Data-Centric Benchmarking of Deep Learning Models for Antarctic Coastline Segmentation using Earth Observation Imagery

Kurzfassung

The Antarctic region plays a crucial role in the earth’s climate system, acting as a key regulator of global sea levels, ocean circulation, and atmospheric dynamics. Accurate delineation of the Antarctic coastline is therefore essential for understanding polar processes and their far-reaching impacts. Advances in satellite remote sensing, particularly Synthetic Aperture Radar (SAR), have enabled continuous and up-to-date observation of this remote and highly dynamic environment. At the same time, recent advances in deep learning have demonstrated how this technology can be successfully used to delineate the Antarctic coastline. However, extracting reliable coastline information from large volumes of earth observation imagery remains a demanding task due to complex and ambiguous semantic patterns. The rise of deep learning is largely due to increased computing power and continuously improved and more sophisticated model architectures. Nevertheless, it is frequently overlooked that the performance of a deep learning model is fundamentally constrained by the quality of its input data, and that the choice of dataset has a decisive impact on the resulting outcomes. For this reason, there has been a growing focus on datacentric deep learning in recent years. This thesis investigates how data-centric deep learning strategies can improve Antarctic coastline segmentation by using a comprehensive and diverse benchmark dataset of multi-sensor SAR imagery. An existing deep learning framework was adopted and further extended to support unified training and evaluation, including accuracy measures for semantic segmentation and the geometry-aware Mean Distance Error (MDE) specifically designed for Antarctic coastline applications. Within this environment, UNet and HED-UNet were benchmarked under identical conditions. Both architectures achieved comparably strong segmentation performance, although, according to the mean MDE, HED-UNet demonstrated more robust and slightly higher geometric accuracy in detecting the coastline. Building on this foundation, the patch-size experiment examines how spatial context, controlled through patch-size selection, affects segmentation accuracy in the environment of Antarctic coastlines. A systematic analysis revealed that performance of the UNet improves significantly with increased spatial context when trained on larger patches. However, an even greater performance improvement viiis observed when trained models of both architectures are applied to 1024-pixel patches instead of 256-pixel patches. Inference with larger patches yields to more stable and consistent segmentation results, reducing artifacts and minimizing iceberg misclassifications. To assess the influence of training data characteristics, two data-centric methods (Error L2-Norm and Data Maps) were adapted for pixel-wise segmentation and used to rank all dataset patches. Easy-patch pruning showed that large portions of the dataset can be removed with only small performance loss. Simultaneously, this can reduce both the computational and time resources during model training. Within hard-patch pruning, both methods were able to identify harmful or noisy samples. Removing these samples results in more consistent training and better performance for delineating the Antarctic coastline. In addition, Error L2-Norm demonstrates that it can reliably detect label errors that were unexpectedly present in the dataset. Overall, the results show that data-centric deep learning can significantly improve the accuracy, robustness, and effciency of Antarctic coastline segmentation. The conducted experiments demonstrate how the full potential of conventional earth observation datasets can be better exploited. Furthermore, this work highlights the potential of applying data-centric methods to this task and provides a solid foundation for further research. Moreover, the findings reveal the strong influence of spatial context on model behaviour and its ability to reliably predict the Antarctic coastline. These insights underscore the importance of high-quality training data for large-scale coastline mapping. They also contribute to a research field that aims to deepen the understanding of Antarctic environmental dynamics and to strengthen the ability to predict global climate processes.