NERSC researchers and intern develop a smart way to classify sea ice types

A new paper published in “Remote Sensing” addresses sea ice type classification in a new and clever way: Combining machine learning and remote sensing. The first author of the paper is a former intern at NERSC who created the central algorithm within the three months he spent in Bergen, and thanks to it, classifying sea ice types might have become faster and easier!

 

Seeing the ice from above

When navigating in polar regions, one of the threats you may encounter is sea ice. Knowing how much sea ice there is, and what ice types are around, is crucial for safety. Satellites provide an amazing way to observe sea ice in the Arctic. The Synthetic Aperture Radar (SAR) works without light and is unaffected by weather – no matter how dark and cloudy it might be, SAR data give us an accurate image of the ocean and sea ice from space. But only having images isn’t enough.

Manual labour in analysis is time-consuming

Those satellite images need to be analyzed and interpreted to give meaning for secure navigation and offshore activities in polar regions. At the moment, this is done manually by ice analysts – people who classify ice types on the images and based on that draw ice charts. That process is extremely labour-intensive; the SAR data provided by the European Space Agency are actually over 100 images for the Arctic alone – every day! Several attempts have been made to assist humans in this job by machine learning algorithms.

Sea ice types, satellite data, and algorithms

A recently published paper, Classification of Sea Ice Types in Sentinel-1 SAR Data Using Convolutional Neural Networks, provides us with a novel solution! Within three months, Hugo Boulze, interning at the Nansen Center, with the help from Anton Korosov and Julien Brajard, wrote an algorithm that has proven to be better than existing ones! An algorithm is basically a recipe that a computer follows to solve a problem, in this case the classification of the four different sea ice types in SAR data. These are old ice, first-year ice, and young ice, with ice free being the fourth category. The distinction between these four types is important for different kinds of ships navigating the Arctic, from heavily armored ice-breakers to light cargo ships unprotected from sea ice impact. Another advantage is that the automated ice charts can be used to initialize numerical models and improve forecasts of sea ice conditions. Difference between an automated and a manual sea ice chart: A: Mosaic of SAR images from the ESA Sentinel-1 satellite for 1 April 2020. B: An automated ice chart produced by the CNN developed at NERSC. C: A manual ice chart produced at the U.S. National Ice Center.Difference between an automated and a manual sea ice chart: A: Mosaic of SAR images from the ESA Sentinel-1 satellite for 1 April 2020. B: An automated ice chart produced by the CNN developed at NERSC. C: A manual ice chart produced at the U.S. National Ice Center.

Machine learning to the rescue

To speed up the process of classifying sea ice, the researchers made use of machine learning and trained computers to recognize ice types from satellite images: They taught so-called convolutional neural networks (CNN) how to identify the four different sea ice types in SAR images. These CNN function like a simplified human brain: they learn to label images correctly during a training phase from a set of examples and then can automatically classify sea ice types in new images. After a series of tests the CNN are performing exceptionally well and now are over 90 % accurate!

Outcomes for the future

Thanks to how accurate and fast the sea ice classification from SAR data has become through the algorithm, the authors have high hopes for its applicability at the Copernicus Marine Environment Monitoring Service. It has enormous potential to improve the production of ice charts for the Arctic Ocean, which are so relevant for navigating these cold waters.

Anton Korosov, a researcher in the Sea Ice Modelling group at NERSC with 15 years of experience in machine learning, sums up the importance of their work: “It is exciting to teach computers to analyze images! Automated ice charts have a bright future, and two very important uses: supporting near real time navigation and improving sea ice models. But artificial intelligence will not replace humans anytime soon – it will rather assist us in tedious tasks and leave more time for interesting research”.

Julien Brajard, a researcher in the Data Assimilation group at NERSC, describes how the inclusive atmosphere at NERSC has contributed to the success of their research: “Geophysical remote sensing is a field where artificial intelligence theoretically has the potential to solve current problems, but concrete applications – beyond the proof-of-concept – such as the one presented in this paper, are rare. The fact that we have been able to produce an algorithm directly applicable in real-life demonstrates nicely that research benefits from a work environment in which scientists from different fields are enabled to work together.”

Isn’t it amazing that the outcomes from a 3-month internship at the Nansen Center can have such far-reaching impacts on Arctic navigation for decades to come?!

 

Reference:

Boulze, H.; Korosov, A.; Brajard, J. Classification of Sea Ice Types in Sentinel-1 SAR Data Using Convolutional Neural Networks. Remote Sens. 202012, 2165. https://doi.org/10.3390/rs12132165