Image Processing Reference
In-Depth Information
seven spectral bands, when classified, would result in typically 10-20 thematic
classes. Moreover, regardless of the degree of sophistication in the classification
methodology, the grouping of “similar” pixels, either based on spectral or spatial
rules, invariably will be highly subjective, strongly dependent on training samples
and class choices, as well as heavily biased by local and
scene-specific conditions (Forster
1985
; Webster
1995
;
Cowan and Jensen
1998
). In all, the resulting classified
image will be a model vastly removed from reality, even
more so when representing urban areas. Why, then, should
multidimensional image data be classified in the first place?
Well, despite wide-ranging caveats, the whole premise
for classification can be distilled to matters related with
communication. More specifically, communication not
only between image data and application objectives but also
between data within integrated GIS databases. Unclassified
image data may be more objective and have a greater range of measurement but only
classified thematic interpretations are acceptable for most applications. Few integrated
projects are capable of handling, let alone communicating, graylevel values or digital
numbers; they instead demand more pragmatic and user-friendly semantic interpreta-
tions that are also consistent with data already present in the GIS (Mesev
1997
).
Classification of multispectral satellite sensor imagery, therefore, involves the
identification and statistical grouping of pixels with similar digital numbers and/or
similar spatial orientation of pixels into meaningful geographical features. In urban
examples, this would entail the recognition and conversion of pixels with similar
multispectral values and/or positional location into thematic categories with mean-
ingful labels, such as buildings, roads, parks, gardens, etc.
The distinction between classifications based purely on
spectral signatures and those that also exploit the spatial
arrangement of pixels is a particularly important develop-
ment in urban applications. Unlike the natural environ-
ment, urban surfaces are typically a complex mix and
intricate arrangement of artificial and natural objects of
irregular size, frequently angular in shape, and exhibiting
variable density. Such unpredictable compositions and
configurations are difficult to replicate from remotely
sensed data if the classification relies solely on spectral
information (Baraldi and Parmiggiani
1990
; Couloigner
and Ranchin
2000
).
classification
usually entails
conversion of
data from
interval to
nominal levels of
measurements
based on spectral
or spatial rules
the principal
difference
between spectral
and spatial
classification is a
matter of whether
similar pixels are
collected in
isolation (spectral)
or as part of
contiguous groups
(spatial)
8.2.1
Urban Land Use from Land Cover
The statistical mechanisms for the classification of images representing urban areas
are almost identical to those used for the natural environment. The fundamental
exception is the inference of land use from land cover (Chapter 6, also see Dobson
