031002-1 CHINESE OPTICS LETTERS / Vol. 9, No. 3 / March 10, 2011
Hyperspectral remote sensing image classification based on
decision level fusion
Peijun Du (杜杜杜培培培军军军)
1∗
, Wei Zhang (张张张 伟伟伟)
2∗∗
, and Junshi Xia (夏夏夏俊俊俊士士士)
1∗∗∗
1
Key Lab oratory for Land Environment and Disaster Monitoring of State Bureau of Surveying and Mapping of China,
China University of Mining and Technology, Xuzhou 221116, China
2
Heb ei Bureau of Surveying and Mapping, Shijiazhuang 050031, China
∗
Corresp onding author: dupjrs@cumt.edu.cn;
∗∗
corresp onding author: cumtwzh@163.com;
∗∗∗
corresp onding author: xiajunshi@126.com
Received Septemb er 9, 2010; accepted November 6, 2010; posted online February 21, 2011
To apply decision level fusion to hyp erspectral remote sensing (HRS) image classification, three decision
level fusion strategies are experimented on and compared, namely, linear consensus algorithm, improved
evidence theory, and the proposed support vector machine (SVM) combiner. To evaluate the effects of the
input features on classification p erformance, four schemes are used to organize input features for member
classifiers. In the experiment, by using the operational mo dular imaging spectrometer (OMIS) II HRS
image, the decision level fusion is shown as an effective way for improving the classification accuracy of the
HRS image, and the proposed SVM combiner is especially suitable for decision level fusion. The results
also indicate that the optimization of input features can improve the classification p erformance.
OCIS co des: 100.4145, 280.4788.
doi: 10.3788/COL201109.031002.
Hyperspectral remote sensing (HRS) is viewed as one of
the most evolving and most promising technologies for
advanced earth observations in the 21st century. Tradi-
tionally, HRS image classification is implemented by a
single classifier with the original hyperspectral data and
other derived features as input. Examples of classifiers
are the support vector machine (SVM), the maximum
likeliho od classifier (MLC), and the back-propagation
neural network (BPNN). These methods have proven
their effectiveness in many applications, however, there
are still some problems. Firstly, each classifier has its own
merits and limitations, and achieving the desired accu-
racy using a single classifier is often difficult
[1,2]
. Sec-
ondly, the adjacent wavebands of HRS data are highly
correlated, thus the simultaneous use of all bands cannot
assure high accuracy. Due to the limitations of both the
classifiers and data, finding new solutions to improve the
classification performance is necessary. Decision level fu-
sion, using a specific criterion or algorithm to integrate
the results of different classifiers, has shown great benefits
in improving the classification accuracy of multi-source
remote sensing images
[3,4]
. After a survey of HRS classi-
fication techniques and decision level fusion algorithms,
some issues on HRS image classification based on deci-
sion level fusion are explored in this letter.
Many decision level fusion algorithms have been
develop ed. After comparing their suitability and per-
formance for remote sensing image classification, we se-
lected three fusion strategies for this study: the improved
evidence theory, the linear consensus, and the SVM com-
biner.
Evidence theory is also known as the Dempster-Shafer
(D-S) evidence theory, which was first applied by Demp-
ster and then developed by Shafer. Compared with the
Bayesian theory, the D-S evidence theory assigns proba-
bility to sets and is able to handle the uncertainty caused
by unknown factors
[5]
. The D-S evidence theory uses the
discrimination framework, the confidence function, the
likeliho od function, and the probability allocation func-
tion to represent and process information. Supposing
that Θ = {C
1
, C
2
, · · · , C
i
, · · · , C
M
} is the discrimination
framework and M is the number of classes, the basic
probability allocation function m is a function from 2
Θ
to [0, 1] meeting the requirements of
m(φ) = 0
A⊆Θ
m(A) = 1.
(1)
If there are two or more different evidences, the or-
thogonal sum can be used to combine those evidences.
Assuming that Z
1
, Z
2
, · · · , and Z
n
are the probability
allocation functions corresponding to evidences F
1
, F
2
,
· · · , and F
n
, the orthogonal sum Z = Z
1
⊕ Z
2
⊕ · · · ⊕
Z
n
is
Z(φ) = 0, (2)
Z(A) = K
−1
×
∩A
i
1≤i≤n
Z
i
(A
i
) , (3)
K =
∩A
i
6=φ
1≤i≤n
Z
i
(A
i
). (4)
When various evidences are inconsistent or contra-
dictory to each other, the combined result of the D-S
evidence theory may be unreasonable. A mo dified ev-
idence combination algorithm was proposed and tested
by Sun et al.
[6]
, which proved superior to the traditional
method in processing evidence that were contradicting
and highly inconsistent. For a remote sensing image,
different classifiers may generate different classification
labels, resulting in generation of evidence with high con-
tradiction, so the modified evidence combination is ap-
plied to the classification integration of HRS images. The
1671-7694/2011/031002(4)
c
° 2011 Chinese Optics Letters