Hierarchical segmentation using a composite criterion for remotely sensed imagery

In this paper, an image segmentation algorithm based upon hierarchical step-wise optimization with a composite merge criterion is presented. In hierarchical step-wise optimization, at each step, the two segments which optimize a criterion/cost function are found and merged. The main innovation proposed in this paper is that different criteria are employed at different stages in the hierarchical process. At the lowest stage, when the segment size is still small, the segment mean is the main information and is used in the merge criterion. For the intermediate stages, with increasing segment size, the mean is no longer sufficient to describe the characteristics of a segment and, therefore, a criterion related to the mean and the variance is considered. At the final stage, additional information, such as the edge information is included in the criterion. In other words, with increasing segment size, more information is required to describe the characteristics of the segments and is incorporated into a composite criterion. Experimental results on a Landsat image show that improved segmentations can result when a composite criterion is employed.