Mapping Antarctic crevasses with machine learning and satellite radar data

Large masses of ice deform under their own weight like a big blob of viscous fluid, flowing outwards towards the sea from their high-altitude centres. Combined with predictions of snowfall and melting, this is the physical theory that allows scientists to simulate the evolution of the Antarctic Ice Sheet and estimate its future contribution to sea-level rise. However, the ice can also crack and fracture on much shorter timescales, leading to the dramatic calving of icebergs and opening of crevasse fields. The effect that these crevasses have on the ice sheet – by changing its material properties, hydrology and the surface albedo – are relatively unknown. The solution to this, as has been true throughout the history of science, will rely heavily on data. Unfortunately, data on the locations of crevasses on the Earth’s ice sheets has historically been sparse, but the proliferation of machine learning for image processing and the deluge of satellite data illuminating the polar regions offers us new opportunities for us to change this.

The Sentinel-1 satellites fly 700km above the surface of the earth emitting pulses of microwave radiation. By measuring the time taken to receive the echo and the strength of the signal they can reconstruct black-and-white images of the Earth’s surface. They can see through clouds, image at night and day and, crucially for us, they can also see through the snow that blankets the Earth’s ice sheets, allowing us to glimpse features invisible to the naked eye.

We developed a series of routines to extract from this data maps of crevasses across the entire Antarctic continent with deep learning. As with many successful applications of deep learning, the method we developed contain a mixture of neural networks and more ‘classical’ algorithms used in combination.

Bootstrapping a neural network

We used two U-Nets, a battle-hardened and widely-utilised architecture nearing a decade in age, to extract two types of crevasse on the ice sheet. These were trained in a weekly supervised way, by feeding tuned versions of the network outputs back in as training data. The outputs were then post-processed using methods we based on those developed in the 1990s to find blood vessels in medical images. Having developed a method that appeared to work, we set about applying it to the entire catalogue of Sentinel-1 acquisitions over the Antarctic Ice Sheet; some 100 000 images, each containing roughly 3Gb of data.

Crevasses on the Antarctic Ice Sheet in June 2021

With these quantitative data at our disposal we can start to understand the role that these beautiful and perilous features have on the Earth’s ice sheets. In particular, how the density of fractures is changing through time. Looking at the Amundsen Sea sector of West Antarctica, where the ice sheet has been experiencing many of its most dramatic changes over recent decades, we have found large increases in the number and density of crevasses in areas where their presence can have a distributed effect on the wider ice sheet.

But the fun doesn’t stop there! In the time since this dataset was created, we have developed methods of assimilating these data with numerical models to understand how fractures change the bulk material properties of the ice, and are working on new ways to use them to learn approximations to the physics of crevasse formation. As always, data leads the way!

Dr Trystan Surawy-Stepney

Paper ref: Surawy-Stepney, T., Hogg, A. E., Cornford, S. L., and Hogg, D. C.: Mapping Antarctic crevasses and their evolution with deep learning applied to satellite radar imagery, The Cryosphere, 17, 4421–4445, https://doi.org/10.5194/tc-17-4421-2023, 2023.