IEEE Geoscience and Remote Sensing Magazine - June 2019 - 130

schemes to select the best band combination for classification. The original bands were first grouped into several clusters based on spectral CCs and then a representative band
from each subspace was selected by ranking the entropy
of the bands in the same subspace. Finally, the best band
combination was searched by maximizing SSI separability.
A band selection framework in [137] was proposed to integrate both clustering- and ranking-based techniques, where
clustering was performed on the band attributes using a
density-based spatial clustering of applications with noise
to remove redundant bands, and representative bands from
all clusters were ranked according to their discriminating capability. In [138], the PSO was used to group all bands into
different clusters, and the impurity-function band-prioritization algorithm was adopted to select the most important
bands. In [139], the manifold ranking algorithm combined
clustering and ranking schemes to select dissimilar bands
here, the k-means algorithm was employed to cluster all
bands, and representative bands of each cluster were chosen
by clone selection and treated as queries. Meanwhile, other
bands were ranked according to the data manifold structure of HSI data, and the dissimilar bands were expected to
be selected by the query set.
Moreover, the clustering scheme can be combined with
a searching scheme. A hybrid band selection strategy [140]
integrated clustering and searching schemes to search for
the best combination of bands with higher classification
accuracy. The band grouping was implemented on a conditional MI between adjacent bands, and then the GA algorithm was adopted to search for the optimal combination
in disjointed groups by maximizing classification accuracy. Additionally, the branch-and-bound search algorithm
prunes irrelevant bands from the aforementioned result,

TABLE 2. CLASSES AND SAMPLES IN THE INDIAN PINES
DATA SET.
LABEL

NAME

SAMPLES

Alfalfa

No-till corn

1,428

Minimal-till corn

830

Corn

237

Grass/pasture

483

Grass/trees

730

Mowed grass/pasture

Windowed hay

478

Oats

No-till soybeans

972

Minimal-till soybeans

2,455

Clean soybeans

593

Wheat

205

Woods

1,265

Building/grass/tree drives

386

Stone/steel towers

Total

130

10,249

and a minimal set of relevant bands was finally obtained.
The clustering scheme was combined with an evolutionary
searching scheme [141], where the spatial clustering generated overall signature variation and determined cluster averages that represented different signature patterns in the
image. The GA algorithm was then applied on the cluster
signatures to extract the bands with the most discriminating information.
The ranking and searching schemes can also be combined. For instance, in [142], the ranking and eliminating
searching schemes were combined to select the optimal
bands for foreign matter classification, where all bands
were first ranked according to minimum redundancy and
maximum relevance. Sequential backward elimination was
then applied to remove bands that contributed the least
to classification.
EXPERIMENTAL RESULTS
This section details some experimental results of band selection for hyperspectral classification using two data sets:
Indian Pines and the University of Pavia, Italy.
The Indian Pines data set was obtained from the Multispectral Image Data Analysis System Group at Purdue
University, West Lafayette, Indiana (https://engineering
.purdue.edu/~biehl/MultiSpec/aviris_documentation
.html). The data set was acquired by NASA on 12 June
1992, using the airborne visible/infrared imaging spectrometer sensor from JPL. It has 20-m spatial resolutions
and 10-nm spectral resolutions covering a spectrum range
of 200-2,400 nm. A subset of the image scene, sized 145
× 145 pixels and covering an area of 6 mi west of West
Lafayette, is used in our experiment. The data set was
preprocessed with radiometric corrections and bad band
removal; 200 bands remained. Figure 1(a) and (b) show
the three-band false-color images and the corresponding
reference samples. The 16 classes of ground objects in the
image scene are detailed in Table 2.
The Pavia University (PaviaU) data set was taken from
the Computational Intelligence Group at the University of
the Basque Country, Spain (http://www.ehu.es/ccwintco/
index.php/Hyperspectral_Remote_Sensing_Scenes). It was
obtained from a reflective optics spectrographic imaging
system (ROSIS)-3 sensor by the Deutschen Zentrum fur Luftund Raumfahrt, the German Aerospace Agency. The ROSIS-3 sensor has 115 data channels and 1.3-m spatial resolutions and covers a spectral range from 0.43 to 0.86 μm. After
removing low SNR bands, 103 bands remained and were
employed in our experiments. The data set contains 640 ×
340 pixels and covers the University of Pavia area. Figure 2(a)
and (b) illustrate the three-band false-color images of PaviaU
data and its corresponding reference data. The image scene
has nine classes of ground objects, as listed in Table 3.
We compared the classification performance of several
typical band selection methods by changing the band subset's size. The comparison methods include unsupervised
ranking-based methods, e.g., MVPCA [23] and FDPC [30],
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

JUNE 2019

https://engineering.purdue.edu/~biehl/MultiSpec/aviris_documentation.html https://engineering.purdue.edu/~biehl/MultiSpec/aviris_documentation.html https://engineering.purdue.edu/~biehl/MultiSpec/aviris_documentation.html http://www.ehu.es/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes http://www.ehu.es/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes

IEEE Geoscience and Remote Sensing Magazine - June 2019

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