anbul 2004 
andard for 
Open GIS 
Group 3D 
2004) is 
which is 
>ometrical- 
rave been 
; extension 
uneras and 
ues for the 
in digital 
2001) was 
traction in 
10t only to 
5, but also 
est site of 
dents have 
ise version 
a textured 
odules for 
hotos and 
nore detail 
duction of 
panies and 
r teaching 
"e 3 
building 
mented in 
of areas in 
gle) inside 
leformable 
statistical 
vironment. 
| a final 
n example 
nd a final 
ier parcels 
particular 
where the 
t direction 
measures 
cular line. 
    
e 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
Existing lines and crossings are snapped and nodes are 
generated, leading to a topologically connected network, which 
is very important in the case of road network. 
The user is supported by a traffic light implementation which 
informs about problems encountered in the extraction process. 
The lines are smoothed and the average width of the line 
segments is computed automatically. The resulting 
topologically connected road network is smoothed and the 
operator can key in or select missing GIS attribute values. 
  
Figure 2 3D VRML model derived with inject of a part 
of Senaatti test site (by A. Novacheva and S.H. Foo) 
   
c) d) 
Figure 3 2D-parcel extraction for a lake feature in an 
IKONOS scene. a) Start position by operator. b) 
Result after first run. c) slight improvement after 2"? 
run. d) Final result with strong generalization. 
An example of linear and area features extracted in an 
orthophoto from an aerial image is given in Figure 4. Existing 
vector data can be imported via a GML2 format using a GIS 
import filter. Besides geometry the imported GML data 
contains complex information about feature semantics, which is 
read by inJECT as well. 
The feature extraction on line and area features can be done in 
digital orthophotos for the capture of 2D GIS vector data as 
described above. In addition, the software is available for the 
capture of 3D features using oriented aerial imagery. Here the 
automation part consists currently of the on-line z measurement 
42 
5 
functionality which automatically derives the height of each 
vertex point of a line feature or the contour of an area feature. 
The automated algorithms have been extensively tested with 
IKONOS 2 and IRS satellite imagery as well as with 
orthophotos from aerial imagery. 
   
    
Figure 4 Linear network and area features in an 
orthophoto derived from an aerial image. Features 
(blue for water, green for fields, red for settlement, 
yellow for roads) have been derived semi- 
automatically. 
If we look at other automation approaches we can see, that new 
feature extraction modules offered for integration into existing 
platforms often lack standards for exchange and/or they lack 
practical feasibility tests on a larger variety of input data. 
The usage of colour imagery is not yet fully exploited. The 
usage of existing ground data, often propagated, is in many 
cases not feasible, as economically too expensive and for 
nationwide applications in federal states much too complicated 
due to too many standards. There are increasing efforts to fully 
automate road extraction from aerial imagery, but there is no 
broad application on the horizon except for some specific 
authorities or agencies. However, if using high resolution 
satellite data, like Quickbird2, Ikonos2 or IRS there is a need 
for automated road extraction and parcel extraction. inJECT has 
been extended to those object types and allow their derivation 
in ortho images in a semi-automatic fashion. There is a need to 
add new modules to those procedures to further increase the 
amount of automation. These could be partly based on 
interesüng research results concerning vegetation extraction 
depending on public acceptance and potential users. 
LESSON 3: Semi-automatic feature extraction from digital 
aerial imagery is introduced to practice. The extension to a 
wider range of features is important. The testing of new 
automation modules for practical applications requires a 
good platform which allows easy implementation and easy 
integration as well as full control and checking capabilities 
in empirical investigations. 
2.5 Feature extraction in digital surface models 
There seems to be a new push for the development of feature 
extraction by using airborne laser scanning data and possibly 
imagery for extraction. However research 1s mainly focusing on 
full automation which is not very realistic in our experience and 
should be directed more to a stepwise semi-automatic approach. 
The usage of existing ground data (footprints of buildings), 
often propagated, has shown extremely good results, despite