International Archives of the Photogrammetry, Remote Sensing and Spatial Information Science 
larly, i.e. they are better correlated with each other. The sur 
face appears smoother and more realistic. The. modelled 
scene radiance (Fig. 2d) can hardly be distinguished from the 
original ortho-image (Fig. 2a). There are, of course, differ- 
ences caused by the limitations of the present program ver- 
sion. The somewhat brighter appearance of Fig. 2a as conr 
pared to Fig. 2d is probably due to a slightly varying BRDF, 
i.e. deviation of the actual reflectance law from Lambert. Fig. 
2a also suggests two distinct regions with different albedo 
factors. By applying only one albedo factor to the image data 
this brightness difference has been evened out thus leading to 
a duller and less bright result image. Further differences are 
caused by the few small shadow areas, since shadows were 
incorrectly treated as non-shadows. 
A crucial point is the optimal value for the penalty factor. We 
processed the data with a series of different values. The one 
which intuitively and subjectively gave the most pleasing res 
ult was considered optimal (0.005). We are, however, still 
unable to comprehend its relationship with the underlying 
stochastic model as given by Equ.(3). Yet, some statistics of 
the differences between SFS-refined DEM and initial DEM 
may give an indication of the significance of the result. Real 
ize that pixel size is 232 m. 
Maximum difference: 279 m 
Mean difference: 0.46 m 
Minimum difference: -265 m 
Standard deviation: 32m 
Fig. 3 exhibits the main processing stages of a 150 by 150 
pixel subsection of an HRSC scene taken on 24 February 
2004 at 266 km altitude in orbit 143. Approximate location is 
137? W longitude and 40? N latitude in the Acheron Fossae 
on the northern slope of Olympus Mons. The topographic 
surface is therefore generally inclined towards North. .Illu- 
mination is somewhat from the south with 48° sun elevation. 
The ortho-image prepared by DLR is illustrated in Fig. 3a. It 
has been radiometrically enhanced for better detail recogni- 
tion. With 100 m pixel size the scene covers an area of 15km 
by 15km. Printing scale is 1:400°000, hence more than 6- 
times larger than that of the MOC image. The corresponding 
DEM is shown as grey scale image in Fig. 3c. It was determe 
ined by the DLR data processing group by image matching 
(SFM) from the HRSC stereo channels at the original 30 m 
resolution level. The modelled scene radiance (Fig. 3b), 
however, shows enormous deficiencies in the DEM caused 
mainly by insufficient surface texture needed for SFM. Real 
ize that Fig. 3b is supposed to appear as Fig. 3a! The contour 
lines of the SFM-determined DEM (Fig. 3d) may give an 
idea of the difficulties SFM has had for matching the image 
data. 
Figs. 3e, 3f, and 3g show the same scene after SFS-refine- 
ment after 10 CGM-iterations. Despite some unwanted (and 
not yet comprehended) boundary effects visible mainly in the 
contour map (Fig. 3g), the resulting scene radiance (Fig. 3e) 
is much more alike the original ortho-image (Fig. 3a). AF 
though there are — probably illumination induced — artefacts 
visible in the DEM-image (Fig. 3f) and other errors due to in- 
complete modelling, a dramatic improvement of the pure 
photogrammetric, i.e. SFM-derived, DEM is obvious. As to 
be expected (intuitively) the larger scale imposes more prob 
lems for SFS concerning details. The shadow region along 
the southern inside slope of the big impact crater exhibits er- 
rors as well. Further improvenients may be expected in due 
course with methodical expansions and refinements of DRS. 
  
Fig. 3b,c,d: 
Scene radiance, DEM im- 
age and contour map prior 
to SFS 
  
Fig. 3e 
Fig. 3e.f.g: 
Scene radiance, DEM ima- 
ge and contour map after 
SFS-refinement 
Fig. 3g 
Figure 3. HRSC narrow-angle scene 15km by 15km 
at printing scale 1:400'000 
1302 
s, Vol XXXV, Part B4. Istanbul 2004 
  
  
Internation 
The main c 
and homog 
derived fro 
outcome o: 
fine-tuned : 
face DEM 
mentary an 
firm the co 
proximate 
main defic: 
with shado 
model, par 
and improv 
The refined 
tion for the 
mosaicked 
may alread 
shown again 
SFS-refinec 
tions. Fig. 4 
south-west, 
in Fig. 4a, : 
timisation o 
  
Figur 
  
ACKNOWI 
This work w 
tion and Re: 
the German . 
The authors 
members M: 
Institute for ] 
and unselfisl 
for our rese. 
Bernd Jeschk 
portant conti 
particularly i 
stitute for E: 
Academy of 
matters of op