A Spectral and Textural Knowledge-Based Approach for Automated 
Extraction of Topographical Factors from Remotely Sensed Images 
Zhang Xiangqian, Li Xiao eng, and Yu Xuchu 
Zhengzhou Institute of Surveying and Mapping, Henan, China 
ABSTRACT: A prototype expert system is developed to demonstrate the feasibility of 
classifying multispectral remotely sensed data on the basis of spectral and textural know- 
ledge. In this paper, the spectral and textural properties of settlement place, vegetation, wat- 
er area, soil, etc. are discussed and a production algorithm of texture image is studied. Ac- 
cording to the relationships of bandto-band and category-to-texture, a knowledge base 
represented by rules and weights is established. The method presented in this paper is of 
fast computation speed and high classification accuracy. 
KEY WORDS: Artificial Intelligence, Classification, Knowledge Base, Image Processing 
arametric pattern recognition techniques are widely 
P used to classify landcover categories in multispectral 
remotely sensed data (Landgrebe, 1981). Parametric 
classifiers are based on the statistical distribution of the 
feature vectors. The limitations of statistical methods are 
that they cannot directly utilize nonparametric know- 
ledge, and have a low classification accuracy. Therefore, 
they cannot fulfil the productive task of the automated 
extraction of topographical map factors. In order to im- 
prove the classification accuracy, some knowledge-based 
classification methods were presented in (Wharton, 1987, 
Mehldau et 1990, Ferrante, 1984). 
ledge-based classification system for high spatial and spec- 
al, A know- 
tral resolution images was pressented in (Wharton, 
1987), and a C-extension for rule-based image class- 
in (Mehldau, 1990). 
Because the two classification systems were developed 
ification system was presented 
in ideal conditions, there are some limitations when they 
are used for remotely sensed images. The multispectral 
image analysis system (MSIAS) presented in (Ferrante, 
1984) is a knowledge-based spectral classification system 
developed by Ferrante et al. The system was designed to 
use a hierarchical decision tree structure with two 
classification levels for TM data. Because the approach 
does not provide a method of recovering from decision 
errors, the final classification is highly sensitive to errors 
at any decision levels. The classification results were con- 
sidered unreliable. 
In this paper, the spectral and textural features of 
map factors in remotely sensed images such as settlement 
place, water, vegetation and soil and the establishment 
of knowledge base are discussed. A classification method 
integrating parametric pattern recognition technique with 
knowledge base is presented here. The method above 
can extract validly topographical factors such as settle- 
ment place, water, forest, soil and road. Experimental re- 
sults demonstrate that the method is of fast conputation 
87 
speed and high classification accuracy. 
RESEARCH SCHEME 
The reflectance-spectral and textural properties of map 
factors such as settlement place, water, vegetation and 
soil are different from each other. They give important 
evidence for visual interpretation. However, for various 
reasons, the different objects of the same category (e.g. 
buildings of different materials) have distinct spectral 
reflectances, and therefore different tones or colors in 
image. In the process of visual interpretation, the re- 
gions of same tone or color are distinguished, then the 
categories of the regions are determined. 
The research scheme in this paper bears analogy to vis- 
ual interpretation process. The first step is to segment 
the preprocessed image with unsupervised classification 
algorithms based on spectral and textural features, and 
compute the averages of every band data and textural 
energy image for each segmentation region. The second 
step is to determine the category of each segmentation 
( input preprocessed. \ 
mul image 
ty 
  
Product textural image and  K-L transform | 
multispectral image into principal component image | 
  
  
Ÿ 
Unsupervised classification is used on first and second | 
| components and textural image to acquire segmented image | 
  
  
Y 
| Compute the averages of every spectral band data and | 
| textureal data for each segmentation region | 
  
: | Decision Making of [Inference Engine | 
' | Classification | 
; | knowledge Base | 
et rr tem 
| Merge same classifications| 
re 
Output ) 
Fig.1 The block diagram of the reseach scheme