ANALYSIS OF CONVERGENT EVIDENCE IN AN EVIDENTIAL REASONING
KNOWLEDGE-BASED CLASSIFICATION
Y. Cohen? and M. Shoshany^
? Institute of Agricultural Engineering, Agricultural Research Organization (A.R.O), the Volcani Center
Bet Dagan 50250 ISRAEL, yafitush@volcani.agri.gov.il
? Department of Transportation and Geo-Information Engineering , Faculty of Civil and Environmental Engineering,
Technion, Haifa, ISRAEL, maximsh@techunix.technion.ac.il
Commission TS ThS 8
KEY WORDS: Remote-sensing, Agriculture, Classification, Knowledge-base, Reasoning, Convergent.
ABSTRACT:
The use of KBSs based on evidential reasoning, for land-cover mapping based on remotely-sensed images is spreading widely. In
recent years, KBS utilizing Dempster-Shafer Theory of Evidence (D-S ToE) were found most successful in wide range of remote
sensing applications. One important feature of the D-S ToE is that it provides a measure for the evidential support (belief)
accumulated for each object class at each pixel. Although cumulative belief values (CBVs) play a major role in classification
decisions, their analysis has received little attention in the literature. The objective of the present study was to investigate and to
characterize the added value of the KBS by the analysis of the CBV. For that purpose we applied a KBS based on D-S ToË to crop
recognition in a wide heterogeneous region and compared its results with those of the application of ISODATA classification. We
investigated the relationships between the distribution of the CBV of the different classes and their corresponding classification
accuracy/reliability. The CBVs were found to be good indicators of levels of classification complexity in both the pixel and the class
scales. In addition to that, levels of two class properties could be analyzed according to the distribution of CBVs of each class:
heterogeneity and uniqueness. Moderate and high correlations (r7=0.69 and r°=0.94) were found between these two properties and
classification efficiency of an unsupervised classification (US). Lower correlations were found between these properties and the
KBS classification efficiency (r=0.59 and r°=0.75). Moreover, US classification was highly affected by heterogeneity and
uniqueness as referred from much higher slope coefficients (5 times higher): US classification efficiency decreased with increasing
heterogeneity levels and decreasing uniqueness levels. These findings are suggesting that in contrast to the US classification the
KBS facilitates identification of a class with little affect of its internal variability (heterogeneity) and its similarity with other classes
(lack of uniqueness).
1. INTRODUCTION
1.1 General Instructions
The use of KBSs based on evidential reasoning, for land-cover
mapping based on remotely-sensed images is spreading widely.
In recent years, KBS utilizing Dempster-Shafer Theory of
Evidence (D-S ToE) were found most successful in wide range
of remote sensing applications (e.g. Wilkinson and M'egier. J.,
1990; Kontoes et al, 1993; Peddle, 1995; Adinarayana and
Rama-Krishna, 1996). One important feature of the D-S ToE is
that it provides a measure for the evidential support (belief)
accumulated for each object class at each pixel. Although
cumulative belief values (CBVs) play a major role in
classification decisions, their analysis has received little
attention in the literature. One important feature of the D-S
algorithm is that it provides a measure of the accumulated
evidential support or cumulative belief value (CBV) for each
recognition class (Ci) inferred at each image pixel (Xi,j). The
advantage of KBSs lies in achieving recognition of a class
despite incomplete, missing and conflicting evidences. The
CBV for Ci at Xi,j depends on the overall applicable evidences
(rules) for Xi,j supporting and/or conflicting Ci. For different
compositions of environmental conditions different
compositions of evidences will be applicable. The CBV
increases with increasing number of supportive evidences and
decreasing number of conflicting evidences and vise versa.
There are few published researches regarding the relationship
between KBS recognition accuracy and reliability of a class and
the level of its CBV. It is to be expected that for classes or sites
with no conflicting and/or incomplete evidences, high reliability
and accuracy will be accompanied by a high accumulation of
supporting evidence. In such cases it is expected that there is
little or no need for the KBS approach. This assumption can be
assessed by comparing the classification results of the KBS
with those of an unsupervised classification (US). In complex
classes or sites there are more conflicting and/or incomplete
evidences and low supporting evidence is accumulated. It is
important to determine how the KBS performs in these complex
situations and whether low CBVs are necessarily accompanied
by low accuracy or reliability. Also, recognition systems
perform differently within a class, ie. the same class in
different sites may gain different CBVs. Analysis of the
distribution of the CBV within a class will facilitate the
determination of how unique and/or heterogeneous a class is.
This in turn, will enable the investigation of whether
heterogeneity and/or lack of uniqueness limit the classification
accuracy and reliability of the KBS. The objective of the
present study was to investigate and to characterize the added
value of the KBS by the analysis of the accumulated supporting
evidence. For that purpose we applied a KBS based on D-S ToE
to crop recognition in a wide heterogeneous region and
compared its results with those of the application of ISODATA
classification. We first describe the study area and ils
heterogeneity. In the subsequent two sections we outline the
principles of the D-S ToE and the GSA and describe the
916
Inte
con
con
Twi
ove
soul
and
leve
500
Agr
area
rela
side
ther
The
leve
Mot
wes
Mot
hori
alon
west
betw
red-
form
varié
crop
3.1
KBS
and,
and
cryst
base
been
S To
(GS
of e
coml
bias
be d
appli
follo
(200
3.1.1
Supp
agric
sumn
Crops
whic
® =
in ©
exhat
in ©
repre:
sumn
of ®
