7. Conclusion
We have shown that the traditional per-pixel approaches were not
very effective in identifying urban land-cover classes. This was proven
by the classification of the entire QuickBird image using the most
widely used classifier namely maximum likelihood rule and spectra of
the selected land cover classes generated from the QuickBird using
discriminant analysis. The discriminant analysis of spectra information
received an overall accuracy of 63.33%. The object-based classifier
produced a significantly higher overall accuracy (90.40%), whereas
the maximum likelihood classifier produced 67.60%. Segmentation
procedures and scale levels employed to identify objects of different
classes were found to be relevant. The same classification procedures
and classification accuracy assessment employed to classify the same
classes using the test image confirms that the object-based approach
(95.2%) outperforms the traditional classifier (87.8%). This study
reveals that it is more difficult to achieve higher accuracies for larger