stanbul 2004
,, XYZ RMS
respectively.
f-GCPs and
d errors (in
tion for the
luals reflect
'rrors reflect
lon error and
accuracy of
curacy of the
justments of
ne order of
nbination of
ror (3 m) in
1ding on the
tial pointing
in all RMS
8 m and 3.3
r (3 to 5 m)
stereo-pairs,
n the least-
propagation
-models, but
1e residuals.
VIS residuals
je order of
nodelling or
of the 3D
ndent check
10% smaller
cause fewer
used in the
AS errors on
d to sensor
across-track
vith in-track
3D physical
final version
hus obtained
cation of the
RMS errors,
error of half-
of the 3-D
features, but
etter, in the
uation of the
:valuation of
ll DEM (120
ige reference
| 4 is a sub-
LIDAR area
he total area)
s and their
e River. The
which were
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004
automatically removed but not interpolated. The full DEM well
reproduces the terrain relief and the different cartographic and
topographic features, which can be seen in Figure 1: such as the
mountains and valleys, the Saint-Lawrence River and its large
island. Even small relief features between the mountains and
the Saint-Lawrence River valley were captured. Much more
topographic details are more noticeable in Figure 4.
(120 km by 60 km; 10 m by 5 m grid spacing). The
black areas are the 5% mismatched areas. The
yellow box represents the HRG DEM (60 km by 60
km) and the green box the Lidar (5 km by 13 km).
SPOT-5 2003 Courtesy SPOT-IMAGE
Figure 3. DEMs extracted from in-track HRS stereo-images
N eo
FI UA , Clpud-
t s
fa
uk » s i5. »
S ("in
i *
a * .
ERU.
Figure 4. Sub-area (5 km by 5 km; 5 m grid spacing) of DEM
extracted from HRS stereo-images. The large black
areas are the mismatched areas and the small black
dots are the blunders, removed but not interpolated.
SPOT-5 2003 Courtesy SPOT-IMAGE
Quantitative evaluation of DEMs was conducted with the
comparison of the LIDAR elevation data in the overlap area and
five to six million elevation points were used in statistical
computations. — Table 2 gives the results computed from
elevation errors for the HRS and HRG DEMSs: the linear errors
with 68% and 90% levels of confidence (LE68 and LE90,
respectively), the bias and the percentage of class over three
times LE68 (in metres).
DEM Area LE68 LE90 Bias Over
Evaluation (m) (m) (m) Three
| | | | LE68
HRS | Total surface | 55 m 10m E 2m 2.2%
HRG Total surface | 65m | 10m T. 2m | 0.796
i HRS , Bare surfaces | 2.7m |: 56m 02m | 4%
| HRG | Baresurfaces | 22m | 50m | 2m | 3%
Table 2. Statistical evaluation of DEMs stereo-extracted from
HRS in-track and HRG across-track stereo-pairs for
the total area and the bare surfaces: linear errors
with confidence levels of 68% (LE68) and 90%
(LE90), bias, and percentage over three LE68.
For HRS DEM, LE68 of 5.5 m was achieved and are good
compared to the stereo bundle adjustment RMS Z-errors on
well-defined ICPs (4.7 m). LE68 corresponds to an image
matching error a little less than £1 pixel (line spacing of 5 m
and B/H of 0.85), which is similar to previous results generally
achieved with different VIR medium-resolution stereo-images
(1-pixel image matching accuracy) (Gülch, 1991). While LE68
(6.5 m) of HRG DEM is a little worse than HRS LE68 due to its
smaller 8/H, the same image matching error of +1 pixel (pixel
spacing of 5 m and B/H of 0.77) is obtained.
The largest errors (three times LE68), although representing
only a very small percentage, are out of tolerance and cannot be
acceptable for DEM in a topographic sense. In order to locate
and understand these largest errors, they were superimposed on
the DEMs or the ortho-images. Most of these large errors
resulted from the clevation comparison of the top of tree versus
the ground due to the different spatial resolutions of SPOT and
LIDAR data and to the different acquisition seasons (deciduous
with or without leaves). These errors are then specific of the
cartographic data and study site largely covered by forests, but
are not representative of the general SPOT stereo-performance
for bald DEM generation. In fact, these DEMS stereo-extracted
from HR data are digital surface models (DSMs), which include
the height of natural and human-made surfaces. The smaller
sensor resolution and the more accurate the DEM, the more
noticeable are the height of some surfaces and the resulting
cartographic features. Consequently, a second elevation
accuracy evaluation was performed only on bare surfaces,
where there is also no difference between the SPOT stereo-
extracted elevation and the LIDAR data.
These results over bare surfaces (Table 2): LE68 of 2.7 m and
2.2 m for HRS and HRG DEMs, respectively are very good
relatively to the pixel spacing. These results are also more
consistent with a priori 3-D restitution accuracy computed from
the stereo-bundle adjustments over ICPs (around 4.7 m and 2.9
m in Z, respectively). The largest percentage of errors over
three LE68 (3-496) is due to isolated trees in the bare surfaces.
Strangely, the multi-date HRG acquisition (5 m pixel spacing
and B/H of 0.77) achieved a parallax error of one-third of pixel,
better than the half-pixel error achieved with the same-date
HRS acquisition (5 m line spacing and B/H of 0.85). It is
