ing process.
Typically, the model parameters which are refined during
model optimization are the coefficients of polynomials which
correct for systematic errors in the nominal trajectory of
the imaging platform. Alternatively, one can optimize over
Fourier series parameters. Adding other parameterizations
would require adding new classes to GEOREG. However, by
virtue of the object oriented design, this requires only a few
lines of code. For instance, if a platform was known a prior?
to change its pitch over time in a saw-tooth fashion, but the
amplitude, frequency, phase and ratio between ascending and
descending segments of the saw-tooth were unknown, then a
class could be added to GEOREG which used the a prior:
information on the shape of the pitch movement and allowed
the unknown parameters like amplitude etc. to be estimated
using GCPs and TIPs. To add this capability to GEOREG
would require less than 10 lines of C++ code.
Moreover, GEOREG is not restricted to refining platform pa-
rameters but can be configured to simultaneously optimize
any model parameter including sensor parameters such as
camera focal length. Furthermore, when tie points are used
GEOREG allows for the simultaneous nonlinear optimiza-
tion of several acquisition models (or bundles) belonging to
different images.
After the acquisition model is constructed, a remapping pro-
cess is initiated which resamples the raw image to a user
selectable map projection using a user selectable resampling
kernel. If a DEM of the region is available, the imagery may
be orthorectified to remove any terrain-induced distortions
in the raw imagery.
These examples clearly illustrate that an object oriented de-
sign is very well suited to providing the flexible structure
necessary in a geocoding facility which has to accommodate
very diverse a priori and a posteriori knowledge.
5 GCP AND TIP ACQUISITION
The model parameter optimization mechanism relies on the
availability of accurate GCPs and TIPs. Currently, there are
two mechanism to generate a posterioriinformation: Manual
marking and automatic correlation with a reference image.
Manual marking of ground control points is accomplished
by marking common features in both the raw image and an
accurate map. Alternatively, the map can be replaced by an
already geocoded reference image. On the other hand, tie
points are marked in multiple raw images.
Usually, only a few GCPs («14) can be economically hand
marked. If more GCPs are required and an orthorectified
reference image is available, automatic correlation is used.
For this purpose, the raw image is resampled using only the
nominal information or information derived from just a few
manually marked GCPs. There are GEOREG utility pro-
grams which allow the user to define a grid of control points
in the raw image or in the reference image and then correlate
the nominally resampled image with the reference image gen-
erating GCPs near the grid points. For example, using this
method, an orthorectified SPOT scene of the Drum Moun-
tains region in Utah was successfully used as a base map
to generate control points for an AVIRIS image of the same
region.
132
6 AIRBORNE AND SATELLITE
DATA GEOCODING
There is a marked difference between the geocoding of op-
tical airborne imagery and satellite imagery. It is possible
to accurately geocode satellite images using only a small
number of GCPs (<10) [3]. In general, this is not true
for airborne imagery which require a much larger number
of GCPs to achieve subpixel accuracy. The reason is that
aircraft are significantly less stable imaging platforms than
satellites and therefore causing geometric distortions in the
image with large spatial high frequency content. Thus, air-
borne data correction becomes a test case for the capabilities
of any geocoding facility.
Airborne images acquired over featureless areas such as some
deserts and snow fields are particularly hard to geocode be-
cause of the difficulty of obtaining a sufficient number of
ground control points. Further, if the airborne image and
the reference image are taken in widely different ranges of
the electromagnetic spectrum, it is exceedingly difficult to
achieve reliable correlation and generate sufficient numbers
of GCPs.
For these types of geocoding problems it is paramount to use
all available information and be able to configure the geocod-
ing software to suit the particular problem at hand. An
object oriented approach provides the flexibility and power
necessary if subpixel accuracy is desired.
A key problem for this type of geocoding is the need to
eliminate erroneous control points which are derived from
false correlation matches. GEOREG can often eliminate this
problem by rejecting GCPs which do not, by a certain mar-
gin, fit the best available image acquisition model. This can
be achieved by ordering the GCPs time-sequentially and let-
ting them drive an extended Kalman filter which for each
instant in the imaging time produces a best estimate of the
platform position and attitude given our estimates of the air-
craft dynamics constraints, the measurement and state un-
certainty covariance etc. If a GCP is inconsistent with this
estimate, it is not allowed to change the state estimate. The
sequence of Kalman filter state estimates then defines the
desired platform trajectory. Even though this is not a global
optimization over all GCPs, it nevertheless turns out to be
a very efficient method for these hard geocoding problems.
As an example, consider the attitude time state for a GERAIS
(Geophysical Environmental Research Corp. Airborne Imag-
ing Scanner) image which is shown in Fig. 2. In this case,
no systematic attitude information was available for the im-
agery. The attitude time state shown was generated using
a previously orthorectified SPOT image as a base map, and
GCPs which were automatically correlated, and input to the
extended Kalman filter. The resulting derived attitude al-
lowed the GERAIS image to be geocoded with «20 meters
RMS error, versus the ca. 300 meters RMS error when no
attitude was accounted for.
7 RESULTS
'To assess the system, an extensive multisensor data set com-
prising of satellite and airborne sensor data over the Drum
Mountains area of Utah was orthorectified and co-registered.
The data
Ol
0.05
-0.05 -
Attitude [rad]
-0.1+
-0.15H-
x
0
Figure 2
solely frc
imaging
SPOT E
(7 ban
AVIRIS
scene (2
bands ea
(24 banc
In gener
known.
tion on
summar;
when trz
forms, it
To begir
area wel
niques ©
were nec
few GC]
image W
The ort]
erence 1
was the
Data
SPO
Land
AVI]
TIM
GEC
GER
Table
ing at 2
there wa
informat
of RCPs
The oth
quency :
GEOSC.
TIMS in
