PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2014
239
Ensemble Learning with Multiple
Classifiers and Polarimetric Features for
Polarized SAR Image Classification
Alim. Samat, Peijun Du, Muhammad Hasan Ali Baig, Sumit Chakravarty, and Liang Cheng
Abstract
Polarimetric
SAR
(
PolSAR
) image processing has become a hot
research topic in
SAR
remote sensing field in recent years.
However, due to the complexity of the image and limited
availability of advanced techniques,
PolSAR
image process-
ing is still a challenging issue. In this paper the suggestion
of ensemble learning (
EL
) is introduced into
PolSAR
image
classification by integrating various polarimetric features
and multiple classifiers. The most popular ensemble learning
methods, including Bagging, AdaBoost, and Rotation Forest
are adopted to combine multiple classifiers and polarimetric
features. The proposed classification scheme is evaluated on
three real
PolSAR
data. Experimental results shows that the
covariance and coherence features can give better perfor-
mance than other polarimetric decomposition features, and
complementary between different polarimetric decompo-
sition features improving the classification performance.
Although a weak classifier gives unsatisfactory classification
accuracy on polarimetric decomposition features, the per-
formance can be highly improved by using
EL
strategies.
Introduction
Land cover classification is one of the primary objectives of
remotely sensed image processing, analysis, and practical ap-
plication. Synthetic Aperture Radar (
SAR
) opens new opportu-
nities for land cover classification owing to its day and night,
all-weather, and high-resolution observation capability to the
Earth’s surface. But due to the low information content of an
individual
SAR
image, single-polarization or single-band
SAR
data cannot provide highly accurate land cover classification
maps (Herold
et al
., 2004; Lee
et al
., 2002). On the contrary,
PolSAR
images contain a large amount of potential information
which is directly related to physical properties of natural me-
dia and backscattering mechanism. Therefore, full polarized
SAR
imagery classification has become a popular and challeng-
ing problem in recent years due to its potential applicability
in different fields. With the goal of achieving high classifica-
tion accuracy, many supervised, unsupervised, and semi-su-
pervised classification methods have been developed. Usually,
the adopted features in classification include scattering matrix
elements, covariance matrix elements and coherent matrix
elements, and polarimetric decomposition features (Cloude
et al
., 1997; Du
et al
., 1996; Krogager, m 1994; Qi
et al
., 2012;
Song
et al
., 2007; Wang
et al
., 2013).
Generally, developing advanced classifiers, or optimiz-
ing the input features or adopting novel pattern recognition
methods are the three common ways get better classification
performance. It is worth noting that, finding the optimal fea-
tures and best classifier is always hard, maybe even impossi-
ble. However, a better classification result may be achieved by
combining multiple classifiers and various features by using
ensemble learning (
EL
) strategies.
Since “strong and weak learn ability are equivalent” was
proofed by Robert E. Schapire (Schapire, 1990), more options
are available to get an improved classification performance
through
EL
, which can integrate the benefits of different fea-
tures, classifiers, and combination methods. As an advanced
machine learning framework,
EL
is able to boost weak learn-
ers, working slightly or obviously better than random guess-
ing to strong learners, thus making accurate predictions.
In spite of the considerable amount of work on the use of
EL
for remote sensing image classification in the past years,
few applications of
EL
to
PolSAR
image classification could be
found in literature (Jiong
et al
., 2008; Loosvelt
et al.
, 2012;
Min
et al
., 2009; She
et al
.,2007). In this paper,
EL
tech-
niques such as Bagging, AdaBoost, and Rotation Forest are
introduced to
PolSAR
image classification based on different
polarimetric features. To evaluate the performance of vari-
ous
EL
methods, some benchmark classifiers such as support
vector machine (
SVM
), logistics decision tree, simple logistic
decision tree, and k-Nearest Neighbor are experimented and
compared. Besides, different ensemble strategies for classifier
combination such as voting, stacking generation, and multi-
Alim.Samat and Peijun are with the Key Laboratory for Satel-
lite Mapping Technology and Applications of State Administra-
tion of Surveying, Mapping and Geoinformation of China, Nan-
jing University, Kunshan Building, Nanjing University Xianlin
Campus, No.163, Xianlin Dadao, Qixia District, Nanjing City,
Jiangsu Province, China 210023 (
).
Muhammad Hasan Ali Baig is with the Institute of Remote
Sensing Applications, Chinese Academy of Sciences, Hyper-
spectral Remote Sensing Laboratory (HRS), No.20, Dutun Lu,
Chaoyang District, Beijing, China, 100101.
Sumit Chakravarty is with the Department of Engineering and
Computer Science, New York Institute of Technology, 903
Tree Creek Parkway, Lawrenceville, GA, 30043.
Liang Cheng is with the Key Laboratory for Satellite Map-
ping Technology and Applications of State Administration
of Surveying, Mapping and Geoinformation of China, Kun-
shan Building, Nanjing University Xianlin Campus, No.163,
Xianlin Dadao, Qixia District, Nanjing City, Jiangsu Province,
China 210023.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 3, March 2014, pp. 239–251.
0099-1112/14/8003–239
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.3.239
