Michael Breuer
In case 1 there are only the hyperspectral image data available but no other auxiliary information. It is obvious that in
such a case geometric correction of the hyperspectral data is impossible. Even for the application of a “rubber-sheeting”
method (see 6.1) ground control points are needed.
In case 2 ground control information can be derived in form of plane coordinates based on the map’s reference system.
If no other information are supplied the non-parametric method (res. “rubber sheeting” or polynomial approach) is the
only possible way to get a corrected image. However, the reachable accuracy depends very much on the shape of terrain
because the relief displacement (see 2.9) is not corrected. Besides this method is a local approach which means that the
correction accuracy may vary significantly within the corrected image. This case is better than nothing but mostly does
not meet the desired results.
In case 3 there is a reference orthoimage
| | s 3
instead of a map. This case is very 28| 5 S o9 recommended
* 7. . c " - ©
similar to case 2 because the orthoimage 20|8|58 SE method for expected
- : ) Lo oz -zfrf n 2% geometric accuracy
provides plane ground control too. But it à a9 cogo z|o 2 £ correction
. t . . © > »- I = Lu © o
contains also image information that o rE5|touo|a|z co
may be used for image-to-image mat- correction
ching (see 6.2). The second makes impossible
possible to reconstruct the flight path
x
1
1
1
1
!
non-parametric 2.30 Pixel
À xi [i xol im is X|-
using a parametric approach. However, method RR
the interior orientation parameters are eli tien S so uic, Prrparametric or iso Pel
unknown and have to be estimated. This param d
influences the correction accuracy A KANN LAK = Petre | - 20 Pixel
directly. m uu ECL
S alix d xad ul a ud ee « 1 - 10 Pixel
« 3 ; method 7
In case 4 image data, the interior and orm
exterior orientation, a DEM and a map e TX[X|HA a < 1 Pixel
are available. The exterior orientation els a SE
parameters are provided with low accu- (LA = low accuracy, HA = high accuracy)
racy. Sometimes position and attitude is
measured by the GPS/INS of the air-
craft's autopilot system only. Normally these systems are neither geometrically connected to the hyperspectral sensor
nor are they able to provide high accurate measurements. But nevertheless the data of such systems are kept as
housekeeping data together with the image data. In this case position and attitude cannot be used for direct
georeferencing but they can serve as starting values that have to be improved by the subsequent algorithm (see 6.3). If
the initial accuracy of the other initial data is sufficient and if the conditions of the site and the mission were good an
accuracy up to one pixel can be reached.
Figure 3. Combinations of initial data and accuracy levels
Case 5 is similar to case 4. Instead of a map there is a reference orthoimage provided here. As mentioned above the
orthoimage offers the opportunity for automated image-to-image matching (see 6.2). If the spatial resolution of the
reference orthoimage is the same or higher than the spatial resolution of the hyperspectral data and if there is sufficient
texture in both image data sets then an accuracy better than one pixel seems to be reachable.
Case 6 can be denominated as the optimal case. Here the position and attitude data are provided with high precision that
means a sophisticated combined GPS/INS measurement unit was mounted directly on the sensor and used during the
mission. With this technique position accuracy in the range of ten centimeters and velocity determination at the level of
a few cm/s is possible. The attitude parameters can be derived with an absolute accuracy in the range of 0,01? (Cramer,
1999, Hutton et al, 1998). Assuming these optimal conditions direct georeferencing can be applied and subpixel
accuracy can be reached.
6 PROPOSED SOLUTIONS
There a several approaches that were proposed during the last years to solve the problem of geometric correction of
hyperspectral image data. Regarding these solutions it becomes apparent that each method starts from a specific initial
situation where the kind of initial data is well defined. The proposed solutions are normally classified to be non-
parametric, parametric or mixed approaches.
6.1 Non-parametric approaches
Non-parametric approaches use polynomial functions or triangulation based methods to correct the disturbances of the
image data but do not take into account any sensor position or attitude data. These methods require an adequate number
98 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.