IEEE Geoscience and Remote Sensing Magazine - June 2019 - 91

making them harder to distinguish from one another. For
these reasons, accurately estimating moisture content,
particle size, or other properties of the underlying materials
requires very accurate emissivity retrieval. As discussed in
the "Emissivity Retrieval" section, uncertainty in the atmospheric parameters and the difficulty associated with TES
make the parameter estimation procedure very challenging for remote sensing applications. In the presence of significant sensor noise, offsets, or shape errors, the retrieved
emissivities may vary significantly from their true values,
further complicating the parameter estimation problem
[70]. These factors and others have made SMA a popular
technique for mapping pure, spectrally distinct minerals in
LWIR hyperspectral data [2], [71].
The combined effects of variations in a material's physical properties, sensor noise, and offsets, as well as uncertainty in atmospheric parameters, lead to the inherent variation observed in LWIR hyperspectral data. Modeling these
natural variations is a key step in understanding the effectiveness of detection, classification, or other processing algorithms. One approach to modeling natural backgrounds
is to treat the scene as a mixture of elliptical distributions.
Each distinct background class is modeled by a single elliptical distribution, but the whole background is a complex
combination of the individual distributions. This approach
has been applied to reflectance target detection in the
VNIR/SWIR and is being investigated for emissivity-based
detection [53], [72]. Compared to reflectance modeling, the
LWIR regime presents a unique difficulty for temperature
modeling. Material temperature, a highly scene-dependent
quantity that is not readily modeled, is a critical component
for algorithm performance estimation.
SCENE CLASSIFICATION
Classification tasks in the LWIR have been focused on mineral and vegetation mapping. Mineral mapping is a popular application for several reasons including the distinct
spectral features that many minerals have, the availability
of ground measurements, and the large areas that minerals cover [4], [58]. Vegetation mapping, identification, and
monitoring is another area of interest, although vegetation
typically lacks the large spectral features seen in many mineral species [73]. The lack of large spectral features makes
remote differentiation of plants extremely difficult in the
LWIR. However, it makes vegetation a good candidate for
blackbody pixels in ISAC algorithms and means that temperature estimates can be very accurate for vegetation.
Because of the relative simplicity of SMA, these techniques are the most popular ways that scenes are broken
into different material classes, especially in mineral mapping applications. Examples include endmember analysis
followed by a spectral angle mapper [74], mapping alluvial
deposits [75], and estimating mineral particle size [54]. A
summary of geologic mapping using HSI is provided in
[76]. Despite the popularity of SMA techniques for mineral
mapping, for general scene classification problems, several
JUNE 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

machine-learning techniques have been applied to both
VNIR/SWIR and, to a lesser extent, LWIR hyperspectral
data. A review of classification techniques can be found
in [77]. Table 3 presents several classification techniques
and associated references where the technique was applied
to HSI.
In contrast to algorithms relying on library spectra, the
training and testing of machine-learning algorithms are
completed using in-scene spectra. An example of the training and testing of classifiers applied to HSI is discussed in
[84]. The training and testing phases, along with some difficulties involved, can be summarized as follows.
1) Training: Training a classifier requires a large number
of labeled pixels for each material. Thus, these types
of methods are not appropriate for target-detection
applications where there are only a few samples of the
target. When classifying background materials, each
training pixel or spectrum must be labeled. Creating
labeled imagery is very time intensive and subject to
human error.
2) Testing: After a classifier has been trained, evaluation and
comparison of classifiers are very difficult. To evaluate
a classifier, a testing data set is constructed from the labeled data. The classifier is run on the test data set, and
the results are compared to the true labels for that data.
In general, it is unclear how the resulting algorithm will
generalize to new scenes, areas, times of year, weather, and
sensors [85].
Other practical problems include the large amounts
of data for modest collection areas, the large number of
spectral bands causing techniques to fail [82], and a lack
of data sets with ground truthing [77]. The first two problems are usually solved by compressing the data in some
way, but this may remove valuable information needed to
separate different materials. The third problem is not easily
addressed because of the difficulty in creating high-quality ground truth, which usually involves setting up wellcharacterized reference articles, taking a variety of atmospheric measurements, and measuring specific materials
on the ground. The recent availability of LWIR data taken
by the HyTES instrument should expose a larger audience

TABLE 3. THE CLASSIFICATION TECHNIQUES APPLIED
TO HYPERSPECTRAL IMAGERY.
TECHNIQUE

SPECTRAL REGION

REFERENCES

Matched filter

LWIR

[2], [71]

Rule-based classification

LWIR

[78]

Linear discriminant analysis

LWIR

[79]

Quadratic discriminants

LWIR

[73]

Morphological techniques

VNIR/SWIR

[80]

Support vector machine

VNIR/SWIR

[81], [82]

Neural networks

VNIR/SWIR

[83]

IEEE Geoscience and Remote Sensing Magazine - June 2019

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