to the particular transformation to be used is made only after a detailed
examination of the input data and depends on the particular recognition
task to be performed.
The final functions to be performed by the preprocessor are
geometric and radiometric corrections for the non-ideal characteristics
of the sensor, sensor platform, and earth-platform relationship. An
ideal image of a portion of the surface of the earth should be distortion
free. The theory of map projections says this is impossible except
as an approximation for small areas. Even if the effect of all other
distortions could be removed the small geometrical errors due to map
projection errors or earth curvature would still remain. Line scan
imagery is taken over a time interval equal to the time it takes to
travel over the ground being imaged making the geometry of line scan
imagery complex due to platform motions as well as sensor distortions.
Achievement of geometric fidelity and precision requires information on
the navigational location and orientation of the platform. Compensation
for sensor distortions and aircraft motions (roll and yaw) is routinely
carried out but to date only incompletely and not entirely satisfactorily.
Radiometric corrections are important not only for sensor induced errors
but also such earth-platform related effects as changes of radiometric
intensity with viewing angle. A discussion of these distortions for
ERTS RBV and Skylab S-190 has been given by Steiner [32] which mentions
ERTS MSS also.
Potentially many problems may exist in each and every data set
which, if not corrected, could significantly reduce the accuracy of
recognition results for that data. Some of these potential problems
are instrument-related while others are associated with the radiation
environment and the scene being scanned. These problems include:
1) level shifts and gain changes in the recorded data resulting from
instabilities in system electronics and tape speed; 2) misregistration
of data between spectral bands due to unequal resolution in all the
bands, the lack of optical alignment (either by design or otherwise) of
the detectors in all bands, or the imperfect alignment of the tape
recorder record and playback heads; 3) noisy data resulting from a
combination of insufficient radiation input and lack of detector
sensitivity; 4) variations in signal levels as a function of scan angle
due to nonuniform angular sensitivity of the scanner, the effects of
atmospheric scattering, and bidirectional reflectance effects; and
5) changes in the scene illumination level during the data collection
mission.
An illustrative example by Nalepka [33] will indicate one approach.
Three segments in the 1971 Corn Blight Watch Experiment intensive study
area of Western Indiana over which multispectral scanner data were being
gathered on a biweekly basis were selected as potential sites. To
support this investigation, ground information, in addition to that
provided by the county agents, was gathered throughout much of the
scanner data collection period [34]. This information included the
location of fields, the crop planted therein, as well as the condition
of the crop.