change that is to be expected. This can be used to roughly
estimate the completeness of a database at some point in
time which may be valuable information for the verification
step.
4 UNCERTAINTY IN DATA FOR ROAD
EXTRACTION
On a first sight one might think that we will have to deal
with the precision of the GIS data and the precision of the
extracted image features. Yet, from the above aspects of
data quality it should have become clear that there are many
more causes for uncertainty.
4.1 GIS data
In fact, positional accuracy of GIS data is only a minor source
of uncertainty. The available data usually permits to outline
the sides of the road surface or the middle of the road within
a few pixels in the image. The (very few) results on updating
road networks by image interpretation are, however, much
worse. Thus, there must be other sources of uncertainty.
Like positional accuracy, attribute accuracy is usually very
high in comparison to the quality of the interpretation results.
Consistency of the GIS data may be considered a more impor-
tant factor. Inconsistent data will yield conflicting evidence
to some hypotheses and thereby can mislead the reasoning
process. Of special interest in GIS data are the topological re-
lations between the features. Egenhofer and Franzosa [1991]
classified eight different topological relationships between two
two-dimensional regions (like meet, overlap, disjoint, etc.).
Winter [1994] argues that such relationships between regions
can not be considered as certain, due to positional inaccura-
cies, however small they are. E.g., even due to the smallest
possible error, two regions that actually meet may be classi-
fied as disjoint or overlapping. Other changes in topology are
less likely. E.g., if one region is actually contained in another,
it is unlikely that they will be classified as disjoint. Winter
[1994] therefore derives conditional probabilities of topologi-
cal relations between regions in a GIS, given their true topol-
ogy. These probabilities very well model the uncertainty in
topological relations. A reasoning process can now take into
account the confidence that has to be given to some relation-
ship and does not have to accept all relations as correct.
Semantic errors in the GIS can also have a large impact on
the image interpretation. Suppose that, according to the
data model, a road database contains the roadsides. This
definition of the data still allows several interpretations. E.g.,
does the road include the sidewalk, or the shoulder? An
incorrect interpretation of the data model can clearly lead to
a large number of errors in the verification step. If the data
model is ambiguous, the verification step should comprise
hypotheses for each of the different interpretations in order
to find the correct one.
Since the purpose of the image interpretation is in updating
the road database, it is obvious that the data completeness of
the GIS is significantly lower than what can be expected for
an up to date GIS. If available, a rate of change may be used
to calculate the expected data completeness at the time of
updating. This number can then be compared to the results
of the verification step.
The information in a GIS is clearly insufficient to automat-
ically solve the interpretation of the aerial images. In this
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
sense, the model completeness of the GIS for the task of au-
tomatic updating is very low. Image interpretation requires
much richer descriptions of the objects than a few vectors in
a GIS. This leads to the problem of object modeling.
4.2 Object models
Describing roads in generic models such that these models
contain sufficient information to recognize all kinds of roads
is an extremely difficult task. Yet, humans have no problems
in recognizing the roads in figure 1 despite the large variety
in shape, size, scale, and pavement.
Figure 1: Variety of road appearances in aerial imagery.
Gunst [1996], after [Garnesson et al., 1990] describes a road
model in terms of geometry, radiometry, topology, function-
ality and context. Many attempts to describe a road only
use geometrical and radiometrical properties. E.g., a road is
defined as two parallel edges that include an elongated ho-
mogeneous area. According to this definition, a side walk,
a single traffic lane, a river, a dike, a beach, and probably
many other objects also can be classified as a road. Some
improvements can be made by including colour or texture
information, but a good result can only be expected if the
912
Car
of 1
tra
roa
son
ext:
eac
knc
the
cer!
Ins:
roa
the
Th:
als
tail
circ
call
of !
Wir
tra
gle
for
ext
ass
bot
sm
or
as
rise
fea
err
sivi
al,
tur
ces
pro
ext
sor
be
the
prc
els
is |
fea
des
ter
noi
int
ed;
the
