is the 
y. To 
about 
is not 
n the 
io the 
result 
  
arked 
|. The 
rea is 
ion is 
0 ( 
pixels 
arious 
e still 
  
S.A.B. Kim 
4 SEGMENTATION SCHEMES 
A problem after applying COG and ECI criterion is that holes in a DEM, arising due to stereo-match failure, still remain 
(Fig. 4b). It is needed to fill the holes. However, a hole-filling scheme should not fill a lake or rivers (for example, 
marked by a rectangle in Fig. 4b). Lakes are identifiable by their large size whereas match-failure regions tend to have 
smaller size. To select a hole surrounded by land, a closed and empty region is segmented. Then segments with areas « 
6000 pixels or 15 km' are regarded match-failure and consequently filled. Empty pixels are grouped in 8-connectivity 
sense. The grouping is performed by scanning four times (each scan starting from each corner of an image). Four scans 
are needed since one scan makes one segment into several fragments (marked by loops in Fig. 5a). The hole 
segmentation scheme fails in preserving rivers (which is a w-shaped feature at the top in Fig. 5a) as an empty area. This 
still remains to be refined. 
  
  
  
  
  
  
Fig. 5 Results of (a) applying hole-fill segmentation scheme and (b) final DEM. In (a) scan is performed only once thus 
the segments are fragmented. 
The second segmentation scheme is necessary to remove noise. The noise is a small land (of several pixels large) 
floating off the coast or off the lakeside. The noise occurs due to imperfect performance of ECI thresholding. To 
remove the noise, land is segmented and segments with sizes « 100 pixels are eliminated. 4-connectivity segmentation 
is employed and 4 scans as in the hole-filling segmentation are applied. The hole-filling and the noise-remove 
segmentation are completed in less than 3 minutes on SGI Octane platform. A DEM after the two segmentation is 
shown in Fig. 5b. The series of intelligent interpolation schemes eliminate 15% of elevation interpolated by simple 
Gaussian method. 
The final DEMs are compared in a different region in Fig. 6. The DEM from this research produces a correct and 
detailed coastline in this region as well. By comparison the coastline of PCI DEM is deteriorated by erroneous 
elevation (spikes) and artificial islands. Intergraph DEM is better than PCI DEM, however the coastline is not as 
detailed as this study's DEM. In addition Intergraph DEM adds substantial error along the coast (large and high island- 
like feature). The relationship between color codes and elevation is different among the three cases. 
5 CONCLUSIONS 
Intelligent interpolation schemes are implemented to improve the quality of a SPOT DEM. To eliminate blunders of 
simple interpolation methods such as those in Franke (1982) or Kim et al. (1999), 
e COG parameter is used: if a COG within an interpolation radius is greater than 0.6, the interpolated value is ignored. 
This eliminates blunders along image boundaries. 
e ECI parameter is used: if ECI within an interpolation radius is greater than 0.8, the interpolated value is ignored. 
This eliminates thread-like features in the result of COG thresholding. 
e Holes in the DEM with sizes less than 6000 pixels or 15 km” are segmented and filled with interpolated values. This 
is necessary because some match-failure regions should be filled but COG and ECI criteria leave them empty. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 499