Fuzzy regions for handling uncertainty in remote sensing image segmentation

Abstract

Increasing availability of satellite imagery is demanding robust image classification methods to ensure a better integration between remote sensing and GIS. Segmentation-based approaches are becoming a popular alternative to traditional pixel-wise methods. Hard segmentation divides an image into a set of non-overlapping image-objects and regularly requires significant user-interaction to parameterise a functional segmentation model. This paper proposes an alternative image segmentation method which outputs fuzzy image-regions expressing degrees of membership to target classes. These fuzzy regions are then defuzzified to derive the eventual land-cover classification. Both steps, fuzzy segmentation and defuzzification, are implemented here using simple statistical learning methods which require very little user input. The new procedure is tested in a land-cover classification experiment in an urban environment. Results show that the method produces good thematic accuracy. It therefore provides a new, automated technique for handling uncertainty in the image analysis process of high resolution imagery. © 2008 Springer-Verlag Berlin Heidelberg.