THE APPLICATION RESEARCH IN ASSISTANT CLASSIFICATION OF REMOTE 
SENSING IMAGE BY TEXTURE FEATURES COMBINED WITH SPECTRA FEATURES 
Y.M. Fang, X.Q. Zuo, Y.J. Yang, J.H. Feng 
Faculty of Resource, Kunming University of Science and Technology, No.253 Xuefu Road, 
Kunming Yunnan, China-fangyuanmin@126.com 
Commission VI, WG VI/4 
KEY WORDS: Image, Assistant Classification, Spectrum Feature, Texture Feature, Gray Level Co-occurrence Matrix 
ABSTRACT: 
With the development and application of high resolution remote sensing satellite, more clear textures occur in the remote sensing 
image. The features of land forms and features reflected by textures are important information in distinguishing ground objects. 
Based on original image, adding texture features can promote veracity and accuracy of classification. If, especially, the spectra 
feature of different objects is nearly similar, texture features will play an important role in distinguishing these objects. For the 
assistant classification of remote sensing image, the methods, that the texture feature images are extracted by the gray level co 
occurrence matrix and the classifications are carried out by combined the texture features with the spectral features, is researched in 
this paper. Extracting the texture feature images are realized by the computer program developed by us. The test results show that the 
assistant classification as the paper mentioned could increase the classification veracity and accuracy of remote sensing images. Also 
the results are analyzed and compared with the traditional ways. 
1. INTRODUCTION 
Automatic recognition of remote sensing image is a great chal 
lenge in the field of remote sensing, computer vision and vague 
recognition. With the rapid development of remote sensing 
technology, the automatic extraction of image information from 
remote sensing images has become a means of interpretation of 
remote sensing images. However, lower accuracy of current 
computer classification is difficult to deal with the map of lar- 
ger-scale and medium-scale. Therefore, enhancing the preci 
sion of special classification is the core of research on remote 
sensing technology and application. 
If the remote sensing image classification applies spectrum- 
feature-based classification method only, the wrong classifica 
tion on the phenomenon of the same thing but different spectra 
and the same spectrum but different things will be created in 
evitably. With the development and application of high resolu 
tion remote sensing satellite, texture features can be represented 
more clearly in the remote sensing image. In theory, the texture 
feature as a band participating in spectrum feature classification 
may effectively avoid the wrong classification caused by the 
phenomenon of the same spectrum but different things and the 
same thing but different spectra, which can promote the reli 
ability and the accuracy of classification. The research in this 
paper confirmed that adding the texture feature information to 
carry on the special classification is an effective way to in 
creases the precision on basis of the remote sensing spectrum 
information. This classification method can extract special im 
age which is significant to the classification application of me 
dium scale or small scale image figure. 2 
we applied has already been pre-processed which is allowed to 
do the following experiment directly. The image has contained 
many land sorts such as the inhabited area, forest, bare land, 
paddies and lakes, which can check the classification effect sat 
isfactorily. 
Figure 1. SP0T5 remote sensing images of this article 
adopted 
3. EXTRACTION OF THE TEXTURE FEATURES 
IMAGES 
3.1 Gray Level Co-occurrence Matrix 
The gray level co-occurrence Matrix, begins with the pixel of 
image (x, y) gray as i, is the probability P (i,j,S,9) that will 
appears simultaneously with pixel(x+Ax, y+Ay) with the 
distance as 8 and the gray as j. It is showed in following figure. 
2. THE BASIC SITUATION OF IMAGE 
Figure 1 is the remote sensing image used in this article which 
is the spot5 image. This image is fused with 2.5 meters resolu 
tion panchromatic data and 10 meters resolution multi-spectra 
image. The size of image is 760x490 pixels. This part of image