rather strong datum resulted in an absolutely parallax-free
stereomodel.
In order to obtain an objective indication of the overall ac-
curacy of the adjusted object points, several tests with varying
numbers of control points were performed. For each test,
unused control was entered as tie point information in the ori-
entation adjustment and could therefore be considered as
check points. The diagram in Fig. 3 exhibits the attained
check point rms-errors vs. the number of control points used.
Obviously, the rms-errors increase with decreasing number of
control points, which my be an indication of still unmodelled
residual systematic errors. The rather constant distribution of
the planimetric rms-errors for tests with higher control density
suggests an influence of a constant error due to misinterpreta-
tion on the ground or in the imagery. Indeed, as (Fraser et al.,
1996) report, such errors in the order of 1-2 pixels cannot be
discounted in matching a feature on the ground with its imag-
ed position due mostly to changes of the terrain surface in the
time span between image acquisition (1993) and ground sur-
vey (1994 and 1995). The vertical rms-values seem not to be
influenced significantly by interpretation errors. This can be
deduced from Fig. 3 and from the fact that, due to the flat ter-
rain, the vertical measurement error is independent of the
planimetric residual error.
Number of check (tie) points
43 40 28 7
RMSE in m
0 6 9 21 42
Number of control points used
Figure 3. RMSE vs. number of control points
(Australia scene)
If the rms-error, p , is assumed to be related to the number of
control points, n, and to a constant interpretation error, Hr,
according to
W =
2 2 2
W7 + po + (k/n)”
then, assuming the vertical interpretation error to be zero and
the ratio k/ug — const. for all three coordinates, the follow-
ing empirical relations in units of m were deduced
14.0)” (47) 72.8
ir, 2 d13| pede 1976
0 36). n*1573
The estimated planimetric interpretation rms-errors in the or-
der of 14 m seem consistent with the previous statements.
Without it and provided 9 control points are available, realistic
measurement rms-errors of 9m, 12m and 7m may be expected
for the X-, Y- and Z-coordinates, respectively. The values
would drop to 7m, 9m and 5m, respectively, if 15 control
points were used. As these error values represent somewhat
absolute quantities, a truly remarkable result, indeed. It shows
that stereophotogrammetric measurements achieve sub-pixel
accuracies, particularly in height.
The high accuracy potential for elevations derived from
MOMS-02 stereomodels was confirmed by independent
stereophotogrammetric measurements along a GPS-controlled
DTM evaluation profile. This profile, situated in the western
part of the Australia scene, is identified by over 16km of a
fence line clearly visible in the imagery and was established
during the 1995 field survey (Fraser et al., 1996). In a preli-
minary investigation, ten reference points established along
the fence line, were measured repeatedly (five times each) in
the Planicomp in due course of the profile measurement. Fig.
4 exhibits the differences between the photogrammetric
heights and the surveyed control elevations. Notice the verti-
cal scale of 2m vs. the 500m in horizontal direction. Several
conclusions may be drawn. First, internal precision is in the
range of 1.2-3.0m standard deviation. Then, external ac-
curacy is biased by unmodelled systematics, e.g. model de-
formation or datum problems, and influenced by interpretation
errors. In the analyzed test data, rms-errors lie in the range of
4.0-9.0m. Finally, if an obvious linear trend is removed, the
remaining rms-values decrease to a range of 2.0-5.0m. A de-
tailled investigation of the entire height profile will be given
elsewhere.
Figure 4. Height differences ( in m) at ten reference points
along DTM evaluation profile
EXPERIMENTAL STEREOPLOTTING
The rather surprising high elevation accuracy could be repro-
duced in a first attempt to generate a "conventional" elevation
contour map. In an area covering 20km by 15km near the
center of the Australia scene (see Fig. 5 as part of the back-
ward looking image) and exhibiting relatively rough topogra-
phy with elevations up to 50m, contour lines with 10m equidi-
stance were measured on the analytical plotter. After some
familiarization with the rather small image scale, the human
stereooperator, until then only acquainted with aerial photo-
graphy, has found it increasingly easy to generate the contour
lines shown in Fig. 6. Continuous lines represent contours
with 10m interval, dotted lines with 5m. Notice that adjacent
542
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
