IEEE Geoscience and Remote Sensing Magazine - September 2021 - 73

supervised, semisupervised, and unsupervised. Several BS
approaches incorporating clustering, ranking, and optimization
have been exploited. In this article, a series of unsupervised
BS (UBS) techniques is presented and discussed
to provide a broader understanding of the field. Based on
conclusions from the review and experiments, a unified
framework for the existing approaches based on opportune
theoretical assumptions is hypothesized, which
may suggest additional clues to possible further research
avenues in BS.
LAND COVER CLASSIFICATION
The acquisition of an HSI of a scene involves recording
multiple spectral bands in a specific range of wavelengths.
Depending on the spectral and spatial resolutions, the
number and size of the images may vary. Processing HSIs
is computationally expensive due to the multiplying effect
of the large number of bands on the size of the data. Focusing
on the context of land cover classification, one more
specific problem can be identified, linked to the very high
dimensionality of HSIs: the well-known Hughes effect, or
the so-called curse of dimensionality. This problem mainly
arises from the limited number of class-specific samples
compared to the number of available bands and the large
spatial variability of the spectral signatures. To tackle such
challenges, dimensionality reduction (DR) approaches are
needed to retain useful spectral details in an image with a
minimal number of bands, hence rebalancing the dimensionality
of the classification problem space with the number
of available labeled samples.
Basically, DR techniques can be categorized into two
types: feature extraction (FE) and feature selection (FS). In
FE, the original higher-dimensional data are transformed
into a lower dimension to reduce their size and increase
the class separability among different objects [1]. Some
of the popular FE approaches include principal component
analysis (PCA) [2], kernel PCA [3], independent CA
(ICA) [4], the minimum noise fraction [5], discriminant
analysis FE, decision boundary FE [6], and nonparametric
weighted FE. By applying these FE techniques, a completely
new set of reduced bands is obtained, totally changing
the original band characteristics and forming a set of synthetic
bands.
For FS and BS approaches, a subset of informative bands
is selected from the original HSI. Although the terms FS
and BS sound similar, there exists a difference with reference
to their " adjacency " and within-band " spatial " information.
The process of FS generally does not rely on correlating
information about adjacent features, while, in the
case of hyperspectral imaging, each band represents an
image acquired from a single instrument across a range of
equally spaced wavelengths, providing strong similarity between
adjacent bands [7] and in-band spatial characteristics.
For HSI dimensional reduction, FE and BS approaches
are often used individually, but it is also possible to use
them together.
SEPTEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
In 1989, Henderson et. al. implemented BS [8] to classify
organic matter in soil. Subsequently, extensive research
into BS approaches was carried out by several investigators.
In recent years, many articles have been published about
BS. Looking at Figure 1, where the number of BS-related
publications is plotted through time, the exponential
growth of interest in BS is clearly visible. Recently, a review
of BS techniques was published in [9], which classifies
BS approaches into six groups based on ranking, searching,
clustering, sparsity, embedded learning, and hybrid
schemes. With this article, we intend to focus on land cover
classification and discuss the
related state-of-the-art BS. We
propose a different categorization
with respect to [9], an
approach based on the phases
of the typical HSI classification
chain, which gives rise
to a different structure. Such
a framework may be useful
in guiding the composition
of a processing chain when a
particular land cover classification
problem is addressed. The article is thus the result
of scanning relevant scientific literature and proposing a
robust scheme for possible BS architectures to be used in a
land cover classification context. In addition, extensive experimental
analysis is carried out to examine the pros and
cons of the algorithms.
Depending on the application and objective, the bands
THE ACQUISITION OF AN
HSI OF A SCENE INVOLVES
RECORDING MULTIPLE
SPECTRAL BANDS IN A
SPECIFIC RANGE OF
WAVELENGTHS.
are selected and may differ in their own respect. Quite often,
BS is performed as a preprocessing technique prior to
classification to select the most informative bands. Other
than classification, BS has been successfully implement -
ed for applications including displays of the most significant
bands for visual interpretation [10]-[12], anomaly
100
120
140
160
180
20
40
60
80
FIGURE 1. The number of papers focused on BS published during
the past three decades.
73
1989-1999
2000-2004
2005-2009
2010-2014
2015-2019
Number of Publications
IEEE Geoscience and Remote Sensing Magazine - September 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2021
