Abstract

Sea ice property retrieval from synthetic aperture radar (SAR) at high resolution has one central problem: there are no comprehensive ground truth data available. One candidate instrument to offer such ground truth is the ICESat-2 altimeter, which has excellent vertical resolution. Due to the ice motion however, both instruments have to be acquiring within a short time difference for the measurements to still be correlated. Because the ICESat-2 altimeter uses a laser at optical wavelengths, cloud free conditions are also needed. As a result, valuable coincident measurements are sparse. Deep learning methods, that have shown promising results for ice classification in the past, particularly suffer from sparse training data. In this paper we propose an architecture agnostic transfer learning and autoencoding approach to overcome the problem of data sparsity. A major component of this approach is using local incidence angle dependencies as a proxy for ice classes. Thus, we can do a majority of the deep learning model training using self-supervised training on data without labels. Using this technique, altimetry derived ice development can be meaningfully extrapolated to Sentinel-1 SAR acquisitions.