OBJECT CLASSIFIERS FOR FOREST CLASSIFICATION 
Gintautas Palubinskas 
Data Analysis Department 
Institute of Mathematics and Informatics 
Akademijos 4, Vilnius 2600, Lithuania 
Commision III WG III/3 
E-mail: raudys%ma-mii.lt.su@fuug.fi 
ABSTRACT : 
The aim of this research is to compare the performance ( probability of misclassification 
) of several object classifiers ( some of them developed by author ) and classical per-pixel 
maximum likelihood classifier for forest ( deciduous, coniferous and mixed ) classification. 
The new analytical method ( derived by author ) is applied for the selection of object classifier 
based on calculating the probability of misclassification. 
Object classifiers are supervised maximum likelihood classifiers incorporating spatial charac- 
teristics of an image during classification based on Markov random field model. 
The research is carried out on Landsat TM data received from IFAG, Frankfurt a.M. 
During investigation the Image Analysis and Classification System IMAX ( developed by 
author and his group ) is used. 
First results show the complexity of the problem and the need for further investigation. 
KEYWORDS : Thematic information extraction, Spatial characteristics of images, Pattern 
recognition, Landsat TM. 
1. INTRODUCTION 
The thematic information extraction from remote sensing 
images is important for solving many practical problems. 
Usually supervised maximum likelihood classifiers which 
assign each pixel of an image to one of m known classes, 
ie. so called per-pixel classifiers, are used ( Jensen, 1986; 
Richards, 1986). But often the quality of such classifiers 
is not satisfactory. 
One of the ways to solve this problem is to incorporate 
spatial characteristics of an image in the process of clas- 
sification. There are some reviews on these questions ( 
Landgrebe, 1981; Swain, 1985; Alfiorov, 1989; Palubin- 
skas, 1990a ). We can see that there are three approaches 
how to use the spatial information: textural, object and 
contextual. This work is concerned with the object ap- 
proach. 
Object classifiers assign the whole object of an image or 
the central pixel of an object to one of m known classes. 
But there are difficulties in realization of object approach 
in general case because of the high dimensionality of the 
vector to be classified. Assumptions about the kind of 
dependence between intensity values of neighboring pixels 
must be made. Object classifiers first were introduced in 
( Ketting, 1976; Landgrebe, 1980 ) in the case of inde- 
pendent pixels of an object. Another assumption is based 
on Markov type dependence, particulary on the separable 
correlation model. All known object classifiers are based 
on this model. 
In ( Palubinskas, 1988a; Palubinskas, 1989 ) the systemati- 
zation of image models based on the separable correlation 
model is made and on this basis some original classifiers 
are proposed. So in total 14 object classifiers ( some of 
them are in the publications of Guyon and Yao, 1987; 
Mardia, 1984; Switzer, 1980 ) were investigated theoret- 
ically. The quality of classifiers is usually measured by 
the probability of misclassification ( PMC ). The new an- 
alytical method to calculate the PMC is proposed in ( 
484 
Palubinskas, 1988b; Palubinskas, 1992 ) which allows to 
compare the performance of several classifiers in the same 
conditions. This method is much more cheaper than the 
traditional method, when the PMC is evaluated on mod- 
eled data. This theoretical analysis helped us to select 5 
object classifiers from 14 for further investigation on real 
data. 
There are some papers ( Landgrebe, 1980; Kalayeh and 
Landgrebe, 1987; Mardia, 1984; Switzer, 1980 and Palu- 
binskas, 1990b ) where some of these object classifiers are 
tested on real remote sensing imagery. These experiments 
allow to select 3 object classifiers from 5 and they can be 
recommended for practical use. However, this conclusion 
is valid not for all situations. 
In this work the above mentioned analytical method of se- 
lection of a classifier is used in the analysis of remote sens- 
ing imagery. At first the statistical characteristics of trai- 
ning data and statistical charateristics which are used for a 
classifier design are calculated. Then the analytical PMC 
of this classifier is calculated. So the analytical method 
allows us to select the object classifier from a set of avail- 
able with a very little computer time expenses. Then the 
selected classifier can be run on the full data set. 
The work of this method is illustrated on the example of 
classifying the Landsat TM image of Frankfurt am Main 
surroundings recorded on 30 July 1984 ( received from 
IFAG - Institut fur Angewandte Geodasie ). 
In Section 2 the object classifiers are described briefly. Sec- 
tion 3 presents the new analytical method for classifier se- 
lection. In Section 4 the experimental results are presented 
and finally Section 5 provides the concluding remarks. 
2. OBJECT CLASSIFIERS 
Consider a two-dimensional multispectral image, where 
the pixels of an image are q-dimensional vectors 
and 
adja 
Let 
is a 
the 
Sup 
but; 
whe 
is + 
COV: 
The 
an | 
pos 
The 
sigr 
whe 
clas 
onl 
The 
In C 
ance 
(Mt + 
The 
resti 
esti 
sam 
cert: 
The 
assu 
Firs 
of a; 
R is 
Kro 
Secc 
mad 
ject 
In t] 
acte 
sum 
dom 
198' 
whi 
this 
num 
All 
Corr