On June 6, 2024, Konrad Heidler defended his thesis entitled Interpretation of Structures in Polar Regions
with Deep Learning Methods, supervised by Prof. Dr. Xiaoxiang Zhu. The defense committee was chaired by Prof. Dr. Jonathan Bamber. We particularly appreciate the engagement of the examiners Prof. Dr.-Ing. habil. Richard Bamler and Prof. Dr. Sébastien Lefèvre from Université Bretagne Sud!

Congratulations, Dr. Heidler!

--

Global climate change is rapidly transforming the polar regions. Remote sensing offers great potential for monitoring the developments in these often inaccessible areas. However, the amount of data is too large for manual analysis. Therefore, in an effort to support monitoring the changes in these regions, this dissertation develops deep learning methods for the remote sensing analysis of targets in these regions. More specifically, the developed models can automatically map the calving fronts of marineterminating glaciers and detect permafrost disturbances in the form of retrogressive thaw slumps (RTS). The fundamental research questions motivating this dissertation are:
1. How can domain knowledge about polar regions be encoded into deep learning models for remote sensing in these regions?
2. Are there more efficient ways of encoding such mapping tasks into deep learning tasks than the standard approaches used in computer vision like semantic
segmentation?
3. How can deep learning models generalise from limited labels to the entirety of the polar regions in a data-efficient way?
4. Will automatically derived observations allow polar science to better understand the developments in polar regions?
In an effort to answer these questions, this dissertation makes the following scientific contributions:
1. A model for mapping calving fronts in Antarctica is developed based on observations in the behaviour of human annotators. The resulting HED-UNet model combines semantic segmentation and edge detection and works on multiple resolution levels [1].
2. Questioning the representation of calving fronts through pixel-wise predictions altogether, a second model is developed. The COBRA model maps calving fronts in Greenland by directly predicting the desired contour lines instead of taking pixel-wise predictions as a proxy [2].
3. In a first downstream study, the COBRA model is applied for calving front detection in Svalbard. Thanks to the automated analysis, it was possible to derive a dataset of more than 100,000 calving front traces. This dataset allows for a better understanding of glacial processes such as the behaviour of surge-type glaciers [3].
4. The feasibility of applying deep learning methodology for the mapping of retrogressive thaw slumps in permafrost regions is established. By evaluating multiple deep learning architectures with regional cross-validation across the Arctic, spatial generalisation is identified as the main challenge [4].
5. In order to address the challenge of spatial generalisation in permafrost disturbance mapping, a data-efficient training routine is proposed. The PixelDINO approach is a method for semi-supervised learning, combining labelled data with unlabelled imagery. In this way, models are trained to generalise well across multiple regions in the Arctic [5].

[1] K. Heidler, L. Mou, C. Baumhoer, A. Dietz, and X. X. Zhu, “HED-UNet: Combined segmentation and edge detection for monitoring the antarctic coastline,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–14, 2022. doi: 10.1109/TGRS.2021.3064606.
[2] K. Heidler, L. Mou, E. Loebel, M. Scheinert, S. Lefèvre, and X. X. Zhu, “A deep active contour model for delineating glacier calving fronts,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–12, 2023. doi: 10.1109/TGRS. 2023.3296539.
[3] T. Li, K. Heidler, L. Mou, Á. Ignéczi, X. X. Zhu, and J. L. Bamber, “A highresolution calving front data product for marine-terminating glaciers in Svalbard,” Earth System Science Data Discussions, pp. 1–28, 2023. doi: 10.5194/essd-2023-396.
[4] I. Nitze, K. Heidler, S. Barth, and G. Grosse, “Developing and testing a deep learning approach for mapping retrogressive thaw slumps,” Remote Sensing, vol. 13, no. 21, p. 4294, 2021. doi: 10.3390/rs13214294.
[5] K. Heidler, I. Nitze, G. Grosse, and X. X. Zhu, “PixelDINO: Semi-supervised semantic segmentation for detecting permafrost disturbances,” IEEE Transactions on Geoscience and Remote Sensing (in review), 2023. doi: 10.48550/arXiv.2401.09271.