IEEE Geoscience and Remote Sensing Magazine - June 2019 - 119

a specific application [13], [14]. One of the most important
preprocessing steps for HSI analysis is dimensionality reduction, which aims to remove spectral redundancy while
preserving critical information for subsequent applications
[15], [16]. Hyperspectral dimensionality reduction can be
achieved through either band selection or feature extraction [17]. Feature extraction transforms the original data
into another feature space with certain criteria [18], [19].
Most feature extraction methods linearly combine all of the
original bands and may complicate the interpretation of
results if the adopted criterion is inappropriate [20]. Band
selection selects a desired band subset from the original
bands and well preserves the spectral meaning of spectral
channels [21]. In this article, we focus on hyperspectral
band selection.
A variety of band-selection methods have been presented in the literature. However, to the best of our knowledge,
there is no study that reviews the current research status

of hyperspectral band selection. In this article, we survey
hyperspectral band selection methods in the literature and
point out challenges and potential future directions.
REVIEW OF BAND SELECTION METHODS
Band selection methods can be categorized into six groups:
1) ranking-based methods, 2) searching-based methods,
3) clustering-based methods, 4) sparsity-based methods,
5) embedding learning-based methods, and 6) hybrid
scheme-based methods. Let an HSI data X = " x i , ! R N # D be
represented as a collection of band vectors B = " b j , ! R D # N,
where x i refers to the spectral signature of the ith pixel, D
is the number of pixels, N is the number of bands with
N % D, and b j corresponds to the jth band. In the supervised methods, a training set with n training samples is
given as data set X tr = "^ x 1, y 1h, f,^ x n, y n h,, where x and
y are a training sample and its class label in "1-of-C" encoding of the C classes, respectively. The problem with
band selection is selecting m representative bands
M = B (:, l) ! R D # m to achieve better or at least as
comparable a performance as using the original
bands in a certain application, e.g., classification,
target detection, and spectral unmixing. Table 1
shows the taxonomy of the hyperspectral band selection methods discussed in this article.
RANKING-BASED METHODS
The ranking-based methods quantify the importance of each spectral band according to a predefined band-prioritization criterion and select
the top-ranked bands in a sorted sequence. The
ranking-based methods can be further divided
into two types, i.e., unsupervised and supervised,
according to whether labeled training samples are
used or not.

©ISTOCKPHOTO.COM/MARCCOPHOTO

UNSUPERVISED RANKING-BASED METHODS
Unsupervised criteria consider the information, dissimilarity, or correlation of bands. For instance, metrics such as variance, first spectral derivative, spectral
ratio, contrast measurement, signal-to-noise ratio
(SNR), third-order statistics (skewness), fourth-order
statistics (kurtosis), kth order statistics, negentropy,
entropy, and information divergence [22], [23] are
often employed to prioritize bands.

JUNE 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

HIGH-INFORMATION CRITERIA
The selected bands are required to have large information capacity. In [14], some classical information measurements, such as information entropy
and information divergence, were used as ranking criteria to select informative bands. In [23],
principal component analysis (PCA) was adopted
to prioritize the energy or variances of band images, and two eigen analysis-based methods were
presented, i.e., minimum-variance PCA (MVPCA)
119

http://www.ISTOCKPHOTO.COM/MARCCOPHOTO

IEEE Geoscience and Remote Sensing Magazine - June 2019

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