International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004
computed stereo-model (Step 3) with 3D least-squares
stereo-intersection; and
6. . Generation of regular grid spacing with 3D automatic and
3D visual editing tools: automatic for blunders removal
and for filling the small mismatched areas and visual for
filling the large mismatched areas and for the lakes.
GCP collection
Least-squares stereo
bundle adjustment
TS
1/ 20,000
topographic maps
Stereo images
Meta-data
3D CCRS multi-sensor
physical model
Stereo model
parameters
Multi-scale mean
normalized correlation
3D stereo intersection
3D automatic and
visual editing tools
Final DSM
Digital terrain
surface model
Statistical evaluation of
elevation errors
Figure 2. Processing steps for the generation of DEMs from
stereo-images and their evaluation with LIDAR
data.
The DEM is then evaluated with the lidar elevation data. About
5 300 000 points corresponding to the overlap area were used in
the statistical computation of the elevation accuracy. Different
parameters (land cover and its surface height), which have an
impact on the elevation accuracy, were also evaluated.
4. RESULTS
4.1 Results on The Stereo-Model Computations
As a function of the number of GCPs used in the stereo bundle
adjustments, two sets of tests were performed for each stereo-
pair. Set 1 was conducted with all the GCPs while Set 2 was
performed with a reduced number of GCPs (10-18) and the
remaining points as ICPs. In Set 2, 10 GCPs were used because
previous results demonstrated that this was a good compromise
with this dataset to avoid the propagation of input data error
(cartographic and image pointing) into the 3-D physical stereo-
models (Toutin, 2004).
GCP RMS ICP RMS Errors
Residuals (m) (m)
X Y Z X Y Zz
Test GCP/
Stereo ICP
|-HRS 98/0 10.1 7.6 3.8 - - -
1-HRG 33/0 2.6 3.1 3.3 - - -
2-HRS 10/88 7.1 6.4 3.1 139 3.7 4.7
2-HRG 10/23 ES 1.4 1:3 2.6 2.2 2.9
Table 1. Results from the least-square bundle adjustment of the
3D physical model for the stereo-pairs (HRS in-track and HRG
across-track): with the number of GCPs and ICPs, XYZ RMS
residuals and errors (in metres) on GCPs and ICPs, respectively.
Table 1 gives for each stereo-pair the number of GCPs and
ICPs, the root mean square (RMS) residuals and errors (in
metres) of the least-square adjustment computation for the
GCPs and ICPs, respectively. GCP RMS residuals reflect
modelling and GCP accuracy, while ICP RMS errors reflect
restitution accuracy, which includes feature extraction error and
thus are a good estimation of the geopositioning accuracy of
planimetric features. However, the final internal accuracy of the
3D modeling will be better than these RMS errors.
Due to the large redundancy of equations in the adjustments of
Set 1, the RMS X-Y residuals are on the same order of
magnitude as the input data errors, being a combination of
image pointing error (one pixel) and planimetric error (3 m) in
addition to the propagation of Z-error (3 m) depending on the
viewing angles. With HRS stereo-pair, the differential pointing
error due to a rectangular pixel is well reflected in all RMS
results On the other hand, the RMS Z residuals (3.8 m and 3.3
m) approximately reflect GCP image pointing error (3 to 5 m)
with B/H of 0.85 and 0.77 for the HRS and HRG stereo-pairs,
respectively. The use of overabundant GCPs in the least-
squares adjustment reduced or even cancelled the propagation
of the input data errors into the 3-D physical stereo-models, but
conversely these input errors are reflected in the residuals.
Consequently, it is “normal and safe” to obtain RMS residuals
from the least squares adjustment in the same order of
magnitude as the input data error; however, the modelling or
internal accuracy is better (less than one pixel).
Set 2 of the tests enabled unbiased validation of the 3D
positioning and restitution accuracies with independent check
data. First, the RMS residuals on GCPs are 20-40% smaller
than the RMS residuals resulting from Set 1 because fewer
GCPs, and thus less equation redundancy, were used in the
least-squares adjustments. On the other hand, RMS errors on
ICPs are 9-14 m and 2-3 m or when compared to sensor
resolution, one-and-half and half-pixel for in- and across-track
stereo-pairs, respectively. The worse results with in-track
stereo-pair are due to the preliminary version of the 3D physical
model for HRS data. Equivalent results with the final version
of the 3D physical model for HRS data should be thus obtained
for the stereo modelling (half-pixel).
Finally, the Z-RMS errors on [CPs are a good indication of the
potential accuracy for the DEMs. However, these RMS errors,
which include the extraction error (image pointing error of half-
pixel) of ICP features, are only an estimation of the 3-D
restitution accuracy of planimetric and elevation features, but
the internal accuracy of stereo-models is thus better, in the
order of sub-pixel.
4.2 Results on DEM Evaluations
The second result is the qualitative and visual evaluation of the
full DEMs and the quantitative and the statistical evaluation of
the DEMs with the LIDAR data. Figure 3 is the full DEM (120
km by 60 km; 10 m by 5 m grid spacing) in the image reference
extracted from the in-track stereo-pair and Figure 4 is a sub-
area (5 km by 5 km; 5-m grid spacing) over the LIDAR area
but in the map reference. The black areas (5% of the total area)
correspond to mismatched areas due to clouds and their
shadows, as well as the lakes and the St. Lawrence River. The
black dots in Figure 4 are the blunders, which were
Inte
autc
repr
topc
mot
isla
the
topc
