Image Processing Reference
In-Depth Information
The high number of potential characteristics of objects leads to a high-dimensional
feature space. For image classification, representative objects for each class are
first identified. This process can be compared to the delineation of training areas
in a supervised classification. These sample objects should be well separated in
feature space by using the most representative and clearly distinguished features.
Algorithms for the optimization of the feature space can support this process. That
means that the entire set of possible feature space dimensions is reduced to a
smaller set of features optimizing separation of image objects. After selecting the
samples and their corresponding feature values, a standard classification algo-
rithm (such as box classifier, nearest neighbor, maximum likelihood) is applied
to all image objects.
Multi-scale segmentation creates a network of image objects connected by vertical
and horizontal relationships. Binnig et al. (
2002
) explain the way in which this
network reflects a hierarchical knowledge structure as a
“self-organizing, semantic, self-similar network.” By oper-
ating with the relations among networked objects and
integrating prior knowledge of the target classes, analysts
can classify objects through a set of rules. Those reflect
both the spectral response and the contextual information.
Integrating knowledge provides a way to overcome the
spectral similarity of different geographical features
(e.g. to separate rural grassland from urban green space
using geometrical properties). Rules are established to
characterize the lateral relationships between objects from
specific classes (e.g. “relative border length with neighbor-
ing class XY”). Moreover, some object-based classification approaches use fuzzy
logic to soften crisp distinctions and characterize the degree of membership of an
object to a certain class. One of the strengths of fuzzy techniques comes from the
way they can be used to describe real-world phenomena with some degree of
uncertainty.
Furthermore it is possible to support any classification with ancillary thematic
data like GIS layers or digital elevation models. Ancillary thematic data can be used
for image segmentation and for class descriptions as additional information in the
object feature database (Fig.
10.3
).
two groups of
classification
methods are used
to delineate
objects from a
segmented image:
sample-based
methods and
rule-based
methods
10.3.3
Post Processing
An object-based classification approach offers the possibility of the spectral, as
well as the semantic grouping of objects. Contiguous objects from the same class
can be merged into one polygon. Objects belonging to the same class on a higher
semantic level can also be merged (e.g. a football field and the surrounding area can
be merged into the target class 'urban'). Objects and classification results can be
stored in either raster or vector formats.
