Christian Wiedemann 
  
In Fig. 4a) and b) the reference network and the extraction result, respectively, are shown for a scanned aerial image 
with a ground resolution of about 0.5m. The extraction was carried out using the road extraction system presented in 
(Baumgartner, 1998). Due to different parameter settings, these results are slightly different to results presented in earlier 
publications, e.g., in (Wiedemann, 1999). The link hypotheses which were sent to the checking module are displayed in 
Fig. 4c) (accepted link hypotheses are drawn broad and in dark gray, rejected hypotheses thinner and lighter; the numbers 
refer to the respective examples shown in Fig. 4d)-f)). 64 hypotheses were sent to the checking module. The check was 
carried out based on the image which was down-sampled to a ground pixel size of 2m. 38 hypotheses were accepted 
and 26 were rejected. As can be seen, most of the gaps within connected components were closed as well as many links 
between different connected components could be inserted. 
The first example (see Fig. 4d)) shows the link hypothesis in the upper right corner. It has been wrongly rejected. The 
reason for this failure is that the missing road is hardly identifiable. A more sophisticated road extraction algorithm, e.g., 
based on context as well as on additional data like high resolution multi-spectral image data, height data, etc. would be 
necessary to extract this kind of (partly occluded) roads. The link hypothesis in the middle of the image is displayed 
in Fig. 4e). In this case, the varying road width prevented the extraction of the missing part. This link hypothesis was 
accepted correctly. In the final result, many accepted as well as rejected link hypotheses lie in the region above and to the 
right of the image part displayed in Fig. 4e). Some of them are wrongly accepted/rejected due to the high complexity of 
this region and the fact that the road extraction algorithm used for verification was designed for the extraction of roads in 
open rural areas only. The third example (see Fig. 4f)) shows the correctly accepted link hypothesis from the lower left of 
the image. It connects two different connected components of the initial road network. 
The evaluation of the initial road network and of the automatically completed one is given in Tab. 1. The completeness 
increases whereas correctness and RMS decrease slightly. This means, most of the added links were correct, but geo- 
metrically not as accurate as the initial result. This was to be expected because the check was performed on imagery 
with a ground resolution which was worse by a factor of four. The decrease of the mean detour factor signalizes that a 
lot of important gaps within connected components could be closed. The connectivity increases significantly, i.e., most 
of the different connected components could be connected by new links. Altogether, the results show that the main 
improvements could be achieved in the topology of the road network. 
  
| | initial | completed | 
  
  
  
  
  
  
  
  
Completeness 80.7% 85.8% 
Correctness 92.7% 90.5% 
RMS 1.05m 1.33m 
Mean detour factor 1.44 1.04 
Connectivity 78.5% 97.9% 
  
  
  
  
  
Table 1: Evaluation results 
6 SUMMARY AND OUTLOOK 
The paper deals with grouping based on functional characteristics of the road network. This is useful to improve the 
results of automatic road extraction where — until now — grouping is mostly performed on a geometric and often local 
level. Major requirements for the road network are to allow for fast, cheap, efficient, and secure transports. The presented 
approach is able to determine hypotheses for new links based on these requirements and to check them using image data. 
Results are presented for a road network extracted automatically from digital imagery. The results are evaluated using 
existing as well as newly defined quality measures. The evaluation demonstrates the feasibility of the presented approach. 
The new quality measures prove to be very useful. 
Future work will be directed towards improvements and extensions of the presented approach. The use of knowledge 
about the scene, e.g., in the form of a digital terrain model or information about land use will improve the calculation of 
the optimal distance. 
The current approach can only propose new links, it cannot be used for the deletion of parts of the road network. Also 
this should be done, based on the function of the roads. A basis might be the approach presented in (Morisset and Ruas, 
1997) which uses multi-agent modeling to determine the importance of roads based on the amount of *road-use". This 
approach has been developed for generalization tasks, but it could possibly be modified for the task of determining parts 
of the network which can be deleted. 
Evaluation is an important research topic, because the quality (geometric as well as topologic) is a decisive factor in the 
introduction of automatic road extraction into practical work. Further investigations will be undertaken to improve the 
proposed quality measures. 
  
984 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.