AI maps icebergs 10,000 times faster than humans

Anne Brackman-Folgman, lead author of the results published today in CryosphereShe conducted her research during her time as a doctoral student at the University of Leeds in the United Kingdom. Now working at the Norwegian Arctic University in Tromsø, she emphasized the importance of large icebergs in the Antarctic environment.

“Giant icebergs are important components of the Antarctic environment. They influence ocean physics, chemistry, biology and of course marine processes. Therefore, it is important to locate icebergs and monitor their extent, to determine how much meltwater they release into the ocean.

By providing images of icebergs regardless of cloud cover and lack of daylight, the Copernicus Sentinel-1 radar mission plays a pivotal role in the innovative approach of using artificial intelligence to map mountains.

In images from satellites carrying camera-like instruments, icebergs, sea ice and clouds appear white, making it difficult to identify actual icebergs.

While in most radar images, as shown by Sentinel-1, icebergs appear as bright objects against a background of dark ocean and sea ice.

However, when the surroundings are complex, it can sometimes be difficult to distinguish between icebergs and sea ice or even from the coast.

“We sometimes struggled to separate icebergs from the surrounding sea ice, which is rougher and older and therefore appears brighter in satellite images,” Dr. Brackmann-Volgemann explained. “The same is true of wind-roughened oceans.

“Also, smaller iceberg fragments, which occur frequently near icebergs because they continually lose bits of ice around their edges, are easily collected with the main glacier by mistake.

“In addition, Antarctica’s coastline may resemble icebergs in satellite images, so standard segmentation algorithms often identify the coast as well rather than just the actual iceberg.”

However, the new neural network approach excels at mapping the extent of the iceberg even in these challenging conditions. Their strength lies in the ability of neural networks to understand complex nonlinear relationships and take the entire image context into account.

To effectively track changes in iceberg area and thickness, which is essential for understanding how icebergs melt and release fresh water and nutrients into the ocean, identifying a specific giant iceberg for ongoing monitoring is crucial.

The neural network presented in this study is highly efficient in identifying the largest iceberg in each image, in contrast to comparative methods, which often select slightly smaller icebergs nearby.

The neural network architecture is based on the famous U-net design. It was rigorously trained using Sentinel-1 images showing giant icebergs in various places, with manually derived outlines serving as the target.

Throughout the training process, the system continually improves its predictions, adjusting its parameters based on the difference between the manually derived scheme and the predicted result. Training stops automatically when the system reaches its optimal performance, ensuring it can adapt and succeed in new examples.

The algorithm was tested on seven glaciers, ranging in size from 54 square kilometers to 1,052 square kilometers, roughly equivalent to the areas of Bern in Switzerland and Hong Kong, respectively.

A diverse dataset was compiled, including between 15 and 46 images for each iceberg, covering different seasons and the years 2014-2020.

One Sentinel-1 image per month was used for each iceberg to ensure diversity of the data set. The results were impressive thanks to an accuracy of 99%.

Dr Brackmann-Volgemann added: “The ability to map iceberg extent automatically with enhanced speed and accuracy will enable us to monitor changes in the iceberg area of ​​many giant icebergs more easily and pave the way for practical application.”

“Satellites are, of course, essential for monitoring changes and understanding processes occurring far from civilization,” noted ESA’s Mark Drinkwater. “This new neural network automates what could be a manual, labor-intensive task of locating and reporting an iceberg.”

“We congratulate the team for introducing this innovative machine learning approach, to achieve a robust and accurate approach to monitoring changes in the vulnerable Antarctic region.”