107
In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Voi. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
should be controlled manually in order to reach a correctness of
95%.
Possible assignment
Result
Correct assignment
76.99%
Manual control necessary
17.87%
Possibly correct assignment
4.96%
Wrong assignment
0.18%
Table 1 : Results Scenario: Elbe
The results are obtained with the threshold parameters t\ =0.5
and ¿2 = 0.001. The variations of the parameters are depicted
in Figure 7. The parameters are responsible for the amount of
road segments which are assigned to the state possibly flooded on
condition that they are classified to the classes water or road. The
decrease of’’Wrong assignment” comes along with the decrease
of’’Correct assignments” and an increase of manual control.
Prameter t 1 = 0.5 Prameter ^ = 0.001
Figure 7: Results dependent on parameter t\ and 12 (red = Wrong
assignment, orange = Possibly correct assignment, yellow = Man
ual control necessary, green = Correct assignment)
P*Jg)
Figure 8: Combination of probabilities and impact of the param
eter ti
In Figure 8 the combination of the probabilities p / (a) and p lJJw (g)
is shown. The grayscale bar indicates the combined probability
Pf(g, a). Every star defines a road segment assigned to the class
water by multispectral classification, the color shows the state as
signed in the reference. Many road segments which are assigned
to the state trafficable in the reference are wrongly classified by
the system to the class water. The reason is the high standard
deviation of the probability densitiy function for the class road
and, therefore, the overlapping of the class road and water. Road
segments in urban areas occluded by shadows are responsible
for this effect. The threshold t\ is depicted in blue which dev-
ide the assignment of the roads to the state flooded and possible
flooded (Figure 8). Shifting this parameter leads to the results il
lustrated on the right plot in Figure 7. Furthermore, the improve
ment of the combined probability is shown in Figure 8. If only
one probability is available, the threshold t\ would be depicted
as a straight horizontal or vertical line. The total required time
to generate the manual reference is about three hours. Compared
to the time needed for the automatic classification (less than one
minute) points out the efficiency of the approach.
The results of the second test scenario are depicted in Figure 9.
A detail of the original TerraSAR-X scene and the assessed road
segments is shown in Figure 10.
Figure 9: Automatic assessment of roads using the classification
system: flooded roads (red), trafficable roads (green) and possibly
flooded roads (yellow)
Figure 10: Detail of original and assessed TerraSAR-X scene
In the second test scenario the real ground truth is available. Hence,
the assignment possibly flooded is not existing in the reference
data. The comparison with the automatic classification system
leads to the result shown in Table 2. After controlling 5% man
ually, altogether over 86% are correctly assigned. The value
of 14% of wrong assignment is caused by mainly two reasons:
Firstly, the resolution of the StripMap mode hardly enables to