IEEE Geoscience and Remote Sensing Magazine - December 2021 - 60

cannot be fully satisfied by HSIs. FCM and FCM-S1 have
a low complexity of O(MNDct) and are relatively efficient,
and thus they are suitable for large hyperspectral data sets,
where t denotes the number of iterations. CFSFDP has a
higher complexity of approximately OMN 2
^^h h and requires
a relatively large memory to store a sizeable pairwise
pixel distance matrix, limiting its suitability for large HSIs.
GMM has a relatively large complexity of
OM ,Nct
2
^^h h
which degrades its suitability for large HSIs to some degree.
By comparison, complete graph-based methods, e.g.,
SC, may perform well, but they are troubled by a large
computational cost. Specifically, SC has a large complexity
of OMNDt
its practicability to a large degree. Comparatively speaking,
abbreviated graph-based methods, e.g., FSCAG and SGCNR,
are very efficient and suitable for large HSIs because
of their lower complexity. The complexity of FSCAG and
SGCNR are O(MNDu) and O(MND log u + MNc2
+ ),c3
respectively, where u is the number of anchors and
v is the number of nearest neighbors, with uv , MN% [43],
[121]. However, their clustering accuracy cannot be guaranteed.
Relative to the above methods, subspace clustering
approaches, e.g., SSC and L2-SSC, may deal better with
the clustering task for HSIs and bring about a competitive
clustering performance. However, such methods generally
have a very large computational complexity of about
OMNt3
^^h h and are very time and memory consuming,
which degrades their attractiveness in real functions and
hinders their applications to large hyperspectral data sets
to a large degree.
In general, clustering is an important and necessary
technique for HSI interpretation, but it has much room for
improvement. Accuracy, efficiency, and intelligence may be
the major lines of development for hyperspectral clustering
in the future. Based on the research status of hyperspectral
clustering, the challenges and possible future research lines
are pointed out as follows.
DEVELOPING EFFECTIVE AND EFFICIENT MODELS
Accuracy and efficiency are both very important in practical
applications. However, most current hyperspectral
clustering algorithms cannot simultaneously consider
these two aspects very well. For example, subspace clustering
may bring about a higher clustering accuracy, but
significant computational complexity generally follows,
which degrades the technique's scalability to large scenes
and limits its practical applications. Centroid-, density-,
and probability-based methods are generally efficient but
with limited clustering accuracy for HSIs. Hence, how to
develop more effective and efficient hyperspectral clustering
models with a high accuracy and a low time cost
is an interesting and attractive topic. Generally speaking,
combining the advantages of different clustering models,
such as hybrid mechanism-based clustering, may be an effective
way to overcome these obstacles. In addition, combining
advanced clustering models with high-performance
60
^^h h and is time consuming, which reduces
2
computing techniques, such as parallel computing, may
greatly enhance the efficiency while guaranteeing a high
clustering accuracy.
+ MNcv
DEVELOPING MULTIFEATURE-BASED METHODS
Hyperspectral remote sensing images generally come with
the serious problem that pixels from the same class have
different spectra, while pixels from different classes have
similar spectra, which greatly degrades the separability
among different classes. Multiple features from different
views/domains, e.g., spectrum, texture, and geometry, describe
ground objects from different views and can provide
complementary information to effectively enhance the
discriminable capability of a clustering model to improve
the clustering accuracy. However, most existing clustering
methods integrate the spatial information by means
of regularization to simply explore the discriminability of
the spectral-spatial information or simply fuse multiple
features through concatenation, which does not fully excavate
the potential of the multidomain information in HSIs.
Therefore, more advanced multiple featured-based clustering
methods should be developed to further improve the
clustering accuracy.
DEVELOPING OBJECT- OR
SUPERPIXEL-BASED METHODS
Hyperspectral remote sensing images contain abundant and
complex spatial neighborhood information; however, they
are seriously influenced by noise during the imaging process.
Most existing clustering methods are pixel-based methods,
which have several inherent shortcomings. First, pixel-based
methods are easily affected by salt-and-pepper noise, resulting
in fragmented cluster maps. Second, pixel-based methods
cannot flexibly model the spatial neighborhoods with
various shapes, which leads to an inadequate exploitation of
the spatial information of HSIs. Last, due to the large number
of pixels, pixel-based methods may be troubled by a
large computational cost, especially for graph-based clustering
and subspace clustering methods. At this point, object/
superpixel-based clustering techniques can effectively overcome
these obstacles. Thus, more object-/superpixel-based
clustering methods should be developed for HSIs to further
improve the clustering performance.
PUSHING DEEP LEARNING INTO THE
HYPERSPECTRAL CLUSTERING FIELD
Hyperspectral remote sensing images generally have a
typical nonlinear structure due to the complex imaging
environment and the influences of many nonlinear factors.
Thus, pixels from different classes are commonly not
linearly separable. On the other hand, low-level spectral
or spatial features have a limited discriminability and
cannot well distinguish various classes with high similarity.
However, most existing clustering methods are based
on linear models to approximate the nonlinearity of HSIs,
which leads to a large systematical error, or simply deal
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021
IEEE Geoscience and Remote Sensing Magazine - December 2021

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