An image segmentation technique to infer the outlines of icebergs, depicted in satellite images, from their shadows and bright sunlit surfaces

Describes a segmentation technique to identify and outline icebergs depicted in Landsat images, when the icebergs are locked in sea ice. Conventional multispectral classification, combined with a simple segmentation process, can be applied when the iceberg is in open water because it is clearly distinguishable from the water. However, when locked in sea ice, it is spectrally very difficult to distinguish from the surrounding ice, and so other cues, such as its shadow and its sunlit facet, must be used to infer the outline of the berg in the image. Thresholding is used to detect the bright side and dark shadow of each iceberg, and these regions are labelled using a simple segmentation technique. They are then linked using spatial reasoning and the characteristics of the linked regions are used to infer the size and shape of the complete iceberg. The technique has been tested on part of a Landsat image depicting a number of well-formed icebergs locked in sea ice and it was capable of correctly delineating approximately 60% of the icebergs depicted. The major limitation was its inability to separate icebergs in contact with each other, and current work is aimed at overcoming this problem