Glacier Calving Front Delineation in SAR Imagery Using Deep Learning
Retreating glaciers are one of the most visible signs of climate change, and modeling how glaciers evolve lets researchers predict future ice loss and its societal impact. But current models for tidewater glaciers lack the calibration data on frontal ablation (ice loss at the glacier’s edge) needed for accurate predictions, and getting that data requires precisely tracking calving front positions over time. Synthetic Aperture Radar (SAR) satellite imagery can monitor glaciers continuously, in any weather, but the resulting data volumes make manually tracing calving fronts by hand infeasible. This thesis tackles automated, deep-learning-based delineation of calving fronts in SAR imagery, establishing a standardized benchmark for the field, assessing where the state of the art actually stands, and then surpassing it with performance close to human accuracy.
The benchmark is built on a public, multi-mission SAR dataset of calving fronts with both binary front segmentation and multi-class landscape labels, using two open-source models as baselines and a test set specifically designed to check how well models generalize to glaciers they haven’t seen before. The thesis first improves on these baselines by adding a Conditional Random Field into post-processing, which connects broken-up front segments and suppresses spurious ocean predictions. Building on this, 22 deep learning models are systematically retrained and evaluated on the benchmark, with performance differences analyzed for statistical significance and effect size. Further analysis looks at how architectural and methodological choices affect performance, and a multi-annotator study reveals a 183-meter gap between human accuracy and the best existing model, pointing to where both humans and models struggle to interpret the imagery.
Guided by these findings, the thesis develops a new state-of-the-art model: a hybrid Transformer-CNN architecture that produces smooth delineations rather than the jagged predictions typical of earlier methods, using larger patches while keeping only the central portion of the output to retain enough context. To handle the mismatch between ImageNet-pretrained weights and SAR data, the model is pretrained on a new, unlabeled SAR dataset of glacier time series, each paired with a single optical image, using two self-supervised multi-modal pretraining strategies that use the optical images as labels. Because these strategies don’t require the SAR and optical images to be captured at exactly the same time, they cut down on manual data curation and let every available SAR image contribute to training. An ensemble of these pretrained models reaches near-human performance, with a 75-meter error compared to the human consensus error of 38 meters.
By pushing automated calving front delineation this close to human accuracy, the thesis lays the groundwork for global-scale automated glacier monitoring, and for generating the frontal ablation calibration data that better glacier models will need.