Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Few-Shot Domain Adaptation with Temporal References and Static Priors for Glacier Calving Front Delineation

Marcel Dreier, Nora Gourmelon, Dakota Pyles, Thorsten Seehaus, Matthias H. Braun, Andreas Maier, Vincent Christlein

TL;DR

This paper tackles the problem of transferring state-of-the-art glacier calving-front delineation from a caffe benchmark to novel study sites, where distribution shift degrades performance. It introduces a three-pronged framework—few-shot domain adaptation, summer reference images, and static rock masks—built on the Tyrion-T-GRU model to improve cross-domain generalization on the Svalbard archipelago. The approach dramatically reduces mean distance error ($mde$) from 1131.6 m to 68.7 m and elevates calving-front segmentation IoU to an ensemble value of 81.1, while also reducing uncertainty across classes. The results demonstrate a transferable, scalable workflow for global-scale calving-front monitoring using SAR time series, with potential for continual learning to extend to new regions and sensors.

Abstract

During benchmarking, the state-of-the-art model for glacier calving front delineation achieves near-human performance. However, when applied in a real-world setting at a novel study site, its delineation accuracy is insufficient for calving front products intended for further scientific analyses. This site represents an out-of-distribution domain for a model trained solely on the benchmark dataset. By employing a few-shot domain adaptation strategy, incorporating spatial static prior knowledge, and including summer reference images in the input time series, the delineation error is reduced from 1131.6 m to 68.7 m without any architectural modifications. These methodological advancements establish a framework for applying deep learning-based calving front segmentation to novel study sites, enabling calving front monitoring on a global scale.

Few-Shot Domain Adaptation with Temporal References and Static Priors for Glacier Calving Front Delineation

TL;DR

This paper tackles the problem of transferring state-of-the-art glacier calving-front delineation from a caffe benchmark to novel study sites, where distribution shift degrades performance. It introduces a three-pronged framework—few-shot domain adaptation, summer reference images, and static rock masks—built on the Tyrion-T-GRU model to improve cross-domain generalization on the Svalbard archipelago. The approach dramatically reduces mean distance error () from 1131.6 m to 68.7 m and elevates calving-front segmentation IoU to an ensemble value of 81.1, while also reducing uncertainty across classes. The results demonstrate a transferable, scalable workflow for global-scale calving-front monitoring using SAR time series, with potential for continual learning to extend to new regions and sensors.

Abstract

During benchmarking, the state-of-the-art model for glacier calving front delineation achieves near-human performance. However, when applied in a real-world setting at a novel study site, its delineation accuracy is insufficient for calving front products intended for further scientific analyses. This site represents an out-of-distribution domain for a model trained solely on the benchmark dataset. By employing a few-shot domain adaptation strategy, incorporating spatial static prior knowledge, and including summer reference images in the input time series, the delineation error is reduced from 1131.6 m to 68.7 m without any architectural modifications. These methodological advancements establish a framework for applying deep learning-based calving front segmentation to novel study sites, enabling calving front monitoring on a global scale.
Paper Structure (15 sections, 4 figures, 1 table)

This paper contains 15 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Tyrion-T-GRU
  4. Benchmark Dataset
  5. Methodology
  6. Few-shot Domain Adaptation
  7. Summer Reference Images
  8. Rock Masks
  9. Experimental Setup
  10. Results and Discussion
  11. Limitations and Future Work
  12. Conclusion
  13. Code and Data Availability
  14. Author contribution
  15. Acknowledgments

Figures (4)

  • Figure 1: Three methodological advancements.
  • Figure 2: Segmentation results for all experiments. Black corresponds to NA, dark gray to rock, light gray to glacier, and white to ocean and ice mélange.
  • Figure 3: Ensemble results for all five test set glaciers. Predicted fronts, ground truth fronts, and their overlap are shown in yellow, blue, and pink, respectively. All calving fronts are dilated for visualization purposes.
  • Figure 4: Uncertainty estimates for ensemble predictions of the baseline experiment (top row) and the final experiment (bottom row). Predicted fronts, ground truth fronts, and their overlap are shown in yellow, blue, and pink, respectively. The calving fronts are dilated for visualization purposes. From left to right, panels show the ensemble prediction and the associated uncertainties for the NA, rock, glacier, and ocean classes.