produced, the ICPs are measured within this image and
compared with the reference. Then the resulting residuals
serve as a measure for the desired geometrical accuracy.
3.3 Practical results
Several tests using MOMS-02 imagery regarding the
dependency on different test sites, different numbers of
GCPs, different pixel measurement techniques as well as
different transformation models have been conducted. The
used data sets of Dubai (U.A.E.), Harare (Zimbabwe) and
Pasajes (Bolivia) show various drawbacks so that each
aspect has to be handled separetely without the possibility
for an absolute comparison.
3.3.1 Test sites: Examinations on different test sites
even in one single datatake can lead to inhomogenious
results due to varying image or topography contrast, differ-
ent target qualities (e.g., road crossings vs. hill tops),
changed weather conditions or unequally accurate refer-
ence material. Table 2 compares three sub-scenes over
Dubai whereby the object coordinates of the GCPs have
always been taken from the same source. There is a conti-
nous decrease in accuracy from sub-scene A ("City Cen-
tre": good contrast, narrow road crossings as targets,
stereo impression possible) over B ("Al Jumayrah": low
contrast due to sensor problems, road-crossings) towards
: Result: ICP-residual
Variable: o's
No. of GCPs
Sx Sy sz
24 3.7m | 39m ! 47m
15 4.3m 3.5m 4.2m
7 38m | 34m | 40m
4 40m | 36m | 47m
Constants: Test site = Dubai-A / Reference = 1 : 10 000 / Pixel sizes =
4.5 m & 13.5 m / No. of ICPs = 26 / Measurement = stereo / Transfor-
mation = collinearity.
Table 3: Influence of different numbers of GCPs
3.3.3 Pixel measurement techniques: Due to the along-
track stereoscopic capabilites MOMS-02 data are suitable
for deriving three-dimensional object information. Table 4
shows that there is a significant increase in accuracy -
especially in the vertical component - if the incoming
image coordinates are measured in stereoscopic rather
than in monoscopic mode because the pointing accuracy
is improved.
Result: ICP-residuals
C (“Bida Saif”: very low contrast, only some very wide road Variable:
crossings). Measurement
sx sy sz
: Result: ICP-residuals Mono 4.6 m 44m 10.9 m
Variable:
Test site
eX sy sz Stereo 4.2m 3.6m 45m
Constants: Test site = Dubai-A / Pixel sizes = 4.5 & 13.5 m/ Reference =
A: Best contrast & targets 3.3m | 46m 6.1m 1:10 000 / No. of GCPs (ICPs) = 25 (30) / Transformation = collinearity
B: Weak contrast 39m | 53m 7.5m Table 4: Influence of different measurement techniques
C: Poor contrast & targets | 100m | 7.6m | 13.4 m Applying an interactive stereo measurement on-screen it
Constants: Pixel sizes = 4.5m & 13.5 m/ No. of GCPs (ICPs) =7 (7) /
Measurement = stereo / Transformation = collinearity / Reference = 1 :
10 000
Table 2: Influence of different test sites
Comparing the test sites Dubai and Harare - both having
good image contrast - by using polynomial transformations
based on the same number of GCPs and ICPs yields dif-
ferences in the planimetric residuals in the order of 0.5 to
1.0 m in favour of the Dubai data set. This leads to the
conclusion that the reference material of Dubai (not gener-
alized orthoimage maps at 1 : 10 000) is superior to that of
Harare (combination of line maps at 1 : 5 000 and 1 : 25
000).
3.3.2 Number of GCPs: The number of used GCPs
affects the reliability of the adjustment process. Experi-
ences with MOMS-02 imagery show that about 10 points
per scene is a sufficient measure. As table 3 indicates,
also the minimum number of 4 points for the specific trans-
formation model (Jacobsen, 1994) will yield acceptable
results if well known and measured points are used.
744
has been found that although a multi-resolution combina-
tion of the HR-channel and the three-times enlarged chan-
nel FW (or BW) decreases the base-to-height-ratio by
factor 2, at the same time the pointing accuracy is impro-
ved - theoretically by factor 3 - due to the use of high reso-
lution data. In fact, the human eyes are so flexible that they
neglect blurring effects coming from the low-resolution
image.
Measurements within the Pasajes data set confirm these
experiences whereby now due to the lack of appropriate
absolute reference material the standard deviation of the
relative orientation (sg) of the corresponding stereo part-
ners serves as the accuracy measure (table 5). This value
is based upon an overdetermination of measured conju-
gate points and is transformed from pixel into correspon-
ding ground values. Both the multi-resolution and the
single-resolution approaches show nearly the same rela-
tive accuracy of about 0.9 pixel which gives evidence
about the same relative pointing accuracy. With that the
absolute accuracy can be improved by factor 3 if the multi-
resolution method with the HR-channel is used. - From
table 5 one can also conclude that the image quality of the
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
