An assessment of the effectiveness of segmentation methods on classification performance

Abstract

Object-based classification approaches have been recently employed successfully in many research studies. These approaches aim to create segments on the image considering spectral similarity of the neighboring pixels, which is known as image segmentation. Segmentation methods use spectral information as well as textural and semantic information of the pixels. It is a fact that parameter setting of segmentation methods is of considerable importance in producing accurate classification results. Therefore, determining optimum values for the parameters is regarded as a critical stage in segmentation processes. In this study, effectiveness and applicability of the segmentation approach was analyzed utilizing a high resolution Quickbird satellite image. Multi-resolution segmentation technique, which has been reported to be a robust method, was employed with its optimal parameters for scale, shape and compactness that were defined after an extensive trail process on the data set. Resulting image object was then used in supervised classification using the nearest neighbor algorithm with fuzzy membership functions. Classification performances produced for different parameter settings were thoroughly analyzed and it was found that parameter setting in segmentation applications produced highly varied classification accuracies. It was also observed that segmentation algorithms could help to improve spectral discrimination, particularly for spectrally similar classes. © 2016 Elsevier B.V., All rights reserved.