Object vs. pixel: A systematic evaluation in urban environments

The objective of this paper is to compare object-based and per-pixel classifiers in a systematic manner using high resolution urban imagery. The prevailing opinion is that object-based methods perform better than single-pixel classifiers, but there has been no formal investigation of this claim using multiple images and identical training samples in a detailed land-cover classification. Furthermore, there has been no standardized study of how different object-based segmentation and scale parameters improve high resolution urban classifications. We used two subsets of QuickBird over Phoenix and Scottsdale, Arizona, to test these issues. Our results show that small-scale segmentation (10) produces higher accuracy. A combination of equally balanced shape and spectral homogeneity (0.5) with compactness parameter of 0.5 is the most effective for image segmentation. The highest overall accuracy was achieved using a per-pixel Minimum distance classifier, but it was only marginally more accurate than the object-based classification.