In this paper we present a feature extraction technique that extracts ice floes for Synthetic Aperture Radar (SAR) sea ice imagery. We investigate two types of ice floes: 1) independent ice floes that collide and meet, and 2) component ice floes that melt and consolidate to form an independent ice floe. To detect independent ice floes, we consider two kinds of edges: 1) clear edges, and 2) blurred edges. We use a spatial enhancement technique combined with thresholding to obtain the clear edges. To detect the blurred edges, we make use of `corners' which are pixels where a considerable change in direction of a floe boundary occurs. To connect the corners appropriately, we incorporate both their geometric and semantic properties into heuristics that choose which pair of corners to connect. To investigate component ice floes, we have developed a technique that combines thresholding, correlation, morphological cleaning, and structural growing. Our technique has been tested successfully on a large number of SAR images, including data from ERS-1.