We present a supervised three-stage mapping (labeling) scheme applied to SAR images of polar regions for detecting frazil/pancake ice and open water. The three-stage labeling procedure consists of: 1. a speckle noise filtering stage, based on a sequence of contour detection, segmentation and filtering steps, which removes SAR speckle noise (and texture information as well) without losing spatial details; 2. a second stage providing Bayesian, maximum-a- posteriori, hierarchical (coarse-to-fine), adaptive (data-driven) and contextual labeling of smooth images featuring little useful texture information, i.e., piecewise constant or slowly varying intensity images that may be corrupted by an additive white gaussian noise field independent of the scene; and 3. an output stage providing a many-to-one relationship between second stage output categories (types or clusters) and desired output classes. Modules 1. and 2., which demonstrated their validity in several applications in the existing literature, are briefly recalled in the current paper. The proposed labeling scheme features some interesting functional properties when applied to sea-ice SAR images: it is intuitive to use, i.e., it requires minor user interaction, is robust to changes in input conditions and guarantees satisfactory classification performances. Application results are presented and discussed for a pair of SAR images extracted from an an ERS-2 scene of the East Greenland Sea, Odden Ice Tongue region, acquired on March 8, 1997, at the time when a field experiment by the research vessel “Jan Mayen” was conducted in the same area.