Breaking the Ice: Video Segmentation of Close-Range Ice-Covered Waters

Abstract

The Arctic Ocean is experiencing significant ice recession, with projections indicating ice-free conditions during summer by 2060. This environmental shift opens new navigation routes, which could serve as a crucial trade route between the Pacific and Atlantic. Current ice navigation relies heavily on subjective decisions by ice experts, highlighting the need for tools that can assist in making objective, data-driven navigation decisions.

This dissertation explores methods for ice condition assessment using ship-borne optical data, focusing on the application of machine learning techniques. We investigate several neural network architectures for the semantic segmentation and classification of ice, specifically aiming to develop a network that is robust to occlusions (e.g., droplets on the lens) and capable of inferring ice condition in occluded regions. We create a novel medium-sized dataset of 946 images with fine annotations and provide our semi-automated approach in order to create large finely annotated dataset. We train an ensemble of traditional convolutional neural networks (CNNs) and show their performance ability on our finely annotated dataset. Finally, we use a modification of SegFlow, integrating features from PWCNet and ResNet, to leverage temporal knowledge and decrease error from lens occlusions. By adjusting the network, we boast improved segmentation performance in both occluded and non-occluded data from baseline approach.

Our methodology demonstrates the potential of neural networks to provide robust ice condition assessments, aiding the pursuit of objective and reliable evaluations of ice conditions. By leveraging machine learning techniques, this research can contribute to safer and more efficient navigation in the increasingly accessible Arctic waters, thereby supporting the development of navigation tools that can keep pace with the changing environmental landscape.