Successes and Challenges of Sea Ice Segmentation on Sentinel-1 Synthetic Aperture Radar

Keywords: sea ice, synthetic aperture radar, machine learning, convolutional neural networks
Abstract Type: Paper Abstract

Abstract

National ice services monitor sea ice (SI) conditions in the Arctic and Antarctic and generate periodic ice charts, crucial for safe marine navigation. The ice charts are largely based on human expert interpretation of remotely sensed images, mainly synthetic aperture radar (SAR). The interpretation is a manual task in which ice analysts define boundaries where different SI types are present. Given the highly-dynamic nature of SI and the need for high-resolution and timely SI charts, machine learning (ML) algorithms have been investigated to aid in the acceleration of the generation of these products. Research groups are developing benchmark datasets to facilitate ML training and evaluation. We used the Extreme Earth (EE) V2 dataset and trained ML models for SI segmentation. EE reflects SI conditions on the East coast of Greenland for twelve Sentinel-1 scenes. We focused on two key objectives of SI mapping: i) separation of SI from water, or “binary classification”; and ii) SI type (stage of development) classification, or “multiclass classification”. We use a customized pipeline incorporating ResNets and DeepLabV3 to develop a convolutional neural network architecture for image segmentation. Binary classification results exhibit test F1 and intersection over union (IoU) generally above 0.95. Multiclass classification remains challenging with test weighted F1 and weighted IoU generally above 0.5, but below 0.85. Despite lower metrics, mixed ice types in the same polygon leave room for subjective interpretation. Our models show potential for automating SI detection at high spatial and temporal resolutions.