Professional Documents
Culture Documents
ABSTRACT
This paper presents a neuro fuzzy clustering algorithm, Fuzzy Kohonen Local Information C-Means (FKLICM),
for classification of remote sensing images. The proposed algorithm is a hybridization of the conventional
Kohonen clustering network and Fuzzy Local Information C-Means (FLICM) to produce a much more efficient
and accurate clustering algorithm. The proposed algorithm first forms a fused image with three Multispectral
bands and pan band of Landsat 7 Enhanced Thematic Mapper Plus (ETMþ) using the Brovey transform. The
fused image is a three band image with higher resolution and better visual perception. The fused image is
reduced to a one-dimensional image using principal component analysis (PCA). The FKLICM algorithm is
applied on the PC-1 image to classify the remote sensing image into different land cover types. Integrating
the neural network with a fuzzy system combines the advantages and overcomes the limitations of both tech-
nologies. The experimental results of the proposed algorithm are compared with two other algorithms, FCM
and GIFP-FCM. The classification results and accuracy assessment show that FKLICM yields better results
than the other methods.
Keywords:
Brovey transform, Clustering, FCM, FLICM, Kohonen clustering network, PCA.
To overcome this problem another clustering algo- winning output neuron is decided by
rithm, fuzzy clustering algorithm with nonlocal adap-
tive spatial constraint (FCA-NLASC) was proposed minfd2ik g for k ¼ 1; 2; . . . ; N and i ¼ 1; 2; . . . c
[13]. FCA-NLASC has a nonlocal adaptive spatial con-
straint term, which is useful in case of noisy image seg-
mentation. A number of image segmentation Step 3. Weight update: the weight of the output neu-
techniques are discussed in [14, 15]. In this paper, a ron is updated by
fuzzy Kohonen local information c-means clustering
(FKLICM) algorithm for classification of remote sens- z1;t ¼ z1;t1 þ Y 1k;t ðxk z1;t1 Þ ð2Þ
ing image is proposed. The algorithm first fuses the
multispectral (MS) bands and pan band using the where Y is learning rate.
Brovey transform to obtain a higher-resolution image.
The fused image contains three bands. Thus, to obtain Step 4. Update the learning rate Y
a single band image, PCA transformation is applied on
the fused image and its first component that is the PC- Step 5. If kz1;t z1;t1 k > e then go to step 2, otherwise
1 image is extracted. The FKLICM clustering algorithm go to step 6.
is applied on the PC-1 image to classify it into different
classes. FKLICM is a neuro-fuzzy model that combines Step 6. Output the final clustering result.
the Kohonen clustering network with FLICM cluster-
ing algorithm. The experimental results show that the 2.2 FCM
proposed hybridized neuro-fuzzy model, FKLICM, is
much more efficient and effective for classification of The FCM algorithm is widely used for image cluster-
remote sensing images. ing [7]. FCM groups the data points into c clusters.
Each cluster has a cluster centre, zi . The cluster
centres are computed by means of optimizing the
2. BACKGROUND value of an objective function. Each point is assigned
In this section, the Kohonen clustering network, fuzzy a fuzzy membership in the range [0, 1]. However, the
c-means and FLICM clustering algorithms are exam- sum of the memberships of a point in all clusters is
ined to identify their advantages and limitations. equal to one. The fuzzy membership values are
placed in the membership matrix. The FCM algorithm
2.1 KCN minimizes the objective function, Jm given by
Equation (3):
The KCN is the simplest neural network, without any
activation function and hidden layer. The network has N X
X c
only two layers, the input layer and output layer. The Jm ¼ 2
ik kzi xk k
um ð3Þ
neuron closest to the input vector in terms of Euclidean k¼1 i¼1
distance is the winner neuron [5]. The weight of the
winner and its predefined neighbours are updated
Clustering is done by iteratively optimizing the objec-
using a learning rule. The operation of KCN is summa-
tive function (Jm ). The steps involved in FCM are as
rized below.
follows.
Step 1. Initialization: initialize the cluster centres
Step 1. Initialize the cluster centres zi ð1 i cÞ, fuzzi-
zi ð1 i cÞ, learning rate Y ð0 Y 1Þ, threshold
ness index mð1 < m 1Þ, fuzzy partition matrix U
e ðe > 0Þ and topological neighbourhood parameters.
and threshold e ðe > 0Þ, number of iterations.
Step 2. Selection of winner: calculate the squared
Step 2. Calculate the fuzzy membership matrix U ¼
Euclidean distance for i ¼ 1; 2; . . . ; c.
[uik ] by Equation (4):
!1
c
X
d2ik ¼ kxk zi k2 for k ¼ 1; 2; . . . ; N and i ¼ 1; 2; . . . ; c kzi xk k 2=ðm1Þ
uik ¼ ð4Þ
ð1Þ l¼1
kzl xk k
Step 3. The cluster centres are updated by The cluster centres are updated by Equation (9):
Equation (5):
PN PN m
um uik xk
ik xk zi ¼ Pk¼1 ð9Þ
zi ¼ Pk¼1 ð5Þ N m
N m k¼1 uik
k¼1 uik
Step 4. If kU tþ1 U t k > e then go to step 2, otherwise On convergence of the algorithm, the pixels of the
go to step 5. image are assigned to the class with highest member-
ship value to obtain the classified image.
Step 5. Output the final clustered image by assigning
the pixel xk to the class c with highest membership 3. PROPOSED ALGORITHM
value.
The proposed algorithm consists of two main steps:
The limitation of FCM is that its performance depends
upon the choice of fuzziness index m. An inappropri- 1. image fusion and dimensionality reduction;
ate value of m leads to unsatisfactory results. The 2. FKLICM.
inability of FCM to consider local spatial information
makes it highly sensitive to noise. Thus, FCM performs These are discussed in the following sections. The
poorly in the case of noisy images. flowchart of the proposed algorithm is shown in
Figure 1.
2.3 FLICM
To further improve the performance of FCM, a new
clustering technique named FLICM was introduced by
Kirindis and Chatzis [11]. It uses the local spatial and
grey-level information in its objective function. The
objective function of FLICM is given by Equation (6):
N X
X c
2
Jm ¼ ik kzi xk k þ aik
um ð6Þ
k¼1 i¼1
X 1
aik ¼ ð1 uij Þm kzi xj k2 ð7Þ
j2Nr
dðk; jÞ þ 1
0 !1=ðm1Þ 11
X
c
kzi xk k2 þ aik
uik ¼ @ A
l¼1 kzl xk k2 þ alk ð8Þ
for 1 i c and 1 k N:
3.1 Image Fusion and Dimensionality Reduction where N ¼ H W. To create the feature vector using
PCA, we proceed as follows. The average vector, x, of
Satellite images consist of multiple number of bands. a vector population can be approximated by
These bands contain complementary information.
Image fusion is done to combine the information in dif-
ferent bands into one image. The new image obtained 1X N
x¼ Ik ð12Þ
is more appropriate for further processing as it has N k¼1 fuse
higher resolution and better visual quality. In this
paper, Brovey transformation [16] is used to fuse the
RGB bands of the MS image with the pan band. Brovey The difference between the each vector and average
transformation is a numerical technique that uses a vector is Dk ¼ Ifuse
k
x. The covariance matrix CI has
ratio algorithm to fuse images. The mathematical eigenvectors ei and corresponding eigenvalues li . The
definition of Brovey transformation is given in covariance matrix CI can be approximated by
Equation (10):
1 X N
tðm 1Þ
mt ¼ m þ ð15Þ
tmax
0 !1=ðmt 1Þ 11
X
c
kzi xk k2 þ aik
uik ¼ @ A ð16Þ
l¼1 kzl xk k2 þ alk
Step 3. The weight of the output neuron is updated by Figure 2: Network architecture of FKLICM.
PN
Y ik;t ðxk zi;t1 Þ
zi;t ¼ zi;t1 þ k¼1
PN ð18Þ
s¼1 Y is;t performance of the proposed FKLICM, it is compared
with FCM and GIFP-FCM. The classification results
Step 4. Update the learning rate Y ik;t . are shown in Figure 3. The various parameters used in
the experiment are c ¼ 3, m ¼ 2, e ¼ 1 106. The
Step 5. Set t ¼ t þ 1: qualitative analysis of the results shows that FKLICM
gives better results than the other two methods. Fur-
Step 6. If kz1;t z1;t1 k > e and t < tmax then go to ther, the accuracy assessment of FKLICM, FCM and
step 2, otherwise go to step 7. GIFP-FCM was performed using ERDASTM software.
An error matrix was used for a series of descriptive
Step 7. Output the final clustered image by assigning and analytical statistical analyses. Overall accuracy
the pixel xk to the class c with highest membership and kappa coefficient were used for the assessment of
value. the proposed algorithm. A total of 256 reference points
were chosen using stratified random sampling.
The use of spatial contextual information makes the
algorithm less sensitive to noise. Also, the inter cluster The error matrices of all the three methods are given in
distance is large enough, avoiding overlapping of Table 2 where classes 13 represent vegetation, land
clusters. The network architecture consists of two and urban area, and water, respectively. The error
layers: input layer and output layer. The input layer matrix shows that the number of misclassifications in
consists of N neurons; the kth neuron represents the FKLICM is minimal while GIFP-FCM shows improve-
kth pixel of the image to be clustered. The output of ment over FCM. The overall accuracy is 96.88 and the
the network is c neurons corresponding to the c cluster kappa coefficient is 0.9485 for FKLICM, which is
centres. The architecture of the network is shown in higher than the other two methods. Thus, the quanta-
Figure 2. tive and qualitative analysis shows that the proposed
FKLICM provides better accuracy and classification
4. EXPERIMENTAL RESULTS results.
In this section, the performance of the FKLICM is eval- The kappa coefficient value lies between 0 and 1. The
uated. The algorithm was implemented in Matlab higher the value of kappa, the higher the classifica-
R2012b. The experiment was conducted on Landsat 7 tion accuracy is. Accuracy and kappa coefficient
ETMþ image with eight bands over Ishinomaki City, values of the three methods are listed in Table 3.
Japan, acquired on 18 February 2003 (image courtesy It can be seen that FKLICM has highest overall
US Geological Survey) [21]. The study image is shown accuracy and highest value of kappa coefficient, con-
in Figure 3(a). The geographical coordinates of the forming that the proposed method outperforms the
area are 38 250 000 North, 141 180 000 East. To test the other two.
Figure 3: (a) Original Landsat 7 ETMþ image; (b) classified image with FCM; (c) classified image with GIFP-FCM, a ¼ 0.7;
(d) classified image with FKLICM.
Authors
Krishna Kant Singh is currently pursuing a Kirat Pal is a professor in the Earthquake
PhD from Indian Institute of Technology, Roor- Engineering Department of the Indian Institute
kee, India. He holds BTech (EI) and MTech of Technology Roorkee. He has authored
(VLSI) degrees. His areas of specialization more than 61 research papers in national and
include digital image processing, VLSI tech- international journals and conferences.
nology and neural networks. He has authored
more than 12 research papers in the above E-mail: kiratfeq@iitr.ernet.in
areas in national and international journals
and conferences.
E-mail: krishnaiitr2011@gmail.com
Madhav J. Nigam is a professor in the Elec- Akansha Mehrotra is currently pursuing a
tronics and Computer Engineering Depart- PhD from Indian Institute of Technology, Roor-
ment, Indian Institute of Technology, Roorkee, kee, India. She has to her credit a number of
India. His areas of research include digital research papers in international conferences
image processing, control systems and neural and journals. Her areas of interest include digi-
networks. He has authored a number of tal image processing, remote sensing and
research papers in the above-mentioned neural networks.
areas.
E-mail: akanshasing@gmail.com
E-mail: mkndnfec@gmail.com