— 
ledge 
frared, 
cales of 
-ontextua] 
imagery to 
re, the 
may differ 
r each 
lationships 
seneral 
ches. Ona 
efined or 
data if 
ar features 
ons and 
ice on the 
pad 
rhbouring 
hen 
e the gap 
ships 
ature 
t the scale 
 CO- 
he near 
pecific area 
ven region. 
m, 7-band 
is dataset is 
t linear 
; and an 
res to be 
should 
ctical terms, 
pability to 
useful for 
or to the 
Gary Priestnall 
“20 ES differemiotion Water; for example, is relatively straightforward to classify from multispectral 
imagery given its reflectance properties in the near infra-red. If fine resolution imagery is available, hedge lines and 
field boundaries may be extracted (for example by segmenting the image). These features, although not necessarily 
required for the final application, provide important collateral information in as much that if linear features extracted by 
f CIO algorithms can be confidently assessed as not being roads or railways, then the classification of the 
remaining linears becomes significantly easier. 
4 CHARACTERISING THE OBJECTS TO BE EXTRACTED 
The first step towards incorporating context, both at a regional and local level, is to characterise the objects under study, 
asenimr properties which include both the typical shape, size and radiometric character but also any contextual clues 
which may be associated with a particular class of object. 
The first stage is to define a generic set of feature properties and then to consider how these can be translated into rules 
for each class of feature under study. These feature classes will include not only those required in the final database but 
any other classes of feature which are deemed important for assisting the feature recognition process. 
For extracting features from imagery the need to consider factors other than the spectral reflectance properties of the 
surface material alone has been clear. By attempting to incorporate more knowledge, techniques become less reliant 
upon assumptions based upon spectral consistency alone. By combining spectral character, with local geometry, and 
knowledge of how that geometry builds up into connected networks, a less local view is taken. Steger et al (1997) 
recognise the need to construct a topologically complete road network. 
Vosselman and de Knecht (1995) characterised roads using five properties: photometric, geometric, topological, 
functional and contextual. 
e Road surface is homogeneous (photometric) 
Road surface has good contrast to its adjacent areas (photometric) 
Roads are elongated (geometric) 
Roads have a maximum curvature (geometric) 
Roads do not end without a reason (topological) 
Roads intersect and form a network (topological) 
Roads are a means of communication between locations (functional) 
Roads may be indicated by a special distribution of trees (contextual) 
The category of function’could be viewed as a contextual property but also as a feature of a topologically complete 
network. In working towards the definition of an object schema, a further property Confidence’can be included. This 
relates to the current belief'in a particular feature labelling in terms of the type of feature and the strength of belief 
Le.: the confidence associated with that labelling. This property is potentially updateable continuously during the 
feature recognition process and would represent the current state of knowledge about one feature in the image. 
Using a structure adapted from the broad categories defined by Vosselman and de Knecht (1995) the characteristics of 
linear features that will be used to discriminate between different feature types are described below. These 
characteristics contribute to the definition of object classes used to store properties and interrelationship information. 
41 Geometric characteristics 
Geometric rules regarding length, curvature, angle, orientation and density have been incorporated in most feature 
extraction studies in recent years, typified by Wang et al (1991). Geometric and photometric rules can be adopted by 
techniques which work on extracted vector primitives, or can operate on-the-fly’ in the form of rule-bases and 
parameters built into road tracking (Grossman and Morlet, 1984) or line following programs such as VTRAK (Laser- 
Scan Ltd.). The geometric characteristics can be defined as: 
® Width: of a linear feature segment, derived during pre-processing and carried forward. 
* Consistency of width: the variation of width along the length of the feature. 
* . Shape, size, and orientation: the degree to which the feature is elongated. Orientation may be common 
to neighbouring roads in certain urban environments (for example gridiron street patterns). 
®  Curvature: sinuosity or fractal indices as discriminators of roads, rivers and railways, assigned to 
features in whole or in part. Typical expected curvatures vary in different environments and terrains. 
* Patterns: such as junctions can be important clues relating to the interrelationship of features and can 
form triggers for spatial searches to complete portions of the network, especially like features 
comprising a network. Section 5 considers a range of junction configurations between different object 
types. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 743