IEEE Geoscience and Remote Sensing Magazine - December 2017 - 39

i-mages, with a size of n 1 # n 2. The values of all of the
pixels in one spectral band shape a gray-scale image
with two dimensions [as shown in Figure 1(a)], which
are both spatial.
Although the greater dimensionality of HSIs compared
with multispectral images improves data information content considerably, it does introduce new challenges to conventional image analysis techniques, which have been specifically designed for multispectral data. Furthermore, it is
almost impossible for humans to visualize spaces of higher
than three dimensions (e.g., red-green-blue images). A misunderstanding of high-dimensional spaces and conventional spaces sometimes leads to incorrect interpretations
of HSIs and the inappropriate choice of the data processing technique. Bearing this in mind, in the next section, we
provide an overview of a few common HSI challenges and
their possible solutions.
MAIN CHALLENGES OF HYPERSPECTRAL
IMAGE ANALYSIS AND POSSIBLE SOLUTIONS
Several factors make the analysis and processing of HSIs
a challenging task. Figure 2 illustrates the main paths in
HSI analysis that have been developed primarily to address
these factors. In this section, we take a closer look at each
of the applications shown in Figure 2. The common understanding of HSIs is that, because such data contain a rich
amount of spectral information, the whole dimensionality
needs to be used to define precise boundaries in the feature
space for a specific application. The increasing spectral resolution of HSIs benefits precision applications (e.g., Earth
observation, precision agriculture, and disease detection).
However, it challenges conventional signal--processing
techniques and, thus, hampers the abilities of HSIs in many
real applications.
Taking classification as an example (because classification is one of the most popular applications for HSIs), we
found in [1] that, when the number of training samples
remains constant, after a few features, classification accuracy actually decreases as the number of features increases. Two solutions have been widely exploited to address
this problem.
1) Dimension (feature) reduction: As mentioned in several
studies, such as [2]-[4], a high-dimensional space is almost empty, and multivariate data can be represented
in a lower-dimensional space, where the undesirable effects of high-dimensional geometric characteristics and
the curse of dimensionality are reduced. This fact has led
to a chain of research on dimension (feature) reduction,
which will be detailed in the "Dimensionality Reduction" section.
2) Robust classifiers: The imbalance between the number of
bands and available training samples has a dramatic influence on supervised classifiers. In this context, HSIs
often demand a vast number of training samples to effectively estimate class parameters. To benefit from the
rich spectral information of HSIs, one possible solution
DECEMBER 2017

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

is based on using effective and efficient classification
approaches that can handle high dimensionality, even
if a limited number of training samples is available.
In addition, along with the detailed spectral information provided by HSIs, it is possible to take advantage
of available spatial information (in particular, for veryhigh-spatial-resolution HSIs) to further improve the
eventual classification map. The "Classification" section elaborates on advances in HSI classification.
Spectral mixing (including both linear and nonlinear models) is another bottleneck for HSI analysis that occurs for a number of reasons, such as insufficient spatial resolution of the sensor and an intimate mixing effect. When mixing takes place, it
is not possible to directly distinguish the materials available in
the pixels from the corresponding measured spectral vectors.
However, detailed spectral information provided by HSIs can
be used to unmix hyperspectral
WHEN THE NUMBER OF
pixels. The "Spectral Unmixing"
TRAINING SAMPLES
section focuses on spectral unREMAINS CONSTANT, AFTER
mixing to address these issues.
A FEW FEATURES,
Spaceborne imaging specCLASSIFICATION ACCURACY
trometers are usually designed
ACTUALLY DECREASES AS
to acquire HSIs with a moderTHE NUMBER OF FEATURES
ate spatial resolution-e.g.,
INCREASES.
a ground sampling distance
(GSD) of 30 m-because of
the inevitable tradeoffs among
spatial resolution, spectral resolution, temporal resolution,
and signal-to-noise ratio (SNR). Spatial resolution enhancement of HSIs is a technology essential to expanding the range
of applications for spaceborne hyperspectral missions. In the
"Resolution Enhancement" section, we discuss techniques for
the resolution enhancement of HSIs.
The degradation mechanisms associated with the
measurement process and atmospheric effects inject undesirable noise that substantially downgrades the quality
of hyperspectral data. The HSI SNR is usually decreased
during the imaging process, depending on different noise
sources. In remote-sensing HSIs, highly corrupted bands
must often be removed before any further processing. Alternatively, HSI restoration can recover those corrupted
bands and also improve the HSI SNR, thereby improving
the effectiveness of any further processing of the HSI. In
this context, the "HSI Denoising and Image Restoration"
section is dedicated to HSI denoising and image restoration techniques that address such effects.
Another emerging research domain in the hyperspectral
community, CD is the process of identifying and examining
spectral-temporal changes in signals. The detailed spectral
sampling and representation in HSIs result in the potential
identification of more subtle spectral variations, which are
usually not easily detected in traditional multispectral images. Accordingly, land cover dynamic monitoring can be enhanced to a finer level. To this end, advanced CD techniques
must be designed to address CD issues in multitemporal
39

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2017