Albert Baumgartner 
  
  
  
  
pairs have to be far away from each other, in order to emphasize the global network characteristics. The last step is to Modu 
calculate the best path between each seed pair. The sum of all best paths is supposed to correspond to the road network. which 
modu 
As this approach was designed for road extraction from multi-spectral satellite imagery, it is able to fuse lines extracted be lin 
from different channels. It can also be applied on a gray scale imagery, however, its capability can not be exploited very decide 
much if only one channel is used. Results for this module, applied to our example image are shown in Fig. 5. for the 
Although the criteria for the selection of these "important places" can be derived from semantically meaningful and Befon 
reasonable parameters, the selection of these points is one of the most sensitive steps within module II. Becau 
For the combination of the two modules the ability of module II to fuse line data from different images or even from foad s 
: : : ; a: road s 
different sensors is essential, because the axes of road segments delivered by module I can easily used as an additional 
input "channel". Another useful feature of module II for our purposes is that different weights can be assigned to the lines ule I 
from different channels. modu 
In Fig 
4 COMBINATION OF LOCAL AND GLOBAL MODULE I7 
Candi 
In this section we show two examples for the combination of module I and module II. In Sect. 4.1 module I and module II turnec 
are applied sequentially. This gives us a rough idea about the use of knowledge about the global connectivity of the road overl: 
network, which was up to now not exploited in module I. The integration of module I and module II described in Sect. 4.2 a bad 
tries to make optimal use of globally best paths which can be found by module II and keeps the good geometric accuracy : 
of module I. Consi 
of the 
4.1 Sequential combination hypot 
this c 
The easiest way to combine the two modules is to combine them 4 
sequentially. For this combination there is no need to change = After 
anything of the internal structures of any module. The output modu 
of module I (cf. Fig. 4) is used as additional input in module II. s paths 
The axes of the extracted road network are fused with the lines 3. least 
extracted in the images at a reduced resolution of about 2 m. By LE resoli 
setting the weight for the axes resulting from module I much B tion I 
higher than the maximum weight of the extracted lines, we en- the se 
sure, that no axes will be lost in the resulting network delivered is kn 
by module II, i.e., we assume that the results of module I are local 
correct. Therefore, in this case the result of the combination 
consists of the axes shown in Fig. 4 and some additional lines 5 E 
which connect the fragmented result that was delivered by mod- 
ule I. Comparing the combination result (Fig. 6) to the previous 
results, then the most significant difference to Fig. 4 is the fact In the 
that almost all parts of the extracted network are connected, and the 7 
that is possible for module II to find a path through the village a que 
at the left side of the image. Compared to the stand alone result fact 
of module II (Fig. 5) there are more side-roads and blind alleys ity m 
connected to the resulting network now, and the use of context road 
information allowed to bridge the shadow in the upper part. Nn 
42 Integrated Combination Figure 6: Results of sequential combination of local | erenc 
and global grouping. plott: 
In this section we describe the integration of module I and mod- refer 
ule II. More exactly, we integrated module II in the extraction process of module I. As mentioned in Sect. 3.1 the elimi- 
nation of wrong initial hypotheses for road segments has a strong influence on the resulting road networking. Basically, The | 
in the elimination steps module I tends to eliminate not only erroneous ones, but correct initial hypotheses, too. Thus, it ness 
achieves good correctness at the cost of completeness. However, in areas where the initial hypotheses for road segments “buff 
are very fragmented and short, e.g., in urban or built-up areas, the elimination of a few correct hypotheses mostly is extra 
equivalent to significantly worsening the chances to extract the road network in this area at all. At this point module II | netw 
is employed to tell module I which of the short segments that are candidates for elimination should be kept due to their | 
importance for the connectivity of the road network. | Com 
Before initial hypotheses are eliminated by module I it sends all its road segments to module II and 
labels them as “good” or “bad” candidates. Module II fuses all segments with the lines extracted The | 
at reduced resolution and starts its road network extraction, using only the “good” segments as possi- lies 
ble seed points. le. the task to select appropriate seed points for module II is solved by module I 
  
62 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
es Se em s ne = € dtt RB AER