IEEE Geoscience and Remote Sensing Magazine - June 2019 - 106

The photos in Figure 5 demonstrate the efficiency of HSI
in rediscovering a suspicious vehicle. A matched subspace
detector [22], [23] was used to determine whether a pixel in
an urban HSI contains the spectra of a suspicious target collected from a previous scene. The matching score is based
on signal energy ratios, which makes the detection invariant to sensor gain and scaling and improves detection in
semishadow areas.
A good understanding of background and object spectral signatures as well as their dynamic behavior in realistic
environments is essential for interpreting hyperspectral imagery. However, spectral measurements from a material of
known composition are insufficient for successfully detecting a small target in a heterogeneous scene [13]. Moreover,
in partial- or deep-shadow areas, the true detection rate will
most likely be low due to the low illumination and detection
of other materials with similar mixed spectra. To overcome
this limitation, complementary data sets are combined to
provide additional insight about a particular scene. Fusing a
hyperspectral data set with other data obtained at different
wavelengths or different times has been shown to improve
detection and classification performance [15], [24].
REAL-TIME PROCESSING SYSTEM
Within defense, the real-time processing of information is
necessary to achieve a high level of situational awareness.
Although there is no doubt that the availability of detailed
spectral information is essential for constructing robust
image descriptors, it also presents a significant challenge
when these high-dimensional data are used to develop
real-time machine-vision systems. When complementary
information must also be included in the identification
process, substantial computational costs are incurred from
disparaging the real-time processing option. To ease computational demands, dimension-reduction techniques are
often applied prior to feature extraction. Although this approach has, to some extent, reduced the size of the spectral

descriptor, data reduction alone proved insufficient in attaining real-time classification [24].
Over the past two decades, extensive R&D has been
invested in hardware and software for big data processing and analysis. These revolutionary computing schemes
do not avert HSI; however, they inspired the development
of several real-time military systems (RTMS), including
War Horse [25], airborne real-time-cueing hyperspectralenhanced reconnaissance [26], the Norwegian Defence
Research Establishment [27], and Hymex [28]. These realtime processing platforms were implemented using various
computing technologies, including vector processing, fieldprogrammable gate arrays, adaptive (i.e., adaptive control
of thought), and artificial intelligence.
DISCRIMINATION BETWEEN TARGETS AND DECOYS
Homing missiles have played an important role in warfare
since the end of World War II [28]. These missiles guide
themselves to intercept targets-mostly aircrafts-that can
maneuver unpredictably or operate a decoy. Military aircraft
often employ pyrotechnic decoys that act as countermeasures for luring away incoming antiaircraft missiles. Most
infrared (IR) decoys are flares, and their purpose is to match
the spectral radiation pattern of the target and the spectral
sensitivity of a homing missile seeker. This cartridge burns
after being fired away (Figure 6) and makes a significant heat
track that is larger than that of a defended object. Most of the
flares in use are made from a solid pyrotechnic composition
of magnesium, polytetrafluoroethylene (Teflon), and viton
(MTV) or similar synthetic rubber binders. These are commonly called MTV flares, and they omnidirectionally radiate
in the 2-5.5-µm wavelengths (Figure 6) [30].
Homing guidance is an onboard sensor (i.e., seeker) that
provides the target data on which guidance decisions are
based [29]. Conventional IR seekers are designed to home
in on the IR signatures of a target aircraft's exhaust in 3to 5-µm wavelengths [32]. A modern IR seeker consists of

1
0.8
0.6
0.4
0.2
0
(a)

(b)

(c)

FIGURE 5. The rediscovery results using forward modeling of a measured reflectance spectrum and constrained subspace matching for a red
car. (a) The targeted vehicle in time T and (b) in time T + 1. (c) The corresponding score map for the target model (0-no detection, 1-detection).
The HSI data were collected during the DUCAS campaign and published by [24].

106

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

JUNE 2019

IEEE Geoscience and Remote Sensing Magazine - June 2019

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