Fuzzy Image Segmentation for Urban Land-cover Classification

In this paper a general fuzzy approach for segmentation-based classification is proposed. Traditional segmentation techniques focus on partitioning imagery into image-objects with well-defined boundaries. Instead, the proposed methodology aims to produce and analyze fuzzy image-regions expressing degrees of membership to different target classes. This approach, called Fuzzy Image-Regions Method (FIRME), is suitable to deal with the spectrally and spatially complexity of urban landscapes. The main stages of the FIRME approach are described, including techniques to produce such regions, alternatives to measure region attributes, and a number of methods for defuzzification. The FIRME method is tested for an urban classification experiment using multi-spectral imagery from Bogota, Colombia. Results suggest that, in complex environments, the FIRME method may be a suitable alternative to hard segmentation as it performs well in discriminating between spectrally mixed geographic objects.