ICIBS: Increasing the Context for Ice Boundary Segmentation

TL;DR

The ICIBS method improves ice boundary segmentation in radar data by integrating spatial and geo-spatial context into deep learning models, achieving state-of-the-art accuracy on benchmark datasets. It significantly reduces errors in ice bottom detection, especially in challenging temperate glacier conditions, and generalizes effectively to new field data.

Why It Matters

Accurate ice thickness mapping is critical for predicting sea-level rise and understanding climate change impacts. This approach enables faster, more reliable analysis of large-scale radar surveys, supporting timely glaciological assessments.

Key Findings

• Incorporating spatial and geo-spatial context into deep learning models improves ice bottom boundary detection in noisy radar data.
• The method achieves mean errors of 10.9 m and 43.4 m on FAU and CReSIS subsets, respectively, outperforming prior approaches.
• It generalizes well to unseen data, reducing error by 49.6% on new Patagonian campaign data compared to baseline models.