ul 2004
used to
eltisa
nd has
X show
caused
terrain.
> image
| scale,
jection
latively
ition Is
0), Tao
. “The
mating
| based
4 lle
point
ven by
(9)
>ctions;
een the
o X, y
sctor of
ojected
V is the
may be
res of a
ssigned
| when
ifferent
n. The
ing the
solving
ig to be
S, eight
) above
‘et al.
m does
to the
ity will
(Fraser
TY
led for
ditions
maging
broom,
es have
cquired
spheric
image
ifferent
ermore,
require
'eration
magery
those
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
different image geometry models that are expected to have
widespread use by interoperable software. OGC (1999b) has
adopted a specification for standardization of image geometry
models. In photogrammetry, the block adjustment and 3-D
mapping often are performed using images acquired by a same
sensor and platform. But when the images are acquired by
different sensors, the block adjustment among different image
geometry models is hard to be implemented due to a
combinatorial overflow. Moreover, while many more imaging
sensors have been launched or will be launched in near future, it
is obviously not convenient for end users and service providers
to constantly upgrade their software to process new sensor data.
As a matter of fact, the software upgrades often fall behind the
availability of the data. This is also expensive and in particular
not necessary for many mapping applications requiring accuracy
at sub-metre level or lower.
Because of the characteristic of sensor independence, the use of
RFM would be a driving force towards the photogrammetric
interoperability among imagery exploitation software. If each
overlapping image comes with a set of RPCs, end users and
developers will be able to perform the subsequent
photogrammetric processing neither knowing the original
sophisticated physical sensor model nor taking account of the
submodels associated with the sensors used to acquire the
images. This is highly beneficial as it makes the
photogrammetric processing interoperable, thus allowing users
and service providers to easily integrate cross sensor/platform
images from multiple data vendors. The different image
resolution and the error estimates associated with the RPCs for
cach image should be processed by appropriate weighting
during the adjustment. For example, the covariance matrix C,, in
Eq. 7 will use different sub-covariance matrixes of misclosures
for the image points measured on different images participating
in the adjustment. Thus many of the difficulties that may arise
from simultaneously adjusting different physical sensor models
can be avoided. This technique is of unique value for users who
require high updating rate and for other applications in which
high temporal accuracy is of essence.
6. PHOTOGRAMMETRIC APPLICATIONS
Many COTS photogrammetric suites have implemented the
RFM and related techniques, including ERDAS IMAGINE (LH
Systems), PCI Geomatica (PCI), SOCET SET (BAE Systems),
ImageStation (Z/1 Imaging), and SilverEye (GeoTango). Using
these systems, traditional photogrammetric processing tasks can
be performed in a unified technical framework. Many mapping
applications using above photogrammetric systems or
proprietary packages have been reported. We will briefly focus
on the photogrammetric applications below.
Kay et al. (2003) evaluated the geometric quality of ortho-
rectifying QuickBird and Ikonos images, for a typical
agriculture area, using GCPs and a DTM derived from the
1:50000 scale map data. Two QuickBird images with Basic and
Standard levels and an Ikonos Geo image, covering an area of
108 km? are rectified. Both results are well with 1:10000 scale
accuracy requirements of the EU Common Agriculture Policy.
Fraser et al. (2002) investigated the application of Ikonos
imagery to 3-D positioning and building extraction. The results
of 2-D and 3-D metric accuracy tests shows a planimetric
accuracy of 0.3-0.6 m and height accuracy of 0.5-0.9 m. Tao et
al. (2004) evaluated the 3-D feature extraction results using two
667
Ikonos Reference stereo scenes at a nuclear plant. The relative
planimetric and vertical accuracies for 3-D features are at the
sub-meter level, and the RFM refinements do not change the
relative accuracy.
Tao and Hu (2004) reported 3-D feature extraction results from
overlapped QuickBird and Ikonos image pairs. The conjugate
points in the QuickBird and the Ikonos images were manually
positioned and were assigned different weighting factors of |
for the Ikonos image and 1/0.6? for the QuickBird image in Eq.
9. When the RPC models are bias compensated using three
GCPs, the object points have the position differences of 1.36-m
RMSE horizontally and 0.84-m RMSE vertically, and the
dimension differences are better than 1-m RMSE.
7. DICUSSION AND OUTLOOK
Extensive tests have been carried using different formulations of
the RFM. These experimental results have revealed that the
third-order RFM is not always the best form in terms of
obtaining highest approximating accuracy (Tao and Hu, 2001a,
2001b; Fraser et al., 2002). Yang (2000) also reported functions
lower than third order were used and the correct order can be
chosen, based on the RMS error analysis, testing aerial
photography and SPOT data. Hanley and Fraser (2001) tested
Ikonos Geo product by first projecting the control points onto
‘planes of control’, to minimize the effect of terrain, and then
transform the image to these points using similarity, affine and
projective transformations. The results show that 0.3-0.5 m
positioning accuracy is achievable from the Geo product
without using the rational function solution. Fraser et al. (2002)
and Fraser and Yamakawa (2003) have extended this work in
two dimensions into three, using similar techniques. They found
that the affine projection, the DLT and relief corrected affine
transformation also can approximate the Ikonos imaging
geometry to sub-meter positioning accuracy in the absence of
high-order error sources. If the most significant coefficients
could be found for each particular imaging sensor heuristically
(e.g., by trial-and-error), then the RFM may be solved with
higher stability in the terrain-dependent approach using a small
number of GCPs, and may be also suitable for replacing
rigorous sensor models as what has been done by terrain-
independent approach.
Fraser and Hanley (2004) found the systematic residual errors in
the along track direction due to perturbations in scan velocity.
The question is then should high order polynomial be used to
compensate for this high-order drift error when the errors are
not well modeled in the physical sensor model.
Furthermore, currently, Digital Globe also provides images each
with multiple sections. Each section is stored in a separate
image file using the same set of RPCs with different line and
sample offsets. Yet, if each image section has a different set of
RPCs as defined in the USM, all the related photogrammetric
processing methods have to be re-formulated.
The characteristics of cross sensor imagery exploitation will
instigate a crossover of images from multiple data vendors into
a new 3-D mapping paradigm. From the viewpoint of imagery
exploitation services providers, the RFM technology enables
extensive interoperability between images from different
sources, regardless of the sensor types and the platforms, due to
its geometric generality. However, new problems arise when we
try to generate DSMs automatically using heterogeneous images
