Thus, we use a stacked generalization (
SG
) (Ting and Wit-
ten, 2002; Hatami and Ebrahimpour, 2007; Sesmero
et al
.,
2015) for change detection of high-resolution remote sens-
ing images. First,
ELM
,
SVM
, and
KNN
are utilized as the base
classifiers at level-0.
ELM
, proposed by Huang (2005), has a
fast learning speed and strong generalization ability, so we
first classify most of pixels by
ELM
.
SVM
, proposed by Vap-
nik (Vapnik and Vladimir, 1995) in 1995, solves non-linear
problems well and can avoid local minima problem.
KNN
is
a simple algorithm which has strong robustness and classi-
fies using analogy, and has a good classifying ability in local
image. Multi-response linear regression (
MRLR
) is then used to
construct the meta-classifier at level-1. Ting and Witten (2002)
have tested four methods - C4.5, IB1, NB, and
MRLR
as meta-
classifier in hybrid ensemble system, and found that only the
MRLR
ensemble achieved satisfying results. Seewald (2002)
also found that
MRLR
can effectively solve the binary classifi-
cation problem. For each sample,
MRLR
utilizes the predicted
values of the
L
base classifiers to construct the input feature
data but ignores the association with neighboring pixels. In
this paper, considering spatial information, the weighted
average of eight neighboring pixels is also taken into account
when constructing the feature data for the meta-classifier. The
fruit fly optimization algorithm (
FOA
) proposed by Pan (2012)
is used to solve the parameters of
MRLR
.
With the increase of spatial resolution and the decrease of
spectral resolution, spectral separability for similar objects
is reduced. Accuracy of change detection is degraded due to
“different spectrum with same objects”. Thus, object-oriented
technique becomes one of the most popular methods for
high-spatial-resolution images (Hao
et al
., 2016; Peng and
Zhang, 2017). However, accuracy of change detection in
object-oriented methods is directly influenced by the initial
image segmentation. Another effective way is constructing
multi-source image features. Li et al. (Li, Huang
et al
., 2017)
proposed a method by integrating macro and micro-texture to-
gether and obtained high accuracy. Volpi
et al
. (2013) stacked
spectral features, texture features, and morphological features
to perform supervised change detection in very high-resolu-
tion images. Peng
et al
. (2017) extracted texture and spatial
features by using
LBP
and Sobel gradient and combined them
with spectral features to obtain the change information for
high-resolution
GF-1
image. These studies have demonstrated
that the inclusion of texture and morphological features can
compensate for the lack of detailed spectral information.
So, in this paper, the statistical texture and structure texture
features (morphological profiles) are utilized to compensate
for insufficient spectral information and the initial pixel
wise change map is combined with the smaller heterogeneity
multi-scale segmentation map to obtain the final change map,
where the influence of over-segmentation in object-oriented
change detection is alleviated and the “salt and pepper noise”
in pixel-wise based change detection is reduced. In experi-
ments, two
ZY-3
and QuickBird datasets are used to demon-
strate the effectiveness of the proposed method.
The remainder of this paper is organized as follows. The
next section introduces the proposed methodology. Then, the
experimental results are discussed and analyzed. Finally, the
conclusion is drawn.
Methodolgy
Stacked Generalization (
SG
)
SG
is a heterogeneous ensemble algorithm, integrating differ-
ent classifiers, followed by a two-level hierarchical structure
(Figure 1). At level-0, each base classifier is trained by the
original training set, and for each pixel, each base classifier
produces a predicted value. These predicted values are used
as the input data for level-1. At level-1, a trainable combiner
integrates the output data of each base classifier at level-0 and
obtains the final prediction. This trainable combiner is also
called “meta-classifier”. The overall effect of the classifier en-
semble system depends on the base classifier used at level-0
and the selection of the meta-classifier at level-1.
Given a training set
S
,
SG
randomly divides the original
training set into
J
sub-training sets of equal size and
K
base
classifiers are used at level-0. The next steps are similar to the
J
-fold cross-validation process: Select a sub-training set (
S
j
) for
training and validation of level-1,
j
= 1, 2,…,
J
. The remaining
sub-training sets
S
(–
j
)
=
S
–
S
j
are used to train the
K
base clas-
sifiers at level-0, and
S
j
is the test dataset for the
K
base clas-
sifiers. The base classifiers at level-0 predict all the samples
in
S
j
. The predicted values of
S
j
and their training labels then
form the training set for the meta-classifier at level-1. Each
sample in the training set has
K
features (
K
predicted values).
After training, the meta-classifier generates a fixed classifi-
cation model, and for a new test dataset, the level-0 model
generates an initial prediction vector as the input values of
the level-1 model. The level-1 model then generates the final
prediction values. The overall description of the
SG
algorithm
is summarized as follows.
Input:
z_train
: the sub-training set (
S
(–
j
)
) of base classifiers;
x_label
: the labels of sub-training set (
S
(–
j
)
) of base classifiers;
z_test
: the test samples (
S
j
) of base classifiers;
y_label
: the labels of training set (
S
j
) of meta-classifier;
z
: the test dataset.
For
all the training samples in set
S
(–
j
)
:
Train the
K
base classifiers at level-0 to estimate the parameter
a
:
x_label = a*z_train
End for
For
all the test samples in
S
j
:
Get the predicted values(
x_value
) of base classifiers at level-0:
x_value = a*z_test
End for
Train the meta-classifier at level-1 to estimate the parameter of
weight
w
:
y_label = w*x_value
Predict a new test dataset:
y_value = w*a*z
Output:
y_value
: the predicted value of the
SG
model.
Figure 1. Sketch map of stacked generalization.
734
November 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
