21
present only a
/iews; if indeed
classification of the region according to some pre-defined classification scheme. Such an overlay might be created
by performing a spectral classification for vegetation (e.g. NDVI) on a satellite image. The overlay has a theme or
purpose, hence the term thematic mapping.
A class represents a partitioning of low level data, often according to some predefined statistical model, such as the
Maximum Likelihood Classifier (MLC). The presence of geographic features is not logically implied by the class
labels and cannot be assumed from any groups of connected pixels that share the same label. Image noise, lack of
resolution and edge effects all conspire to ensure that the spectral response of a pixel does not always equate to class
membership and that the class membership of a pixel does not always imply that the pixel is part of a specific
its, each with a
el assumes that
all points, so a
It, quantisation
eyed point will
scales) a point
ler station may
mples must be
)•
feature, whose type is the class in question. The conversion from classes to features is therefore more complex than
connected component labelling, since it must take account of the natural variance within features, which may mean
that they ultimately contain pixels with ‘foreign’ class labels. For certain applications, the final step of feature
extraction from classified coverages is simply too restrictive and artificial. This commonly occurs where the results
of a classification are not broadly homogeneous or spatially clustered, so that the formation of features is likely to
introduce an unacceptable lowering of precision, since the production of significantly sized spatial representations
would involve too great a generalisation of the data.
Other operations, common to computer vision and image processing, produce data at the thematic level of
abstraction. For example, an edge detector labels the components of an image with values related to the likelihood
of a component being part of an edge. This results in a thematic layer representing a (continuous) classification of
edge strength. Again, individual objects are not explicitly recognised, and further processing is required to form a
meaningful segmentation of the scene.
•del is given by
>cribed by each
2.4 FEATURE MODEL OF SPACE
; not explicitly
y may proceed
k with since its
used properties
Current GIS typically operate a ‘feature oriented’ view of space, where the user interacts with delineated and
labelled spatial objects. These objects are, to all intents and purposes, uniform in nature. In fact, all of the spatial
properties of the feature have been reduced to just two components; namely a geometric description (shape) and a
class or feature label. A weakness inherent in this treatment of data is that all of the variance and subtlety in the
original data from which the features were formed is disregarded, or at best may be statistically summarised in a
coarse fashion. Our pre-occupation with the production of homogeneous regions seems to originate from the desire
lising has been
imposed upon
to work with clear and visually appealing maps, rather than a desire to analyse data in the most accurate way that we
are able. Any heterogeneity in the underlying data is replaced by exactly one spatial description that has exactly one
class label and is useful at one spatial scale, and perhaps only for a specific number of tasks (Burrough, 1986, p.
137). Worse still, current feature descriptions do not convey all of these attributes, so the user may not be aware of
what they are!
jach geometric
e original data
iel regards the
5 a monothetic
This state of affairs biases GIS to the production of maps, i.e. the output stage of any exercise. Much of the
analytical work involved in forming the underlying regions used in mapping is currently carried out by external
functionality, such as Remote Sensing Systems (RSS), from which the data is imported into the GIS.
Whilst there is a loss of precision in the formation of features, this is often regarded as being offset by the increased
utility that feature descriptions can provide. Features are conceptually simple entities, which aids the user since they
appear to be both unambiguous and homogeneous. Their homogeneity also ensures that manipulation and
imeter of which
a set of
, and used here,
sat TM), or
e sensors that
tt and Curtis,
data before it is
combination is straightforward in an algorithmic sense, and that results can be readily interpreted. GIS operations
take as their operands geographic features that could not easily be described in terms of ‘raw’ image data. It is not
suprising then that most GIS operations require that data be in feature form before they may be carried out.
2.4.1 Extraction, Abstraction and Uncertainty
As the process of feature formation is carried out (termed extraction by Smith & Park, 1992) the level of meaning
jrtion, to re-
3f the image.
that is encapsulated by a particular region increases. For example, a region extracted from an image to represent a