Supervised Segmentation by Region Merging

Summary

A region merging segmentation method is combined with supervised classification to delineate and categorize objects in the terrain on the basis of remotely sensed imagery. The segmentation algorithm is a hybrid between region growing and split and merge. The algorithm makes a recursive, bottom up quadtree traversal, which starts at single pixels (or larger quadtree leaves in which the pixel values are constant) and recursively merges adjacent regions, forming irregularly shaped segments at all stages. Order dependency problems are solved by performing several iterations, while slowly relaxing the homogeneity criteria until a user defined degree of segmentation is reached. By making the algorithm output segmentations at several threshold levels, a segmentation pyramid is created. Combining this with supervised classification allows for the selection of segments from different pyramid levels, which yields a partitioning of the space into objects, while preventing object fragmentation and object merging.