Glacier Calving Front Delineation Achieves 68.7m Error Reduction with Domain Adaptation

Scientists are tackling the critical challenge of accurately mapping glacier calving fronts, a key indicator of ice loss and sea level rise. Marcel Dreier, Nora Gourmelon, and Dakota Pyles, from the Pattern Recognition Lab and Institut für Geographie at Friedrich-Alexander-Universität Erlangen-Nürnberg, alongside Seehaus et al., demonstrate a significant advance in this field. Their research addresses the common problem of machine learning models failing when deployed at new locations, showing how incorporating temporal references and static spatial information can dramatically improve performance. By reducing delineation errors from over 1100m to just 68.7m without altering the core model, this work establishes a robust framework for global glacier calving front monitoring and unlocks more reliable data for climate change research.

Their research addresses the common problem of machine learning models failing when deployed at new locations, showing how incorporating temporal references and static spatial information can dramatically improve performance.

Domain adaptation improves glacier calving front mapping

Scientists have achieved a significant breakthrough in monitoring glacier calving fronts, overcoming limitations in applying existing Deep learning models to new geographical locations. This advancement addresses the challenge of transferring models trained on benchmark datasets to real-world scenarios, specifically focusing on the Svalbard archipelago, which presented an out-of-distribution domain. This minimises the need for extensive, time-consuming manual annotation of new study sites. Furthermore, the researchers incorporated spatial static prior knowledge through the use of rock masks, providing crucial contextual information about local glacier geometry.

These masks support the discrimination between rock, glacial ice, and challenging features like ice mélange, a mixture of sea ice and icebergs that often confounds segmentation algorithms. A key innovation lies in the inclusion of summer reference images within the input time series. By integrating images captured during periods when ice mélange is minimal, the model’s performance is significantly improved, particularly in scenes heavily affected by this phenomenon. Tyrion-T-GRU, the state-of-the-art model used in this work, processes time series of Synthetic Aperture Radar (SAR) acquisitions to produce semantic segmentation maps identifying glacier, ocean, rock, and areas with no available information. The calving fronts are then derived from these segmentation maps, and the refined methodology substantially enhances the accuracy of this process. Experiments reveal that this combination of techniques, few-shot adaptation, static priors, and temporal referencing, effectively mitigates the domain shift between the benchmark data and the Svalbard environment.

Svalbard Domain Adaptation for Calving Front Delineation utilizes

Scientists addressed limitations in applying state-of-the-art glacier calving front delineation models to novel environments by pioneering a domain adaptation strategy focused on the Svalbard archipelago. Initial testing revealed that Tyrion-T-GRU, a model achieving near-human performance on the CaFFe benchmark dataset, yielded a substantial delineation error of 1131.6m when deployed at this new study site. Researchers recognised this inaccuracy stemmed from an out-of-distribution domain, necessitating methodological improvements to facilitate accurate calving front monitoring. To mitigate this, the study compiled a new Svalbard dataset comprising one manual annotation per glacier, supplementing the existing CaFFe data for additional training.

Experiments employed a temporal approach, utilising acquisitions from three years prior to the training set to rigorously evaluate performance. The team innovated by constructing time series consisting of annual composites of images, rather than strictly consecutive acquisitions, and strategically incorporated summer reference images. This technique aimed to reduce ambiguity caused by ice mélange, a challenging feature due to its similar backscattering characteristics to glacial ice, particularly when it dominates scenes. Tyrion-T-GRU already shares information across a sequence of eight images, but this refinement further enhances its ability to discern glacial features.

Furthermore, scientists harnessed spatial static prior knowledge by integrating a rock mask as an additional input modality. This mask, representing stable rock locations, provides crucial contextual information regarding local glacier geometry, aiding discrimination between rock, glacier ice, and ice mélange during the segmentation process. The system delivers semantic segmentation for each acquisition, classifying areas as glacier, ocean, rock, or areas with no available information.

Domain adaptation halves calving front delineation error

Scientists have achieved a significant reduction in calving front delineation error through a novel domain adaptation strategy. Initial tests using the state-of-the-art Tyrion-T-GRU model on the Svalbard archipelago yielded a delineation error of 1131.6m, insufficient for precise scientific analysis. Researchers addressed this limitation by incorporating spatial prior knowledge and summer reference images into the input time series, successfully reducing the error to 68.7m without altering the model’s architecture. The study focused on the calving front, the critical boundary between a glacier and the ocean, utilising Synthetic Aperture Radar (SAR) for year-round observations.

While near-human performance has been attained on the “CAlving Fronts and where to Find thEm” (CaFFe) benchmark dataset, transferring these results to real-world scenarios revealed limitations when encountering data from new sensors and previously unrepresented study sites. Specifically, the Svalbard archipelago presented an out-of-distribution domain, differing in geometry, surface conditions, and climatic influences from the glaciers within the benchmark dataset. To overcome this, the team compiled a new dataset comprising one manual annotation per glacier in Svalbard, supplementing the existing CaFFe data for training purposes. Experiments demonstrated that even a single label per glacier provided sufficient guidance for domain adaptation, contrasting with approaches utilising multiple images.

A key challenge identified was ice mélange, a mixture of sea ice and icebergs, which exhibits similar backscattering characteristics to glacial ice, causing segmentation ambiguity. Tyrion-T-GRU mitigated this by sharing information across a time series of eight images, but persistent errors remained when ice mélange dominated the scenes. To further refine performance, the researchers introduced an inference strategy leveraging summer acquisitions, when ice mélange is typically absent, creating an annual composite time series. Furthermore, the incorporation of a static rock mask as an additional input modality provided crucial spatial prior information. These rock locations, remaining stable over time, aided in discriminating between rock, glacier ice, and ice mélange during the segmentation process.

Domain adaptation improves glacier calving delineation by transferring

Scientists have demonstrated that a state-of-the-art model for delineating glacier calving fronts, while performing well in benchmark tests, exhibits insufficient accuracy when applied to a new and previously unseen study site. This limitation arises because the model was trained exclusively on the benchmark dataset, representing an out-of-distribution domain for the novel location. Through these methodological advancements, the delineation error was substantially reduced from 1131.6m to 68.7m, crucially, without requiring any modifications to the model’s underlying architecture. The authors acknowledge that the current study is limited to Svalbard archipelago S1 images, and further evaluation with other regions and sensors is necessary. Future research could benefit from continual learning approaches, allowing the model to progressively assimilate new data and adapt to changing conditions without complete retraining.