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.

The incidence angle dependence of synthetic aperture radar (SAR) backscatter from sea ice has been observed to differ between ice types. This is indicative of its complex dielectric and backscattering properties and cannot be analytically determined or easily modelled. Without assumptions this dependence can only be measured from SAR through multiple observations of the same ice. Due to the sea ice drift, doing so at high resolution is time inefficient and to some extent only feasible for land fast ice. Thus, the incidence angle dependence can currently not be easily related to ground measurements or be fully exploited for classification tasks. By reformulating the problem of incidence angle dependence to a domain transfer task and making use of known physical relations, we can train a Generative Adversarial Network (GAN) to predict the incidence angle dependence from a single Sentinel-1 satellite scene patch-wise using no only local data. The network’s predictions are validated for multiyear and first-year ice backscatter using coincident ICESat-2 altimeter measurements. The network suggests a diverse incidence angle dependence in the HV channel, that so far have been observed only scarsely. The results of this study are a novelty in sea ice remote sensing insofar that artificial intelligence manages to solve a task that would be nearly impossible for a human observer.