Feature Representation for Sea Ice Mapping

Abstract

Sea ice monitoring is essential for climate research, Arctic navigation, and operational decision-making. Synthetic aperture radar (SAR) imagery is the primary sensing modality used by national ice services because of its independence of atmospheric and lighting conditions, and sensitivity to surface structure. However, SAR-based sea ice classification remains challenging due to spatially non-stationary statistics caused by incidence angle effects, seasonal transitions, and strong within-class variability. These factors complicate feature extraction and limit the robustness and transferability of conventional deep learning models.

This thesis investigates feature representation learning for sea ice classification in SAR imagery through both supervised and self-supervised paradigms. The first contribution introduces a supervised semantic segmentation framework that integrates convolutional neural networks (CNNs), transformers, and unsupervised region segmentation. The proposed Irregular Tokens on Transformers (ITT) architecture forms multi-scale, homogeneous tokens using Iterative Region Growing on Semantics (IRGS) and applies self-attention to capture long-range spatial dependencies. A multi-task training scheme combines pixel-level and region-level loss functions, encouraging region-consistent feature representations while preserving fine-grained boundaries. Experiments on multi-season RADARSAT-2 scenes demonstrate improved overall accuracy, sharper boundary delineation, and reduced predictive uncertainty compared to a CNN baseline. An expert audit conducted by the Canadian Ice Service further supports the operational relevance and stability of the approach across freeze-up and melt conditions.

While supervised learning delivers strong performance when annotations are available, SAR labeling remains costly and domain specific. The second contribution explores self-supervised pre-training toward a SAR foundation model for sea ice classification. By leveraging masked representation learning and multi-task objectives, the proposed framework learns transferable representations from unlabeled SAR imagery. The study evaluates whether large-scale pre-training alone is sufficient to address domain shifts across sensors and seasons, or whether task-specific adaptations remain necessary. Results show that self-supervised pre-training substantially improves downstream performance and generalization, but optimal accuracy is achieved when combined with structured fine-tuning aligned with sea ice semantics.

Overall, this thesis demonstrates that robust sea ice classification fundamentally depends on how feature representations are learned, and provides principled strategies for improving scalability, generalization, and operational viability in Arctic SAR applications.