ertainty 
ertainty 
alues of 
type of 
type of 
> of this 
es, and 
tudy is 
, Image 
es. For 
e stated 
to the 
used by 
certain 
has its 
acy and 
jility is 
d in all 
ertainty 
ssity of 
analysis 
and the 
ethods 
to the 
ith this 
n of the 
matical 
nd the 
lved in 
a case 
ite the 
1. A 
on the 
m into 
ved. 
AINTY 
Image 
ison is 
inexact or uncertain in some respect or other. As 
analyzed by Frost (1986), this is due to several factors: 
a) the universe of discourse is truly random; b) the 
universe of discourse is not strictly random but for 
some reason there is insufficient data; c) available 
knowledge represents a 'gut feeling! and such 
judgmental knowledge can be useful when more 
sound knowledge is not available; d) available 
knowledge is couched in terms which are themselves 
vague (e.g. the word 'usually' in 'Canary grass usually 
will not follow canola; and e) the knowledge source is 
imperfect. 
Among these factors, b) is a typical situation in 
spectrally-based remote sensing image classification. 
For example, suppose crop types are to be identified 
only based on spectral information. A commonly used 
supervised classification method is to compute the 
likelihood that a field grows a type of crop using 
probabilistic reasoning, based on the evidence obtained 
from training areas. 
Furthermore, ancillary information, such as soil types 
and digital terrain models, may be used to improve 
classification accuracy. This is achieved through the 
representation of relationships between ancillary data 
and crop types using certainty values such as 
probabilities and certainty factors, and the 
incorporation of these certainty values into the 
probabilistic reasoning. These certainty values are 
usually estimated from two sources, i.e. databases and 
human experts. Databases are used as samples to 
compute probability values, while the statements may 
be expressed in different ways by experts. For example, 
an expert may state: "Oats usually grow well on the 
land with elevation between el and e2, soil types t1, t2, 
and t3, and slope ranging from sl to s2". This 
statement is an empirical rule. It is judgmental; a 
vague term (usually) is included in the statement; and 
maybe only part of the ancillary themes of concern are 
addressed, hence being imperfect or incomplete. Thus, 
in addition to the uncertainty situation b), situations c), 
d), and e) may all be encountered in the classification 
of remote sensing image based on multiple knowledge 
Source reasoning. 
The UIU Problem in Remote Sensing Image 
Classification 
A way to examine the UIU problem in remote sensing 
image classification is to look into the sources where 
related knowledge for the classification is generated. 
These sources can be generalized into three types: one 
is non-time-serial databases, such as spectral image 
databases for sampling training areas; the second is 
historical databases or time-serial databases which are 
used in the elicitation of ancillary knowledge; the last 
is human experts who provide expertise related to the 
ancillary information of concern. Figure 1 outlines the 
major sources that cause the UIU problem. 
Beneath the probabilities generated from non-time- 
serial databases, there exist at least two types of 
uncertainty. One is database accuracy, which deals 
939 
KNOWLEDGE SOURCES 
DATABASES HUMAN EXPERTS 
   
   
  
    
  
  
  
  
  
  
’ 
TIME SERIAL NON-TIME SERIAL ÁN 
DATA DATA i : 
+ 2 * 7 I A . 
4 X ^i 
Fh x + : 4 “i 
4508 NT Si 4 A P 
, vs 1 a4‘ A ' S i i A ; 
AR ' | . 
TRIS Satins pti 
- 
- 
35s , "ue Set 
Figure 1 Major Sources Causing the UIU Problem 
with data quality. The other is the sufficiency of the 
sample size available in the databases for statistics 
purpose. For example, a database with ten thousand 
records indicating the relationship between soil types 
and vegetation distributions may have over thousand 
records addressing soil type A, but only a few records 
addressing soil type B. Thus, the probabilities 
representing the relationship between soil types and 
vegetation distributions would be more reliable or 
certain for soil type A than for soil type B. 
For time-serial databases, there are even more 
uncertainties existing in the probability values 
generated from the databases. The uncertainties of 
database accuracy and sample sufficiency also apply to 
time-serial databases. In addition, two other factors 
affect the probability values based on this type of 
databases. One is the number of time periods (e.g. the 
number of years), since statistics based on few time 
periods may be seriously biased, especially for the 
themes that are closely related to socio-economic 
situations. The other is the standard deviation of an 
event's occurrences during different time periods, 
since a large standard deviation may suggest the effect 
of some factors (e.g. socio-economic factors) that are 
not of concern in the knowledge elicitation. This can 
be depicted through an example, as shown in Figure 2. 
The height of the bar represents the number of fields 
that grew flax in that corresponding year. The large 
difference of flax field occurrences between 1986 and 
other years, which causes a large standard deviation, 
suggests a possibility that the high occurrence of flax in 
1986 results from social economic factors such as the 
crop price. If statistics based on such a database aim to 
generate crop rotation rules, the result would probably 
be biased. 
      
  
  
Year 1986 1987 1988 1989 
Figure 2 Occurrences of Flax Fields in An 
Experimental Area