A Double-Strategy-Check Active Learning
Algorithm for Hyperspectral Image Classification
Ying Cui, Xiaowei Ji, Kai Xu, and Liguo Wang
Abstract
Applying limited labeled samples to improve classification
results is a challenge in hyperspectral images. Active Learn-
ing (
AL
) and Semisupervised Learning (
SSL
) are two promising
techniques to achieve this challenge. Combining
AL
with
SSL
is an excellent idea for hyperspectral image classification.
The traditional method, such as the Collaborative Active
and Semisupervised Learning algorithm (
CASSL
), may in-
troduce many incorrect pseudolabels and shows premature
convergence. To overcome these drawbacks, a novel frame-
work named Double-Strategy-Check Collaborative Active
and Semisupervised Learning (
DSC-CASSL
) is proposed in
this paper. This framework combines two different
AL
algo-
rithms and
SSL
in a collaborative mode. The double-strategy
verification can gradually improve the pseudolabeling ac-
curacy and facilitate
SSL
. We evaluate the performance of
DSC-CASSL
on four hyperspectral data sets and compare it
with that of four hyperspectral image classification meth-
ods. Our results suggest that
DSC-CASSL
leads to consistent
improvement for hyperspectral image classification.
Introduction
Hyperspectral images (
HSIs
) are characterized by hundreds
of bands acquired in contiguous spectral ranges and narrow
spectrum intervals.
HSIs
have been widely implemented in
a diverse range of applications, including change detection,
image fusion, urban planning, as well as classification (Stein,
Zheng, and Kikkinidis 2013; Bigdeli, Samadzadegan, and
Reinartz 2013; Govender, Chetty, and Bulcock 2007).
HSI
clas-
sification is a crucial processing in many applications. Hence,
extensive research efforts have been foc
image classification. A lot of excellent s
tion models, such as the support vector
madzadegan, Hasani, and Reinartz 2017)and neural networks
(; Li, Huang, and Liu 2017) have been proposed in the last
decades. However, supervised models have shown to perform
well on applications that have large available data sets and
depend on the assumption that the data applied in the train-
ing process. The numbers of available labeled samples are
scarce and very costly to collect. Hence, as to hyperspectral
image classification, a foremost assignment is how we can la-
bel limited samples to improve the generalization capabilities
of the model and accurately classify the images.
AL
and (SSL)
are two promising techniques to address the problem.
Active learning, which comes from the machine learning
field, provides many useful tools to select the high-quality
samples to train effective classifiers (Yang 2011; Albalooshi
et al.
2018).
AL
framework has been successfully applied in
remote sensing applications and speech recognition (Tu
et al.
2013; Crawford, Tuia, and Yang 2013; Tuia, Volpi, and Copa
2011; Settles 2010). Active learning obtains the high-quality
samples by selecting the informative samples with a query
function for human labeling iteratively (Wang and Ye 2015).
Therefore, the primary task of
AL
is to design query functions
that should contain a series of criteria for selecting the most
informative samples. Initially, data are divided into two sets:
labeled and unlabeled. Then, beginning the iteration loop,
it utilizes a query function to search the unlabeled data and
select the most informative samples for manual labeling.
Simultaneously, the labeled set and the unlabeled set are
updated by adding the informative samples to the labeled
set and removing them from the unlabeled set. The classi-
fiers are retrained with the appended labeled samples. The
step of training and the step of assigning labels are iterated
alternately until stable and approved classification results are
generated. In this way, the time consumption and the cost of
human labeling can be greatly reduced. Hence, active learn-
ing is deeply researched in classification of the hyperspectral
images. Mitra, Shankar, and Pal (2004) proposed an active
support vector learning algorithm for supervised classification
in hyperspectral images. It initially applied a small number
of labeled samples, then refined by querying for the labels of
samples from an unlabeled data pool. The label of the most
interesting unlabeled samples is queried at each step. In this
method, active learning is exploited to minimize the number
of labeled data implemented by the
SVM
classifiers. Demir,
Persello, and Bruzzone (2011) presented different batch-mode
AL
techniques for the classification of
HSI
with support vector
machines. The combination of uncertainty criterion and di-
versity criterion contributes to the selection of the potentially
most informative set of samples at each iteration. This method
will reduce redundancy among the selected samples.
Another approach, which is semisupervised learning, is
earning. It utilizes both the labeled data
a for the model learning. Moreover,
ng pays more attention on the unlabeled
data in unsupervised approach. It stimulates the supervised
model by increasing the quantity of the training samples, and
improves the generalization ability of the classifier by using
the labeled data and the unlabeled data. Alhichri
et al.
(2015)
proposed a semisupervised classification approach for remote
sensing images based on a hierarchical learning paradigm.
This approach is composed of multiple layers feeding into
each other with spectral and spatial information. Bruzzone,
Chi, and Marconcini (2006) proposed progressive semisuper-
vised
SVM
approach. It enhances the
SVM
classifier by assign-
ing pseudolabels to the unlabeled data that are closest to the
margin bounds. As mentioned before, active learning solves
the problem by improving the quality of the training samples,
while semisupervised learning solves the problem by increas-
ing the quantity of the training samples. Combining active
learning with semisupervised learning becomes an inevitable
choice. Recently, many researches have demonstrated that
College of Information and Communications Engineering,
Harbin Engineering University, Harbin, 150001, China,
(
;
; xukai@
hrbeu.edu.cn;
).
Photogrammetric Engineering & Remote Sensing
Vol. 85, No. 11, November 2019, pp. 841–851.
0099-1112/19/841–851
© 2019 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.85.11.841
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2019
841
