Bayesian sparse representation model
for SAR image classification
Kaiyan Dai,
a
Wentao Lyu,
a,
* Shuyun Luo,
a
and Qingjiang Shi
b
a
Zhejiang Sci-Tech University, School of Information Science and Technology,
Hangzhou, China
b
Tongji University, School of Software Engineering, Shanghai, China
Abstract. We present a Bayesian sparse representation (BSR) model for synthetic aperture
radar (SAR) image classification. It involves the design of the sub-categorization scheme for
the samples, the formulation of the BSR model, and the implementation of the classification
decision. First, the training samples of each target category are sub-categorized into multiple
subclasses. The discriminative features are then extracted from each sample within one sub-
class. These same features from all samples are gathered to construct a sub-dictionary. After
collecting all sub-dictionaries from all features, the sparse reconstructions are performed for all
features. A BSR framework is employed for such purposes. Fin all y, a fusi on strategy is applie d
to the residuals to predict the class label of the test sample. By sub-categorizing the samples
into multi-clusters, the proposed model decreases the intra-class variations between the sam-
ples and thus improve the representation ability of the features to different targets. The test
results using real field data demonstrate that t he proposed method has s uperiority to some
state-of-the-art methods.
© 2019 Society of Photo-Optical Instrumentation Engineers (SPIE)
[DOI: 10.1117/1.JRS.13.046517]
Keywords: synthetic aperture radar image classification; subcategorization; Bayesian sparse
representation; residual fusion.
Paper 190421L received Jun. 3, 2019; accepted for publication Nov. 26, 2019; published online
Dec. 11, 2019.
1 Introduction
Image classification plays an important role in addressing synthetic aperture radar (SAR) image
processing tasks. Usually, image classification includes two stages. In the training stage, mul tiple
features of the samples are first extracted from different categories. These features are then used
to train a specific classifier. In the testing stage, the features of each test sample are extracted and
then fed into this trained classifier to predict the label.
So far, many features have been presented for SAR image classification,
1–5
such as the gray-
level co-occurrence matrix (GLCM) feature,
1
the local binary pattern (LBP) featu re and its
extended types,
2
the histogram of oriented gradient (HoG) feature,
3
the visual feature,
4
the fuzzy
feature,
5
and other statistical features, just name a few. In the field of the classifier design, the k-
nearest neighbor (kNN),
6
the linear discriminant analysis,
7
the support vector machine (SVM),
8
and the neural network
9
are the representative classifiers. The combinations of these features and
classifiers have good performance based on the selected data for image classification. However,
these features may rely on some specific applications. This means that the selected features have
weak representations to the targets in the complex scenes, including inhomogeneous regions,
seriously speckled environments and weak intensity differences. In such cases, the features of the
targets cannot be effectively extracted. As a result, the representation abilities of the features to
the target degrade. Some works, accordingly, have shown that a salient kind of feature incorpo-
rated with a sound classification scheme can be expected for target classification.
Sparse representat ion (SR) is an emerging technique that represents the test samp le as a linear
combination of training samples and then determines the class of the test sample by the
*Address all correspondence to Wentao Lyu, E-mail: alvinlwt@zstu.edu.cn
1931-3195/2019/$28.00 © 2019 SPIE
Journal of Applied Remote Sensing 046517-1 Oct–Dec 2019
•
Vol. 13(4)
