IEEE Geoscience and Remote Sensing Magazine - June 2019 - 153

Endmembers of Change Class 1
Reflectance

0.8
0.6
0.4
0.2
0

100
150
200
250
Stacked Spectral Channels

300

Endmembers of Change Class 3

1
0.5
0

Hyperion data set is shown in Figure 17). By taking advantage
of the endmembers extraction and change and no-change
spectral compositions within a pixel, the identification of
the number of changes is performed by identifying the distinct endmembers, and the
IN THE LITERATURE,
discrimination of multiclass
THERE ARE FEW WORKS
changes is addressed by linear unmixing and abundance
ADDRESSING CD AT
analysis. Reliable CD outputs
THE SUBPIXEL LEVEL.
are obtained, and subpixellevel spectral variations not
detected with the pixel-level
CD techniques are identified. An example of subpixellevel multiclass CD results on the Benton County data set
is shown in Figure 18. The MSU approach identified seven
change classes, whereas the aforementioned pixel-level
methods identified six classes. The new change (i.e., class 7)
is associated with the change of roads surrounding irrigated
agricultural fields. Because the road width is less than a full

Reflectance

two approaches are supervised. Taking advantage of the
selected bands (i.e., BS-SVM) or transformed features (i.e.,
S 2MCD) and the robust classifiers, they result in better CD
performance in terms of higher accuracies.
These works mainly focus on the detection of multiclass
changes at the pixel level. In the literature, there are few
works addressing CD at the subpixel level. In [84], CD is implemented at the subpixel level based on nonlinear spectral
unmixing and decision fusion, where the inner-pixel subtle
changes are analyzed and multiple land-cover compositions are combined. Although it was originally designed for
CD in MS images, this technique can also be applied to HS
images with the available reference training data. An unsupervised MSU model is proposed in [38]. It investigates
the spectral-temporal mixture properties in multitemporal
HS images, and the CD problem is solved by analyzing the
abundances of different distinct multitemporal endmembers (including both change and no-change classes) at
the subpixel level (an example of extracted multitemporal
endmembers in the MSU approach on the Benton County

100
150
200
250
Stacked Spectral Channels

300

Endmembers of Change Class 2

0.8
0.6
0.4
0.2
0

0.5
0

Reflectance
300

Endmembers of Change Class 7

0.8

Reflectance

Reflectance
Reflectance

0.4

100
150
200
250
Stacked Spectral Channels

0.6
0.4
0.2
0

100
150
200
250
Stacked Spectral Channels

300

Endmembers of Change Class 6

0.6

300

Endmembers of Change Class 4

Endmembers of Change Class 5
0.8

0.2
0

100
150
200
250
Stacked Spectral Channels

300

1
0.5
0

0.8

100
150
200
250
Stacked Spectral Channels

300

Endmembers of No-Change Class

0.6
0.4
0.2
0

100
150
200
250
Stacked Spectral Channels

300

FIGURE 17. The endmembers extracted by the MSU approach on the Benton County Hyperion data set. Seven unique change classes are
identified and shown in different colors; the no-change endmembers are in black. The endmembers present some variability within the
same class.
JUNE 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

153

IEEE Geoscience and Remote Sensing Magazine - June 2019

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