Glacier retreat is one of the most visible indicators of ongoing climate change. Modelling glacier evolution allows for the prediction of future ice mass changes and their potential impacts on society. Current models for tidewater glaciers lack calibration data on frontal ablation, which limits the accuracy of their predictions. Estimating frontal ablation requires precise, time-resolved calving front positions, which can be derived from satellite imagery. Synthetic Aperture Radar (SAR) imagery offers all-weather, year-round monitoring capabilities at high temporal resolution, but the manual delineation of calving fronts is infeasible due to the increasingly large data volumes. Therefore, this thesis addresses the deep learning-based automated delineation of glacier calving fronts in SAR imagery, establishing standardised practices in the field through a benchmark framework, determining the state of the art, and surpassing it with near-human performance. The benchmark framework is based on a publicly available, multi-mission SAR dataset of calving fronts, enabling reproducibility and comparability of deep learning models. The dataset provides binary calving front segmentation and multi-class landscape segmentation labels, supporting diverse methodological approaches. Two open-source deep learning models serve as baselines, and the test set design allows evaluation of model generalisation to previously unseen glaciers. Quantitative and qualitative improvements over the baselines are demonstrated by integrating a Conditional Random Field (CRF) into the post-processing of the baseline model trained on zone labels, connecting disjoint front segments and suppressing spurious ocean predictions. Building on this foundation, 22 deep learning models are systematically benchmarked. All models are adapted to, retrained on, and evaluated with the benchmark dataset, and performance differences are analysed in terms of statistical significance and effect size magnitude. Further statistical tests are conducted to investigate the impact of architectural and methodological factors on model performance, and a multi-annotator study reveals a significant performance gap of 183 m between humans and the best-performing model. Interpretation challenges for both humans and models are examined to identify avenues for future research. Guided by these insights, a new state-of-the-art model is developed. A hybrid transformer–Convolutional Neural Network (CNN) architecture provides smooth delineations in contrast to the jagged predictions of the previous state of the art. Larger patch sizes and retaining only the central part of the output allow sufficient contextual information to be incorporated. To address the domain shift between ImageNet-pre-trained weights and SAR data, the model is pre-trained on a novel unlabelled SAR calving front dataset of glacier time series, each paired with a single optical image. Two self-supervised multi-modal pre-training strategies are introduced, both using optical images as labels. These strategies eliminate the need for temporal alignment of SAR and optical images, reducing manual data curation and allowing all available SAR images to contribute to training. An ensemble of the pre-trained models achieves near-human performance, with an error of 75 m compared to the human consensus error of 38 m. By advancing automated calving front delineation to near-human performance, this thesis lays the foundation for global-scale automated monitoring of calving fronts, thereby enabling the generation of frontal ablation calibration data necessary for improved glacier model predictions.
Nora Gourmelon (Thu,) studied this question.