IEEE Geoscience and Remote Sensing Magazine - December 2017 - 66

data set into subsets for efficient cluster-based processing.
In spectral-domain partitioning, a single pixel vector (spectral signature) may be stored in different processing units,
and communications would be required for individual
pixel-based calculations, such as spectral angle computations [311]. In spatial-domain partitioning, every pixel vector (spectral signature) is stored in the same processing unit
[312]. These concepts have
been explored not only in
clusters [313], [314] but also
in heterogeneous workstation
RECONFIGURABILITY IS
networks [315].
NO LONGER A PROMISE
BUT A REALITY, AND IT IS

HARDWARE
ACCELERATORS
SEVERAL ALGORITHMS
In addition to clusters, other
IMPLEMENTED ON
kinds of high-performance
THE SAME BOARD.
computing architectures have
been used. Specifically, onboard processing of HSIs is
required in certain contexts.
This is particularly the case in time-critical applications
[316], [317]. For this purpose, other solutions have been
explored. In particular, reconfigurable computing provides
higher performance (throughput and processing power)
compared to other specialized hardware processors [318].
Reconfigurability is no longer a promise but a reality, and
it is feasible to have several algorithms implemented on the
same board and dynamically select one out of a pool of algorithms from a control station [319]. A field-programmable gate array (FPGA) is a chip in which there is a matrix of
blank cells called configurable logic blocks. This device can be
used to implement any circuit (provided there is a sufficient
number of logic blocks). FPGAs have been widely used to
accelerate hyperspectral imaging algorithms for onboard
processing [320]-[322].
Another high-performance computing architecture that
has provided excellent performance when accelerating hyperspectral imaging computations is the graphical processing unit (GPU) [323]. GPUs are now fully programmable
using high-level languages such as NVIDIA CUDA (http://
developer.nvidia.com). The GPU specializes in computerintensive, massive-data-parallel computation (which exactly describes graphics rendering). Therefore, more transistors can be devoted to data processing rather than to data
caching and flow control. The fast-growing video game
industry exerts strong economic pressure for constant innovation. This has motivated the extended use of GPUs for
accelerating many different hyperspectral imaging-related
tasks [316], [324]-[335].
FEASIBLE TO HAVE

CLOUD COMPUTING
Recently, more sophisticated high-performance architectures have been used to process hyperspectral data. For
instance, cloud computing platforms are increasingly employed for processing hyperspectral data in distributed
66

architectures. In this sense, cloud computing offers advanced capabilities for service-oriented and high-performance computing [336]. Furthermore, using cloud computing to analyze large repositories of hyperspectral data
can be considered a natural solution resulting from the
evolution of techniques previously developed for other
types of computing platforms [311]. In particular, using
GPUs within distributed scenarios has been radically extended worldwide, thanks in part to the increasing development of deep learning-based frameworks (e.g., Apache
Spark, Caffe, Theano, Torch, and TensorFlow), which also
find application in the hyperspectral analysis community
[64], [65], [84], [337], [338].
However, the recent literature still provides few examples of the use of cloud infrastructures to implement hyperspectral analysis techniques in general and to perform
supervised classification of hyperspectral data in particular. This may be due to the lack of open repositories of
HSIs available for public use, a situation that is expected to
change in the near future, as large distributed repositories
of hyperspectral data for open use become available to the
scientific community.
CHALLENGES IN FAST COMPUTING
The most important challenge related to consolidating fastcomputing techniques to analyze hyperspectral data (particularly in the context of real-time platforms) is still the
high energy consumption required by high-performance
computing architectures, which reduces their applicability in real scenarios for onboard operation. Currently, the
power consumption required by devices such as GPUs is too
high for their incorporation into satellite platforms. Another
shortcoming is the fact that these platforms are often subject to radiation tolerance issues. Future developments in
hardware instruments for onboard operation are required
for efficient real-time processing of HSIs, particularly in the
context of satellite missions.
CONCLUSIONS
The role of HSI analysis cannot be underestimated for a
plethora of applications, especially those related to CD and
scene classification. Without a doubt, the use of such valuable data has been well established in the remote-sensing
community, and the precise investigation of such data is increasing significantly. To this end, the considerable number
of airborne and spaceborne hyperspectral missions as well
as the increasing number of scientific publications on this
particular subject demonstrate that the area of HSI analysis
is substantial, dynamic, and vibrant.
The field of hyperspectral imagery is extremely broad,
and it is impossible to investigate it comprehensively in
one literature review. This article focuses particularly on
algorithmic approaches that have been developed, adapted, or proposed since 2013, covering a number of key research areas, such as DR, classification, spectral unmixing,
resolution enhancement, image restoration, CD, and fast
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2017

http://http:// http://developer.nvidia.com

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