Sea Ice Concentration Estimation and Ice Type Classification from Dual-Frequency Satellite Synthetic Aperture Radar

Licentiate thesis, 2017

The sea ice cover in the Arctic has undergone dramatic changes in recent years. The perennial sea ice extent is decreasing by 12.2 % per decade while annual mean sea ice thickness has decreased by more than 2 m for the central Arctic Basin from 1975 to 2012. High resolution information of the ice cover is necessary for a better understanding of the involved processes. Furthermore increased economic, scientific and touristic activities in the Arctic demand ice information for safer navigation in ice infested waters. Satellite synthetic aperture radar facilitates year round monitoring of the sea ice cover with high spatial and temporal coverage. High resolution is a requirement to capture small scale sea ice features like leads and the dynamics of the ice cover driven by the atmosphere and ocean. This thesis presents investigations on sea ice characterization from multi-spectral SAR imagery. Dual-polarization C- and L-band images from Sentinel-1 and ALOS PALSAR-2 have been used to derive sea ice concentration, for creation of ice-water maps and ice type classification. The developed algorithms for sea ice concentration estimation and ice/water classification use spatial autocorrelation as a texture feature to improve the discrimination of ice and water. The mapping between image features and the output variable is realized with a neural network. The proposed algorithms show good performance when evaluated against manually derived ice charts and radiometer data. We demonstrate that C- and L-band contain complementary data and a combination of these frequencies could achieve more robust classification results. Furthermore the separability and signatures of ice types in different ice regimes, i.e. marginal ice zone, pack ice and areas containing fast ice, have been investigated. Classification only based on backscatter intensities has been carried out by means of a support vector machine on selected examples of the same C- and L-band dataset. The results indicate that also for ice type classification a combination of frequencies can improve the classification accuracy.