tep is to 
etwork. 
>xtracted 
ited very 
gful and 
en from 
dditional 
the lines 
1odule II 
the road 
Sect. 4.2 
accuracy 
  
of local 
1e elimi- 
asically, 
Thus, it 
egments 
nostly is 
iodule II 
to their 
II and 
xtracted 
: possi- 
odule I. 
Module II then returns all parts of its road network to module I 
which are not covered by the "good" segments. From this result 
module I gets information about the good segments which might 
be linked according to global grouping criteria, and it can easily 
decide which of the “bad” segments should be labeled as important 
for the connectivity of the road network. This procedure is invoked 
before every elimination step of module I. 
Because the local grouping of module I bridges gaps between the 
road segments, relaxes its grouping thresholds, and can rate other 
road segments as "good" or "bad" ones, also the seed pairs for mod- 
ule II can change. Therefore, it is necessary that module I asks 
module II before every elimination step again. 
In Fig. 7 this process is exemplified for an early elimination step: 
The road segments displayed in white are considered to be “good”. 
Candidates for elimination are displayed in black. The paths re- 
turned by module II are dotted white lines. In this example the 
overlap with a global path prevents some short road segments with 
a bad rating from being eliminated. 
Considering, that module I is quite good at detecting correct parts 
of the road network, and that module II is able to provide good 
hypotheses for the global connectivity, both modules benefit from 
this combination. 
After the last elimination step all paths which were provided by 
module II are added to the network extracted by module I. For those 
paths which can not be verified using the methods of module I, at 
least the position of the path is adapted to the features in the high 
resolution image. This optimization is done, because the connec- 
tion hypotheses from module II are known to be quite reliable, if 
the seed pairs are well selected, and only the geometry of the paths 
is known to be less accurate. The final result of this integration of 
local and global grouping modules is given in Fig. 8. 
5 EVALUATION 
In the previous section we made some qualitative statements about 
the results. Now, to get more imdependent statements we apply 
a quantitative evaluation. For this evaluation we compare the ex- 
tracted road networks with a reference network. The internal qual- 
ity measures of module I and module II about parts of the extracted 
road network are not taken into account. In the following we give a 
short description of the external evaluation procedure we use. For 
more details see (Heipke et al., 1998, Wiedemann, 1998). The ref- 
erence data for this evaluation is shown in Fig. 9. It was manually 
plotted at a resolution of 0.25 m. The width of the roads in the 
reference network ranges from about 3 m to 8 m. 
The evaluation scheme allows for statements about the complete- 
Albert Baumgartner 
  
i a. IA |a. fos d 
Figure 7: Detail: Global link hypotheses (dot- 
ted, white) and road segments (start segments for 
global grouping are white, others, black). 
  
Figure 8: Results of integrated combination of lo- 
cal and global module 
ness and the correctness of the extracted roads by matching the extracted data to the reference data using the so-called 
"buffer method". For the correct parts of the extracted roads it provides further an RMS-value of the position of the 
extracted axes with respect to to manually plotted reference. The completeness indicates how much is missing in the 
network. whereas the correctness is related to the probability of an extracted linear piece to be indeed a road. 
Completeness is defined as the percentage of the reference data that is explained by the extracted data, i.e., the percentage 
of the reference data which lies within the buffer around the extracted data. 
The correctness represents the percentage of correctly extracted road data, i.e., the percentage of the extracted data that 
lies within the buffer around the reference network. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 63