberinfrastructure and NASA Earth Sciences Division.