KNOWLEDGE BASED IMAGE CLASSIFICATION IN A GIS 
Eugene Derenyi, David Fraser and Richard Pollock* 
Department of Surveying Engineering 
University of New Brunswick 
Fredericton, N.B. Canada 
Magdy Halim 
Universal Systems Ltd. 
Fredericton, N.B. Canada 
ISPRS COMMISSION III 
ABSTRACT 
A geographic information system (GIS) equipped with both vector and r 
incorporating a-priori knowledge, in the form of segmentation by polygons 
data. This paper discusses the design and implementation of this system and illustrates the analysis process on an example. 
KEY WORDS: Knowled 
Integrated System. 
  
ge base, Image analysis, Image classification, Segmentation, A-priori knowledge, GIS, Image statistics, 
*Presently with the Department of Computer Vision, University of British Columbia, Vancouver, B.C., Canada. 
INTRODUCTION 
Airborne and especially spaceborne imaging sensors collect a 
vast volume of data on the state of the earth's environment 
and its resources. These data are processed and analyzed in 
digital image analysis systems where the digital number of 
each pixel is examined individually and then assigned to a 
theme or class based on its compliance with values found in 
training samples of a particular class. This is the well 
known supervised classification. The decision is essentially 
made on the bases of the spectral response value represented 
by the digital number of a pixel. This per-pixel spectral 
classification is rather simple to implement in computers. It 
yields acceptable results when the classes in question have a 
unique spectral response pattern and the training areas are 
truly representative samples. 
This technique is, however, very simplistic compared to the 
art of visual image interpretation. It is analogous to the case 
of covering a photograph with a blind, and then attempting 
to sort the thematic information by observing the grey levels 
or colours through a small slit which is sequentially scanned 
over the image. Human interpreters consider not only tone 
and colour, but also properties like texture, shape, size, 
pattern, site and association. An IF-THEN deductive 
reasoning process leads to the final answers. Numerous 
attempts have been made to incorporate some of the above 
elements into the digital image classification process, but 
only texture could be dealt with successfully. Even in this 
case, a so-called ‘texture image” is derived and used as a 
data layer in a per-pixel classification. Perhaps the heart of 
the problem is that all other elements express spatial 
relationships which cannot be modelled in a per-pixel 
observation. 
Another shortcoming of current digital image 
classification procedures is that the image data is usually 
analyzed in isolation from other pertinent information. The 
data sets are, of course, often multi-layered, incorporating 
multi-spectral, multi-temporal and multi-level acquisitions. 
Non-image data such as topography, geophysics, geology 
may also be utilized as additional raster layers. Information 
such as property and land use boundaries, infrastructure, 
prior inventories and statistics are rarely utilized. As yet, 
such information could serve as a knowledge base on which 
to build subsequent analyses. Such information is, 
however, stored as cartographic and attribute data in 
geographic information systems (GIS), and the vector and 
alpha-numeric representation used is not compatible with the 
976 
raster organization of digital images. Therefore, a necessary 
prerequisite to knowledge based image analysis is to store 
and manipulate image, cartographic and attribute data in a 
single system. 
2. THE HYBRID GIS 
Researchers at the Department of Surveying Engineering, 
University of New Brunswick (UNB), among others, have 
realized the need for data integration and developed 
CARIS/RIX, a Raster Image Extension (RIX) to the 
Computer Aided Resource Information System (CARIS) 
[Derenyi and Pollock, 1990; Derenyi, 1991]. 
CARIS is a comprehensive GIS, developed and marketed by 
Universal Systems, Ltd. (USL) in Fredericton, N.B., 
Canada [Masry, 1982; USL, 1991]. It runs under both the 
UNIX and VMS operating systems and supports a wide 
range of input and output devices. The user interface is 
based on X-windows and MOTIF standards. Originally 
CARIS was built as a vector based system and was recently 
equipped with the capability of raster representation of 
geographic features and extended to accommodate the 
processing and analysis of digital images. 
CARIS/RIX provides the means for digital map to image and 
image to map geometric correction and registration. It 
incorporates numerous image processing functions such as 
various contrast enhancement routines, arithmetic 
operations, thresholding, spatial filtering, . principal 
component transformation and image processing in the 
frequency domain. Geographically registered images can be 
displayed as a backdrop to digital cartographic data. On- 
screen digitization in the raster image backdrop and 
interactive cartographic editing is possible. Provision is 
made for polygon based image statistics generation, which is 
a key element in knowledge based image classification and 
will be elaborated on in the next chapter. 
3. KNOWLEDGE BASED IMAGE ANALYSIS 
3.1 The Concept 
Knowledge based image analysis is viewed by the authors as 
the incorporation of non-spectral information into the 
process. On an elementary level, the inclusion of non-image 
raster data layers in the per-pixel classification could be 
aster data processing capabilities sets the stage for 
and as lexical data, into the analysis of remotely sensed