International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
\ 
slope. Together with the moderate correlation eith the CBVD 
and low slope it is suggested that the KBS facilitates 
identification of a class beyond its internal variability 
(heterogeneity) and its similarity to other classes (lack of 
uniqueness). 
  
  
100% |, oio EI QM 
z | 1 
2 = € | | 
m R? - 0.7218 sil 
Q = » | | 
> = n | | 
o | | 
5 75% " 2 bas | 
o | | 
= o e | 
S y = 0.7358x + 0.2597 | | 
= 50% | R? = 0.9416 E 
= | | 
2 | o US-CEM (%) | 
S oo m KBS-CEM (%) | 
25% — ES 
| 
0% 25% 50% 75% 100% / 
Uniqueness Index - PHBL | 
| 
Figure 4: Relationships between uniqueness index (PHBL) 
and classification efficiency measures (CEM) of US and 
KBS classifications. 
6. SUMMARY AND CONCLUSIONS 
The use of KBSs based on evidential reasoning, for land-cover 
mapping based on remotely-sensed images is spreading widely. 
Secondary products of such classification techniques are the 
CBVs which are unique features of the Dempster-Shafer 
algorithm. However, despite the major role of CBVs in KBS 
classification decisions, their analysis has received little, 
attention in the literature. In the present study relationships 
between CBVs of the different classes and the 
accuracy/reliability of their corresponding classifications were 
investigated. The CBVs were found to be good indicators of the 
level of classification complexity on both the pixel and the class 
scales. In this framework we added two new parameterizations 
for the CBV distribution: PHBL and CBVD, two parameters 
which contribute to the analysis of the heterogeneity and the 
uniqueness of a class. 
Correlations were found between US and KBS classification 
efficiency and levels of heterogeneity and uniqueness of a class. 
However, US classification efficiency was much more affected 
by the heterogeneity and uniqueness levels of a class as referred 
by five times higher slopes of the trend-lines. In other words, 
the KBS facilitates identification of a class beyond its internal 
variability (heterogeneity) and its similarity with other classes 
(lack of uniqueness). Finally, contrary to the intuitive 
expectation, CBVs do not indicate the reliability of 
classification. Low CBVs are indicative of complex situations 
or difficulties but do not necessarily imply that they cannot be 
resolved by the KBS. 
7. AKNOWLEDGEMENT 
We thank Arik Solomon and Eyal Sarid for realizing the 
Gordon-Shortliffe algorithm. We thank also to Dr. Victor 
Alchanatis for his helpful comments. 
8. REFERENCES 
Adinarayana, J., and Rama-Krishna, N. .1996. Integration of 
multi-seasonal remotely sensed images for improved land use 
classification of hilly watershed using geographical information 
systems. International journal of remote sensing, 17, pp. 1679- 
1688. 
Cohen, Y., 2000. Knowledge-based crop recognition by data 
sources integration in a Mediterranean environment. PhD 
Dissertation. Israel: Bar-Ilan University. 
Cohen, Y., and Shoshany, M., 2002. A national knowledge- 
based crop recognition in Mediterranean environment. 
International Journal of Applied Earth Observation and 
Geoinformation, 4, pp. 75-87. 
Gordon, J., and Shortliffe, E. H., 1985. A method for 
managing evidential reasoning in hierarchial hypothesis space. 
Artificial Intelligence, 26, pp. 323-357. 
Kontoes, C. C., Wilkinson, G. G., Burrill, A., Goffredo, S., and 
M 'egier, J., 1993. An experimental system for the integration of 
GIS data in knowledge-base image-analysis for remote sensing 
of agriculture. /nternational Journal of GIS, 7, pp. 247-262. 
Peddle, D. R., 1995. Knowledge formulation for supervised 
evidential classification. Photogrammetric engineering and 
remote sensing, 61, pp. 409-417. 
Shafer, G., 1976. A mathematical theory of evidence. Princeton, 
New Jersey: Princeton University Press. pp. 302. 
Wilkinson, G. G., and M'egier. J., 1990. Evidential reasoning 
in a classification hierarchy a potential method for integrating 
image classifiers and expert system rules based on geographic 
context. International journal of remote sensing, 11, pp. 1963- 
1968. 
920 
OB 
D 
ABS! 
Three 
for qu 
just b. 
these 
in col 
more 
3D sti 
and si 
of the 
for tw 
recogi 
The ic 
(objec 
operat 
for r 
(Balts: 
Brunn 
et. Al. 
Jaynes 
Brenn: 
metho 
and ju 
more 
extract 
vaguer 
makes 
intensi 
The o 
expres: 
proced