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

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.

💡 Research Summary

The paper addresses a critical gap between benchmark performance and real‑world applicability of deep‑learning models for glacier calving‑front delineation. While the state‑of‑the‑art Tyrion‑T‑GRU model achieves near‑human accuracy on the CaFFe benchmark, its Mean Distance Error (MDE) jumps to 1,131.6 m when applied to the Svalbard archipelago, an out‑of‑distribution domain with different glacier geometries, surface conditions, and climatic influences. To bridge this gap without altering the network architecture, the authors propose three complementary strategies: (1) few‑shot domain adaptation, (2) inclusion of summer reference images, and (3) incorporation of static rock masks as an additional modality.

Few‑shot domain adaptation – The authors collected a single manual calving‑front annotation for each of the 145 glaciers in Svalbard (all from summer 2019). These annotations were transformed into the same zone‑label format used in CaFFe (glacier, ocean, rock, NA) and merged with the original CaFFe training set. By training Tyrion‑T‑GRU jointly on this combined dataset, the model gains exposure to the specific spatial patterns of Svalbard while still leveraging the large, diverse CaFFe corpus. Validation on a small set of 2016 images confirmed that a single label per glacier is sufficient to guide adaptation.

Summer reference images – Ice mélange, a mixture of sea ice and icebergs, often obscures the true calving front in SAR imagery, leading to systematic segmentation errors. The authors mitigate this by augmenting each eight‑image time series with three additional summer acquisitions (July, August, September) where ice mélange is typically absent. The central four images remain the focus of analysis, but the presence of clean summer frames provides temporal context that helps the recurrent units disambiguate glacier‑ice from mélange.

Static rock masks – Rock outcrops are temporally invariant and delineate the lateral boundaries of glaciers. The authors generated rock masks by intersecting OpenStreetMap coastlines with glacier polygons derived from the Randolph Glacier Inventory and recent Landsat‑8 front mapping. These masks are fed to the model as a separate input channel, effectively supplying spatial priors that improve discrimination between rock, glacier ice, and ocean/ice mélange.

The experimental pipeline consists of four incremental experiments: (a) baseline Tyrion‑T‑GRU trained on CaFFe only, (b) baseline + few‑shot Svalbard data, (c) (b) + summer references, and (d) (c) + rock masks. Five independent models from the final configuration are ensembled, and class‑wise uncertainties are estimated as the standard deviation of logits across the ensemble.

Results – Each added component yields a substantial reduction in MDE: few‑shot adaptation alone cuts the error by 686.3 m (to ~445 m), summer references shave another 240.7 m (to ~204 m), and rock masks halve the remaining error (to ~103 m). The final ensemble achieves an MDE of 68.7 m and an overall zone IoU of 81.1 %. Notably, rock IoU reaches 99 %, glacier IoU 98.4 %, and ocean IoU 97.7 %, while NA IoU stays at 97.7 % (the latter is less critical because NA regions are pre‑masked). Uncertainty analysis shows a marked decrease in variance across all classes when the three strategies are applied, indicating more confident predictions, especially near the calving front and coastline.

Limitations and future work – The study is confined to Sentinel‑1 SAR data over Svalbard; transferability to other sensors (e.g., SAR from different platforms) or regions remains to be demonstrated. The authors suggest continual learning as a promising direction to enable incremental updates without full retraining, preserving performance on previously learned domains while adapting to new ones.

Conclusion – By leveraging a minimal amount of site‑specific annotation, temporal augmentation with summer images, and static spatial priors, the authors demonstrate that a state‑of‑the‑art glacier front segmentation model can be adapted to a new domain with a >95 % reduction in error, all without architectural changes. This framework paves the way for scalable, global monitoring of marine‑terminating glaciers using SAR time series, supporting downstream scientific analyses such as frontal ablation estimates and sea‑level rise contributions.