An evaluation of an object-oriented fuzzy analysis for land cover classification on Macquarie Island

Macquarie Island has been severely threatened by invasive species such as rabbits and rodents, which were first introduced by sealers in the 1880s. Because of this impact the eradication of rabbits and rodents has been planned for the near future. In evaluating the progress and success of the eradication program on Macquarie Island, spatial information in the form of a land cover map on Macquarie Island from remote sensing data should be useful due to its inaccessibility. This study therefore proposes a new approach for mapping Macquarie Island's vegetation communities based on very high resolution QuickBird imagery acquired on 18 March 2007 and object-oriented classification technique. The VHR satellite data has increased in geometric detail and accuracy, but it has some problems in traditional pixel-based classification. With rich spatial and spectral information from the spatial refinement, the internal variation in a class increases. On the other hand, the object-oriented classification first segments the image in to meaningful segments or objects, and then assigns classes to those objects based on their spectral and spatial characteristics. Moreover, the object-oriented classification can consider geographical features of objects, topological entities, and spectral statistical features. Three settings of object-oriented classifications were applied to the image and results were compared to three pixel-based classifications (Minimum distance to mean, Maximum likelihood and Support vector machine (SVM) classifications). The first object-oriented classification was carried out based on only spectral bands. The second was operated with spectral bands and the hillshade layer which was obtained from DEM. The third was conducted with spectral bands, the hillshade and NDVI layers. The highest accuracy result was yielded by the third object-oriented classification. The next highest was the result of the SVM classification. The third highest was the first object classification. Object-oriented classification could achieve relatively higher accuracy results. Despite its high accuracy, the result of the SVM classification still remained noisy. The best setting for object-oriented classification could not be achieved, and although it was more suitable to create maps in this study, this classification has huge possibilities.