This study employs Synthetic Aperture Radar (SAR) imagery from the Marginal Ice Zone Experiment (MIZEX) 1987 to identify an optimal set of statistical descriptors that accurately classify three types of ice (first-year, multiyear, odden) and open water. Two groups of statistics, univariate and texture, are compared and contrasted with respect to their skill in classifying the ice types and open water. Individual statistical descriptors are subjected to principal component analysis and discriminant analysis. Principal component analysis was of little use in understanding features of each ice and open water group. Discriminant analysis was valuable in identifying which statistics held the most power. When combined, univariate and texture statistics classified the groups with 89.5% accuracy, univariate alone with 86.8% accuracy and texture alone with 75.4% accuracy. Range and inertia were the strongest univariate and texture discriminators with 74.6% and 50.8% accuracy, respectively. Despite the use of a non-calibrated SAR, univariate statistics were able to classify the images with greater accuracy than texture statistics.
http://archive.org/details/seiceclassificat1094527745
Lieutenant Commander, United States Navy
Approved for public release; distribution is unlimited.