Session: 07-01-01 Arctic Environments and Frontier Regions

Submission Number: 180747

Towards Standardization of Labeling Training Images for Sea Ice Characterization Using Machine Learning 

Problem Statement: 

Many organizations are working to interpret images of sea ice conditions collected by ship-mounted optical cameras using modern machine learning techniques. While there is widespread agreement on some classes, there is also significant disparity in classes, labelling instructions, and annotation approaches, which limits the interoperability of manually labelled datasets and between labelling systems. Machine learning benefits from large data volumes and standardization of the labelling process; object detection, image classification, and image segmentation accuracy and robustness are expected to improve with harmonization of the processes and establishment of best practices.

Abstract 

Machine learning approaches are a cost effective and labor efficient means to extract engineering and research data from inexpensive optical cameras. Semantic segmentation of sea ice images from shipborne cameras is an emerging area of research which supports improvements in automation, route optimization, and performance monitoring objectives in ice covered waters. Several studies have presented scene segmentation and classification models in this domain, however there are a range of considered labels, labelling tools and labeler instructions. We summarize labelled datasets (published and unpublished) from X organizations (NTNU, NRC-OCRE, Aalto, Waterloo & MUN + others TBD) along with information related to preprocessing, labelling approaches, considered classes and labelling instructions. Additionally, we present a proposed set of labelling instructions for key ice types and ice features which are of interest to vessels operating in ice covered waters. The intent of this work is to compare and harmonize labelling methodologies from historical work, discuss and recommend best practices, identify gaps in current labeling applications, and support standardization for future work in this domain. An exploration of different use cases for characterized ice data and the implications for labelling approaches will be discussed.

Presenting Author: Andrei Sandru Aalto University

Presenting Author Biography: Dr. Andrei Sandru received PhD degree from Aalto University, Finland. Currently works at the Department of Energy and Mechanical Engineering.

Authors:

K Andrea Scott University of Waterloo
Andrei Sandru Aalto University
Ekaterina Kim Norwegian University of Science and Technology
Jeffrey Brown National Research Council Canada
Mikko Suominen Aalto University
Nabil Panchi Norwegian University of Science and Technology
Oscar De Silva Memorial University of Newfoundland
Richard R. J. Duan University of Waterloo
Rocky S. Taylor Memorial University of Newfoundland