radiometric preprocessing steps which are
described in more detail in the following
sections.
Sensor Performance
The PMI sensor consists of five cameras, each with
an area detector array consisting of 385 pixels by
288 colour elements (bands). Each of the cameras
provide eight profiles for which each profile
element is an average over 10 spatial pixels with
a 38-pixel gap between profiles. Due to problems
with the first camera during the overflight, the
data from this camera could not be incorporated
for further processing. Furthermore, each third
profile of the remaining four cameras was lost
with respect to the transcription of the data from
the high density digital tape to CCTs. The
remaining 28 profiles, consisting of 773 lines,
were examined for bad data in the spectral and
spatial domains. Figure 2 shows some examples of
spectra with bad data. Bad bands were located in
profiles 9 (bands 113-137), 14 (279), 16 (281-
288), and 21 (268) and replaced with the average
of the adjacent bands. A correction, however, was
not applied for profiles 9 and 16 which had 5 or
more contiguous bad bands. These particular bands
were not included for data interpretation
purposes. The next step was to replace bad data
lines within a specific camera by the average of
the adjacent lines. In addition, the first 20
bands of each of the remaining profiles were
eliminated from further processing due to bad
data, leaving 268 bands for subseguent analysis.
An estimate of the random noise which is image
independent resulted in a signal-to-noise ratio
(SNR) for band 247 (751.21 nm) of 140:1 for a low
signal level (DN=500; 100.4 x 10" 8 W cm" 2 nm" 1 sr" 1 )
and of 1380:1 for a high signal level (DN=15000;
3013.1 x 10“ 8 W cm" 2 nm" 1 sr" 1 ). Similar SNRs were
obtained for the other bands. A "laboratory"
method was used for SNR calculations as described
in Borstad et al. (1985). Preliminary investiga
tions indicate that fixed pattern noise could not
be detected within the image, especially for bands
with low signal amplitude as reported in the
aforementioned reference. A further study,
however, will be conducted in the future using a
fast Fourier transform for each band to reveal
potential noise pitch and 'notch filtering' in the
frequency domain to remove major noise components
if necessary (Curran and Dungan, 1988; Hlavka,
1986).
The central wavelength (band) locations were
checked using the position of the atmospheric
absorption features calculated with the 5S
radiative transfer code (Tanre et al., 1986).
Major and minor absorption features, for example
at 590 nm (0 : ,, H 2 0), 690 (H^O, O*), 720 (H*0), and
760 nm (HzO, Oa), could be detected in terms of
the wavelength position at 589.53 nm, 687.32,
718.61, and 760.34 nm for the different PMI
cameras. The disagreement in the location of the
absorption features is due to the minimum 5 nm
wavelength interval allowed in 5S compared to a
1.3 nm sampling interval for the PMI sensor.
Calibration
An integrating sphere is used as a standard
calibration source for the PMI sensor, resulting
in a set of multipliers used to convert DN values
into radiances in terms of W cm" 2 nm" 1 sr“ 1 . A
total of 288 radiometric calibration values are
provided (one for each band) along with the image
data. Besides the removal of the dark current
from the PMI data, a uniformity (radiometric)
correction, assuming detector linearity, was also
applied for the alignment of the different cameras
(Borstad et al., 1985).
The compensation for the different sensitivities
of the cameras was not fully achieved with the
uniformity technique because the detector
responses are, in general, non-linear. Further
corrections were necessary using a polynomial
regression approach for a total radiometric
adjustment of the cameras. This method is based
on using targets which cover two or more cameras.
Such a so-called standard target should be homoge
neous. At least 10 such homogeneous, standard
targets encompassing the DN range of the scene to
be corrected were selected in order to generate
the polynomial equation for each band (Figure 3).
Thus, a separate correction was applied to each of
the 16-bit DN levels for a compensation of the
non-linear behaviour of the DN differences between
cameras in roost of the bands. A second degree
polynomial was used to remove the DN differences
between the 'nadir' camera 3 and cameras 2, 4, and
5 for each of the bands. Satisfactory results
could be achieved with this method as shown in
Figure 4 for a sugarbeet field, located within
cameras 4 and 5, before and after radiometric
alignment. The effectiveness of this adjustment
technique depends on the ability to select enough
standard targets for the generation of the
polynomial curve as described in detail by Staenz
(1990).
This camera adjustment technique also addresses
some of the viewing angle effects which are quite
severe for a scan angle range of approximately
72.5 degrees. This effect is demonstrated in
Figure 5 using corn data acquired with a
Spectrascan spectrometer on the ground over a
viewing angle range of ±24.3 degrees. The varia
tion caused by the viewing angle in one specific
camera (14.5 degrees FOV), however, could not be
removed with the correction of the radiometric
misalignment of the cameras. An assessment of the
eight profiles within a specific camera show
various degrees of data variation (especially in
camera 4), caused not only by the viewing angle
phenomenon, but also by sensor related effects
(CCD responsivity). An algorithm will be imple
mented in the ISDA software package in the future
to deal with these within-camera effects.
The removal of scene-related effects, such as
those due to the atmosphere or viewing geometry,
are important preprocessing steps for a
normalization of the image data in an absolute
sense. Such steps are necessary for a spectro
scopic analysis of the data, especially to
incorporate the generated spectra into a database
for comparison with laboratory spectra, field
spectra and spectra derived from other imaging
spectrometers. An absolute normalization
procedure consisting of an airborne version of the
5S radiative transfer code was applied to the PMI
data set (Tanre et al., 1986; Teillet, 1989).
This procedure begins with a conversion of the raw
data into radiances using the radiometric calibra
tion values delivered with the data set. A
comparison with ground-based spectra of similar
targets, however, indicate that the resulting
reflectances are, on average, approximately a
factor of two too high. The performance of this
procedure is satisfactory for Airborne Visible/
Infrared Imaging Spectrometer (AVIRIS) data as
demonstrated in Staenz and Goodenough (1989).
Based on the results usually achieved with the
atmospheric model and extensive investigations
