554 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 14, NO. 4, APRIL 2017
Particle Swarm Optimization-Based Band Selection
for Hyperspectral Target Detection
Yan Xu, Student Member, IEEE, Qian Du, Senior Member, IEEE,
and Nicolas H. Younan, Senior Member, IEEE
Abstract— This letter proposes particle swarm optimization
(PSO)-based band selection (BS) approach for hyperspectral
target detection. Due to lack of training samples in a detection
problem, it is more difficult than classification-purposed BS. The
objective function, called maximum-submaximum-ratio (MSR)
gauging target-background separation, is proposed for target
detection during PSO searching. Typical target detectors such
as target-constrained interference-minimized filter and adaptive
coherence estimator are studied. Experimental results demon-
strate that the proposed MSR-based objective function in con-
junction with PSO-based searching can select a small band set
while yielding similar or even better detection performance than
using all the original bands, sequential forward search-based BS,
or BS relying on detection map similarity assessment.
Index Terms— Band selection (BS), hyperspectral imagery,
particle swarm optimization (PSO), target detection.
I. I NTRODUCTION
T
ARGET detection is one of major tasks of hyperspectral
imaging. A hyperspectral image cube contains hundreds
of spectral bands, leading to a high computation burden on
target detection. Thus, it is necessary to do dimensional reduc-
tion on the original data for efficient analysis. Typically, there
are two kinds of methods for dimensional reduction. The first
one is transform-based methods (such as principal component
analysis [1]), and such methods may not be preferred since
they alter the physical meaning of the data. The second one
is band selection (BS), which is to select a subset of bands
while still generating satisfactory results using the selected
bands [2]–[4].
BS algorithms have been applied to hyperspectral data
analysis, and they can be implemented in either unsupervised
or supervised. Unsupervised BS is to select the subset of
bands without any prior information. For instance, a simple
yet efficient similarity measure method was developed in [5]
for BS. Supervised BS algorithms intend to select the bands
producing maximum class separability with prior knowledge
of training samples. Many methods (see [5], [6]) have been
developed to measure the class separability for BS. However,
most BS algorithms are focused on classification problems,
and few have been proposed for target detection. It is more
difficult to separate targets and nontargets with selected bands
Manuscript received September 27, 2016; revised December 31, 2016;
accepted January 20, 2017. Date of publication February 22, 2017; date of
current version March 3, 2017.
The authors are with the Department of Electrical and Computer Engi-
neering, Mississippi State University, MS 39762 USA (e-mail: yx131@
msstate.edu; du@ece.msstate.edu; younan@ece.msstate.edu).
Color versions of one or more of the figures in this letter are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LGRS.2017.2658666
due to lack of training samples for target and background
modeling.
The search strategy is also an important issue. To avoid
exhaustive search, which is computationally prohibitive to
hyperspectral BS, sequential forward search (SFS) and sequen-
tial floating forward (SFFS) methods can be used [7]. Their
basic idea is to select the best band for maximizing an
objective function, then one additional band combining with
the existing selected band or bands is selected to maxi-
mize the objective function. The process continues until the
desired number of bands is reached. Recently, particle swarm
optimization (PSO), invented by Eberhard and Kennedy [10],
has been applied to hyperspectral BS with the objective of
maintaining classification accuracy [8], [9]. The PSO imitates
the social behavior of flocks. Each particle is a solution, and
the positions of particles are randomly initialized. Then the
particles fly over the problem space until certain criteria are
reached. Compared with SFS and SFFS, PSO offers two major
advantages: it can provide better high-dimensional solution
due to global search, and it can be easily implemented in
parallel. Although several other evolutionary algorithms are
developed recently, such as ant colony optimization [11] and
firefly algorithm [12], which have been applied to hyperspec-
tral BS, we limit our discussion with PSO in this letter.
Here, we focus on the task of target detection where targets
with known spectral signatures are to be detected from an
unknown background. Targets, as small man-made objects, are
often sparsely populated [13]. No training or labeled samples
are available for both target and background [14]. The key
of PSO-based selection is to design an effective objective
function during searching. More often, labeled samples are
available when evaluating an objective function. However,
it is impossible for target detection. Thus, a detection-specific
objective function is required. Intuitively, we can compare the
detection outputs for band searching with a certain criterion,
such as Euclidean distance or correlation coefficient (CC).
However, a large similarity does not necessarily mean satisfac-
tory performance, because the number of target pixels is much
smaller than background pixels and similarity assessment is
dominated by background. Thus, we prefer the metric that
can well gauge target and background separation. Specifi-
cally, we propose the maximum-submaximum-ratio (MSR)
to quantify such separation. The typical target detectors, i.e.,
target-constrained interference-minimized filter (TCIMF) and
adaptive coherence estimator (ACE), are studied. Experimental
results using the proposed MSR-based objective function in
conjunction with PSO-based searching can select a small band
set while yielding similar or even better detection performance
than using all the original bands. The parameter setting of
MSR is discussed based on the performance of detectors.
1545-598X © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.