Geographically-Weighted Weakly Supervised Bayesian High-Resolution Transformer for 200m Resolution Pan-Arctic Sea Ice Concentration Mapping and Uncertainty Estimation using Sentinel-1, RCM, and AMSR2 Data

Researchers Mabel Heffring and Lincoln Linlin Xu have introduced a groundbreaking AI system for high-resolution Arctic sea ice monitoring. Their paper presents a Geographically-Weighted Weakly Supervised Bayesian High-Resolution Transformer that generates 200-meter resolution sea ice concentration maps across the entire Arctic region. The model addresses four critical challenges in remote sensing: subtle ice feature detection, inexact training labels, uncertainty quantification, and data heterogeneity from multiple satellite sources. By combining Sentinel-1, RADARSAT Constellation Mission (RCM), and Advanced Microwave Scanning Radiometer 2 (AMSR2) data, the system provides operational sea ice charting with unprecedented detail.

The technical approach features several innovations: a Transformer architecture with global and local modules for detecting subtle ice patterns like cracks and leads; a geographically-weighted loss function that supervises at region level rather than pixel level; Bayesian parameter treatment for uncertainty estimation; and decision-level fusion of three satellite data types. Evaluated under pan-Arctic minimum-extent conditions in 2021 and 2025, the model achieved 0.70 overall feature detection accuracy using Sentinel-1 data while preserving pan-Arctic SIC patterns with R²=0.90 correlation to the ARTIST Sea Ice product. This represents a significant advancement in operational sea ice monitoring, providing both high-resolution mapping and reliable uncertainty estimates crucial for climate research and navigation safety.