Self-supervised representation learning of radar altimeter waveforms: advancing sea ice surface classification as a downstream task

This thesis investigates whether self-supervised representation learning can be used to extract compact, informative representations of radar altimeter waveforms that improve the downstream classification of sea ice surface types.
In the context of polar research, satellite radar altimeter data are used for the retrieval of sea ice thickness, an important input to climate models. This could be substantially improved by a classification of surface types such as new ice (NI), first-year ice (FYI), multi-year ice (MYI), and open water from the same data, allowing a better parameterization of the retrieval process. However, traditional waveform representations are limited to a small set of parameters, potentially discarding richer information about sea ice surface properties contained in the radar altimeter waveforms. At the same time, the effective use of this information is hindered by high-dimensional signal representations, noise, and the limited availability and reliability of labeled training data in polar regions. This motivates the exploration of self-supervised representation learning to distill the relevant information as the elementary step towards a novel multi-category sea ice and ocean surface classification directly from satellite radar altimeter data.
Following a data-driven progression from simple to more advanced methods, the study begins with an exploratory analysis using established dimensionality reduction techniques, including Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP), to assess class separability in traditional waveform parameter spaces and in the full waveform domain. These analyses demonstrate that while waveform parameters reliably distinguish leads from sea ice, the separation of finer-grained ice classes remains challenging.
Building on these insights, self-supervised representation learning approaches are systematically developed and adapted to the specific characteristics of radar altimeter data. Auto-encoders and variational auto-encoders are first employed to learn lower-dimensional latent representations directly from waveforms, showing that reconstruction-based objectives capture additional information beyond traditional waveform parameters and yield improved classification performance when combined with those parameters.
To further enhance class separability, the thesis introduces and evaluates a contrastive learning framework tailored to radar altimetry, including a novel application-specific definition of positive pairs and a custom batch sampling strategy. The results demonstrate that self-supervised contrastive learning produces well-structured, low-dimensional embeddings that significantly outperform traditional waveform parameter baselines in the task of sea ice surface classification, increasing the accuracy by 4%. Combining learned embeddings with traditional waveform parameters yields an additional performance gain of 7%. Furthermore, numerical experiments identify clear bounds for the optimal dimensionality of the embedding space.
The findings confirm that radar altimeter waveforms encode discriminative information beyond established waveform parameter representations and that self-supervised learning provides an effective means to access this information despite limited and uncertain labels. Overall, this thesis advances the state of the art in radar altimetry based sea ice classification and establishes self-supervised representation learning as a powerful tool for extracting physically meaningful information from complex geophysical signals, with direct relevance for future sea ice thickness retrieval pipelines and broader applications in marine and climate data science. At the same time, it opens up a new applicational domain for self-supervised learning and, in particular, contrastive learning approaches.