inbul 2004
ge graph-
inction. of
'rification.
1g a strict
> of false-
ids will be
> the latter
ion inside
from the
work. All
t network
, but with
n.
n refer to
h to road
1 context
sment of
system in
proaches
place to
trial area
rried out,
ment and
uses a
7), it was
; acreage
riculture,
learn the
ions, the
s and an
"solution
rom the
a texture
segments
ter files.
solution
1ess and
train the
te set of
assisted
lassifies
: of the
up the
training
xture of
as with
am the
ses of a
gh. The
nportant
ction 5
ased on
ructures
ssumed,
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
thus a shadow has to exist next to a building. The azimuth and
zenith angles of the sun position are taken as a priori
knowledge. First hypotheses for shadows and roofs are
generated using two different image segmentation operators. To
validate these hypotheses instances of roofs are grouped with
one or more instances of shadow. Shadow hypotheses of
buildings are derived with a threshold decision in the image.
Since shadows are not well visible in a green colour channel,
the green colour has been masked during shadow detection.
Additional shadows have a limited area, so shadows e.g. near a
forest can be excluded. Roof hypotheses are generated in a more
complex procedure. Here the so-called colour structure code
(Rehrmann and Priese 1998) is used to segment the entire
image. Additionally greenish areas are masked out. Only roof
hypotheses of a plausible size are selected, additionally the
compactness and orthogonality of roof labels are validated. In
the last step the grouping of instantiated shadow and roof labels
to validated buildings is performed. The neighbouring position
of a shadow to a roof has to fulfil the illumination model. The
resulting building hypotheses are divided into houses and
industry halls using the area of the objects as criterion.
3.2.3 Evaluation of an ATKIS region The results of the
structural and textural approach are combined to verify or
falsify the checked ATKIS regions. The two approaches lead to
different measures of quality. The texture classification, which
is a holistic operator, leads to a pixel-wise assignment of
classes. The structural approach identifies complex objects by
using a combination of different clues and the structure of
objects. The considered ATKIS regions are evaluated with an
evaluation catalogue, that was designed by use of the ATKIS
catalogue and with use of the experience of human operators.
Both the structural and textural conditions have to be passed for
a validated ATKIS region. Settlement and industrial areas are
verified by means of the found buildings. Vegetation classes are
falsified if houses or buildings are found in the region. This
decision is based on the definition in the ATKIS objects
catalogue. Within the vegetation classes forest can be identified
and verified. Agriculture are not supposed to contain forest.
4. COMPLETE WORKFLOW
The main feature of the complete workflow for automated
quality control is the interaction of automatic procedures with
the interactive steps or the decisions taken by the human
operator. Orthoimages and the ATKIS DLMBasis are available
for the automatic procedures as well as for the human operator.
The results of the automatic procedures are passed to the human
operator in the form of a traffic light diagnostics. A speed-up of
the automated workflow in comparison to a purely interactive
workflow is attained by relying upon the objects accepted by the
automatic procedures, i.e. the ATKIS objects highlighted with
green colour. The human operator has to concentrate
exclusively on his final decision for the rejected ATKIS objects,
ie. the red ones. The consequences of this approach are
described in Table 2. The percentage of corresponding
acceptance when comparing an ATKIS object to the orthoimage
indicates the efficiency of the system. Objects that have been
accepted by the automatic procedure, but would have been
rejected by the human operator, will result in undetected errors
when using the workflow as depicted in Figure 2. Their
percentage has to be as small as possible since only very few
errors should remain undiscovered. For avoiding false alarms,
ie. false negatives according to Table 1, all ATKIS objects
rejected bv the automatic procedure have to be checked finally
by the human operator.
739
Automatic Green Red
Human Operator
Green Efficiency Interactive
Final Check
Red Undetected Interactive
Errors Final Check
Table 2. Confusion matrix of decisions: human operator vs.
automatic procedure, effects (cf. Table 1).
Visualisation
à; (ArcGIS)
Final Decision
by Operator
Traffic Light
Knowledge-Based s
Component more Diagnostics
# o.k.
Figure 2. Workflow of quality control (q.v. text)
5. EXAMPLES AND RESULTS
We have compared the diagnostics of the automatic system with
the results that have been obtained by a human operator without
automatic image interpretation. Figure 3 shows an example of a
situation that has been detected by the automatic road
verification. The paths in the scene have been acquired
improperly. Paths are of secondary interest. Nevertheless they
are analysed by means of the automatic procedure
demonstrating its ability to detect errors. Table 3 subsumes the
results of the road verification.
Figure 3. Example for the road verification. Orthoimage,
ATKIS DLMBasis, and automatic diagnostics with detected
error.
Whereas road extraction is currently restricted to the road
objects stored in the ATKIS DLMBasis, the work on built-up
areas is based on classification and analysis of the whole
orthoimage. Thus the analysis of built-up areas is intended to
verify existing built-up areas, to determine their changes and
expansions, and to detect development areas and objects that
have been overlooked during data acquisition.
