Globally scalable glacier mapping by deep learning matches expert delineation accuracy

Area of Science:

• Earth Science
• Glaciology
• Remote Sensing

Background:

• Accurate global glacier mapping is essential for climate change research.
• Automated, large-scale glacier mapping using satellite data is underdeveloped.

Purpose of the Study:

• To develop an automated deep learning model for multitemporal global glacier mapping.
• To assess the model's generalization capabilities and performance across diverse regions and data types.

Main Methods:

• Proposed Glacier-VisionTransformer-U-Net (GlaViTU), a convolutional-transformer deep learning model.
• Developed five strategies for multitemporal global-scale glacier mapping using open satellite imagery.
• Incorporated synthetic aperture radar (SAR) data (backscatter, interferometric coherence) to enhance accuracy.

Main Results:

• The best strategy achieved intersection over union (IoU) >0.85 on unseen data, with performance varying by region (>0.75 for debris-rich areas, >0.90 for clean ice).
• GlaViTU performance approached or matched human expert-level delineation.
• SAR data integration improved accuracy across all applicable regions.
• Calibrated confidence scores were reported for glacier extents, enhancing reliability.

Conclusions:

• GlaViTU offers a robust solution for automated multitemporal global glacier mapping.
• The model demonstrates strong generalization capabilities and near-expert performance.
• The released benchmark dataset will aid future research in automated glacier monitoring.