1371
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
particular emphasis for the second. With respect to the former,
the accuracy of the DTMs were also verified for geocoded
rDTMs (geocoded after the generation process), while for the
latter it was deeply studied the influence of the GCPs (number
and the geometric distribution) on the DTM’s accuracy.
Parameter
Value
Number of tie points
60
Search window size
800 pixel
Moving window size
30 pixel
Correlation coefficient
>0.8
Terrain detail
level 6 (over 7 levels)
Table 1. Optimization of the DTM’s automatic extraction
procedure.
43.90000
43.85000
43.80000
■§ 43.75000
■
« 43.70000
I 43.65000
3 43 60000
43.55000
4.80000 5.00000 5.20000
Decimal Longitude
5.40000
Figure 2. Ground control points distribution over the study area.
One of the major factor determining the accuracy of the
generated DTM, as well as the processing time needed for its
generation, is the ‘terrain detail level’ parameter. Since the
DTM extraction is carried out by means of automatic image
matching to find homologous features on the Aft and Fore
images of the stereo pair, the terrain detail level determines the
number of image pyramids used during the image matching.
The use of the ‘minimum terrain detail level’ stops the process
after the coarsest level of the image matching, while the use of
the ‘maximum terrain detail level’ iterates until the image
matching is performed at the highest resolution as possible.
In this study it was observed that the use of terrain detail level 6,
instead of the maximum level (level 7), reduced the computing
time by a factor of 4.6 without any sensible decrease in the final
DTM’s accuracy.
3.3 Ground control points selection
Because the satellite’s pointing and ephemeris information are
often inadequate for their use in applications involving high-
resolution imagery, DTMs must be referenced to a map
coordinate system using GCPs (Lang, 1999). The use of GCPs
in the DTM generation process brings to the so called ‘absolute
DTMs’, where ‘absolute’ means that the terrain’s elevation
values are referred to a geodetic datum.
When no GCPs are available it is still possible to produce
DTMs from stereo pairs .by means of automatic image matching
algorithms,. In this case the so called ‘relative DTMs’ will be
produced, where ‘relative’ means that the terrain’s elevation
values are not referred to a geodetic datum but to an arbitrary
plane (e.g., the lowest value in the scene).
This study investigated the DTM’s generation for both the
relative (rDTM) and the absolute (aDTMs) methods, with
4. RESULTS AND DISCUSSION
4.1 Medium-resolution DTM generation
4.1.1 The influence of GCPs in the generation of relative
DTMs:
When generating relative DTMs, the output is not correctly
georeferenced and elevation data are referred to an arbitrary
plane. Thus, without properly geocoding the error (in the z
coordinate) of the Cartosat-1 rDTM was observed to be many
hundreds of meters. After georeferencing the rDTMs using 5
GCPs of the original C-SAP dataset, the vertical accuracy
drastically increased to less than 20m (LE90).
The DTM’s accuracy was tested at two different scales: i) at
local level using as Independent Check Points (ICPs) the
remaining 17 GCPs of the original C-SAP dataset, and ii) at
global scale using the MNTDBTOPO® DTM as ground truth
for the whole test site.
For geocoded rDTMs, results showed no significant difference
in the mean value of residuals (p) for both the comparison to
the ICPs and the MNT DBTOPO® DTM, respectively 4.33m
and 4.24m. On the contrary, this is not true for standard
deviation (a), RMSE and LE90, which showed higher values
when computed with respect to MNT DBTOPO® DTM
(<r=13.55m, RMSE=14.19m and LE90=19.00m), indicating that
the statistics of the ICPs (cj=8.09m, RMSE=9.02m and
LE90=16.10m) could not be assumed as a correct term of
comparison for the complete elevation range.
4.1.2 The influence of GCPs in the generation of absolute
DTMs:
When generating absolute DTMs, the number and geometric
distribution of the GCPs have a great impact in the final DTM’s
accuracy. For this reason, the original set of 22 GCPs was
divided into a first subset of j GCPs used for the DTM’s
georeferencing and a second subset of 22-j ICPs used for the
evaluation of results.
Several tests have been done, varying j form 2 to 22 to find the
optimum number of GCPs to be used. Table 2 shows a
summary of results.
With the exception of test #5, the comparison of the Cartosat-1
aDTM with the ICPs showed a RMSE between 1.41m using 9
GCPs (for test #14) and 5.17m using 2 GCPs (for test #3), while
the LE90 was between 1.91m using 6 GCPs (for test #11) and
8.91m using 2 GCPs (for tests #2 and #3). Even if the observed
mean values of residuals were small and between -0.12m (for
test #13) and -2.80m (for test #3), they were not null, thus
indicating a small bias in the output data. Regarding their
standard deviation, the observed values were between 1.36m
(for test #11) and 4.49m (for test #1).
The best overall performance was obtained for test #14 using 9
GCPs (p=0.47m, cr=1.39m, RMSE=1.41m and LE90=2.77m),