284 
LAND COVER CLASSIFICATION BY IMPROVED FUZZY C-MEAN CLASSIFIER 
ZHAO Quan-hua, SONG Wei-dong, Bao Yong 
School of Geomatics, Liaoning Technical University, Fuxin 123000,China 
zqhlby@163.com 
Commission VI, WG VI/4 
KEY WORDS: land cover; improved fuzzy c-mean classifier; principal component analysis; TM data; confusion matrix; 
ABSTRACT: 
Result from uncertainty and mix pixel of most remote sensing information, spectral information of some pixels are fuzzy in the 
classification process of land cover. In most time, we can’t ascertain some pixels to be one or the other, because of it, it is hard to get 
satisfied classification results by traditional hard classification method. Then we can adopt soft classification method such as fuzzy 
clustering in the classification process of land cover. With Satellite Liaoyang panchromatic image, improved fuzzy c-mean Classifier 
was used to land cover classification in this paper. First samples of three kinds of features were selected from the panchromatic image 
considering the spectral information of features. Then each average of contrast, entropy and variance of histogram characteristics 
were computed as clustering center of fuzzy clustering. 2.5 was selected as weight index,3 as clustering number, Mahalanobis 
distance scale which can detect characteristic space for data muster of super ellipse structure was applied in the fuzzy clustering. The 
three kinds of features were ascertained through priori knowledge after fuzzy clustering. At last, confusion matrix was introduced in 
precision evaluation. The result shows that the classification method based on fuzzy c-mean classifier is a high precision method for 
land cover classification. 
precision method for land cover classification. 
1. INTRODUCTION 
Land cover classification of remote sensing image is a process 
of detecting and identifying types of object based on the 
characteristics of spectrum, texture and geometric. The 
classification is a important step of image recognizing and 
interpreting because it is necessary for information extraction 
from remote sensing image, detecting of dynamic 
transformation, making thematic map and establishment of 
remote sensing database. So land cover classification has a 
bright future. Satellite images data is always enormous 
because of the characteristic of multi-band, multi-spatial 
resolution, multi-time resolution. Some pixels of image data 
are always fuzzy and uncertain because of the complexity of 
the feature spectrum and the diversity of disturbing factors. So 
in the process of land cover classification, it difficult to 
ascertain some pixels to be one or the other. In order to 
improve the accuracy of classification, soft classification is 
necessary to ascertain the membership for each pixel. 
Traditional classification method is not perfect because it is 
based on the statistical relations between statistical 
characteristics and training sample datas. Fuzzy classification 
is based on the hypothesis that each pixel involved many 
features, but membership of each feature is different. 
Membership of training samples is ascertained in the process 
of training classifier. So fuzzy clustering method can get high 
precision for land cover classification. In the paper, improved 
fuzzy c-mean classifier was used to land cover classification 
based on Matlab. At last, confusion matrix was introduced in 
precision evaluation. The result shows that the classification 
method based on improved fuzzy c-mean classifier is a high 
2. IMPROVED FUZZY C-MEAN CLUSTERING 
Fuzzy c-mean clustering was proposed on the basis of k-mean 
method and fuzzy theory. Euclidean distance is adopted in 
Routine fuzzy c-mean classifier. It’s a method of all-direction 
uniform. But the pixels of remote image do not distribute like 
that showed by scatter diagram of pixels. So it always can not 
get ideal result expected. In order to bring the advantages into 
full play of fuzzy c-mean classifier, a new defined distance 
should be introduced in it. So improved fuzzy c-mean 
classifier was introduced in this paper. Its basic theory is 
described as 
follows: 
n is the number of pixels for clustering, 
X = fo,* 2 , 
represent the muster of all pixels, in 
X = {x,,* 2 . 
the muster, 
vector , c is the classification number. 
is the clustering center for each class, 
is the muster of clustering center of all classes. 
, p is the number of feature 
v , =(v!,vf,...v' ) 
V = {v l ,v 2 ,...,v c ] 
represent the membership of pixel 
belong to 
number 1 . The membership array is defined as follows: 
U = [u ik ] 
L IK Jcxn /1 \