Sar Ice Classification Using Fuzzy Screening Method

A semi-automatic SAR sea ice classification algorithm is described. It is based on combining the information in the original SAR data with those in the three ‘image’ products derived from it, namely Power-to-Mean Ratio (PMR), the Gamma distribution and the second order texture parameter entropy, respectively. The latter products contain information which is often useful during the manual interpretation of the images. The technique used to fuse the information in these products is based on a method called Multi Experts – Multi Criteria Decision Making fuzzy screening. The Multiple Experts in this case are the above four ‘image’ products. The two criteria used currently for making decisions are the Kolmogorov-Smirnov distribution matching and the statistical mean of different surface classes. The algorithm classifies an image into any number of predefined classes of sea ice and open water. The representative classes of these surface types are manually identified by the user. Further, as SAR signals from sea ice covered regions and open water are ambiguous, it was found that a minimum of 4 pre-identified surface classes (calm and turbulent water and sea ice with low and high backscatter values) are required to accurately classify an image. Best results are obtained when a total of 8 surface classes (2 each of sea ice and open water in the near range and a similar number in the far range of the SAR image) are used. The main advantage of using this image classification scheme is that, like neural networks, no prior knowledge is required of the statistical distribution of the different surface types. Furthermore, unlike the methods based on neural networks, no prior data sets are required to train the algorithm. All the information needed for image classification by the method is contained in the individual SAR images and associated products. Initial results illustrating the potential of this ice classification algorithm using the RADARSAT ScanSAR Wide data are presented and its possible extension to fuse the information in these data with the ENVISAT ASAR image products is also discussed.