IEEE Geoscience and Remote Sensing Magazine - June 2019 - 9

SOURCES OF TARGET SPECTRAL VARIABILITY
I don't like sand. It's coarse and rough and irritating and
it gets everywhere.
-Anakin Skywalker
The most obvious and immediate cause of variability in a
target spectrum is the intrinsic variability in the target material itself. Unless the material is a pure substance, there is
likely some variability in its composition; for instance, we
can refer to "sand" as a single material, but there are many
kinds of sand, and any given kind of sand is composed of
varying ratios of mineral components [2]. "Pure" materials
can undergo chemical variation (e.g., due to oxidation or
hydration), and even a chemically pure material with fixed
optical properties can exhibit spectral reflectances that vary
with material morphology. In addition to these intrinsic
variability sources, the effective reflectance at a pixel can
depend on extrinsic properties of the material, such as concentration, thickness, or size relative to the pixel. Environmental effects, although not affecting the target reflectance
itself, can have a large impact on observed at-sensor spectral radiance; these effects include atmosphere, illumination, adjacent materials, and issues within the sensor itself.
The ultimate consequence of all these sources of variability is that the observed target spectrum in the field often differs from spectral reflectance or emissivity measured
in the laboratory and archived in spectral databases, such
as the U.S. Geological Survey Spectral Library [3]. The first
step in bridging this gap is to identify and catalog these individual sources, while at the same time recognizing that
JUNE 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

their effects occur together and superimpose on each other.
If the sources are understood, then models can be formulated, and from these models algorithms can be designed
to achieve some measure of robustness to this variability.
INTRINSIC TARGET VARIABILITY
The most direct source of spectral variability in a target material is due to the chemistry or morphology inherent to the
material. When in powdered form, "pure" solid materials
become particularly complicated because their spectra are
highly variable [4] depending on particle size (e.g., see Figure 1), particle shape, and packing density; this can be further complicated in the case of fine-grained (or "intimate")
mixtures of multiple materials. Under ideal conditions,
however, the fundamental reflectance properties of a pure
solid material are directly characterized by its optical constants (refraction index n and extinction coefficient k) [5],
[6]; from those values, one can derive modeled reflectance
signatures based on morphology (e.g., for particle size [7],
[8] or packing density [9], [10]). As sensor technology has
improved in recent years, there has been an increasing need
for such reflectance models in order to generate simulated
spectra, because obtaining comprehensive spectral measurements under the multitude of possible morphologies
is intractable [11], [12]. For solid materials, both pure and
mixed, this reflectance also depends on the angle from
which the material is illuminated as well as the angle from
which it is observed. Those dependencies are encapsulated in the bidirectional reflectance distribution function
(BRDF) [13]-[15], which is a function of four variables that
define how light reflects off of an opaque surface [16]:
r BRDF =

L (i O, z O) -1
[sr ].
E (i i, z i)

(1)

0.8

0.6
Reflectance

(VIS), near infrared (NIR), and short-wave infrared (SWIR) or
by the spectral emittance in the long-wave infrared (LWIR)].
Nonetheless, the notion of a single predetermined
spectral signature for each material is an ideal concept
not observed in real-world applications. Even in precisely
controlled laboratory experiments, the measured spectral
reflectance of the same material exhibits some variability.
Matters get even worse in hyperspectral remote sensing
where target materials are imaged by a sensor mounted on
a platform, because many environmental and geometric
aspects are involved, and they unavoidably affect the measured spectral radiance for the given material. All of these
sources of variability will alter the spectral intensity and
shape of a measured spectrum and will affect the performance of hyperspectral image exploitation, particularly in
the case of target detection.
The purpose of this survey is to examine the major
causes of spectral variability in target materials, survey
some associated physical models, and discuss how hyperspectral analyses can be made more robust to this variability. Although there are many kinds of hyperspectral-image
exploitation algorithms (e.g., classification, clustering, and
unmixing) [1] and virtually all of them are affected by spectral variability, we will concentrate on the problem of target
detection and, more specifically, the detection, identification, and location of particular materials of interest.

0.4

0.2

0.5

1
1.5
Wavelength (µm)
<5 µm
74-250 µm

2.5

<74 µm
250-1,200 µm

FIGURE 1. Four spectra of the mineral actinolite, based on four
ranges of particle sizes, with the smaller particles generally
more reflective. (Data taken from the U.S. Geological Survey
database [3].)

IEEE Geoscience and Remote Sensing Magazine - June 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2019