Extrapolating Airborne Laser Scanner derived Sea Ice Classes to Synthetic Aperture Radar Measurements

Kurzfassung

Retrieval of sea ice types using Synthetic Aperture Radar (SAR) has the potential for near real-time and high resolution monitoring of the polar regions throughout the year. The core challenge of the idea has, since its inception, been the unavailability of in-situ measurements of the ice. Not only does this impede the rate at which progress can be made, progress itself becomes hard to recognise as we do not have extensive in-situ data to validate the results. Using helicopter-borne Airborne Laser Scanner (ALS) data and near coincident SAR acquisitions from the MOSAiC mission we can alleviate this problem with twenty matched product pairs. We compare five different deep learning classifiers and can establish their performances for ice types directly derived from the ALS measurement. We find a clear discrepancy between methods which are able to learn from the class distributions (segmentation) and those that cannot (classification). These findings infer a limit of the validity of classifiers trained from e.g. ice-charts, where the spatial class distributions are not covered well by the data. For the first time we can compute accuracies of modern classifiers for SAR-based retrieval, which are truly representative of their real-world performance.