.ines and 
ms for a 
skeleton 
closing, 
‚are of 
nes may 
re trace 
ed. One 
3) is to 
irichlet 
image. 
ages and 
thinning 
skeleton 
A2). 
he line 
irection 
(by in 
in [16]. 
s to be 
cludes a 
e system 
and line 
skeleton 
ines to 
rpolated 
1lation. 
)f inner 
| taneous 
.richlet 
IT image 
me the 
1 value 
. of the 
polygon 
lelaunay 
sponding 
ng the 
sk. The 
| when 
Ir parts 
Ss not 
om four 
gure 8 
les are 
ines to 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
mask 4 pixels topology scan on 
tesselated image Thiessen polyg. 
1 2 
3 4 
using 2 confused 4 intersecting 
diagonals topology triangles 
4e—— 2 Bede b 
PAS PARE 
3—— 4 C d 
using 1 unique 2 adjacent 
diagonal toplogy triangles 
F2 à ——e D 
A i 2i 
3——4 i-a 
  
  
  
  
  
  
  
  
  
Fig. 7: 
VE TN TR TS X EN S X ny 
Mask and result of topology scan 
  
  
  
  
  
and skeleton 
Fig. 8: TIN generated from contours 
lines 
  
Fig. 9: DTM from contours and skeleton 
959 
The topology of the network is documented by two 
tables. The first relates every triangle to its 
nodes by identification number; the second relates 
every triangle to its neighbours. 
The TIN provides a better data structure than the 
original contours for deriving various . DTM 
products, and without loss of information. Based 
on TIN, Alias [1] developed a set of programs for 
computing a grid DTM (in ILVIS format), generating 
contour lines and slope models. The improvement of 
the grid DTM, compared with figure 1, is 
illustrated in figure 9. Linear interpolation is a 
good choice for rough terrain with many 
breaklines. Other interpolation methods are 
adequate, if smooth surface representation is 
required (see [11]). 
CASE STUDY 
How does the automatically derived skeleton 
compare with a photogrammetrically sampled one? 
We selected mountainous terrain in southern France 
(near Bonnieux) as a test site. Contours of an 
area 3 by 3 km were digitized from a 1:50,000 
topographic map with a contour interval of 20m. 
After interactive editing, the contours were 
rasterized vith a 6.5m pixel size. Figure Al shows 
the contours overlaid on the relief-shaded DTM 
which was produced directly from the contours. 
Figure A2 shows the automatically-derived skeleton 
lines, overlaid again on the same relief shades as 
in figure Al. The skeleton lines were produced as 
illustrated in figure 3. 
Photogrammetric selective sampling was done on the 
Planicomp C120 1:60,000 photographs. Digitizing 
the ridge and drainage lines, peaks, pits and 
passes took some 4 hours. Figure 10 shows the 
portion of the overlaid skeletons that corresponds 
two-colour version of the entire 
[10]). The grid DTM generated 
and skeleton 
to figure 9 (a 
area is given in 
through triangulation from contours 
(see figure A3) has lost its plateaus! 
    
« C» 
Em 
ANA 
of skeletons from manual 
'fringes' 
  
    
Fig. 10: Superimposition 
and automatic extraction (the 
belong to the automatic one) 
As shown in figure 10 (although it is difficult to 
discern which lines belong to which skeleton), the 
overall agreement of the two sets of lines is 
good. In addition to the differences caused by 
insufficient registration (the usual problem vith 
control points), there are obvious differences in 
what is extracted by the two methods. The