aida nd ie A dd
Michael Breuer
advantageous regarding effective data storage whereas regular grids are preferred if grid interpolation is needed. This is
especially the case in the following context. The spatial resolution of the grid should not be higher than ten times the
spatial resolution of the image data. The DEM must be based on the same reference system in which the geometric
correction is performed. If this is not the case an appropriate transformation has to be applied before other processing
steps. The DEM is needed to correct the relief displacement (see 2.9).
5 THE CONCEPT OF A GENERAL APPROACH
A general concept for geometric correction must be oriented towards a holistic philosophy that takes into account all
available auxiliary information. The aim is to get an optimal solution even if the initial information is sketchy.
Generally a geometric correction algorithm is divided into three main processing steps. That is first the preparation,
second the establishment of a three dimensional coordinate frame based on the reference system and third the re-
sampling of the hyperspectral image data. Sometimes purists prefer to do without the resampling because they want to
keep the radiometric information unchanged. But in this case the results that are gathered from the hyperspectral image
data have to be transformed instead. Therefore a third processing step is always needed.
5.1 Preparation
The preparation step is necessary to prepare all initial data in order to make it compatible with the geometric correction
software. It sometimes appears that the initial data are affected by blunders. In general every initial data set can be
involved here. An effect that occurs quite often is the occurrence of registration gaps during the image data capturing.
This arises especially when attitude and position data are stored together with the image data. In such a case some lines
may appear as "black" lines without any information. To fill the gaps one practical method is to take the image
information from neighboring pixels. Another problem arises if flight position and attitude data are disturbed. In this
case the data have to be filtered to separate the noise from the signal. Sometimes the position and attitude values have to
be derived from related measurements that are stored as housekeeping data together with the image information.
5.2 Georeferencing
During the second step the coordinates for each pixel of the hyperspectral data set are determined. To do this a three-
dimensional coordinate frame is established and filled during the subsequent processing. This can be done in different
ways. The appropriate algorithm depends on the availability and the accuracy of the available auxiliary information (see
4). These yields different accuracy levels that will be described later on (see Fig. 3). The best case would be if the
position and the attitude of the flight path were directly measured with high precision during the mission. Then “direct
georeferencing" can be performed. The worst case is if there are neither attitude data, nor positions data and even no
DEM. To get a solution that may be better than nothing it would be the only way to apply a non-parametric
interpolation method (res. "rubber sheeting"). Combinations between different proposed solutions seem to be useful.
Some proposed solutions are discussed in section 6.
5.3 Resampling
During the resampling step the image data are transformed to the regular grid of the reference frame. A desired spatial
resolution and a method for interpolation have to be selected. Some well known interpolation methods are "nearest
neighbor" (the pixel value of the nearest original pixel is taken), “bilinear interpolation” (the pixel value is derived from
the adjacent original pixels using a first order interpolation), and "cubic convolution" (the pixel value is derived from
the adjacent original pixels using a third order interpolation). Other methods are possible such as "interpolation using
weighted means" (were the interpolation is done due to the distances of the adjacent pixels). Because hyperspectral
image data are designated for classification purposes based on the inherent hyperspectral information it is sometimes
not desired as mentioned above to change the original pixel values during the resampling. In this case the classification
is run on the original hyperspectral data and the classification results are resampled afterwards instead of the
hyperspectral data.
5.4 Accuracy Levels
The aim is to define accuracy levels in which the geometric correction results can be classified. Such a classification is
useful for everybody who wants to estimate the expected correction accuracy that may be reachable in an individual
case. However, it is impossible to postulate a scheme that is universally valid because of the variety of conceivable
initial situations. But the presented scheme may play the role of a guideline or scale.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 97
