liminating
component
es over the
which has
described
out by the
formation
netrists to
Baltsavias
ge that by
corrected
| provides
weighted
r forms of
d can also
0.3 pixel,
med on a
shift and
e number
ns is 3 to
does not
scarded.
rocedures
computed
meters in
ar within
ages, and
lue to the
n terrain
rotation
ch region
arameter
a certain
carded as
ween the
vill cause
evations.
a certain
e of the
a certain
which are
factor of
mally be
€ terrain
lange in
is being
used as a
This test
a digital
this test
d ortho-
E
'ount in
. speed;
. accuracy of the generated DEM.
Regarding the user interface, the use of icons, windows,
point and click selections, makes the whole package easier to
use. Response time is also important. This aspect has been
borne in mind for any procedure that is run interactively. For
example, two different drawing algorithms co-exist for
drawing a triangulation network, one for the overview, and
another for enlargements which explores and draws the
. network locally. The way the windows are refreshed has also
been carefully considered.
The most time consuming part of the software is by far the
least square matching package. For a window size of 21 x 21
pixels and for a mean number of 4 to 5 iterations, there are 50
000 double precision floating point multiplications. It has
been estimated that with a 20 Mflops processor it is possible
to compute a DEM or a complete SPÔT stereo pair in 6 to 7
hours. The software has therefore recently been ported to a
HP730 RISC processor.
It has been necessary to pay attention to the way the
information flows through the different loops. Some
examples are: limiting the number of conversions between
bytes (pixel intensity values) and floating point numbers;
suppressing unnecessary multiplications (during iterative
loops it is a lot faster to add a step than to multiply the index
of the loop by the given step); to optimize file transfers, image
files should be rewritten in blocks, each block being a small
window of the image with size matching that of the smallest
block transfer provided by the operating system. This facility
can improve the performance of the system by up to 60%.
Initial accuracy tests have been made by comparing the
computed elevations with a DEM that has been observed on
the Wild-Leitz BC2 analytical stereoplotter with the SATMAP
software. 55,000 points at 250m interval were manually
observed over the Sydney region on an image pair with a B/H
ratio of 1.0, with an accuracy, tested against maps at a scale
of 1:4,000, of 5 to 6m. The location of points in this DEM
did not match those derived by computation which was made
at an interval of approximately 120 m. A weighted average
elevation was derived at the coordinate positions of the
manually observed DEM from the nearest 3 computed DEM
points. This method of interpolation may lead to slight errors
in interpolation, but since most of the terrain elevations vary
gradually in the Sydney area, the errors are likely to be small.
À total of 9000 points in the computed DEM were used for
the comparison, corresponding to 1700 points in the manually
observed DEM. The RMS variation between the two DEM's
was then computed but large residuals greater than 27m (3 x
RMS) were discarded in this computation. The total number
of discarded residuals was about 5% of the total number
computed. The RMS derived was 9m with an average value
of 1 m. This accuracy estimate is.influenced by the accuracy
of the observed DEM and the effects of the interpolation. If
the manually observed DEM is assumed to have an accuracy
of 5m, the accuracy of the computed DEM is of the order of
7m. Further tests will be carried out with new scenes
currently under investigation. These results compare with
those published by Theodossiou and Dowman (1990) which
indicate accuracies varying from 6m to 25m depending on the
terrain slope, with considerable variation in the performance
for a given terrain slope.
The stereopair tested were recorded with a 6 weeks time
difference between the two images and some specularity
problems. Furthermore, the Sydney area is a difficuit
environment because the density of buildings and roads can
result in brightness values, while forest areas appear very
dark. As well, the images were taken in 1986 when the early
column noise of SPOT had not yet been corrected. Despite
these problems, the Moravec-Least Square approach has been
robust. But even if more than 95% of the points generated are
correct matches, some gross non-matching errors still occur.
Methods are currently being developed for the identification
973
of some parameters derived from the Moravec operator and
the least square adjustments that would allow the software to
automatically discard these errors. This will involve a
learning process to calculate the influences of the different
parameters and is a first step towards the use of artificial
intelligence techniques in the domain of automatic DEM
computation. The inclusion of the matched linear features
into the software will also be an important source of feature
information for determining matched points.
References
Butler, N. (1992). "Linear Based Matching of Stereo SPOT
Satellite Images". Paper submitted to ISPRS Congress,
Barnard, S.T. and Thompson, W.B. (1981). "Disparity
Analysis of Images. IEEE Trans. PAM I - 2, pp. 330 -
Gruen A.. & Baltsavias E.P. (1985). Adaptive Least Squares
Correlation with Geometric Constraints. SPIE, Vol. 595,
Rosenholm, D. (1987). "Multipoint Matching Using Least
Squares Technique for Evaluation of Three-Dimensional
Models". Photogram. Eng. & Rem. Sens. Vol. S3. DD.
Sloan, S.W., (1986). "A Fast Algorithm for Constructing
Delaunay Triangulation in the Plane", Report from the
Department of Civil Engineering and Surveying, The
Theodossiou, E.I. and LJ. Dowman 1990 'Heighting
Accuracy of SPOT' Photogram. Eng. & Rem. Sens. Vol.
56 pp 1643-1649.
Trinder, J.C., Donnelly, B.E. & Kwok, L.K. (1988).
"SPOT Mapping Software for Wild Aviolyt BC2
Analytical Plotter". International Archives of Photo. &
Rem. Sens., Vol. 27, Pt B4, pp. 412 - 421.
