Professional Documents
Culture Documents
29 3B Paper2 Fizazi Hanane Amir
29 3B Paper2 Fizazi Hanane Amir
29 3B Paper2 Fizazi Hanane Amir
S ; i=1,2,,l.
) ]( 1 , 0 [
i i i i
b B U b s + =
(1)
U [0,1]: a random number drawn uniformly in the interval [0,1].
The second operator (equation (2)) generates a point s' in the neighborhood of s taking into account an A
amplitude. The amplitude A equal to A
site
in case of creation of hunting site and A
locale
in case of exploration of a
hunting site.
s
'
=(s
i
'
)
S ; i=1,2,,l.
) ]( 5 . 0 , 5 . 0 [
'
i i i
i b B AU b s + =
(2)
U [0.5,0.5]: random number drawn uniformly in the interval [0.5,0.5].
It should be made sure that s
i
'
[B
i
-b
i
]. The second operator also has a uniform behavior in the interval
[s
i
-0.5A(B
i
-b
i
), s
i
+0.5A(B
i
-b
i
)].
Figure 2: Exploring the hunting site s
1
respecting an A
locale
amplitude.
2.2 API Algorithm
Nest posed in search space
T0 ;
While (stop condition is not checked)
Do
For ( any ant of the population) do
While (Patience< P
locale
)
create sites of hunting
explore the hunting sites
end while
end for
if (nest must be moved) then
move the nest
end if
T T +1
End while
Return the position of the nest
3. Coding and objective function
Several steps are necessary to use the API method for satellite image classification: coding of ants, definition
of the objective function and the application of exploration operators.
3.1 Objective function
An ant population is encoded as strings of representatives of classes (C
1
, C
2
, ..., C
n
) representing a problem
solution (Figure 3). The coding can be done by real or integer numbers. The number of classes is fixed a priori,
therefore the size of ant is known. In this research, an ant is encoded by positive real numbers. Consider the
following case as an example: Suppose a satellite image of three bands, the configuration of the ant i is as
follows:
Figure 3: Configuration of the ant i.
Ants of the population moving in the search space using an objective function measuring the fitness of each
ant. The aim of this population is to find the best representatives of classes for a better classification.
The Euclidean distance (see equation (3)) between learning samples and representatives of classes has been
used as objective function in various researches:
( ) ( )
=
=
n
i
i i
y x y x D
1
2
,
(3)
Several other classification validity indices have been developed to determine an optimal classification, for
example the Separation Index, "SI", the Davies-Bouldin "DB" index and the Xie-Beni Index, "XBI" [27]. The
DB index is widely used in remote sensing image classification [28], [29]. The DB index can be calculated as
follows:
( ) ( )
=
=
<
non si
k C x D
j i
n j
ki
; 0
, min arg ; 1
1
(4)
Where
x
i
: The gray level of pixel i.
M
k
: Number of elements in class k.
ki
: Membership function of pixel i to class k.
X
k
: All elements of class k.
Then the average (v) and standard deviation (S) of each class are calculated as follows:
( )
( )
k
X x
i
M
i
ki
M
i
i ki
k
M
x
x
v
k i
k
k
=
=
= =
1
1
(5)
2 / 1
2 1
|
|
\
|
=
k
X x
k
k
k
v x
M
S
(6)
Now calculate the distance between the averages classes with equation (7):
t
j k kj
v v d =
(7)
d
kj
is the Minkowski distance of order t between the k
th
center and the i
th
center. We set t=2.
Then the R
k
value of k
th
center is calculated by equation (8).
+
=
kj
j k
k j j
k
d
S S
R
,
max
(8)
The DB value is defined as the average R of all classes (equation (9)).
=
=
n
k
k
R
n
DB
1
1
(9)
To arrive at a good classification of an image, it is necessary to minimize the DB value. A minimum value of
DB corresponds to classification with the smallest inner-class scatter and the largest class separation. In our
work the equation (4) is not used, because the classes of the training samples are known.
In the following example we explain how we used the DB index and the Euclidean distance as an objective
function for the supervised satellite images classification. Let A
1,
A
2,
A
3
a partition of A as:
A
1
= (1, 2, 3, 4), A
2
= (5, 6, 7, 8), A
3
= (9, 10, 11, 12).
We add an element to each class and subsequently calculate the DB value. This manipulation is summarized
in Table 1.
Test Class 1 Class 2 Class 3 DB value
1 A1 {3.5} A2 {8} A3 {12.5} 0.5802
2 A1 {3} A2 {7.5} A3 {12} 0.5352
3 A1 {2.5} A2 {7} A3 {11.5} 0.5051
4 A1 {2} A2 {6.5} A3 {11} 0.4926
5 A1 {1.5} A2 {6} A3 {10.5} 0.5051
Table 1: search of representatives of classes using DB value.
In Table 1 we see that the minimum value of DB is obtained when the elements 2.5, 4.5 and 6.5 are added to
classes A
1,
A
2
and A
3
respectively. It is also noted that the DB value decreases as the added elements are close to
the additions of test 4.
The minimal value of DB is acquired when not only the elements added represents well the classes but also
when they are far between them.
We propose to consider the added elements in Table 1 as representatives (C
1
, C
2
, C
3
) of the classes A
1
, A
2
, A
3
then we calculate the Euclidean distance (ED) between the representatives of the classes and the elements of
each class. That is summarized in Table 2.
Test C
1
C
2
C
3
ED value
1 4 6 8 6
2 3.5 5.5 7.5 5
3 3 5 7 4
4 2.5 4.5 6.5 4
5 2 4 6 4
Table 2: search of representatives classes using ED value.
The minimum value of ED in Table 2 is repeated in several tests. Tests 3 and 5 dont represent the true
representatives of the classes but have the same value as the test 4. We can say that the results are not
guaranteed when ED is use as objective function.
3.2 Classification
The classification of objects using the Euclidean distance still has limitations because a given object is not
necessarily assigned to the appropriate class. Figure 4 show a real example of our problem, where the element to
be classified will be assigned to the small class because it is closer to the point representing of this class. But it is
more appropriate to be assigned to the large class.
Figure 4: problem of classification
In this work we apply both ED and DB index for the classification stage. For the ED function the principle is
simple it is enough to affect the pixel want to be classified to the class where the representative is closest. For the
DB function, we add the pixel wants to classify on the training samples of first class and, we add in the other
class samples training the optimal centers found during the training and calculate thereafter the DB value. We
repeat this procedure for each class. The pixel will be assigned to the class where the DB value is minimal. Our
principle of classification is summarized in Table 3.
Suppose that we have three training samples (E
1,
E
2,
E
3
) of three classes. E
1
= (1, 2, 3, 4), E
2
= (5.5, 6, 6,
6.5), E
3
= (9, 10, 11, 12)
The representatives of classes found during the training stage using the DB value are: 2, 6, 11, for three
classes respectively. Our problem is: classification of the element 8.
DB value
E
1
{8} E
2
{6} E
3
{11} 0.9227
E
1
{2} E
2
{8} E
3
{11} 0.4625
E
1
{2} E
2
{6} E
3
{8} 0.4121
Table 3: classification using DB value
In the Table 3 the minimum DB value is obtained when we added the element 8 in sample 3, so this element
will be assigned to class 3. If we use in the ED function, the element 8 will be assigned to Class 2 because it is
closer to the point representing this class.
4. Experimental results
This work has been implemented using the C++ Builder 6 programming language. To test our approach, we
used an image of the Oran region (West Algeria) of 10241024, acquired by the LANDSAT satellite in 1993.
This image has been chosen for the diversity of themes within it. The bands used are: red, infrared and blue.
In our application we used a portion of the study area of size 12088 (area shown in Figure 5). This latter is
made up of 4 topics: sea, undertow, forest and sand.
In the goal to make a comparison between the Euclidean distance and Davies-Bouldin index, we use them as
objective functions in algorithm API. We note API-ED when we use the first function and API-DBI when we use
the function proposed.
Figure 5: Study Area.
The algorithm API has several parameters must be determined a priori. After several tests, the parameters are
fixed as follows: n=18, p=60, A
site
=100, A
locale
=20, P
locale
=10, T=5 ;
Figure 6: Classification of the study area by: (a) API-ED. (b) API-DBI.
From Figure 6 we see that both approaches have different classification results. The API-ED method could
distinguish the 4 classes of the study area, but there are pixels in the forest class (Figure 6 (a)) misclassified.
Mthode Mer Ressac Sable Foret
Mer 100 17.5 0 0
Ressac 0 82.5 0 0
Sable 0 0 95 2.5
API-ED
Foret 0 0 5 97.5
Mer 100 7.5 0 0
Ressac 0 92.5 0 0
Sable 0 0 100 2.752
API-DBI
Foret 0 0 0 97.248
Table 4: Confusion matrix of the classification
The API-ED approach has produced confusion because pixels in the forest class were classified as pixel of sea
class (circle in Figure 6 (a)). Visually API-DBI method has a better classification. In Figure 6 (b) all classes are
not confused.
To compare quantitatively the results of the two methods, we examine the Overall Accuracy (OA) and the
Kappa coefficient.
Table 4 shows the confusion matrix of API-DBI and API-ED approaches. The API-ED method produced
confusion between the sea class and the undertow class, because the pixels belonging to the sea class have been
classified as pixels of the undertow class. This approach has affected also pixels of the sand class to the forest
class.
API-ED API-DBI
OA 93.125% 97.437%
Kappa Coefficient 0.9159 0.9655
Table 5: Performance of the classification.
The API-DBI approach could make a good classification except a little confusion between the sea and the
undertow classes and the forest class with sand class.
Table 5 shows that API-DBI method produces the best OA, that is to say, the best percentage of pixels well
classified. We note an increase in OA from 93,125% for API-ED to 97,437% for API-DBI so an improvement
of 4.312%.
It is recognized that the OA is not sufficient to measure the performance of a classification technique [30]. To
measure this performance, we also used the Kappa coefficient. In Table 5, we note that the Kappa coefficient
increased from 0.9159 for API-ED method to 0.9655 for API-DBI method thus an improvement of 0.0496.
4. Conclusion
Thus, we used the behavior modeling of the ant Pachycondela apicalys for satellite images classification using
an objective function called Davies-Bouldin index.
To evaluate the performance of this function we have compared it with the Euclidean distance function. By
examining either the OA or the Kappa coefficient the experimental results show that the DB index function
could give good performances. We noted a large increase in the OA and also an improvement on Kappa
coefficient.
The large number of parameters and the difficulty of their choices stand as major obstacles for the use of the
API algorithm. As perspective, would not set the parameters of the ant in time but to make dynamic.
3. References
[1] Richards J., Remote Sensing Digital Image Analysis, Springer-Verlage, Berlin, 1994.
[2] Lukin, V., Ponomarenko, N., Kurekin, A., Lever, K., Pogrebnyak, O., Fernandez, L.P.S. : "Approaches to
Classification of Multichannel Images". J.F. Martnez-Trinidad et al. (Eds.): CIARP 2006, LNCS 4225, pp. 794
803, 2006.
[3] J. Ediriwickrema, S. Khorram, Hierarchical maximum-likelihood for imroved accuracies, IEEE
Transaction on Geoscience and Remote Sensing, Vol.35, No.4, pp.810-816, Jul. 1997.
[4] Christian Darken, and John Moody, Fast adaptive K-means clustering: some empirical results, Proceedings
of International Joint Conference on Neural Networks, San Diego, pp.233-238, Jun. 1990.
[5] Kaichang Di, Deren Li, and Deyi Li, Remote sensing image classification method study on spatial data
mining, Journal Wuhan Technical University of Surveying and Mapping, Vol. 25, No. 1, pp.43-47, Feb. 2000.
[6] S. Chitoub, A. Houacine and B. Sansal, Principal component analysis of multispectral image using neural
network, IEEE International Conference on Computer System and Application, Beirut, pp.89-95, Jun. 2001.
[7]: Izabatene, F.H., Hannane, A.M.: "Optimisation de rseaux de neurones par algorithmes de colonies de
fourmis pour la segmentation des images satellitaires". 4
me
symposium international IMAGE, pp 124-129,
Guelma, Algrie, 2008.
[8]: Amghar, Y.T., Fizazi, H. : "Optimisation du rseau RBF par un algorithme neuronal". Proceeding ISOR,
2008.
[9] Iftene, T., Mahi, H., Isabaten, H., "Approche de classification par rseaux de Kohonen pour tablir des cartes
doccupation du sol de la rgion doran (Algrie)". Tldtection, vol. 3, n 2-3-4, pp 361372, 2003.
[10] : Bandyopadhyay, S., Pal, S.K. : "Pixel Classification Using Variable String Genetic Algorithms with
Chromosome Differentiation". IEEE transaction on geoscience and remote sensing, vol. 39, n. 2, pp 303-308.
February 2001.
[11] Yanfei Zhong, Liangpei Zhang, Jianya Gong, and Pingxiang Li, A Supervised Artificial Immune Classifier
for Remote-Sensing Imagery. IEEE Transactions on Geoscience and Remote Sensing, Vol. 45, No. 12, Dec
2007.
[12] M. Dorigo, V. Maniezzo and A. Clorni, The ant system: optimization by a colony of cooperating agents,
IEEE Transaction on Systems Man and Cybernetics, Vol.26, No. 1, pp.1-26, Feb.1996.
[13] Qin Dai and Jianbo Liu, Application of Ant colony optimization algorithm to remote sensing image
classification, Proceeding of SPIE, Vol. 6788, pp.67881A1-6, Nov. 2007.
[14] Xiaoping, L., Xia, L., Lin, L., Jinqiang, H., Bin, A., "An Innovative Method to Classify Remote-Sensing
Images Using Ant Colony Optimization". IEEE transaction on geoscience and remote sensing, vol. 46, n. 12, pp
4198-4208. December 2008.
[15] S. Le Hegarat-Mascle, A. Kallel and X. Descombes, Application of ant colony optimization to image
classification using a Markov model with non-stationary neighborhoods, Image and Signal Processing for
Remote Sensing XI, Proceeding of the SPIE, Brugge, Vol 5982, pp.59820C1-10 Sep. 2005.
[16]: Khedam, R., Belhadj-Aissa, A. : "Classification dimages satellitaires multispectrales : de lapproche
gntique vers lapproche immunitaire". 4
me
symposium international IMAGE, pp 22-29, Guelma, Algrie,
2008.
[17] Zhang, L., Zhong, Y., Li, P. : "Applications of artificial immune systems in remote sensing image
classification". Geo-Imagery Bridging Continents XXth ISPRS Congress.p. 397 ff, Istanbul, Turkey, 2004.
[18] Yang, M.D.: "A genetic algorithm (GA) based automated classifier for remote sensing imagery". Can. J.
Remote Sensing, Vol. 33, No. 3, pp. 203213, 2007.
[19] Yang, Y., Lohmann, P., Heipke, C.: "Genetic Algorithms for the Unsupervised Classification of Satellite
Images": IntArchPhRS. Band XXXVI/3. Bonn, 2006, S. 179-184.
[20] Yang, Y. F., Lohmann, P., Heipke, C.: "Genetic algorithms for multi spectral image classification". In:
Schiewe, J., Michel, U. (Eds.): Geoinformatics paves the Highway to Digital Earth, Festschrift zum 60.
Geburtstag von Prof. M. Ehlers (2008), Nr. 8, S. 153-161.
[21] Dorigo, M. and Stutzle, T. "Ant Colony Optimization". MIT Press, 2004.
[22] Monmarch, N., Venturini, G., and Slimane, M. On how the ants Pachycondyla apicalis are suggesting a
new search algorithm. Presented at the First International Workshop on Ant Colony Optimization, Brussels,
Belgium, 1998.
[23] Picarougne, F., Recherche d'information sur Internet par algorithmes volutionnaires. Thse de doctorat.
Universit Franois Rabelais, Tours, France 2004.
[24] Aupetit, S., Contributions aux modle de Markov cach: mtaheuristiques d'apprentissage, nouveaux
modles et visualisation de dissimilarit. Thse de doctorat. Universit Franois Rabelais, Tours, France 2005.
[25] Fresneau, D., Individual foraging and path fidelity in a ponerine ant . Paris, France 1985.
[26] Fresneau, D., Biologie et comportement social dune fourmi ponrine notropicale: Pachycondyla
apicalis. Laboratoire dEthologie Exprimentale et Compare, universit de Paris XIII, France 1994.
[27] Bezdak, J.C., Pal, N.R.: "Some new indexes of cluster validity". IEEE Transactions on Systems, Man, and
Cybernetics - Part B: Cybernetics, Vol. 28, pp. 301315. 1998.
[28] Bandyopadhyay, S., and Maulik, U.: "Genetic clustering for automatic evolution of clusters and application
to image classification". Pattern Recognition, Vol. 35, No. 6, pp. 11971208. 2002.
[29] Maulik, U., Bandyopadhyay, S.: "Fuzzy partition using a real-coded variable-length genetic algorithm for
pixel classification". IEEE Transactions on Geoscience and Remote Sensing, Vol. 41, No. 5, pp. 10751081.
2003.
[30] R. G. Congalton, A review of assessing the accuracy of classifications of remotely sensed data Remote
Sens. Environ., vol. 37, no. 1, pp. 3546, Jul. 1991.