Towards representation learning of radar altimeter waveforms for sea ice surface classification

Abstract. Satellite radar altimeters provide crucial insights into polar oceans and their sea ice cover, enabling the estimation of sea level, sea ice freeboard, and thickness. These retrieval algorithms depend on accurate discrimination between radar altimeter waveforms from sea ice and ocean surfaces in heterogeneous and dynamic surface conditions. A further and less mature step is classifying different sea ice types in addition to the ice/ocean discrimination. We aim to develop new methods for a novel multi-category sea ice and ocean surface classification directly from satellite radar altimeter data to improve sea ice climate data records. Traditional waveform representations are limited to a small set of parameters, leading to information loss. Moreover, machine learning models for sea ice classification often depend on supervised training, which is vulnerable to uncertainties in labeled data, especially in polar regions. To address these limitations, we explore self-supervised learning methods to optimize waveform representations, which can capture more detailed information for a classification with finer granularity. Furthermore, they do not require labeled data, which is not available at the spatial coverage and resolution of radar altimeter waveforms. We apply these techniques to SRAL data from the Sentinel-3 mission. We show that the information preserved in the latent space of an auto-encoder enhances the feature space of traditional waveform parameters, improving the subsequent classification process, when comparing our results to available sea ice charts and other remote sensing products. Our results demonstrate better generalization compared to supervised approaches.