IEEE Geoscience and Remote Sensing Magazine - June 2019 - 120

120

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

JUNE 2019

Sparsity-based methods
formulate band selection into a sparsity-constrained optimization
problem.

D. Sparsity
based

Graph-based criteria: spectral clustering [102], an MDPP [104], graph clustering [105], dominant-set extraction [106], and a complex network [107]

C.3. Graph based

SVM-classifier criteria: An EFS, [123], an MR-SVM [124], and KFS [125]
MLR-classifier criteria: MLR [127], and SMLR [128]
Application model criteria: sparsity-promoting priors [130] and SCBS [131]
Deep-learning model criteria: deep CNNs and distance density [132], segmented autoencoders [133], attention-based CNNs
[134], and CM-CNNs [135]
Combined ranking and clustering criteria: an SSI [136], PSO and IFBP [138], and manifold ranking [139]
Combined ranking and searching criteria: conditional MI and GA [140], spatial clustering and GA [141], and minimum redundancy/
maximum relevance [142]

E.2. Other learning
based

F.1. Hybrid scheme
based

Sparse regression criteria: A LASSO [118], a sparse HSIC and a surrogate-kernel-based algorithm [119], a group-sparsity-based
algorithm [121], and a spectral-spatial hypergraph model [122]

E.1. Classifier learning
based

D.3. Sparse regression
based

D.2. Sparse representa- Sparse representation criteria: SpaBS [110], multitask sparsity pursuit [111], a collaborative sparse model [112], ISSC [113], SWLRSC
tion based
[114], FRSR [115], DWSSR [116], and SSR [117]

SNMF criteria: an SNMF [108], and an SNMF-EMD [109]

Unsupervised criteria: an exemplar-based AP [96], an FMAP [97], and a DFM [98]
Supervised and semisupervised criteria: discriminative kernel clustering [99], semisupervised adaptive AP [100], and an NTMI [101]

C.2. AP based

D.1. SNMF based

Unsupervised criteria: k-means [90], kernel k-means [92], and dual clustering [93]
Supervised and semisupervised criteria: a prototype space algorithm [94], and semisupervised k-means [95]

C.1. k-means based

B.3. Eliminating search- SBS criteria: BSC CEM [81], VGBS [82], ACMBS [83], a minimum-noise band selection algorithm [84], and a joint-skewness band
ing based
selection algorithm [85]

taheuristic algorithm [80], and a chaotic binary-coded gravitational search algorithm [80]

B.2. Updated searching Classifier-independent criteria: CSFS [69], an immune clonal strategy [70], PSO [71], GA [72], TMI, and STMI [74]
based
Classification-dependent criteria: gravitational search and BA [76], FODPSO [77], a probabilistic memetic algorithm [79], a me-

Unsupervised-searching criteria: LP [45], spectral rhythm [46], and a geometry-based algorithm [47]
Supervised-searching criteria: BAO, ADM, and MDM [48]; MEAC [49]; NPFS [50]; MMAIQ [51]; a simple, equal-width interval binning [53]; and VPNRS [54]

Classification-aimed criteria: MMCA [23], MI [32], feature weighting [33], phase correlation [34], and NHMC [35]
Spectral-unmixing-aimed criteria: OSP [23], VNVBS [36], ICA [37], and an alternative separability index [38]

A.2. Supervised ranking based
B.1. Incremental
searching based

High-information criteria: MVPCA and MSNRPCA [23], information entropy [14], and the covariance-based method [24]
Low-correlation criteria: CBS [25], CMBS [26], optimal band index [27], and curve fitting [28]
Large-dissimilarity criteria: ECA [29], and E-FDPC [30]

TYPICAL EXAMPLES

A.1. Unsupervised
ranking based

SUBCATEGORY

NOTE: SNR: signal-to-noise ratio; PCA: principal component analysis; MVPCA: minimum-variance PCA; MSNRPCA: maximum SNR PCA; CBS: constrained band selection; CMBS: constrained multiple-band
selection; FDPC: fast density peak-based clustering; ECA: exemplar component analysis; E-FDPC: enhanced fast density peak-based clustering; MMCA: minimum misclassification canonical analysis; NHMC:
nonhomogeneous hidden Markov chain; VNVBS: variable-number variable-band selection; OSP: orthogonal subspace projection; ICA: Independent component analysis; MI: mutual information; TMI: trivariate MI; STMI: semisupervised TMI; NTMI: normalized TMI; LP: linear prediction; BAO: band add-on; ADM: average distance method; MDM: minimum distance method; MEAC: minimum estimation abundance covariance; NPFS: nonlinear parsimonious feature selection; MMAIQ: maximal minimal associated index; VPNRS: variable-precision neighborhood rough set; PSO: particle swarm optimization; GA:
genetic algorithm; CSFS: clonal selection feature selection; BA: Bat algorithm; FODPSO: fractional-order Darwinian PSO; SBS: sequential backward selection; CEM: constrained energy minimization; BSC:
band selection for CEM; VGBS: volume-gradient-based band selection; ACMBS: autocorrelation matrix-based band selection; AP: affinity propagation; FMAP: feature-metric-based AP; DFM: discriminative
feature metric; MDPP: multigraph determinantal point process; SNMF: sparse nonnegative matrix factorization; EMD: earth's mover distance; SpaBS: sparse representation-based band selection; ISSC:
improved sparse subspace clustering; SWLRSC: squared-weighted low-rank subspace clustering; FRSR: fast and robust self-representation; DWSSR: dissimilarity-weighted sparse representation; SSR: symmetric sparse representation; LASSO: least shrinkage and selection operator; HSIC: Hilbert-Schmidt independence criterion; EFS: embedded feature selection; SVM: support vector machine; MR-SVM: modified recursive SVM; KFS: kernel-based feature selection; MLR: multinomial linear regression; SMLR: sparse MLR; SCBS: sparse-constrained band selection; CNNs: convolutional neural networks; CM: contribution map; SSI: intraclass separability index; IFBP: impurity-function band prioritization.

Hybrid scheme-based
methods integrate
two or more different
schemes into select
bands.

Clustering-based methods select representative bands from band
clusters.

C. Clustering
based

F. Hybridscheme based

Searching-based methods search the aimed
band subset, which
can optimize a given
criterion function.

B. Searching
based

Embedding learningbased methods
optimize the objective
function of a learning
model while selecting
proper bands.

Ranking-based methods quantify the importance of each band and
choose the top-ranked
bands.

A. Ranking
based

E. Embeddinglearning based

BRIEF DESCRIPTION

CATEGORY

TABLE 1. THE TAXONOMY OF HYPERSPECTRAL BAND SELECTION METHODS.
IEEE Geoscience and Remote Sensing Magazine - June 2019

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