IEEE Geoscience and Remote Sensing Magazine - December 2021 - 37

For hyperspectral clustering, in the case of c different
classes, the core task is to partition the pixels into c different
groups based on a certain clustering model, with
each group corresponding to a certain class. Different
methods deal with the internal structure and the complexity
of HSIs, with various model assumptions, which
determines their clustering effect to a large degree. A
taxonomy of the hyperspectral clustering methods considered
in this article appears in Table 1.
CENTROID-BASED CLUSTERING METHODS
Centroid-based methods are the most classical and representative
clustering approach, and they were also the earliest
to be introduced to HSI analysis [45], [46]. Such techniques
TABLE 1. THE TAXONOMY OF HYPERSPECTRAL CLUSTERING METHODS.
CATEGORY
MECHANISM
Centroid
Assumes the cluster has a ball-like structure in the
feature space; clusters HSIs by iteratively minimizing
the overall partition error
Density
Assumes clusters are density point sets separated
by sparse areas in the feature space; clusters HSIs
based on the local density and relative distances
of pixels
Probability
Bionics
Intelligent computing
Assumes pixels from the same class satisfy a probability
distribution model; clusters HSIs based on a
probability rule
Simulates the complex internal structure of HSIs
with a certain biological model; clusters HSIs
through a biological evolution algorithm
Based on other clustering models; utilizes advanced
intelligent computing algorithms to search
for the global optimal solution to the clustering
model
Graph
Models the similarity among pixels with an
adjacency matrix; clusters HSIs with a graph cut
algorithm
SUBCATEGORY
Hard partition
REPRESENTATIVE METHODS
k-means [47], ISODATA
[49], NC-k-mean [52]
Soft partition
-
FCM [45], FCM-S1 [64],
FLDNICM [69]
CFSFDP [71], DAE [72],
SSDL [77]
-
GMM [79], ICAMM [80],
CLDD [86]
-
Single objective
Multiple objective
Complete graph
Bipartite graph
Abbreviated graph
Subspace clustering
Models the internal complex structure of HSIs via
the union of subspaces; explores the underlying
adjacency between pixels through self-representation
learning; groups HSIs by applying spectral
clustering (SC) to the adjacent matrix induced by
the coefficient matrix
Deep learning
Relies on deep neural networks to learn more
discriminative features for clustering and more
accurately simulate the nonlinearity of data
-
Hybrid mechanism
-
Deals with the clustering task by combining two or
more clustering models
Spectral-spatial subspace
clustering
Multiple-view subspace
clustering
Kernel subspace clustering
Autoencoder
Separated network
Generative network
-
SOM [88], UAIC [42],
UADSM [39]
FCIDE [92], MoDEFC [31],
PSO-GMM [93]
AFCMDE [32], AFCMOMA
[94], MOPSO [38]
SC [105], SENP [106],
NLTV [107]
SSCC-BG [115], S-SC [116],
BGP-CJS [117]
FSCAG [43], SGCNR [121]
S4C [40], L2-SSC [41], SSC3DEPF
[128]
SSMLC [139], k-SSMLC [140],
p-SSMLC [141]
KSSC-SMP [142],
KSLRSC [143]
DCN [147], DMC [148],
DSCNet [151]
CCNN [155], DBNC [156],
JSL [159]
CatGAN [162], DAGMC [164],
VaDE [166]
k-GMM [168], k-FDPC [169],
SDCR [174]
ISODATA: iterative self-organizing data analysis technique algorithm; NC-k-mean: neighborhood-constrained k-means; FCM-S1: FCM with mean filtered spatial information; FLDNICM:
fuzzy local double neighborhood information c-means; CFSFDP: clustering by the fast searching and finding of density peaks; DAE: density analysis ensemble; SSDL: spectral-spatial
(SS) diffusion learning; GMM: Gaussian mixture model (MM); ICAMM: independent component analysis MM; CLDD: clustering based on the latent Dirichlet distribution; SOM: selforganizing
map; UAIC: unsupervised artificial immune classifier; UADSM: unsupervised spectral matching classifier based on artificial deoxyribonucleic acid (DNA) computing; FCIDE:
fuzzy clustering (FC) using improved differential evolution (DE); MoDEFC: modified DE FC; PSO-GMM: particle swarm optimization-based GMM; AFCMDE: automatic FC based on
multiple-objective DE; AFCMOMA: adaptive multiple-objective memetic FC algorithm; MOPSO: multiple-objective PSO; SC: spectral clustering; SENP: Schroedinger Eigenmap with
nondiagonal potentials; NLTV: graph-based nonlocal total variation; SSCC-BG: SS coclustering based on a bipartite graph (BG); S-SC: sequential SC; BGP-CJS: BG partition-based
coclustering with joint sparsity; FSCAG: fast SC with anchor graph; SGCNR: scalable graph-based clustering with nonnegative relaxation; S4C: SS sparse subspace clustering; L2-SSC:
2
, -norm regularized sparse subspace clustering; SSC-3DEPF: SSC based on 3D edge-preserving filtering; SSMLC: SS-based multiple-view low-rank SSC; p-SSMLC: parallel SSMLC; DCN:
deep clustering network; DMC: deep multiple-manifold clustering; DSCNet: deep SSC based on an autoencoder network; CCNN: clustering based on a convolutional neural network;
DBNC: deep brief network nonparametric clustering; JSL: joint unsupervised learning; CatGAN: categorical generative adversarial network; DAGMC: deep adversarial Gaussian mixture
autoencoder clustering; VaDE: variational deep embedding; k-GMM: hybridization of the k-means and the GMM; k-FDPC: hybridization of the k-means and fast finding of density peaks
clustering; SDCR: sparse dictionary-based anchor regression.
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
37
IEEE Geoscience and Remote Sensing Magazine - December 2021

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