International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
still required substantial editing efforts. The major problem
with this approach is the dependency on map data which is in
many cases not feasible, as already discussed above.
It should be also noted, that airborne laser scanner data has
limited XY resolution and requires reliable automatic break-line
detection for usage in urban areas, if no ground information 1s
available. Such break line detection can be combined with
analysis of image data to support the hypothesis when checking
for edges in the aerial imagery. As classical film cameras are
not practical to fly in parallel to laser flights which require
usually flying heights below 2000m above ground, the focus is
on the increasing usage of small to medium format digital
airborne cameras on the same platform. A seldom discussed
issue is also the usage of the intensity “image” of the laser
system, which has of course a low planimetric resolution, but
which can provide substantial support for the feature extraction
(Arefi et al., 2003).
LESSON 4: The usage of airborne laser scanner data can
increase the automation, however, currently depends on
existing map data or tedious manual delineation of
building footprints. Using laser scanner data alone is
regarded as not sufficient to solve the derivation of
features in complex areas in practice, the current
development aims at sensor fusion with digital images.
First examples of automated breakline detection are
available.
2.6 GIS Interface
For the handling of the GIS vector data an interface between
inJECT and the GIS software packages Dynamo and GeoMedia
(Intergraph) has been developed in a special version (Ohlhof et
al., 2004) based on the GML2 format standard from the Open
GIS Consortium (OGC). With this interface the vector data and
the associated XML schemes can be automatically imported
and exported. Imported features can be edited within inJECT,
the user can modify the geometry and can select or key in the
attribute values of each GIS object.
For the import of GML data into GeoMedia the existing GML
data server can be used, whereas the export of GML data from
GeoMedia can be carried out with Intergraph's GML export
module. The resulting GML can be transformed to DXF or
VRML using XSLT stylesheets. In case of a 3D extraction
GML3 is offered by inJECT without a specific interface to GIS.
GML can be applied for interoperability between different
organisations and companies, which has already been
successfully tested.
LESSON 5: There is an uncertainty on how to model
buildings and how to store this data in a GIS or CAD
system. It is certainly advantageous to have a defined
standard for this, which does not exist yet. However, the
GML 3 standard defined by OGC seems to be a vehicle for
a broad range of users.
2.7 Empirical investigations and comparisons
There is still a great lack of knowledge on performance of such
systems in various image scales and object densities. EuroSDR
has started an initiative to investigate the current status of
automated building extraction in this respect (EuroSDR, 2004).
Four data sets with aerial images, laser scan data and building
footprints have been provided to test feature extraction from
aerial images and digital surface models.
c
426
ISPRS has several working groups dealing with those issues
(e.g. Sithole and Vosselman, 2003). ISPRS WGIII/8 works on
the reliability and performance of algorithms and which has
prepared comprehensive test data sets with ground truth
information (ISPRS WGIIUI/8, 2004)
There is also an observation, that many users simply want
visualization results and do not aim at the highest possible
geometric quality, which in turn is not supporting the need to
really look at those geometric aspects in more detail. But there
are not even comparisons on the efforts needed to measure
buildings or parcels with different tools. Requests are made to
measure 100 buildings, but a description of the level of detail is
missing, or there is no information on complexity of buildings
or density. Sometimes the information about the number of
measured buildings per second is only an academic issue, as it
might take much less time to measure those 100 buildings in a
stereo model, compared to the time to get access to the
orientation data of those images or the GCP's to orientate the
images in a digital photogrammetric workstation.
LESSON 6: The investigations conducted by organizations
like ISPRS or EuroSDR should be supported much more
by agencies and companies. It is vital for further success,
that there is a thorough evaluation of the performance on
an objective basis.
3. DISCUSSION
Some few software developments in automated feature
extraction have been implemented as commercial systems.
They have shown that there is a potential on a small but
growing market. It is recognized, that there is a need for further
development adapting to currently developed standards. inJECT
as one example has been substantially extended with new
automation modules and a strong GIS interface. Not only
inJECT, but also other systems have shown, that the concept of
semi-automation is excellent for practical applications, as there
is always an editing option, if some automation fails due to low
image quality, disturbances and other effects. The extraction of
complex polyhedral roof structures in inJECT is now highly
supported. The possibility to measure roads and parcels with a
high level of automation has substantially increased the
applicability for a wide range of users. The automated 2D
extraction modules used for that purpose has been tested with
several types of satellite imagery having a ground pixel size of
0.8 to 5 m as well as aerial orthophotos of about 0.5 m ground
resolution. The operator is well supported during the measuring
phase. In case of the road tracker, a traffic light approach is
used as guidance. The software has been proven to be an
excellent platform to add and test external software modules
that increase the automation level significantly. By adapting the
OGC-defined GML standards the field for future applications is
wide spread. Compared to many GIS systems inJECT has
major advantages in data capturing and editing especially in
3D. Since there is in 2D and even in 3D extraction no need for
stereo viewing, the system is more open also to professionals
outside the photogrammetric world which have here an easy-to-
learn tool at hand to carry out basic feature extraction tasks
with high accuracy and reliability.
4. CONCLUSIONS
We can summarize thc experiences during the last decade with
the following statements:
e The semi-automatic approach has proven to be the
right way for practical applications. Also non-
Intern
There
feature
new s
resour
goal «
imaget
Arefi,
classif
Comm
Integrz
2003.
Baltsa:
Procee
Extrac
Image:
Swetsé
Cyber(
EuroSI
Gülch,
feature
Photog
©
Science
