IEEE Geoscience and Remote Sensing Magazine - December 2021 - 54

the advantages of different schemes to further improve the
clustering performance. A typical example is the combination
of the centroid-based clustering scheme with other
approaches, such as the k-GMM [168]. The k-GMM combines
centroid- and probability-based clustering. Taking
advantage of the k-means and the GMM, k-GMM obtains
better clustering accuracy for HSIs. In addition, in [169],
an improved fast density peaks-based clustering (k- FDPC)
algorithm was proposed, which is a hybridization of the
k-means and the CFSFDP. Based on the CFSFDP approach,
this algorithm calculates the local density based
on an adaptive bandwidth pdf, and it searches cluster centroids
by fitting the density and distance decision graph.
Subsequently, it infers the pixel assignment through the
k-means. As a result, k-FDPC outperforms both k-means
and CFSFDP.
In addition, bionics-based clustering can also be combined
with other approaches. For example, in [170], a
fuzzy Kohonen local
information c-means clustering
(FKLICM) method was put forward. It employs the Kohonen
NN to model the complexity of remote sensing images
and integrates the discriminative rules of the FLICM to
enhance the discriminability of the model. Consequently,
more accurate clustering results are obtained. In addition,
by combining the advanced artificial bee colony (ABC)
model with MRF, a novel ABC-MRF clustering algorithm
was developed for HSIs [171]. The ABC model is utilized
to better search cluster centroids and optimize the objective
function, with the MRF utilized to incorporate spatial
neighborhood information to further improve the clustering
accuracy.
Moreover, the graph-based clustering scheme can be
flexibly combined with other clustering schemes as well.
For example, in [172], a Gaussian SC model (GSC) was
constructed by integrating the powerful information extraction
ability of the graph model into the GMM framework.
In [173], by combining the k-means with the graph
model, a graph-based k-means (G-k-means) technique
was developed. It utilizes the graph model to estimate the
parameters and initializations for the k-means, which effectively
improves the clustering performance to obtain
an accurate segmentation result for HSIs for Mars exploration.
In addition, by combining anchor graph-based clustering
with subspace clustering, a sparse dictionary-based
anchor regression (SDCR) algorithm was introduced for
HSIs [174]. It constructs a more representative dictionary
through dictionary learning with double sparsity constraints,
and it utilizes the anchor subspace regression to
efficiently evaluate the similarity between hyperspectral
pixels. With the help of SC, the final clustering result is obtained.
By integrating the advantages of the anchor graph
and the subspace, SDCR achieves good performance.
Generally speaking, by comprehensively taking advantage
of two or more different clustering schemes, hybrid
mechanism-based clustering methods can overcome the
defects of both techniques and may bring about better
clustering performance. In theory, hybridization can be
extended to any two or more clustering schemes, and progressively
more attractive hybrid clustering methods may
be developed through future research.
EXPERIMENTS
In this section, the performance of some popular and representative
hyperspectral clustering algorithms is evaluated,
including FCM (https://github.com/wwwwwwzj/
fcm) [45], FCM-S1 [64], CFSFDP (https://github.com/DesperadoZ/
Density_Peak_Clustering) [71], GMM (https://github.com/
AdamaTG/Matlab_GMM) [79], SC (https://github.com/jhliu17/
SpectralClustering) [105], FSCAG [43], SGCNR [121], SSC
(http://vision.jhu.edu/code/) [40], [122], and L2-SSC [41].
Specifically, FCM is one of the most representative centroidbased
clustering methods, while FCM-S1 is a classical improved
version of FCM, achieved by incorporating spatial
information. CFSFDP is a representative density-based
Centroid-Based Model
0.04
0.02
Combination
(a)
Density-Based
Model
Graph-Based Model
(b)
FIGURE 10. The hybrid mechanism-based clustering scheme. (a) The HSI. (b) The hybrid clustering model construction. (c) The hybrid
model optimization. (d) The clustering result.
54
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021
20
10
0 0
Probability-Based
Model
(c)
(d)
10
20
Optimization

https://www.github.com/wwwwwwzj/fcm https://www.github.com/wwwwwwzj/fcm https://www.github.com/DesperadoZ/Density_Peak_Clustering https://www.github.com/DesperadoZ/Density_Peak_Clustering https://www.github.com/AdamaTG/Matlab_GMM https://www.github.com/AdamaTG/Matlab_GMM https://www.github.com/jhliu17/SpectralClustering https://www.github.com/jhliu17/SpectralClustering http://vision.jhu.edu/code/

IEEE Geoscience and Remote Sensing Magazine - December 2021

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