ibul 2004
satellite
g image
ojection
eference
nce the
rallel to
ojection
; study
sometric
y image
e space,
process.
which in
<elihood
^ometric
ct space
bing the
in image
pixels is
escribed
to many
ner and
d works
'cting it,
es of the
| y2) for
n image
gy. The
rrelation
ice point
ates and
asure is
between
^ which
n as the
repeated
strategy
rrelation
(2)
ister and
e of the
in object
selected
pace and
natching
covering
ere then
ates and
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004
5. RESULTS
Two tests were carried out on an area of 10km by 10km centred
on the Melbourne central business district. The height
difference throughout the test area was only about 60m (except
for large office buildings), with the terrain being as indicated by
the DTM shown in Figure 2.
Figure 2. DTM of 10 x 10km Melbourne test area.
The 100km? test area was selected since it coincided with a
reference DTM and 80 precisely GPS-surveyed GCPs. The
difference between the two tests was in the number of GCPs
used to calculate the parameters of the affine projective model.
Although only four points are required to determine the
parameters, a larger number gives a degree of redundancy.
Therefore in the first test 10 GCPs were used, and in the second
20 were employed.
Table 1 shows the number and percentage of successful
matches for each test. In this study, successful and unsuccessful
matches have been differentiated from each other according to
the values of the cross-correlation coefficient.
model, and ground surveyed check points. These results are
presented in Tables 2 and 3 below. There were 60-70 surveyed
check points and around 110 DTM check points used.
y 7 0.8 y 0.0
10 GCPs x
RMS ;
(m) c (m) RMS (m) c (m)
Difference:
DSMvs. 4.11 4.14 3.91 3.94
check points
Difference:
DSM vs. 5.59 5.62 4.96 4.98
DTM
Table 2. Height difference between DSM and reference data for
test with 10 GCPs.
y» 0.8 y» 0.9
20 GCPs
RMS (m) o (m) RMS (m) o (m)
Difference:
DSM vs. 3.18
check points
2.96 3.01
LI
N
—
Difference:
DSM vs. 4.78 4.80 3.96 3.98
DTM
Nistof Number of
GCPs matched y? 0.8 vis
Pel points
34352 21119
43572
10 43572 (78.84%) (48.27%)
; . 34597 21359
20 43894 (78.82%) (48.66%)
Table 1. Number of successful matches.
It can be seen from Table 1 that there is no significant
difference between the proportion of successful matches for
each test. This is a very encouraging result since it indicates
that reducing the number of GCPs from 20 to 10 does not
impact negatively on the results. From a practical point of view
this is very important, since collection of high quality ground
control can be a very time consuming process. À further point
to notice is that the test results endorse the use of the matching
strategy used in this study: in both cases nearly 50% of the
matched points have a correlation coefficient greater than 0.9.
Results of each of the tests were further analysed by comparing
the generated surface model with both a reference terrain
Table 3. Height difference between DSM and reference data for
test with 20 GCPs.
Tables 2 and 3 show the RMS height differences, and the
standard deviations (c) of those differences, between the
derived surface models and the reference data, for two different
groups of triangulated points: those with cross-correlation
coefficient values greater than 0.8, and those with values
greater than 0.9. As would be expected, the groups of
triangulated points with higher cross-correlation coefficients
give a better surface representation than those with lower cross-
correlation values. Even so, the differences between the groups
are quite small, with this result being consistent for both tests.
The points used for the comparison between the surface model
and the reference data were chosen carefully so that the height
differences could be measured in regions unaffected by ground
features such as buildings and vegetation cover. The fact that
the differences between the surface models and the check points
are less than the differences between the surface models and the
DTM is difficult to explain, but is most likely due to errors in
the reference DTM (either relative or absolute), which was
created from stereo aerial photography and required significant
manual editing.
Since the GPS surveyed check points are obviously the most
reliable reference, they have provided the best yardstick against
which to assess the SPOT 5 surface model. With just 10 GCPs
to calculate the parameters of the affine projective model, the
RMS differences were of the order of 4m, which equates to 0.8
pixels. When 20 GCPs were used to calculate the affine
projective parameters, the RMS height differences were around
3m, or 0.6 pixels.