571
The Ramifications of Sampling Methods
for Imaging Spectrometers
J. Douglas Dunlop 1 & John R. Miller 2
1. Earth Observations Laboratory, Institute for Space and Terrestrial Science
Department of Geography, University of Waterloo
Waterloo, Ontario, Canada, N2L 3G1
2. Department of Physics, York University
Earth Observations Laboratory, Institute for Space and Terrestrial Science
4700 Keele Street, North York, Ontario, Canada, M3J 1P3
ABSTRACT
Imaging spectrometers are capable of generating data at
prodigious rates which can exceed the bandwidth of either the
recording device or the transmission channel. This is a direct
result of the spectrometer's potential for collecting high
resolution both spectrally and spatially, simultaneously. Some
form of sampling is commonly employed to reduce the volume
of data, but this can have profound effects on the information
content of the resulting signal. This paper will discuss the
ramifications of some of the methods of sampling typically
used on spectrometer data, vis-à-vis the loss of information.
It will also consider new methods of sampling which have the
potential of preserving more of the generic information content
of the data, independent of any specific application.
Key Words: Fourier, imaging, spectrometers, sampling.
INTRODUCTION
The decade of the environment is upon us. To identify pollu
tants and to monitor their effects it will become increasingly
important to use sophisticated remote sensors. With its ability
to measure the spectrum in contiguous bands for each pixel,
the imaging spectrometer promises to become the preeminent
environmental sensor of the 1990's. Before the end of this
decade spacebome imaging spectrometers will become opera
tional and the number of spectral bands that are available to the
satellite data user will increase by over an order of magnitude
from what is currently available. The wealth of data produced
by an imaging spectrometer provides both an opportunity and
a challenge. The opportunity is to more accurately monitor
terrestrial processes by remote sensing than has previously
been possible using multispectral scanners. The challenge is to
effectively cope with the prodigious volume of data so that we
might tap its wealth of information.
For example the NASA imaging spectrometer HIRIS due for
launch in 1996 is capable of generating 500 Mbits/s of data.
Currently it is not technologically practical to continuously
digitize, transmit and record at 500 Mbit/s data rates.
Colvocoresses (1977) and Schowengerdt (1980) both
suggested that perhaps a mixture of critically placed low
spatial resolution narrow spectral bands and a single high
spatial resolution band may be the optimal way to reduce the
data rates with minimal information loss. The CASI
instrument implements this strategy in what is called the track
recovery channel (Babey & Anger, 1989). More commonly,
imaging spectrometers (both CASI and FLI) utilize modes of
operation which operate with either high spectral resolution or
high spatial resolution, but not both simultaneously (Buxton,
1988).
Each of the above represents a different sampling strategy, but
other strategies are possible and may turn out to be more
effective. What artifacts are introduced by the sampling
process? This paper attempts to shed light on the ramifications
of the various sampling procedures so as to provide insight
that may allow improved analysis. Finally, an alternate
method of sampling is suggested that might be more appro
priate for imaging spectrometers due to the strongly correlated
nature of the signals they produce.
BACKGROUND
The list of known and planned spacebome imaging spectrom
eters include HIRIS, HRIS, MODIS, and MERIS. Already in
use as airborne systems are AVIRIS, ROSIS, AIS, FLI and
CASI with the earliest flights dating back to 1983. Imaging
spectrometers have high resolution in all domains; high spatial
resolution, high spectral resolution and high radiometric
resolution. The problem of how to effectively filter and reduce
the volume of data and yet to retain the desired information has
been considered in the context of multispectral imagery for
many years.
Image Sharpening
A number of researchers have considered the problem of how
to use information from a high spatial resolution band to
sharpen the edges in other lower resolution bands. It was first
suggested by Colvocoresses (1977) and first implemented
when Schowengerdt (1980) presented his technique for
restoring the high spatial frequency information to Landsat
MSS imagery.
Roller and Cox (1980) studied the possibility of improving
classification accuracies and the interpretability of Landsat
Multispectral Scanner (MSS) imagery by increasing its spatial
resolution using Return Beam Vidicon (RBV) imagery.
Hallada and Cox (1983) compared the methods of
Schowengerdt with those of Roller and Cox using data from
an airborne Daedalus 1260 multispectral line scanner. Hallada
and Cox note that the combination of mixed spatial and
spectral resolutions is very adaptable to the concept of imaging
spectrometers. They state "On-board data compression would
be a simple matter of integrating the signals across detectors in
the spatial and/or spectral domains".
Simard (1982) was first to propose that the SPOT 20 m multi
spectral data could be sharpened by using the 10 m
panchromatic band to modulate the brightness. Cliche et al.
(1985) describe algorithms for integrating the SPOT panchro
matic band with the multispectral bands to improve image
sharpness to yield results resembling colour infrared
photography.
Tom and Carlotto presented a Least Mean Squares (LMS)
approach to edge sharpening (Tom et al., 1984). It relies on
the fact "that at sufficiently small resolution multispectral
bands are strongly correlated", so if the digital levels of the
any two bands are plotted against each other in a scatterplot
then all the data will fall approximately along a straight line
Any particular spectral band will generally not be globally
correlated with the reference band (or bands), so the regres
sions are performed adaptively within a sliding window which
moves across the image.
