Automated Detection of Sea Ice and Open Water from RADARSAT-2 Images for Data Assimilation

Spaceborne synthetic aperture radar (SAR) high resolution (at or below 50 m) images from the ongoing satellite missions such as Canadian RADARSAT-2 currently provide the most reliable information on sea ice conditions. Dual-polarization RADARSAT-2 HH-HV ScanSAR imagery is the main source of data for operational production of Ice Charts at the Canadian Ice Service (CIS). The upcoming Canadian RADARSAT constellation mission equipped with three SAR platforms will further increase the amount of SAR data over the Arctic region. One of the key applications of the large amount of SAR images is assimilation of observations over sea ice into coupled ocean-sea ice-atmosphere numerical models, such as the Regional Ice-Ocean Prediction System being implemented at the Environment and Climate Change Canada.

Automated detection of sea ice and open water from SAR data is very important as an initial step to assimilate SAR data into numerical models. However, a high variability of the normalized radar cross-section (NRCS) from sea ice and open water often leads to overlapping their NRCS signatures, which makes it challenging to unambiguously separate sea ice and open water. Conventional classification approaches based on various learning techniques are found to be limited by the fact that they typically do not indicate the level of confidence for ice and water retrievals. Meanwhile, only ice/water retrievals with a very high level of confidence are allowed to be assimilated into sea ice models to avoid propagating and magnifying errors into the numerical prediction system.

The main objective of this study is to develop a new technique for ice and water detection from dual-polarization RADARSAT-2 HH-HV images which provides the probability of ice/water at a given location.

We collected many hundreds of thousands of SAR signatures over 100% sea ice concentration areas with various sea ice types (i.e. new, grey, first-year, and multi-year ice, and their mixtures) and over 0% sea ice concentration areas (open water) from all available RADARSAT-2 images and the corresponding Canadian Ice Service Image Analysis products over the period from November 2010 to May 2016. Our analysis of the dataset revealed that ice/water separation can be effectively performed in the space of SAR-based variables independent of the incidence angle and noise floor (such as texture measures) and auxiliary Global Environmental Multiscale Model parameters (such as surface wind speed). Results of the conducted analysis and the choice of the predictor variables will be specifically discussed in the presentation. An ice probability empirical model as a function of the selected predictor variables was built in a form of logistic regression, based on the training subset from November 2012 to May 2016. The developed ice probability model showed very good performance on the independent testing subset (November 2010 – October 2011). With the ice/water probability threshold of 0.95 reflecting a very high level of confidence, 87.4% of the testing ice samples were classified with the accuracy of 99.1%, and 70% of the testing open water samples were classified with the accuracy of 99.5%. The developed technique for detecting ice and open water could be further extended to automatically detect sea ice types (e.g. first-year and multi-year) from RADARSAT-2 images.