Jamet - 8 
mountainous areas, even if the contours keep an acceptable cartographic aspect, the 
process won't fit the requirements (as expected), due to the presence of a dense forest 
coverage. On the three other areas, the prognostic is better. The overall accuracy seems 
compatible with the BDTopo content (complementary quality controls on the manual 
contour lines of the BDTopo are being performed to confirm this point). Moreover, as 
shows fig. 1, the remaining errors, that are mainly due to the presence of dense 
vegetation in the thalwegs, seem easily detectable from their characteristic shapes, and 
should represent an acceptable editing time (manual digitisation of the contour represent 
from 10 to 15 hours works on an equivalent area of an average landscape). 
4. Comparison of the two processes 
4.1. Weaknesses of semi-automatic methods. 
The first experience lead on road extraction can temperate much enthusiasm about the 
introduction of automatic tasks in an interactive process. It appears that the addition of 
new tools in the operator environment, be they quite efficient by themselves, also brings 
more complexity in the capture task. If the time for actual data capture decreases much, 
this benefit keeps far lesser on the overall time (which is the only of interest for the 
producer). 
Unfortunately, the time losses due to the automation itself (that is to the errors of the 
road follower) keep low in the present experience. This means that we cannot count 
with an amelioration of the process based on a simple amelioration of the robustness of 
the automatic road follower. The major part of this "unproductive" time is spent in 
waiting... for the operator reaction. This keeps quite affordable in rural areas where long 
roads in a simple context can be handled as single objects. But this makes a fully 
manual capture as efficient in urban areas, where the road sections are quite short, and 
the interaction level keeps very high. Theses results may be significant at a more 
general level: data capture isn't reduced to plotting, but includes a logical definition of 
objects (through topology handling in our case). When the density of relations between 
objects is high, the management of the representation becomes the bottleneck of the 
process.