Gary Priestnall
objects can be regarded as contextual information (Baumgartner ef al, 1997). Tonjes (1998) describes a knowledge
based approach for the interpretation of aerial images that combines context from multiple sensors (visual, infrared,
SAR). The integration or fusion of the contextual information (regional classifications based upon different scales of
imagery) from different sensors allows multi-resolution extraction schemes to be devised. The derivation of contextu
clues of a spatial nature such as containment can be achieved using relatively coarse resolution multispectral imagery to
reclassify the image into broad land cover regions, for example using the SPOT image in Figure 2 below. Here, the
urban region has been classified. The characteristics of linear features within this region (roads for example) may differ
from respective linear features in rural areas. If the rules defining these characteristics can be encapsulated for each
regional context area, the identification and differentiation of linear features could be made easier.
SPOT XS Image Urban Context Region
(© SPOT Image Copyright 1998)
Figure 2. Imagery suitable for derivation of regional contextual information
Spatial context can be considered at several 'levels', from the regional (above) to the more local, where the relationships
between an object and its immediate surroundings become important. Baumgartner et al, (1997) discuss the general
distinction between the different 'levels' of spatial context, using the terminology of contextregions and sketches. Ona
regional level three kinds of context regions are introduced: urban, rural and forest. Context regions can be defined or
classified by segmentation and texture analysis. Results of texture analysis can be combined with other GIS data if
available, for example urban boundary data sets. For each context region special relations exist between linear features
and background objects. For example buildings are more likely to run near and parallel to roads in urban regions and
shadows from trees are likely to affect forest regions. Background objects can therefore have a strong influence on the
characteristics of linear features. Background objects on the one hand support and on the other hand hinder road
extraction.
On a more local level, context sketches can be used to model the relationship between linear objects and neighbouring
objects. Context regions and sketches may potentially be defined by any contextual clues used by humans when
interpreting images. An example is the declusion shadow’ where a hypothetical road part is required to bridge the gap
left by a shaded area. Another example can be seen in Stilla and Michaelsen (1997) where the local relationships
between roads and linear groupings of buildings are discussed.
The ALFIE project considers the use of regional context both in terms of limiting the choice of acceptable feature
recognition solutions and in effecting the parameters used by feature extraction algorithms. Local context, at the scale
of the 'sketches' discussed above, is introduced in the form of a matrix of feature interrelationship rules. The co-
operative use of other collateral information in a feature extraction strategy is discussed below.
3.2 Target imagery and collateral information
Despite the increasing number of remote sensing satellites currently in operation and planned for launch in the near
future, there still remains a risk that the most appropriate imagery may not be available worldwide, for the specific arc
of interest. This may be due to cloud cover or simply that a particular sensor has yet to acquire data in the given region.
As such, optimal and minimal target datasets have been determined. The minimal target dataset includes; 30m, 7-band
multispectral imagery; 10m panchromatic imagery; and an elevation model at around 100m post spacing. This dataset is
assumed to be available for all parts of the world and will provide the basic source data from which to extract linear
features. The optimal dataset includes; 4m, 3-band multispectral optical imagery; Im panchromatic imagery; and an
elevation model at around 30m post spacing. In essence, the minimal dataset should allow major linear features to be
extracted and the object schema populated to the object class level (see section 5), while the optimal dataset should
allow the extraction of fine detail linears and allow the population of the object schema in its entirety. In practical terms,
a mixture of these datasets is likely to be available. The key is to ensure that the extraction system has the capability to
work within the bounds of these two target datasets.
As shown in Figure 2, the coarser resolution multispectral imagery (defined in the minimal dataset) will be useful for
determining the urban/rural context regions. A further use will be the extraction of collateral information prior to the
742 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
actu
ima
fiel
requ
the
rem
The
assi
The
for «
any
For
surf
upo
kno
IeCC
Vos
func
The
netv
relai
ie:
feati
Usir
line:
char
4.1
Geo
extr;
tech
para
Scai