Fuzzy Image Segmentation for Urban Land-Cover Classification

A main problem of hard image segmentation is that, in complex landscapes, such as urban areas, it is very hard to produce meaningful crisp image-objects. This paper proposes a fuzzy approach for image segmentation aimed to produce fuzzy image-regions expressing degrees of membership of pixels to different target classes. This approach, called Fuzzy Image-Regions Method (FIRME), is a natural way to deal with the inherent ambiguity of remotely sensed images. The FIRME approach comprises three main stages: (a) image segmentation which creates fuzzy image-regions, (b) feature analysis which measures properties of fuzzy image regions, and (c) classification which produces the intended land-cover classes. The FIRME method was evaluated in a land-cover classification experiment using high spectral resolution imagery in an urban zone in Bogota, Colombia. Results suggest that in complex environments, fuzzy image segmentation may be a suitable alternative for GEOBIA as it produces higher thematic accuracy than the hard image segmentation and other traditional classifiers.