Escolar Documentos
Profissional Documentos
Cultura Documentos
ABSTRACT
Remote sensing image classification is an important and complex problem. Conventional remote sensing image
classification methods are mostly based on Bayesian subjective probability theory, but there are many defects for its
uncertainty. This paper firstly introduces evidence theory and decision tree method. Then it emphatically introduces the
function of support degree that evidence theory is used on pattern recognition. Combining the D-S evidence theory with
the decision tree algorithm, a D-S evidence theory decision tree method is proposed, where the support degree function is
the tie. The method is used to classify the classes, such as water, urban land and green land with the exclusive spectral
feature parameters as input values, and produce three classification images of support degree. Then proper threshold
value is chosen and according image is handled with the method of binarization. Then overlay handling is done with
these images according to the type of classifications, finally the initial result is obtained. Then further accuracy
assessment will be done. If initial classification accuracy is unfit for the requirement, reclassification for images with
support degree of less than threshold is conducted until final classification meets the accuracy requirements. Compared
to Bayesian classification, main advantages of this method are that it can perform reclassification and reach a very high
accuracy. This method is finally used to classify the land use of Yantai Economic and Technological Development Zone
to four classes such as urban land, green land and water, and effectively support the classification.
Keywords: evidence theory, decision tree, support degree, remote sensing classification
1. INTRODUCTION
The classification technique of remote sensing images is a branch of pattern recognition techniques in remote sensing
field. It aims to the identification of remote sensing images, i.e. recognizing and classifying ground cover information in
remote sensing images thereby distinguishing the corresponding ground truth and extracting the required information [1-
2]. The classification of remote sensing data is important. The uncertainty of remote sensing data is that the value of the
attribute has a confidence level, which comes from the acquirement, transmission, storage of remote sensing data.
Dempster-Shafer evidence theory (D-S evidence theory) [3] is the extension of probability, which constructs the one-to-
one relationship between proposition and aggregation. D-S evidence theory is an uncertainty theory through
transforming the uncertainty of proposition to the uncertainty of aggregation. D-S evidence theory is applied in the
description, processing and deduction of uncertain, incomplete, unreliable data or information in recent years [4-6].
Classification is an important task of data mining, which is to construct models to classify the data into different classes.
Decision tree classifier [7-8] is a supervised classification method, which is nonparametric and does not need the data in
normal distribution. It depends on the classification rules, which can learn from classification process or predefinition, to
classify the data. There are many decision tree algorithms such as ID3, C4.5, CART, etc., which are effective and widely
used in classification field, but they could not deal with uncertain data in the construction and classification of the
decision trees.
As for the limitation of traditional decision tree algorithms, a D-S evidence theory decision tree method is proposed,
which combines the D-S evidence theory with decision tree classifier. The method can deal with the uncertainty of
Multispectral, Hyperspectral, and Ultraspectral Remote Sensing Technology, Techniques, and Applications III,
edited by Allen M. Larar, Hyo-Sang Chung, Makoto Suzuki, Proc. of SPIE Vol. 7857, 78570Y 2010 SPIE
CCC code: 0277-786X/10/$18 doi: 10.1117/12.869544
Downloaded from SPIE Digital Library on 20 Apr 2011 to 159.226.100.156. Terms of Use: http://spiedl.org/terms
remote sensing imageries. To utilize evidence theory to decision tree algorithm, support degree is proposed, which is
classification rules of the decision tree. When compared with statistical classification methods, the decision tree method
using support degree shows great superiority. Experimental results demonstrate the proposed method is effective and can
improve the classification accuracy.
m( ) = 0 , m( A) = 1
A D
(1)
Pls ( ) = 0 ,
Pls ( A) =
B I A
m( B) , A D, A (3)
Pls( D) = 1 , (5)
D-S theory evidence provides an explicit measure of ignorance about an event A and its complementary A as the length
of the internal [ Bel ( A) , Pls ( A) ] (called belief internal). It can also be interpreted as the imprecision on the true
probability of A . The mass assigned to D can be interpreted as the global ignorance since this weight of evidence is not
discernible among the hypotheses. In summary, as for probability theory, using numerical values in [0, 1] allows us to
represent uncertainty, but using the two functions Bel and Pls , D-S evidence is also able to represent imprecision.
Downloaded from SPIE Digital Library on 20 Apr 2011 to 159.226.100.156. Terms of Use: http://spiedl.org/terms
If masses are assigned only to simple hypotheses ( m( A) = 0 for | A |> 1 ), then the three functions m , Bel and Pls are
equal and are a probability, called Bayesian mass function. Otherwise, there is no direct equivalence with probabilities.
x = ( x1 , x2 ,..., xm )T , y = ( y1 , y2 ,... ym )T .
Euclidean distance || x y || presents the similarity of x and y :
m
|| x y ||= (x y )
i =1
i i
2
.
If || x y || is smaller, the difference between x and y in every feature is smaller. Otherwise, the difference is bigger.
A pixel x belongs to class A , which means x is more nearer to the average vector of class A center.
Downloaded from SPIE Digital Library on 20 Apr 2011 to 159.226.100.156. Terms of Use: http://spiedl.org/terms
1 m j
yj = xi , ( j = 1, 2,...m) ,
n i =1
Average vector y = ( y , y ,... y ) of class A is constructed. The pixel x belongs to class A , which means the distance
1 2 m
Suppose i represent x Ai (i = 1, 2,...n) , so let D = {1 , 2 ,... n } be a recognition frame. Plausibility function ( Pls )
is deducted on the recognition frame.
C
Pls ({i }) = , (i = 1, 2,...n) and C is a constant.
|| x Ai ||
C min || x Ai ||
Pls( A) = max( Pls({i })) = max = i(1,..,n ) , A D
i A i A || x Ai || min || x Ai ||
i A
So we can get the support degree function S ( A) on the recognition frame D in S-D evidence theory:
min || x Ai ||
S ( A) = 1 Pls ( A) = 1 i(1,..., n ) .
min || x Ai ||
i A
The bigger the value of S ( A) is, the more similar x should belong to class Ai . So the support degree function is the rule
of the classification.
Downloaded from SPIE Digital Library on 20 Apr 2011 to 159.226.100.156. Terms of Use: http://spiedl.org/terms
(4) Execute the decision tree to classify the data, and produce the classification image of one ground object based on
different support degree.
(5) Produce the other round object classification images according to steps from (2) to (4).
(6) Choose proper threshold to produce binary images of different ground object classification. The pixels whose support
degree is less than the threshold are assigned to 0, while others are assigned to 1.
(7) Overlay the binarization images of ground object classification.
(8) Appraise the classification accuracy of the final overlaid image. If the accuracy is lower than need, go to (6);
otherwise, classification is finished.
4. APPLICATION EXPERIMENT
The experiment chooses 2006s Landsat 5 / Tm images of Yantai Economic and Technological Development Zone. The
D-S evidence theory decision tree method is used to classify the land cover of Yantai Economic and Technological
Development Zone to four classes such as urban land, farmland, forest land and water, and effectively support the
classification. The following are the steps:
(1) Select TM 5-4-3 spectral bands, do geometry correction, and subset the images of Yantai Economic and
Technological Development Zone.
(2) Choose the common and representative data as the training samples. Classification accuracy depends on the quality
and quantity of the samples.
(3) Calculate the maximum, minimum and average values of the interesting spectral bands. These values can be used to
calculate support degree of the ground object classification.
(4) Construct decision tree with different support degree according to the three classes.
(5) Choose zero as the threshold to produce binary images of different ground object classification, execute the decision
tree algorithm to classify the data, and produce three ground object classification images (figure 1(a)-(c)).
(6) Overlay the binarization images of ground object classification, and produce the classification image of the three
ground objects (figure 2(a)).
Downloaded from SPIE Digital Library on 20 Apr 2011 to 159.226.100.156. Terms of Use: http://spiedl.org/terms
Figure 1. Classification images of support degree, (a) water, (b) urban and (c) green land.
Figure 2. The result of image classification based on evidence theory, (a) the first classification and (b) the second
classification.
Downloaded from SPIE Digital Library on 20 Apr 2011 to 159.226.100.156. Terms of Use: http://spiedl.org/terms
5. APPRAISAL OF CLASSIFICATION ACCURACY
Compared to Bayesian classification, main advantages of this method are that it can perform reclassification and reach a
very high accuracy. From figure 1(a)-(c), three classification images based on D-S evidence theory decision tree method
is showed. Through the comparison with original remote sensing images, the accuracy of water (figure 1(a)) is high,
while the accuracy of the other two classification result is low. Binarization operations with the three classification
results are done. The pixels that support degree is zero belong to one class, and the others belong to another class. Four
classes such as water, urban land, and green land is numbered to 1, 2 and 3. Then the three binarization images are
overlaid to one result image (figure 2(a)). We randomly choose 320 points from figure 2(a) and compare with the
reference and original images, so we get the classification error matrix and the accuracy assessment report (table 1).
Table 1. Classification error matrix and the accuracy assessment report.
Class Water Urban Green land Number of samples Classification
accuracy
Water 69 1 5 75 0.9200
Urban 1 71 23 95 0.7474
Green land 3 46 101 150 0. 6733
Total number of samples: 320, correct classified samples: 241, and overall classification accuracy: 0.7531.
From table 1 the total appraisal result of the accuracy is 0.7531, which is similar to the result of six times classification of
the maximum likelihood classification method (the result is 0.7312). Because the accuracy of green land is lower, we can
adjust the threshold to reclassify until get the ideal accuracy. Because the area of green land is large, we choose the
support degree of the three cover classes: urban land more than 0.4, water more than 0.5, green land more than 0.3 and
overlay the three images second times. The figure 2(b) is the result. We randomly choose 320 points from figure 2(b) and
compare with the reference and original images. The total classification accuracy reaches 0.9023 and the accuracy of the
three classes is 0.9600, 0.8532, and 0.9233. The classification accuracy meets our demands.
6. CONCLUSION
Remote sensing data has uncertainty and plausibility result from the data acquirement, transmission, storing, handling
etc.. D-S evidence theory is a powerful tool that can be applied to express the uncertainty of the data. Decision tree is a
classification algorithm, which is a non-parametric, multi layer method, free of data distribution hypothesis, decision tree
is more robust and flexible for data analysis and interpretation in the application. Its time complexity is low and has fast
classification speed. Combining the D-S evidence theory with the decision tree algorithm, a D-S evidence theory
decision tree method is proposed, where the support degree function is the tie. The method is used to classify the classes,
such as water, urban land and green land with the exclusive spectral feature parameters as input values, and produce
three classification images of support degree. Then the proper threshold value of support degree is chosen to each
classification image and binarization handling is executed. Then overlay these images according to the type of
classifications, and the initial result is obtained. Finally further accuracy assessment will be done. If initial classification
accuracy is unfit for the requirement, reclassification is conducted through re-choosing the support degree threshold of
the images of every ground object classification, until final classification meets the accuracy requirements. Compared to
Bayesian classification, main advantages of this method are that it can perform reclassification and reach a very high
accuracy. This method is successfully used to classify the land cover of Yantai Economic and Technological
Development Zone to three classes such as water, urban land and green land. The experiment effectively supports the
classification method and has precise classification result.
Downloaded from SPIE Digital Library on 20 Apr 2011 to 159.226.100.156. Terms of Use: http://spiedl.org/terms
REFERENCES
Downloaded from SPIE Digital Library on 20 Apr 2011 to 159.226.100.156. Terms of Use: http://spiedl.org/terms