An efficient image segmentation algorithm for landscape analysis

Abstract

Widespread development and use of object-based GIS in the environmental sciences has stimulated a rapid growth in demand for parcel-based land cover data. Despite the fact that image segmentation techniques applied to remotely sensed data offer the most effective and direct approach to generating such data their use is still restricted to specialist applications. This paper describes a general purpose segmentation algorithm capable of creating parcel boundaries from a wide range of image types. A brief review of image segmentation in a range of disciplines identifies key elements of a successful segmentation algorithm. The structure and implementation of the algorithm is then described and its performance is illustrated using Landsat ETM imagery of Eastern England. Comparison of the segmentation product generated by the algorithm with those generated by independent human analysts demonstrates that the computer algorithm and the manually derived products have just less than eighty percent correspondence. Most of the differences stem from the more detailed results achieved by the segmentation algorithm.