Image Processing Reference
In-Depth Information
facets in both anthropogenic and natural settings. Accordingly, we are able to detect
more categories of geographical features and the principle set of classes addressed
is being extended. With refined image resolution there likewise is an increasing
demand for adequate image analysis and interpretation
techniques. A significant amount of structure-related infor-
mation is revealed by high-resolution imagery (see a related
review in Chapter 7). This additional information can be
used to describe classes with properties beyond their spec-
tral reflectance. By complementing the spectral behavior of
geographical features with their spatial characteristics we
are bridging remote sensing techniques and GIS functional-
ity. The term 'semantic class definition' has been suggested
to signify structural and geometrical properties of image
features. This term helps us to conceptually distinguish between 'target classes' and
'ancillary classes'. The former addresses the final set of classes the user wants to
report as a result from image analysis. The latter comprises all intermediate catego-
ries used during the processing of an image, which are used to describe the internal
structure of the target classes. This kind of semantic class definition of geographical
features requires (1) representations of the scene over several scales and (2) rules to
integrate heuristics about these representations. A toolset supporting this approach
is implemented in the image analysis software eCognition from Definiens GmbH,
Munich (eCognition User Guide 2004). This chapter explains the rationale for
object-based image analysis, presents and discusses image segmentation techniques,
and finally demonstrates examples from urban scene classifications.
a definition of
semantic classes
combines the
spectral behavior
of geographical
features with
their spatial
characteristics
10.2
Image Segmentation
10.2.1
Limitations of Pixel-Based Classification Techniques
Most of the methods for image processing developed since the early 1970s until
present are based on classifications of individual pixels clustered in a multi-dimen-
sional feature space (Chapter 8). Although a range of
sophisticated and now well established techniques have
been developed, the current demand from the remote
sensing community is not fully met due to different char-
acteristics of high resolution imagery. These new sensors
significantly increase the within-class spectral variability
and, at the same time, decrease the potential accuracy of
a purely pixel-based approach to classification (Tadesse
et al.
2003
). Consequently, traditional image processing
algorithms are being complemented and sometimes replaced by novel classifica-
tion methods. The currently evident trend towards image segmentation suggests
exploring 'segments' as spatially contiguous and spectrally homogeneous groups of
image segmentation
suggests exploring
spatially
homogeneous
groups of pixels
rather than single
pixels
