ndlin- 
rank: 
an 
- an 
92. 
linea 
artoc 
Jerlir 
es fc 
; Sep 
undeg 
GEOCLASS, A NEW APPROACH TO THE THEMATIC CLASSIFICATION OF MULTISPECTRAL IMAGERY 
R.Steffensen 
President, Geostudio Consultants Limited 
525 St. Laurent, Suite 24, 
Ottawa K1K2Z9 - Canada 
Commission III 
ABSTRACT : 
The use of digital multispectral imagery for thematic classification has become standard practice. In a conventional 
approach only color attributes of individual pixels are used in classifying terrestrial objects. Since many terrestrial 
features possess similar color attributes, thematic classifications solely based on color attributes are often 
inaccurate. A new approach has been implemented that in addition to color takes also into account the spatial 
relationship of individual pixels to their neighbours. The use of spatially related information greatly enhances the 
analyst's ability to make improvements to the type of thematic classes that can be extracted and to their accuracy. 
This was demonstrated by a successful application of this approach to the classification of crop types using TM 
imagery. 
KEY WORDS: Algorithm, Image Segmentation, Image Analysis, Image Interpretation. 
1. INTRODUCTION 
Although twenty years have passed since the first 
Earth Resources Satellite (Landsat 1) was launched, 
the utilisation of imagery data provided by this type of 
satellite for thematic mapping has not yet fulfilled the 
early expectations raised by this new type of 
environmental data (Ryerson, 1989). While the quality 
of the data both in terms of spatial and spectral 
resolution has improved during this time span (e.g. 
TM vs. MSS), the same cannot be said for the 
computer-based methodologies applied to the 
extraction of thematic information from the physical 
data. In particular, the approaches and techniques 
used in these methodologies have not progressed 
substantially since that time. The conventional 
approach assumes that individual picture elements 
(pixels) represent actual classification objects by 
ignoring any other pixel attribute but color. As 
pointed out by Sijmons (1987): “treating scene elements 
as independent objects is an incorrect model that 
ignores structural features which explicitly consider 
the spatial relationship between neighbouring 
elements”. If the spatial element is not taken into 
consideration, “mixture” pixels, those pixels where 
two or more terrain cover types mix, cannot be 
distinguished from “pure” pixels that represent a 
single land cover type. In addition, many land 
features characterized by a typical spatial 
configuration, such as roads, cannot be distinguished. 
Usually, roads generate spectral signatures very 
similar to all other man-made objects found in built- 
up areas. Clearly, the discriminating algorithm to be 
applied in any attempt to identify the linear pattern of 
roads will have to take into consideration both the 
spectral and spatial attributes of pixels. The addition 
of the spatial configuration of pixels to multispectral 
color in the analysis of remotely sensed imagery data 
is not only useful for classifying specific thematic 
categories, such as roads, but should be considered an 
essential element for improving the accuracy of any 
type of thematic classification. In fact, using spatial 
attributes it is indeed possible not to be any longer 
bound by the erroneous concept of representing pixels 
903 
as classification objects. Through a segmentation 
process a terrestrial scene can be partitioned into 
areal and linear structural elements. characterized by 
the association of pixels having similar spectral and 
spatial properties. These structural elements will 
have a much closer link to land features than 
individual pixels, bringing about significant 
improvements to the entire process of image 
classification. This paper describes the design and 
development of GEOCLASS, the first computer-based 
methodology commercially available capable of 
analysing multispectral imagery in both the spectral 
and spatial domain. 
2. CONVENTIONAL APPROACH 
To illustrate the advantages to be gained by the 
GEOCLASS methodology, a short review of the 
shortcomings of the conventional approach based on 
pixel processing is required. The basic premise in 
computer-assisted multispectral classification of 
remotely sensed imagery is that terrestrial objects 
display sufficiently different reflectance properties in 
different regions of the spectrum to allow specific 
colors to be associated to specific objects. This is in 
general a valid assumption. However, this 
assumption should not be meant to imply that every 
picture element of two different objects is different in 
color. For instance, if it is true that the average color 
of a wheat field in a satellite summer image is 
different from that of a field of canola, it is not true 
that every pixel belonging to a wheat field is 
necessarily different in color from every pixel 
belonging to a field of canola. Ignoring this fact leads 
to a number of problems that permeate throughout the 
entire conventional classification process. However, 
there is even a more serious drawback. This has to do 
with the most importantaspect of image classification, 
which is the choice of training samples. Training is 
the initial stage of the pattern recognition process 
followed in remote sensing. It consists of determining 
a valid sample for the spectral signatures of all objects 
existing within a selected scene. A valid sample 
means that the sample must best represent an entire 
population of a specific land feature. Since all 
classifiers commonly used in remote sensing rely on