573
Figure 2. (a) Sampling from the CCD array in spatial
mode, (b) Sampling from the CCD array in
spectral mode. These are as is done by the FLI
and CASI instruments.
array of the FLI and CASI. The rectangular area represents
the 2-dimensional CCD array and the masks cover the
elements of the array included in the summations that produce
the bands. Conversely, the spectral mode data is recorded
with full spectral resolution and is reduced spatially. Figure 2
(b) shows the method used by the FLI and CASI. Robust
models of the expected spectral signal's shape can be used to
compensate for information lost during sampling, but the
spatial signal does not yield to such models readily.
A common method of spatial mode sampling, which has been
used in multispectral scanners since their inception, is to
average over discrete, strategically placed, spectral bands.
This is done on the premise that the spectra will be well-
behaved between the bands, and even more importantly,
within the bands, so that the shape of the entire spectra can be
correctly inferred.
To reconstruct the spectrum a model is required, which can be
as simple as linear interpolation or as complex as an inverted
Gaussian (Miller et al., 1990). The model can provide the
missing component of the signal that was not measured, yet is
well understood, to aid in the reconstruction. For example
absorption features can be avoided based their well known
spectral position and width. This is done because the
absorption feature contributes high frequency of, known
phase, whose content is undesirable. Conversely, the inverted
Gaussian model rejects the high frequencies of the signal to
precisely parameterize the phase and the dominant frequency
that characterizes the subtle changes in slope and position of
the vegetation reflectance red edge. The drawback of under
sampling comes when the assumptions of the model break
down, or when the expected spectrum is not known, so no
model is applicable. The aliasing that occurs as a result of
undersampling makes it impossible to distinguish between
high and low spectral frequencies.
Alternatively, the spectral range can broken into uniform
contiguous spectral bands that sample the signal optimally
according to linear system theory. No assumptions are made
about the well-behavedness of the signal, but no use is made
of the spectra's characteristic structure either. If only a small
number of bands are allowed and the spectral range is great
then the high frequency portion of the signal will be filtered
out and the resulting spectrum will be featureless and uninte
resting. However, in terms of sampling theory, this is the best
approach because the power in the aliased frequencies is
minimized and therefore the signal can be accurately recon
structed, albeit at low resolution.
Spectral mode only became possible with the advent of
Unsampled
Spectral
Spatial
Overlap
- max 0 max
Spatial Frequency
(cycles/element)
Figure 3. Representation of the spectral and spatial modes
in the Fouier domain. Note the Unsampled
region contains the high-spectral/high-spatial
frequencies.
imaging spectrometers, but the concept of spatial averaging to
change the resolution of an image is not new. The spatial
signal is not well behaved and does not yield to models that
would help with interpolation of an undersampled signal. The
only tenable model of the signal is given by linear systems
theory which requires that the sampling be done in accordance
with the Nyquist criterion.
Spectral mode undersampling is quite different than for spatial
mode because the digital levels of the pixels in a scene,
although strongly correlated, are essentially random events.
The spatial location of any arbitrary feature cannot be
predicted, since no model of how the signal will vary across
the scan-line is possible. When the signal is undersampled
then the only valid way that the signal can be analyzed is using
probability theory.
FOURIER DOMAIN REPRESENTATION
The two data sets can be overlaid in the Fourier domain to
elegantly demonstrate what information the data samples do,
and do not, contain. Consider the discrete 2-dimensional
Fourier transform of the 2-D signal measured by an imaging
spectrometer for a single image line. The name given to the
Fourier spectrum of the spectral signal will be the spectral
frequency spectrum and is analogous to the Fourier spatial
frequency spectrum. The use of the term spectrum in
reference to the Fourier spectrum must be distinguished from
the spectrum used to describe intensity of light within a range
of electromagnetic wavelengths.
The spatial mode data has high spatial resolution and low
spectral resolution and so represents full coverage of the
spatial frequency spectrum but only covers the lower range of
spectral frequencies (see figure 3, "Spatial"). The spectral
mode data is the converse with high spectral resolution and
low spatial resolution, so it covers the full spectral frequency
spectrum but only the lower range of spatial frequencies
(figure 3, "Spectral"). As is evident in Figure 3, overlaying
the two spatial/spectral frequency spectrum plots results in two
regions which are uniquely sampled by the individual modes,
a region that is sampled by both modes and a region that is not
sampled at all. The region which is labeled “Overlap” repre
sents the low spatial and low spectral resolution image that
could be formed either by averaging the spatial data in the
spectral direction or the spectral data in the spatial direction.
Thus the overlap region represents truly redundant data. The
area labeled “Unsampled” represents the portion of the signal
that contains high frequency both spatially and spectrally.