Discrimination of different sea ice types from CryoSat-2 satellite data using an Object-based Random Forest (ORF)

Sea ice type is one of the most sensitive variables in Arctic sea ice monitoring, and it is important for the retrieval of ice thickness. In this study, we analyzed various waveform features that characterize the echo waveform shape and Sigma0 (i.e., backscatter coefficient) of CryoSat-2 synthetic aperture radar altimeter data over different sea ice types. Arctic and Antarctic Research Institute operational ice charts were input as reference. An object-based random forest (ORF) classification method is proposed with overall classification accuracy of 90.1%. Accuracy of 92.7% was achieved for first-year ice (FYI), which is the domain ice type in the Arctic. Accuracy of 76.7% was achieved at the border of FYI and multiyear ice (MYI), which is better than current state-of-the-art methods. Accuracy of 83.8% was achieved for MYI. Results showed the overall accuracy of the ORF method was increased by ?8% in comparison with other methods, and the classification accuracy at the border of FYI and MYI was increased by ?10.5%. Nevertheless, ORF classification performance might be influenced by the selected waveform features, snow loading, and the ability to distinguish sea ice from leads.