Determining the number of classes for segmentation in SAR sea ice imagery

In this paper, we describe a segmentation technique for SAR sea ice imagery that determines the number of classes in the image without a priori knowledge of the characteristics of the image. Image segmentation is important to sea ice research such as classification, and floe and lead analyses. In SAR sea ice imagery, however, backscatter characteristics vary for different seasons, temperatures, wind activity, and geographical locations, etc. As a result, image processing techniques that pre-determine the number of classes could generate segmentation that contains erroneous merging of classes and/or unnecessary separation of a class leading to unrecoverable mistakes during the classification phase. We have designed an image segmentation technique that combines image processing and machine learning methodologies. It computes spatial and textural statistics from the image and determines the number of classes by conceptually clustering these statistics. We have also tested this technique on a large database of sea ice imagery, and it has shown successes in determining the number of classes without human intervention