mixed pattern (Ghosh
et al.
, 2011). It tries to find the
best label for every pixel based on a fuzzy measure to
represent a degree of a pixel belonging to one class.
The final classification result can be estimated when an
objective error function is minimized.
3. Expectation Maximum-Markov Random Field (
EM-MRF
)
Following Bruzzone and Prieto’s framework (2000), an
EM-MRF
framework is constructed. This method first
characterizes the density function of changed and un-
changed classes after EM clustering. The final change
mask can be obtained when the general energy reaches
the minimum based on MRFs modal. To minimize the
energy term, we use the same
ICM
algorithm in our
proposed procedure.
Plate 2 shows the change-detection maps from three previous
methods and our proposed method over the whole interested
area for qualitative comparison. It is clear that
PCA
and
FCM
,
as context-insensitive methods, both caused a certain amount
of salt-and-pepper noise; among the three methods,
FCM
performed the worst as it labeled an almost unchanged region
of water as the change.
EM-MRF
and our method, as context-
sensitive methods, could obtain similar results with a low
noise level.
Figure 4 shows the examples of five subsets with different
change types and the detection results from three unsuper-
vised method and our proposed method. We crop five sub-
images with representative area of 50*50 pixels for each from
original images and different detection maps in Plate 2. From
the result, our proposed method can generally outperform the
other three methods over the different change types. Espe-
cially for the noisy type of
local reflectance change
and
water
quality change
,
PCA
,
FCM
,
EM-MRF
all easily over-detect falsely,
while our proposed method can keep them out for the final
results (Figure 4b and Figure 4c). The only exception among
all the examples is the case of
barren land<->built-up area
(Figure 4e). This is because we think the transition of
built-up
area
to
barren land
with extremely smooth surface usually
fails to hold distinguishing change on our saturation level, as
their materials are similar, which affects the performance.
Table 2 is the result of the quantitative evaluation for the
detection results from four methods using Im and Jensen’s
(2005) evaluating framework. A total of 800 sample points
were randomly created within the study area. The reference
data are acquired from Google Earth
™
with the help of expert
interpretation and field survey. Each subset (or pixel) is first
spatially matched with the corresponding high spatial resolu-
tion image. The change type included in each pixel is then
checked by manual interpretation. Based on the reference
data, 141 sample pixels are categorized into “changed” and
659 are labeled as “unchanged.” To compare the change de-
tection accuracy of four techniques, the error matrix and the
corresponding overall accuracy and Kappa statistic as well
as user’s and producer’s accuracy are calculated (Story and
Congalton, 1986; Congalton, 1991). Table 2 lists the error ma-
trix derived for each method. The overall accuracy and Kappa
statistic for our proposed method are 95.1 percent and 83.3
percent, both are the highest among four methods.
EM-MRF
is
ranked the second and
FCM
and
PCA
perform the poorest based
on the overall accuracy. When we look at the accuracy for
individual classes, the performance of our proposed method
is also the best among four methods for both “change” and
“no-change” classes. However, compared with the individual
user’s and producer’s accuracy of “no-change” class, the ac-
curacy of “change” class from our proposed method are much
Figure 4. The subsets and change-detection results of four unsupervised for different subset scenes (The detected changes from differ-
ent methods are shown in pure white): (a)
vegetation<->built-up area
; (b)
local reflectance change and vegetation<->built-up area
; (c)
unchanged
water body with different quality
; and (d)
water<->barren land
; (e)
barren land<->built-up area
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2015
643
