ABSTRACT
Object-based image analysis (OBIA), as the name indicates, operates on objects representing real-world entities as the constituents of our geographic reality. Spatial key features (basically size and shape) are intrinsic parameters to be considered in the information extraction process, making spatial characteristics an additional feature domain in classi¦cation. Using regionalization techniques for image segmentation, there is a complementarity (one could say a trade-o§) between spectral and spatial similarities, that is, between color and neighborhood. In other words, the spatial constraint balances the spectral behavior, which leads to (scalable) generalization and a reduction of the so-called salt-andpepper e§ect (Bischof et al. 1992, Blaschke and Strobl 2001). In OBIA, the category that a group of picture elements is assigned to depends on both spatial and spectral characteristics. Ÿis may sound like a restriction narrowing down classi¦cation power, but it actually extends the set of target classes to be addressed. More precisely, it opens another dimension of potential target classes (Lang 2008), on two additional semantic levels next to the level of spectral classes (see Figure 15.1). Ÿis comprises, ¦rst, all subcategories or instances of a spectral class that are de¦ned by shape features on individual object level: a class <water> may split into <lake | river> depending on its length/width ratio. A class <builtup> may split into several subcategories of a village typology based on footprint physiognomy. Second, and even more crucial, we ¦nd spatial properties in terms of relations among objects. We will come back to this issue in Section 15.2, when discussing the term class modeling. In a nutshell, the integration of object relationships (including relative coverage, distance, and in particular topological features) allows for addressing complex, composite target classes on high semantic level.
