Globally Scalable Glacier Mapping by Deep Learning Matches Expert Delineation Accuracy

TL;DR

Globally scalable glacier mapping is addressed by introducing GlaViTU, a convolutional-transformer U-Net, and five strategies to generalise across regions and sensors. The method fuses optical, SAR, and DEM data and includes uncertainty calibration to quantify confidence in predictions. It achieves high average $IoU$ around 0.89 on unseen data, with InSAR data providing consistent gains and thermal data offering limited benefits, and it approaches expert delineation. The work provides a public benchmark and an end-to-end workflow for automatic multitemporal glacier outlines, enabling scalable global inventories and long-term change analysis.

Abstract

Accurate global glacier mapping is critical for understanding climate change impacts. Despite its importance, automated glacier mapping at a global scale remains largely unexplored. Here we address this gap and propose Glacier-VisionTransformer-U-Net (GlaViTU), a convolutional-transformer deep learning model, and five strategies for multitemporal global-scale glacier mapping using open satellite imagery. Assessing the spatial, temporal and cross-sensor generalisation shows that our best strategy achieves intersection over union >0.85 on previously unobserved images in most cases, which drops to >0.75 for debris-rich areas such as High-Mountain Asia and increases to >0.90 for regions dominated by clean ice. A comparative validation against human expert uncertainties in terms of area and distance deviations underscores GlaViTU performance, approaching or matching expert-level delineation. Adding synthetic aperture radar data, namely, backscatter and interferometric coherence, increases the accuracy in all regions where available. The calibrated confidence for glacier extents is reported making the predictions more reliable and interpretable. We also release a benchmark dataset that covers 9% of glaciers worldwide. Our results support efforts towards automated multitemporal and global glacier mapping.

Figures (16)

• Figure 1: Tile-based dataset and results overview. The glacier outlines and tidewater glacier areas are based on RGI7.0rgi. Debris coverage is adapted from Herreid and Pellicciottiherreid_rgi_debris_2020. The IoU values are presented for the GlaViTU model trained globally with Optical+DEM data. The statistics for Central, South West and South East Asia are aggregated.
• Figure 2: Semantic segmentation results for the independent acquisition test data as derived using GlaViTU with regional encoding and bias optimisation:a, b the Swiss Alps, c, d Southern Norway, e, f, g Alaska and h, i Southern Canada. The satellite images are presented in a false colour composition (R: SWIR$_{\approx2.2\mu m}$, G: NIR, B: R). Landsat images courtesy of the U.S. Geological Survey. Copernicus Sentinel data 2019.
• Figure 3: Reliability diagrams for confidence derived with Monte-Carlo dropout:a before and b after predictive confidence calibration. ECE stands for expected calibration error, lower values indicate better calibration. Bins and orange curves depict the actual accuracy versus confidence as evaluated on the validation subset, and green lines show the ideal calibration case. After calibration, confidence aligns more closely with the actual accuracy, which enables interpreting predictive confidence in more absolute terms. In this particular case, one can expect that $X\%$ of the pixels predicted with confidence $X\%>60\%$ are predicted correctly. Without this step, one can only compare confidence levels of predictions to each other.
• Figure S1: Closeups demonstrating decadal glacier changes mapped with GlaViTU: a before, b after and c both overlaid. Three upper rows: Aletsch complex (1998 vs 2023). Three lower rows: Hardangerjø kulen (1996 vs 2022). Polylines represent glacier outlines and semi-transparent polygons indicate 95% confidence bands. Landsat images courtesy of the U.S. Geological Survey. Copernicus Sentinel data 2022.
• Figure S2: Comparison of the results derived with DeepLabv3+/ResNeSt-101 and GlaViTU on the tile-based test dataset and Optical+DEM data:a--d debris-covered ice, e a water body, f surrounding rocks and g ice mélange. The satellite images are presented in a false colour composition (R: SWIR$_{\approx 2.2 \mu \text{m}}$, G: NIR, B: R). Landsat images courtesy of the U.S. Geological Survey. Copernicus Sentinel data 2019.