forest receives a more precise (specialised) definition and associated behaviour from the underlying data model than
does the unlabelled region in the original image.
Feature formation is often closely coupled to abstraction-, the newly-formed features are described in simple terms
within the data model, with internal complexity hidden from the user by structural abstraction. Also closely coupled
to the notion of extraction is that of in(tro)duced error or uncertainty. As the definition of a region in space is made
more precise (certain), the confidence with which the definition applies must decrease. This is due to the
generalisation that has been applied to the data in order to perform extraction. It is interesting to note that the
opposite is often implied: that is, the definition is precise therefore the accuracy must be high. The amount of
uncertainty that is introduced is determined by the degree to which generalisation is required. For example in
landuse classification, highly homogeneous and well-defined regions (such as paddocks) will not incur the same
uncertainty overhead as poorly defined heterogeneous regions (such as remnant vegetation).
2.4.2 Multiple Spatial Representations
The description of a region as a feature of interest should be regarded as only one element in a set of possible
descriptions that could be formed of the feature, for example by changing the data source, or the algorithms and
control parameters used to interpret the data. The full range of possible interpretations of a feature may offer some
insight into the uncertainty inherent in the extraction process, since the effects of applying different classification
strategies and object formation methods may be analysed and compared.
Existing GIS are built on the false premise that a feature can be precisely and unambiguously defined by some
process, and this definition then placed in a data store. Unfortunately, such an approach is a gross simplification of
the real situation. The idea that a single feature can be accurately described by a single spatial representation is often
inappropriate and indeed can be very limiting, for any of three possible reasons outlined below.
(i) The science of image processing does not produce exact results and it is often advantageous to include more
than one interpretation of a scene fragment within the database. That which defines a feature as distinct from its
surroundings is often unclear and, in some cases, a single interpretation of a scene upon which experts would agree
is often not realisable 4 . Indeed, it would seem that the task for which a feature is to be used has a great influence on
how that feature might best be described spatially.
(ii) A variety of remote sensing platforms can provide source data, each one having different sensing capabilities,
including spectral response and spatial resolutions. So, the representation that might be constructed will vary
according to the source of the image data, giving rise to problems related to cartographic generalisation (Muller,
1991), in that different representations of the same feature can be derived from data of different resolutions.
(iii) Geographic data is by nature dynamic. Image data is collected on a regular basis over the same location (every
16 days for LANDS AT TM, 2.5 days for SPOT). This poses the immediate problem that several 'versions' of each
recognised geographic feature may be created even if the sensor and feature extraction policy remain fixed. If the
effects of temporal change are to be studied it becomes necessary to keep these versions as time-separated
representations of the same feature (e.g. Worboys, 1992: Langran, 1992).
In summary, the representation of a feature is affected by: (i) the feature extraction techniques used; (ii) the platform
or source of data used and (iii) the time from which the data is drawn.
2.4.3 Distinction of Thematic Data and Object-Classes
An important distinction must be drawn between classes defined by data classification procedures and classes as
defined by the object data model. In the object model, a class is used to describe a set of objects of the same logical
type, and is composed of two parts. The intension is the definition of a distinct type of object, the extension is the
4 For example, if there are n geologists, there are at least n +1 opinions as to how the data should be interpreted.