that
d to
wing
(5)
tion
pled
with
'elief
ross-
cient
The
jrey-
nage
line
G4,
Sr
tar-
coef-
(6)
ol, a
era
'ela-
nd-
ilat-
'acy
igh-
In order to reduce the time needed to match corre-
sponding points, a multiresolution or pyramid hierar-
chical approach is applied, whereby a sequence of im-
ages of the same object is presented at successively re-
duced resolutions (Li, 1991). First, the image at level k
is filtered by a low-pass filter, thereby removing high-
frequency noise. This filtered image is afterwards re-
sampled with a reduced sample density to obtain the
image at level k+1 (Wong and Hall, 1978). In each step
the images are reduced by a factor 2 (fig.4). Once the
pyramid is created, matching is performed on the
Target Image Search Image
Target array Search space
119 |122 [113 | 1149 |125
121-1124. |115. |113 | 125
122 ]125
122
120 |116 |118 |123 |121
1417 \121 |120 |125 |433
M search array MS
Fig.3 : Detailed view of the Matching Process
coarsest level (highest k-value) according to a regular
grid superimposed on the target image. The positions
of the matched points then serve as an estimate for
the positions on the next level (k-1).
The matching algorithm is further enriched by intro-
ducing a priori knowledge (heuristics) : (i) at each
level, the parallax displacement cannot exceed a prede-
fined value, according to the maximum terrain height
and the angle of incidence of the image. (ii) In view of
the epipolar constraint there exists a maximum dis-
parity in the y-direction (which increases with both
viewing angle and latitude). (iii) If the position of a
new matching on the image plane k-1 lies to far from
the estimated value of level k, matching is rejected.
(iv) Since improper matches can occur on snow sur-
faces (saturation) and shadows (dependent on the ele-
vation and angle of incidence of the sun), each of
these texture classes are identified on the basis of the
mean and variance of the greyvalues in the target
array. According to their matching accuracy they will
be rejected or not.
Once matching is performed at the highest level (k-1),
parallactic displacements are converted into heights
through spatial resection. Corresponding points are
transformed to the final cartographic reference system
and an accurate position is calculated by finding the
midpoint of the parallax of the projecting rays (fig.2).
475
k=1
Fig. 4 : Multiresolution Image Library
4. EXPERIMENTAL RESULTS
4.1. Data Acquisition
A set of two panchromatic stereoimages covering the
central part of the Ser Rondane Mountains (72°S,
25°E), Dronning Maud Land, Antarctica were obtained
(table 1). This mountain range, situated 200 km south
of the Princess Ragnhild Coast, belongs to a series of
mountains surrounding the East Antarctic continent.
Eight geodetic points were used as reference points for
the geometric correction model. From a topographic
map at scale 1:50,000 of the same region 28 control
points were collected.
view
Scene centre angle
Date G.R.S.
20.02.91] 151/690] 71954'S / 25*57E | 22.9E
24.01.91| 152/690| 71°50'S / 25°10'E | 18.1W
Table 1: Ser Rondane scene specifications
Both images are of good radiometric quality, since no
clouds occur and a B/H ratio of 0.75 only favourites
the matching accuracy. A higher B/H ratio for polar
images is not advisable, because this would hamper
the epipolar registration. Due to the late acquisition
date and the steep relief of the mountain range
(heights between 1000 and 2500 m.a.s.l.) shadows be-
come indeed very large. Since both images are scanned
a month apart the areal extent of shadowed areas also
differs. A higher detector saturation was noticed on
the east-looking scene.
4.2. Model Accuracy
In order to test the accuracy of the geometric correction
model, the stereoimages were geocoded using the ref-
erence points. Parallaxes were calculated for the con-
trol points, from which three dimensional coordinates
were obtained and afterwards compared with the orig-
inal coordinates. The results of the accuracy test is
shown in table 2.
