semisupervised classification approach for
HSI
was presented
based on a hierarchical learning paradigm, which is com-
posed of multiple layers feeding into each other with spectral
and spatial features. Ji
et al.
(2014) proposed a semisuper-
vised hyperspectral image classification method to address
both the pixel spectral and spatial constraints. In this method,
the relationship among pixels is formulated in a hypergraph
structure. In the constructed hypergraph, each vertex denotes
a pixel in the hyperspectral image.
Advances on Combining AL and SSL for
Hyperspectral Image Classification
Combining
AL
and SSL is a promising choice. Many existing
works have introduced this combination. Samiappan and
Moorhead (2015) proposed a semisupervised approach which
adapts active learning to a cotraining framework in which the
algorithm automatically selects new training samples from
unlabeled samples. This proposed approach is validated by
applying a probabilistic support vector machine classifier.
CASSL
combines
AL
and SSL to invoke a collaborative label-
ing process by both human experts and classifiers. In
CASSL
,
an
AL
-based pseudolabel verification process is performed for
improving the pseudolabeling accuracy to facilitate SSL. The
unlabeled data with low pseudolabeling confidence in SSL
will become the query candidates in
AL
.
Proposed Methodology
As stated above, the framework of integrating
AL
and SSL to
enable collaborative labeling by both the human experts and
classifiers can obtain more accurate results in the hyperspec-
tral image classification. Assigning pseudolabels to unlabeled
data and retraining the classifier with both the labeled data
and pseudolabeled data are crucial issues. A good threshold
that can avoid introducing incorrect pseudolabels, while
the high threshold will result in fewer samples meeting the
condition at each iteration, and may lead to an “empty loop”
phenomenon. A low threshold can assign incorrect pseudo-
labels to unlabeled data and thus deteriorate classification
performance. Moreover, the effectiveness of the pseudolabel-
ing procedure may heavily depend on the initial labeled data
set. If the initial training set does not match the underlying
class distributions, it is difficult to train
fier at the very initial stage and the judg
samples is constantly changing. If a sing
old is selected for pseudolabeling proce
not be able to adapt to the changing model. Therefore, the
subsequent pseudolabeling procedure may invoke so many
wrong pseudolabels that deteriorate the performance of the
final classifier.
Discriminative Information Mining and Multiple Verification
DSC-CASSL
integrates two different
AL
and SSL in a collab-
orative manner for hyperspectral image classification. This
framework enables a collaborative labeling procedure by both
human experts and three classifiers (two check classifiers and
one base classifier) to obtain more confident labeled samples
to improve the classification performance. In
DSC-CASSL
, we
apply
nEQB
and
MCLU
as
ESAL
partly to select the informative
samples for manual labeling. The
nEQB
, query by committee,
considers the uncertainty of the samples by the maximum
disagreement among the committee of learners. It can select
the most informative samples from the uncertain samples
pool.
MCLU
selects the most informative samples according to
the confidence values. These methods are widely applied in
hyperspectral image classification.
Discriminative information, which represents the qual-
ity of the training data, is vital to improve the generalization
ability of the classifiers. It is widely acknowledged that the
training data are limited, using the unlabeled data to enhance
the training data is an inevitable choice. The discriminative
information will be improved by adding the newly labeled
data in the active-learning process and the assigning pseu-
dolabels to unlabeled data in
DSC-CASSL
. Meanwhile, the
generalization ability of classifiers is improved gradually at
each iteration by adding a batch of informative samples to the
labeled data set with two different query function.
Detailed Steps of the DSC-CASSL Framework
In this section, we describe the details of the proposed
DSC-
CASSL
framework.
DSC-CASSL
framework is divided into two
parts. ESAL (Cui, Kai, and Zhongjun 2018) is the base of the
verification part and also reinforces the performance of base
classifier. It is similar to the usual supervised active learning
method, but the process of selecting informative samples has
been changed. We integrate
MCLU
and
nEQB
into a collabora-
tive sample selection strategy. At each iteration, we assume
that the number of samples to be labeled is Q, the samples
contributed by the
MCLU
is
q
1
, and the samples contributed by
the
nEQB
strategy is
q
2
. It is worth noting that
q
1
and
q
2
repre-
sent not only the number of samples, but also the information
of samples. Designing an appropriate function is the most im-
portant and challenging task for
EASL
. In this paper, the fitness
function of the
EASL
is defined as follows.
q Q w
q Q w
A
B
1
2
= ×
= ×
(7)
Q q q R
=
+
1 2
(8)
where
w
A
and
w
A
are the assigned weight parameters for
MCLU
and
nEQB
, respectively. A range of parameters (
w
A
= 0.1,
w
A
= 0.3,
w
A
= 0.5,
w
A
= 0.7,
w
A
= 0.9) were considered in Cui,
Kai, and Zhongjun (2018). We can observe that
w
A
= 0.5 per-
forms better than the other parameters at each iteration. This
phenomenon can be attributed to the fact that complementing
nEQB
and
MCLU
for each other will achieve better performance.
Hence, we set the value of
w
A
and
w
B
are the same, both of
them are 0.5. If
q
1
∩
q
2
≠
Φ
, it indicates that the two strategies
select the same number of valuable samples. So there will be
q
2
vacancies, and the value of
Q
–
q
1
∪
q
2
f R. R is a random factor. The function
lect valuable samples within the thresh-
nts. It should be underlined that R does
not exist at each iteration, but according to the results of the
EASL
to determine R. The pseudocode for ESAL is illustrated
in Algorithm 1 (see next page).
Initially, we utilize the initial labeled samples and the
pseudolabeled samples to train base classifiers. And base clas-
sifier is utilized to predict the unlabeled data. At the begin-
ning of the iteration, the pseudolabeled data set is empty.
Next, we select
q
1
the most informative unlabeled samples
by utilizing MLCU technique and select
q
2
the most informa-
tive unlabeled samples by utilizing
nEQB
algorithm. And then
these selected unlabeled data are labeled by human experts.
We denote the
q
1
newly labeled samples as
L
Q
1
and the
q
2
newly labeled samples as
L
Q
2
. It should be underlined that
DSC-CASSL
doesn’t increase labeling cost because
CASSL
and
DSC-CASSL
has the same numbers of the unlabeled samples
that need to be labeled by human experts at each iteration. We
suppose both
CASSL
and
DSC-CASSL
select Q unlabeled samples
at each iteration. When the
MCLU
technique and
nEQB
tech-
nique simultaneously select the same samples, we will utilize
the random factor R for supplement. Simultaneously, the
labeled set and the unlabeled set are updated by adding the
newly labeled samples to the labeled set and removing them
from the unlabeled set. The updated labeled set are applied to
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2019
843
