This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1
SAR Target Small Sample Recognition Based
on CNN Cascaded Features and
AdaBoost Rotation Forest
Fan Zhang , Senior Member, IEEE, Yunchong Wang, Jun Ni, Student Member, IEEE,
Yongsheng Zhou
, Member, IEEE,andWeiHu
Abstract— Automatic target recognition (ATR) has made great
progress with the development of deep learning. However, the tar-
get feature in synthetic aperture radar (SAR) image is not
consistent with human vision, and the SAR training samples
are always limited. These hard issues pose new challenges to
the SAR ATR based on convolutional neural network (CNN).
In this letter, we propose an improved CNN model to solve the
limited sample issue via the feature augmentation and ensemble
learning strategies. Normally, the high-level features that are
more comprehensive and discriminative than the middle-level and
low-level features are always employed for category discrimina-
tion. In order to make up the insufficient training features in
the limited sample case, the cascaded features from optimally
selected convolutional layers are concatenated to provide more
comprehensive representation for the recognition. To take full
advantage of these cascaded features, the ensemble learning-
based classifier, namely, the AdaBoost rotation forest (RoF),
is introduced to replace the original softmax layer to realize a
more accurate limited sample recognition. Through the AdaBoost
RoF method, not only are these features further enhanced by
the rotation matrix but also a strong classifier is constructed
by several weak classifiers with different adjusted weights. The
experimental results on MSTAR data set show that the cascaded
features and ensemble weak classifiers can fully exploit effective
information in limited samples. Compared with the existing
CNN method, the proposed method can improve the recognition
accuracy by about 20% under the condition of ten training
samples per class.
Index Terms— AdaBoost, convolutional neural network (CNN),
ensemble learning, rotation forest (RoF), synthetic aperture
radar (SAR), target classification.
I. INTRODUCTION
S
YNTHETIC aperture radar (SAR), as an active imaging
sensor, has the advantages of collecting all-weather, day-
and-night high-resolution images with multiple flexible imag-
ing modes [1]. SAR automatic target recognition (ATR) is
one of the important contents of remote sensing information
extraction, and it has been playing an important role in Earth
Manuscript received July 19, 2019; revised August 23, 2019; accepted
August 31, 2019. This work was supported in part by the National Natural
Science Foundation of China under Grant 61871413 and Grant 61571422.
(Corresponding author: Yongsheng Zhou.)
The authors are with the College of Information Science and Technology,
Beijing University of Chemical Technology, Beijing 100029, China (e-mail:
zhyosh@mail.buct.edu.cn).
Color versions of one or more of the figures in this letter are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LGRS.2019.2939156
system monitoring and other applications [2]–[4]. Effective
feature extraction method can not only improve the accuracy
of target recognition but also can reduce the requirement of
the sample amount. Since SAR target characteristics do not
conform to human vision system, target feature extraction has
always been a hot and hard topic in ATR community. With the
development of deep learning in recent years, the high-level
semantic feature extraction based on the convolutional neural
network (CNN) opened up a new era for high-precision SAR
target recognition [5]–[10].
In the case of limited samples, feature augmentation proves
to be one of the effective ways to solve the small sample issue.
As a straightforward feature augmentation, the combination of
the features extracted from different CNN layers will improve
the application performance. The features from different layers
are combined to realize a highly accurate road extraction,
which outperforms the existing neural network methods that
only consider the high-level features [11]. The complementary
advantage of different layers is explored to achieve a high-
quality image retrieval [12]. The multilevel feature-based
segmentation neural network is utilized to achieve the high
accurate cloud and cloud shadow detection [13]. The feature
combination could improve the recognition performance, but
how to combine different layers and which layers should be
combined are also worthy of being further studied.
In order to further improve the discriminative performance
of CNN for target recognition, the softmax layer can be
replaced with other classifiers, e.g., support vector machine
(SVM), random forest (RF), Hash forest, and so on [14]–
[16]. It can be seen that the replacement of the softmax
layer with a supervised classifier will improve the recognition
performance of CNN. However, these supervised classifiers
will require more training samples than the softmax classifier.
In order to balance the precision and sample amount, the linear
combination of many weak classifiers, namely, the ensemble
learning, maybe a compromise solution for the limited sample
learning case. The ensemble learning-based target recognition
methods are widely studied [17]–[20]. The combination of RF
and AdaBoost algorithm has achieved good results in remote
sensing image classification [18]. In addition, the rotation
forest (RoF)-based method enhances the diversity of features
by the feature transformation and feature reduction in SAR
image classification [19], [21]. The RoF is further combined
with the AdaBoost algorithm to reduce the deviation of a
1545-598X © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.