CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
106 
In Figure 4 a schematic overview of the used classification sys 
tem is depicted. A multispectral classification is carried out to 
assign the road objects to the different classes. The results of the 
multispectral classification combined with the membership func 
tion leads to the assignment of the road objects to the different 
states. 
Multispectral Classification 
r~l r~I 
water forest clouds roads 
Figure 4: Schematic overview of the classification system 
4 RESULTS AND EVALUATION 
The presented system has been exemplarily tested with two sce 
narios representing flood disasters. In both cases roads are as 
sessed concerning their trafficability. The first scenario is the Elbe 
flood in the year 2002 near Dessau, Germany. Three different 
data sources are used for the assessment: Firstly, an IK.ONOS- 
Image with four channels (red, green, blue and infrared), cf. Fig 
ure 5. The ground-sampling distance of the panchromatic chan 
nel is 1 meter and the color-channels is 4 meter. As second 
source a digital elevation model with a resolution of 10 meters 
is used. Finally, the objects to be assessed are taken form the 
ATKIS (German Official Topographic Cartographic Information 
System) database. The test scene covers an area of 33 km 2 , which 
contains 5484 line segments. In the following investigations only 
the road objects are studied. 
The second study area is located in Gloucesterhire Region in 
Southeast England. In July 2007 the record flood level at Tewkes 
bury was measured. During the flooding a TerraSAR-X scene in 
StripMap mode with a spatial resolution of 3 meter was acquired. 
The polarization is HH, which is more efficient than HV or VV 
to distinguish flooded areas (Henry et ah, 2003). The test scene 
covers an area of of 9,5 km 2 . Additionally, linear membership 
functions from the original rivers are derived and an automati 
cally extracted flood mask is used. As GIS-objects 522 roads 
from OpenStreetMap are assessed. 
The test scenarios are very appropriate to test the classification 
system due to their diverse global context and the different kinds 
of roads. The roads vary from paths to highways. Both test sce 
narios are evaluated using manually derived reference data. The 
availability of reference data describing the real status of roads 
during the flooding is very difficult caused by the fast changes of 
the water level and the accessibility of the roads. One possibil 
ity is to derive the reference data from the image itself, which is 
done for the Elbe scenario. This kind of reference data does not 
describe the ground truth, but the information which is possible 
to get from the studied image. In the case of the Gloucesterhire 
scenario high resolution airborne image with a resolution of 20 
cm are available. This imagery which was acquired half a day 
later than the studied TerraSAR-X scene was used to infer the ex 
act ground truth. To draw conclusions from the following results, 
it is important to consider the kind of used reference data. 
The result of the Elbe scene is visualized in Figure 5. The red 
lines refer to flooded roads, green lines to trafficable roads and 
the yellow lines point out, that no decision is possible by the auto 
matic system. In Figure 6 a detail of the original IKONOS image 
and the assessed roads is shown. 
Figure 5: Automatic assessment of roads using the classification 
system: flooded roads (red), trafficable roads (green) and possibly 
flooded roads (yellow) 
Figure 6: Detail of original and assessed IKONOS scene 
Comparing the result with the manually generated reference leads 
to the numerical results shown in Table 1. "Correct assignment” 
means that the manually generated classification is identical with 
the automatic approach. In the case of "Manuel control neces 
sary” the automatic approach leads to the state possibly flooded 
whereas the manual classification assigns the line segments to 
flooded or trafficable. The other way around denotes the expres 
sion "Possibly correct assignment". "Wrong assignment”’ means 
that one approach classifies the line segment to flooded and the 
other to trafficable. With the current implementation of the sys 
tem the approach achieves a correct assignment for 78% of the 
road objects. Only a very small value of false assignments is 
obtained. This result is deteriorated due to the 5% of "Possibly 
wrong assignments”. Less than 1/5 of all road segments (17%)