nbul 2004 
> test, our 
as, while 
d. Fig. | 
] network 
ipe of Fig, 
ts. 
site Thun, 
] junction 
> presented 
e lines and 
] junctions 
st by our 
have been 
> correctly 
  
ration with 
rland. The 
the VEC25 
extraction 
y L+T at an 
hieves very 
very high. 
0.5 m both 
irements of 
Fig. 1. 
International Archives of the Photogrammetry, Remote Sensing 
  
A higher completeness has been achieved by our partner when 
spring photography is used. This is because of less tree 
occlusions. Even some roads along forest borders can be 
extracted in spring images. In addition, some roads in fields, 
invisible in summer images, are visible in spring images and are 
extracted, thus further contributing to a higher completeness of 
the extracted road network. 
We also tested our system on unknown data provided by the 
National Geographic Institute, Belgium. Although the images 
are black and white with quite poor radiometric quality, and no 
DSM is available, the performance of our system is also quite 
good in flat open rural areas. By comparing the results with the 
manually measured reference data for ca. 13 km roads, the 
achieved completeness and correctness are 97.6% and 98.1% 
respectively (Zhang and Baltsavias, 2002). 
The system has been modified to work even with orthoimages, 
whereby the 3D information is not extracted by image 
matching, but by overlaying the 2D information on the DSM 
and DTM. Extensive tests conducted by our Swiss project 
partner using various resolution orthoimages (0.20 m ~ 0.60 m) 
have shown that the results are quite similar to that from the 
stereo imagery, and that an increase of the pixel size leads to a 
proportionally much smaller deterioration of the geometric 
accuracy of the extracted roads. With this development, the 
Dutch Ministry of Transport, Public Works and Water 
Management (MTPWWM) awarded to us, after an evaluation of 
various research systems, a project for a feasibility study of 
semi-automated updating of the Dutch road database, using 
color orthoimages of 0.5 m pixel size from aerial images of 
1:25,000 scale with 15 cm focal length. The study site is 
situated near the city of Weert in the province Limburg (in the 
south of the Netherlands), covering an area of 12 * 12 sq. km. 
The landuse changes gradually from open rural to urban, with 
the complexity of the scenes increasing correspondingly. The 
images were taken in June, 2000. The images do not have good 
quality; they are too green and noisy. In many cases the roads 
show very poor contrast with surroundings. The image edges 
are poorly defined; also color shifts between bands are 
observed. In addition, trees at road sides usually occlude roads 
very much in these summer images; some roads are even totally 
occluded. We also observed that the roadmarks on roads are 
very weakly represented in such images. The old road databases 
are created by digitizing 1:10,000 topographic maps, with an 
RMS error of about 10 m. The database allows distinguishing 
national roads and a small part of the provincial roads in the 
Netherlands, and provides the number of lanes for them. The 
other roads are in a single class. There are no clues that can be 
used to infer the approximate road width. Available height data 
are from laser scanning (raw and filtered heights). Both datasets 
have points regularly distributed with a 5m x 5m spacing. 
During the test, our system is only applied in open rural areas. 
Fig. 3 shows a portion of the test results. The roads in rural 
areas are correctly and reliably extracted by our system. In Fig. 
4, the details of road extraction and junction generation for this 
dataset are presented in several examples. 
Reference data for the Dutch dataset is not available at moment. 
The accuracy of the extraction result cannot be accessed. In 
each test image we computed the ratio of the length of the 
tracted roads to the length of rural roads in the existing 
database (the total length of the rural roads in the old database 
I5 ca. 500 km). The ratio values range from 80% to 92%, 
depending on the complexity of the scene. Generally, the 
Performance is worse compared to the performance on the 
and Spatial Information Sciences, Vol XX XV, Part B4. Istanbul 2004 
  
Swiss dataset. This is mainly caused by: (1) the poor image 
quality, (2) insufficiently road information in the existing road 
database, especially the lack of the road classes or road width, 
(3) the images are taken in summer, many roads are occluded by 
trees, (4) the worse spatial resolution of 0.5m compared to 
0.22m of the Swiss data. 
  
Fig. 3. Extracted 3D roads and road network in the test site in 
the Netherlands superimposed on image as black lines. 
  
Fig. 4. Details of road extraction and junction generation in the 
Netherlands dataset. The extracted roads are shown in 
black lines and the outdated roads in white lines. 
6. DISCUSSION AND CONCLUSION 
In this paper, we have presented a practical automated system 
for road extraction from stereo and ortho-images focusing on 
rural areas. The roads should have a minimum width of about 3 
pixels in order that edges on both road sides are extracted. The 
system has several advantages over other approaches. It uses 
existing knowledge, image context, rules and models to restrict 
the search space, treats each road subclass differently, checks 
the plausibility of multiple possible hypotheses, therefore 
provides reliable results. The system contains a set of data 
processing tools to extract various cues about road existence, 
and fuses multiple cues and existing information sources. This 
fusion provides not only complementary information, but also 
redundant one to account for errors and incomplete partial 
results. Working on stereo images, the system makes an early 
transition from 2D image space to 3D object space. Road 
hypotheses are generated directly in 3D object space. This not 
only enables us to apply more geometric criteria to generate 
hypotheses, but also largely reduces the search space, and 
speeds up the process. The hypotheses are evaluated in images 
using accumulated knowledge information. Whenever 3D 
features are incomplete or entirely missing. 2D information 
from stereo images is used to infer the missing features. By 
incorporating multiple knowledge, the problematic areas caused 
by shadows, occlusions etc. can be often handled. Based on the 
extracted roads, the road junctions are generated and modeled, 
thus the system provides an up-to-date and complete road 
network for practical uses. We also present in this paper the 
results. of road extraction in benchmark tests conducted 
independently by our project partner. The quantitative analysis 
using accurate reference data is also presented. The comparison 
of the reconstructed roads with such data shows that more than 
9476 of the roads in rural areas are correctly and reliably 
1057