IEEE Geoscience and Remote Sensing Magazine - June 2019 - 133

bands increases. MVPCA takes the least amount of time;
MMCA takes the second-least amount of time. ISSC is faster than FDPC, LP, and WaluDI, although it is slower than
MVPCA and MMCA. The WaluDI and LP take longer than
the four other methods, and the time consumed by the
WaluDI is the longest.
SUMMARY AND DISCUSSION
Numerous narrow bands of hyperspectral data provide
the potential for discerning the subtle spectral difference of ground objects but bring about high computational costs and information redundancy. Therefore, in
recent last decades, a variety of band selection methods
has been proposed for selecting proper bands and reducing the dimensionality of HSI data. In this article, we
have symmetrically reviewed state-of-the-art band selection methods grouped into six categories, i.e., ranking
based, searching based, clustering based, sparsity based,
embedding-learning based, and hybrid-scheme based.
From our research, we found that the recent development of band selection methods is unbalanced. The
achievements of ranking-based, searching-based, and
clustering-based methods are relatively more remarkable, particularly the former two. Sparsity-based methods have achieved more attention due to the popularity
of sparsity theory.
Furthermore, we compared the classification performance of using bands chosen by several popular band selection approaches on two widely used hyperspectral data
sets. Although a band selection method might behave
divergently on different HSI data depending on the level
of scene complexity and classification difficulty, using selected bands can yield better accuracy than using all the
original bands.
The ranking-based methods have low computational
complexity and are suitable for larger hyperspectral data

sets, but the classification accuracies from their selected
bands are unsatisfactory. The searching-based, clusteringbased, and deep-learning-based methods usually require
higher computational costs and, accordingly, are more suitable for smaller hyperspectral data sets. Sparse representation-based band selection methods, e.g., ISSC [113], FRSR
[115], and SSR [117], may bring about high classification accuracy while taking lower computational costs in practice.
Specifically, the ISSC and SSR methods have clear physical
meanings for selected bands. Additionally, when dealing
with noisy hyperspectral bands, the MSNRPCA [23], OSP
[23], and minimum-noise band selection [84] methods can
be considered for real-world applications.
Band selection is an important technique for hyperspectral dimensionality reduction and has room for further improvement. Based on our current research status, we recommend the future directions of hyperspectral band selection
as follows.
DEVELOPING SEMISUPERVISED METHODS
Hyperspectral data typically come with a small number
of labeled samples, with a majority of samples unlabeled.
Unlabeled samples may assist with the mathematical
modeling of band selection; for instance, they may help
accurately quantify spectral redundancy. However, most
existing supervised band selection methods do not utilize
the information from unlabeled samples; therefore, more
semisupervised methods should be developed to promote
band selection performance.
DEVELOPING ACTIVE LEARNING-BASED
AND DEEP-LEARNING-BASED METHODS
Current applications of active learning in hyperspectral
data alleviate the act of labeling for classification. Active
learning can absorb the most informative and representative training samples in low-density areas and ensure

TABLE 5. THE CLASSIFICATION ACCURACIES (%) ACHIEVED USING 30 SELECTED BANDS
FROM THE PAVIAU DATA SET.
MVPCA

MMCA

WaluDI

FDPC

ISSC

ALL BANDS

Asphalt

86.54

90.45

90.06

93.4

91.62

Meadows

96.22

97.74

97.8

97.74

96.93

98.28

98.16

Gravel

72.53

70.99

71.2

79.62

74.17

79.3

77.27

Trees

83.79

91.81

93.33

93.76

91.52

92.97

89.75

Painted metal sheets

99.09

98.6

99.17

99.59

97.77

98.26

98.95

Bare soil

52.19

85.95

84.49

89.48

84.38

89.66

90.14

Bitumen

79.78

82.37

86.8

85.63

77.78

86.3

85.38

Self-blocking bricks

83.43

90.86

91.85

88.05

84.82

89.86

90.2

Shadows

99.65

100

99.53

99.88

99.06

100

99.89

ACA

83.69

89.81

90.51

91.86

88.50

91.26

OCA

86.04

92.42

92.70

93.76

91.32

94.14

93.56

81.11

89.91

90.27

91.70

88.45

92.20

91.42

NOTE: ACA: average classification accuracy; KC: Kappa coefficient.

JUNE 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

133

IEEE Geoscience and Remote Sensing Magazine - June 2019

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