The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
282 
To analyze the reliability of self-diagnosis, we matched the in 
ternally evaluated results to a manually plotted reference (see 
(Hinz & Wiedemann, 2004)). The comparison showed that al 
most every road section of the green category is a correctly ex 
tracted road (above 90%). The self-diagnosis also detects “false 
alarms” in the extraction with high reliability (80% - 90%). 
Considering the evaluation of yellow-labeled road sections one 
can state that these parts of the road network should indeed be 
investigated by a human operator because the correctness values 
are generally lower and vary to a notable extent (50% - 75%). 
It is furthermore interesting to observe what would happen if an 
operator had checked exclusively the yellow-labeled road sec 
tions. Under the assumption that a human operator is able to 
discern correct and wrong detections without any error, the cor 
rectness of the overall result would remain in the range of 95%, 
while the amount of editing drops down significantly: only 25% 
to 50% of the whole road network need to be checked. 
However, it is important to note that one can improve only the 
correctness through employing this scheme for internal evalua 
tion. Completeness can only be increased when identifying po 
tential gaps in the extraction and closing them. To this end, the 
system provides user-assisted tools for road extraction. 
3.4 User-assisted extraction: Manual interaction vs. quality 
of the results 
Semi-automatic, i.e., user-assisted, tools have the advantage that 
the quality of the results is guaranteed, because a human opera 
tor controls the data acquisition process and prevents errors on 
line. Yet the overall benefit of such systems depends not only on 
their sophisticated algorithms but also on adequate tools for 
editing. Quite a lot of promising approaches for semi-automatic 
road extraction have been presented and analyzed in the last 
decades. Two groups of approaches can be distinguished: Road 
trackers and path or network optimizers. Road trackers need a 
starting point on the road and a second point to define the direc 
tion of the road. These approaches have quite a long history, see 
e.g. (Groch, 1982; McKeown & Denlinger, 1988; Vosselman & 
de Knecht, 1995; dal Poz et al., 2000; Baumgartner et al., 2002, 
poggio et al., 2005). Path optimizers are designed to find an 
optimum path between two points on a road. Typical methods 
include dynamic programming (Fischler et al., 1981) or active 
contour models, so-called “snakes”, (Kass et al., 1988; Neuen- 
schwander et al., 1995; Grün & Li, 1997). In (Butenuth, 2006, 
2007) the path optimization is extended to full networks with a 
predefined topology, which has strong advantages for user- 
assisted road extraction, see for instance the results shown in 
(Butenuth, 2008). 
Road trackers and path optimizers are characterized by 
complementary properties: Road tracking is usually based on a 
road profile selected by a user for a particular road to be traced. 
In this way, the specific radiometry of this road is included into 
the procedure. This is in particular helpful when different roads 
of varying appearance should be extracted. Such appearance- 
based constraints are commonly not included in path 
optimization. Snake algorithms, for instance, need to be fed with 
a generic image energy, which is derived through more or less 
complex filtering operations like a gradient amplitude map, 
Laplacian map, distance transform, etc. On the other hand, road 
tracking algorithms do not include any topological information 
about the connectivity of the road network. Disturbances due to 
background objects or noisy images (like SAR images) lead 
often to very wiggling tracks or even useless results. In such 
situations, snakes and in particular network snakes show clear 
advantages over tracking procedures, since the geometric and 
topologic constraints involved in the optimization process act as 
regularization for the noisy data. 
Figure 6 shows an example for user-assisted tracking of a main 
road in an IKONOS image taken during the Elbe flooding in 
Germany, 2002. The yellow parts were traced automatically, 
while simple user clicks were asked at the blue positions. Here, 
the operator had to decide whether tracking should just continue 
or interactive editing is necessary. Tracking could continue at all 
interruptions except the one shown in the cut-out of Fig. 6. It 
illustrates a situation, where cutting-off the tail of the track and 
manual digitizing of a short road section is necessary due to the 
occlusion of a small cloud. A detailed description of this 
algorithm and the variety of options for user interaction can be 
found in (Baumgartner et al., 2002). Figure 7 shows the same 
tracking algorithm applied to a RADARSAT SAR image. As 
can be seen, much more interactions were necessary due to the 
worse image quality. Yet the importance of introducing 
topologic constraints comes clear, when applying snake 
algorithms and network optimization procedures to such images. 
Although, compared to road tracking, more initial user clicks 
were necessary to roughly digitize the paths and set up the 
correct topology of the network (see Fig. 8), the rest of the 
optimization of the whole network was achieved completely 
automatically. This appeared to be much more convenient than 
Figure 6. User-assisted road tracking in an IKONOS image taken 
during the Elbe flooding in Germany, 2002. Yellow: tracked road 
sections; blue: user clicks.