How Ciklum AI Team Built an Algorithm to Detect Areas of Sea Ice to Prevent Maritime Accidents

Author: Oleg Panichev, Senior Research Engineer at Ciklum. Oleg has 5+ years of experience in machine learning, deep learning and data science, with a background in biomedical signal processing. He took 5-th place with his team in the Epileptic Seizure Prediction competition organized by Melbourne University.

The Ministry of Economy, Trade and Industry launched Japan’s first satellite data platform “Tellus” intended for industry use on 21 February 2019. The Tellus Satellite Challenge is a data analysis contest aimed at promoting the use of Tellus, namely in the form of producing a visual representation of satellite data use case models, uncovering outstanding human resources specializing in data analysis and disseminating information and educating others on the types of satellite data and its various formats.

The First Regional Coast Guard Headquarters currently provides information on sea ice formations to prevent maritime accidents caused by sea ice around Hokkaido. To do this, in addition to visual observations of sea ice from the air and patrol boats, the First Regional Coast Guard Headquarters also analyses satellite observation images taken of sea ice.

The Challenge

Develop an algorithm to detect areas of sea ice with a high degree of accuracy from synthetic aperture radar (SAR) data to prevent maritime accidents.

Duration: 2 months

The Solution

Observation data from PALSAR-2 onboard the Advanced Land Observing Satellite 2 “Daichi-2” (ALOS-2) taken of the sea area around Hokkaido was used for this contest. PALSAR-2 is a SAR sensor that retrieves data by bouncing radio waves down to the Earth’s surface and receiving the radio waves that are reflected back.

The Ciklum team, consisting of two research engineers, developed a pipeline to work with high-resolution satellite images and train a deep learning model to perform binary semantic segmentation. One of the specifics of work with satellite images is that they are quite large and should be preprocessed and cropped properly to be handled by a neural network. Also, predictions made by the neural network should be merged properly to reduce artefacts on the edges of crops.

The Result

The final result was evaluated using the intersection over union (IoU) metric. A perfect model would have a performance of 1. The Ciklum team developed a model that detected sea ice with an IoU metric of 0.78539 and ranked 10th of 115 teams.

Originally published at https://www.ciklum.com.