Image Processing Reference
pixels (rather than dealing with individual pixels), supporting geometrical, topological
and/or textural properties (Antunes et al. 2003 ). The need for image segmentation is
especially important in high-frequency images of urban and sub-urban settings since
we are dealing with very significant scene complexity and detailed structures.
Moreover, many urban features are primarily defined by their spatial inter-relationships
instead of spectral characteristics.
Traditional image analysis uses individual pixels as the basic unit in classification
of remotely sensed imagery. Pixels carry an integrated spectral signal and are charac-
terized by their ground sampling cell size. Essentially, the set of pixels making up an
image defines its entire amount of information. Nevertheless, working from individual
pixels is a limiting strategy, since pixels are not treated as building blocks within
their spatial context, but rather as independent samples (which they in fact are not!).
The alternative approach is to consider image objects made of homogeneous clusters
of adjacent pixels with meaningful geometric and other spatial properties. These
clusters promise a much richer and more powerful working environment throughout
the classification process (Blaschke and Strobl 2001 ). Spatial relationships describing
hierarchical ('vertical') or lateral pixel neighborhoods can be fully considered during
classification. Other strategies are implemented through the 'Expert Classifier' by
ERDAS Imagine ® /Leica-Geosystems or the 'Feature Analyst' from Visual Learning
Systems (Erdas Imagine ® 2004 and Visual Learning Systems 2004). The former
provides a rule-based approach to characterize pixels through the integration of
evidence and previous knowledge. The latter extracts features by inspecting certain
spatial arrangements of pixels representing the target features (Lang et al. 2003 ).
Reasons for Object-Based Classification
As pointed out above, shortcomings of pixel-based classification have stimulated new
classification concepts to be investigated. With pixels grouped into homogeneous
regions (image objects) by segmentation, these aggregates may reflect semantic objects
of interest from the real world in a more appropriate way than purely spectrally defined
agglomerations of single pixels. Classification starting from image objects rather than
individual pixels can utilize spatial and geometrical properties as well as relationships
among objects. Blaschke and Strobl ( 2001 ) make a case for the use of segmentation
algorithms to delineate objects based on contextual information in an image.
Using textural information and additional characteristics like the size or shape
of the objects the per-pixel classification process is complemented by a new image
processing technique. In particular, information from high resolution images is
being aggregated at multiple levels of detail, resulting in hierarchical sets of homog-
enous regions according to the given application semantic (Fig. 10.1 ). Different
scales of an image lead to vertically and horizontally interconnected objects that
could potentially enlarge the set of target classes.
Overall, object-based classification techniques are a new and innovative
approach, especially for applications handling human-made features. In the realm