Image chips were used in an early study for automatic
satellite image rectification using ground control points
(Malmström 1986). The chips and their object space
coordinates were generated using scanned aerial photo-
graphy, and were subsequently stored in a database. Area
based matching was used to locate the chips in Landsat
MSS imagery. To the knowledge of the author no follow
up studies have been published. Possibly this is due to the
sensitivity of area based matching to radiometric changes
of the grey values. In multiscale, multitemporal, and mul-
tisensor matching the corresponding assumptions are
easily violated.
The approach of Drewniok and Rohr (1995; 1996) was
developed for urban large scale imagery. The authors use
manhole covers as control points. Three-dimensional
coordinates are available in a sewage cadastre. These
manhole covers have a diameter of approximately 6 pixels
and are assumed to exhibit a rotationally symmetric
brightness pattern described by three parameters. These
parameters are determined in a so-called learning phase,
and the result is a template for the manhole covers.
Detection of the manhole covers is then performed via
least squares estimation between the templates and the
image grey values. The used images show a large number
of manhole covers, thus the aspect of redundancy can be
exploited to advantage. In order to find the corre-
spondence between the candidates for manhole covers in
the image and the actual manhole covers in object space,
a common scale factor is assumed, and a relational des-
cription based on relative distances is constructed. Sub-
sequently, relational matching is performed, followed by
a spatial resection for computing the orientation parame-
ters. A critical point in this well designed approach is the
question how well the assumed radiometric model des-
cribes the actual appearance of the manhole covers in the
images.
Relational matching was also used by Vosselman, Haala
(1992). Large scale colour images were registered to a
scanned map. The images were pre-processed using
standard image processing routines (classification follo-
wed by binarization, edge detection and line following),
the maps were digitized manually. Subsequently, a rela-
tional description of the image and map content was set
up. The correct match was found using tree search me-
thods. The method was tested with roads, rivers, and land
parcels and yielded promising results. The third dimen-
sion was not considered in the approach.
308
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Dowman et al. (1995; 1996, see also Newton et al. 1994)
use a similar approach. Their study on map registration
of aerial and satellite imagery is based on polygons ex-
tracted from the map and the image by means of edge
detection and segmentation. The matching is carried out
by dynamic programming using a description of the po-
lygons based on shape, orientation, and area. Examples
are given for forest areas, and large buildings. Also this
method only works in two dimensions.
Christmas et al. (1995) use probabilistic relaxation for
matching road networks extracted from maps and im-
ages, respectively. Both road networks are represented
as attributed relational graphs. The nodes of these graphs
are line segments. Only unary and binary relations are
considered. Beside other findings the authors successful-
ly show the rotation invariance of their sophisticated
approach. Again, the third dimension is not dealt with.
In an interesting approach Holm et al. (1995) use lakes
and small islands for a completely autonomous orienta-
tion for satellite imagery taken over Finland. This work
has its roots in a prior publication by Holm (1991). Due
to the specific Finish landscape an abundance of control
data is available, and the third dimension plays a marginal
role only. Water bodies of the images to be orientated are
mainly extracted by thresholding the histogram of the
near-infrared band, and are described in terms of area,
perimeter, compactness etc. After using the satellite orbit
data for a coarse geo-coding feature based matching is
carried out using the descriptions of the water bodies.
This step is followed by a robust estimation of the para-
meters of a two-dimensional affine transformation. Suc-
cessful tests are reported using Landsat TM, SPOT XS
and NOAA AVHRR images. The system has been instal-
led at the National Land Survey of Finland and the
Environmental Data Center of Finland.
Schickler (1992; 19952) has developed a module for ab-
solute orientation for the production of orthophoto maps
at the Landesvermessungsamt Nordrhein-Westfalen.
The images are approximately of scale 1:12.000. Three-
dimensional wire frame models of houses are used as
control information. The work builds upon prior studies
by Forstner (1988) and Sester, Forstner (1989). Assu-
ming good approximate values for the orientation para-
meters (+/- 50 m for the centre of projection and +/- 1
degree for the angles) the wire frame models are projec-
ted into the image, and a search space is defined. In this
area str
ching b
cluster
ted in 2
the ma
diagno:
images
inform:
ever, w
mensio
increas
as wire
and ob
main €
therefo
In sun
shown
single
and rel
role he
detecta
ching
values
ces. F
have
Redun
object:
signals
be ens
and th
In thi
orient:
shown
have |
tation
Modu
matic
day (E
cia et
in di
1995a
can a
perfor
dules
