Discrimination of First Year Sea Ice Features Using Polarimetric SAR Data

In this study, we evaluate the capability of spaceborne fully polarimetric C-band synthetic aperture radar (SAR) image data for discrimination of First Year Sea Ice (FYI) features. We introduce the combination of supervised and un-supervised classification approach for studies with limited ground truth information. ISODATA unsupervised classification method is used for spectral segmentation, which subsequently serves as training and validation resource for supervised Maximum likelihood (ML) classification and accuracy assessment, respectively. The study uses high resolution; quad-polarized RADARSAT-2 data acquired during the early spring period in April, 2009 on Hudson Bay, Churchill. Our results show a successful discrimination of seven spectral classes using unsupervised approach, grouped together to achieve four ice feature classes namely smooth ice, rough ice, deformed ice and open water mixed with leads. Using the supervised Maximum Likelihood approach, the SAR data unambiguously separates between the same four ice classes with a significant accuracy. We observe that co-polarized channels produce better classification accuracy when used in single or in combination to each other than used in combination with cross-polarized channel.