28/03/2018 | Return to Latest News

Ship or iceberg, can you decide from space?

In November 2017, St. John’s-based applied R&D organization C-CORE and international energy company Equinor challenged the machine-learning community to find a better way to classify an object seen in a satellite image as a ship or an iceberg. Three months later, more than 3500 teams worldwide were working on the problem.

“The response has been terrific,” says C-CORE’s Vice President for Remote Sensing, Desmond Power. “Unmatched, in fact. Equinor’s prize purse of $50,000 certainly helped. But similar competitions with larger prizes have attracted less interest. I guess everyone is fascinated by icebergs. And everyone remembers the Titanic. Helping reduce that threat is a big motivator.”

During winter and spring, drifting icebergs can threaten navigation and marine operations in the North Atlantic, particularly in the area offshore Newfoundland and Labrador known as Iceberg Alley. Currently, the primary method for monitoring ice conditions and assessing iceberg risk is aerial reconnaissance, supplemented by platform- and vessel-based monitoring. However, satellite sensors and data analysis are increasingly being integrated into monitoring processes, decreasing fuel consumption and reducing the carbon footprint and cost of iceberg monitoring.

Synthetic Aperture Radar (SAR) satellites are not light-dependent and can “see” through darkness, cloud, fog and most weathers, taking snapshots over very large areas (up to 250 X 250 km). C-CORE has been working with satellite data for over 30 years and has developed a computer vision system that analyzes SAR data to automatically detect and classify icebergs and vessels. Equinor, an active player offshore Newfoundland, is a frontrunner in developing and applying new technologies to ensure a safe, secure and cost-efficient supply of energy and to reduce carbon emissions. Equinor is already extensively utilizing information technology and digitalization, but the rapidly growing digital technology ecosystem creates new opportunities. C-CORE and Equinor, both organizations dedicated to innovation, asked if there’s a better way to automatically detect and classify icebergs and vessels by using machine learning.

C-CORE and Equinor leveraged Kaggle, a platform for predictive modelling and analytics competitions in which companies and researchers post data, and statisticians and data miners compete to produce the best models for predicting and describing the data. This crowdsourcing approach relies on the fact that a wide variety of strategies can be successfully applied to any predictive modelling task and that competition is an efficient method for incenting and identifying the most effective technique.

C-CORE created a dataset of 5000 satellite SAR images containing either a ship or an iceberg. The SAR images are from the Sentinel-1A/B, a satellite constellation operated by the European Space Agency under Europe’s Copernicus program. Satellite radar works in much the same way as blips on a ship or aircraft radar: it bounces a signal off an object and records the echo, then that data is translated into an image. An object will appear as a bright spot because it reflects more radar energy than its surroundings, but strong echoes can come from anything solid – land, islands and sea ice, as well as icebergs and ships.

Competitors were challenged to build an algorithm that automatically identifies whether the object in the image is a ship or iceberg. The competition has concluded. Four winning teams have been selected, who share the $50,000 prize provided by Equinor.

C-CORE and Equinor are working toward validating the winners’ classification algorithms in real-word conditions. The improvements made in detection and classification are expected to decrease the cost and carbon footprint of iceberg monitoring and increase the safety of marine operations in iceberg-prone regions.