IEEE Geoscience and Remote Sensing Magazine - June 2019 - 152

Ω

L0
Ωc
ω1

L1
ω3

L2

Binary CD

ωn

Multiple CD

ω2

ω4

ω5

ω6

Root Node (SCVs)
Intermediate Node (Partial SCVs)
Leaf Node (Detected Change Classes)
FIGURE 15. A three-level hierarchical tree for modeling the multiclass CD problem (Benton County Hyperion data set) with S2CVA.

(a)

(b)

(c)

(d)

(e)

(f)

FIGURE 16. Multiclass change maps obtained on the Benton
County Hyperion data set using (a) HSCVA [35], (b) S2CVA [36], (c)
BSCVs [71], (d) band selection and support vector machine [28],
(e) S2MCD [37], and (f) multitemporal spectral unmixing [38] approaches. The detected change classes are represented in different
colors, and the no-change classes in white.

TABLE 4. MULTICLASS CD RESULTS OBTAINED WITH
STATE-OF-THE-ART TECHNIQUES ON THE BENTON
COUNTY HYPERION DATA.
METHODS

OA, %

K

OE, PIXELS

CE, PIXELS

TE, PIXELS

HSCVA [35]

95.26

0.8800

1,715

1,918

3,633

S2CVA [36]

95.36

0.8822

1,878

1,676

3,554

BSCVs [71]

94.72

0.8662

1,935

2,138

4,073

BS-SVM [28]

96.37

0.9089

1,245

1,469

2,714

S2MCD [37]

96.25

0.9055

1,323

1,519

2,998

152

heterogeneous change classes are present in the 2D feature
space. In the example, six changes (i.e., ~ 1 - ~ 6) were detected by following the three-level (i.e., L 0 - L 2) tree hierarchy shown in Figure 15, where the leaf nodes correspond to
the six change classes. The multiclass change map is shown
in Figure 16(b).
In [71], the authors propose an unsupervised multiclass
CD method based on binary SCVs (BSCVs), which converts
the original SCVs into binary code words that highlight the
change information. The binarization of the SCVs moves
to a simpler domain that can be exploited to discriminate
the different kinds of change. The BSCVs are then represented into a dendrogram (i.e., a tree structure) to separate
the types of change, taking into account the hierarchical organization of the change components. The binary representation is effectively exploited to construct the dendrogram
because it represents the change information in a simpler
and more explicit way compared with the real value representation (i.e., the SCVs).
In [28], the authors analyze in detail the CD performance on HS images by selecting the most informative
band subset from the original high-dimensional HS data
space. Several issues are investigated, including 1) the estimated number of multiclass changes, (2) binary CD
performance, (3) multiple CD performance, (4) estimated
optimal number of selected bands, and (5) computational
efficiency. In particular, both supervised and unsupervised
CD approaches are studied. Experimental results obtained
on different data sets demonstrate that it is feasible to perform the considered CD task for HS images in a reduced
feature space without losing the discrimination capabilities
of a change detector.
In [37], a semisupervised technique [termed semisupervised multiclass change detection (S 2 MCD)] is proposed to
enhance the multiclass CD performance in HS images. It
takes advantage of the unsupervised change representation, i.e., the compressed adaptive SCV representation in
the 2D polar domain [36], to provide prior knowledge of
the multiclass change information. Then, the pseudotraining samples are generated for the no-change class and for
each change class based on a stratified random-sampling
strategy. The multiclass CD task is completed with high
detection accuracy using the generated pseudotraining
samples and advanced supervised classifiers, such as SVM
and random forest for classifying the original SCV or transformed features.
The results of a qualitative and quantitative experimental comparison of some state-of-the-art multiclass CD
methods applied to the Benton data set are provided in Figure 16 and Table 4. We considered the following methods in
the experiments: HSCVA [35]; S 2 CVA [36]; BSCVs [71]; BSSVM, where the 18 most informative bands were selected
[28]; S 2MCD (using the RF classifier on the first 10 MNF
components) [37]; and MSU [38]. The first three approaches
are unsupervised, which is useful in practical applications
when the ground reference data are not available. The last
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

JUNE 2019
IEEE Geoscience and Remote Sensing Magazine - June 2019

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