wed the 
plied by 
ological 
ess. 
|Ccuracy 
her land 
s related 
" further 
? spatial 
, Cotton, 
lent for 
1 winter 
ative for 
nplitude 
en tree 
1S cause 
specific 
‚ask and 
endence 
> Israeli 
nverted 
ore, soil 
oth crop 
ion was 
and the 
one o ne 
relationships 
dure 
Cohen, Yafit 
  
checked and defined between certain cluster and land-cover. If a checked cluster was found distinctive, i.e. relates to one 
land cover type, (one-to-one relationships) a merging rule was formalized. If a checked cluster was found mixed, i.e. 
relates to more than one land cover type (complex relationships), a hierarchic splitting rule generation process was 
applied. Primarily, splitting potential, which lies in imagery properties was examined. Splitting criteria that could be 
explained by environmental or agricultural effect were formalized for further classification. Only if splitting criteria based 
on imagery data wasn’t sufficient were spatial properties examined. Actually, most splitting criteria combined imagery 
and spatial constraints. By this mechanism each “mixed-cluster” was splitted into more distinguished sub-clusters, which 
were then merged respectively with suitable land cover classes. Knowledge engineering was involved in order to make 
retrieved relations concept a computer-usable knowledge. They were translated into formal binary rules. Basic rule format 
used here is as follows: 
IF <condition: C; and C, and …C,> Then «conclusion: Li? 
where C; are logical constrains on one or more database features and L; are land cover types. Rule condition referred 
firstly to a mixed cluster and then to a set of constraints. In the conclusion part each pixel was labeled with certain land 
cover type whenever a set of conditions was true. Actually, the condition and the conclusion parts applied the splitting 
and merging actions respectively. Most 
knowledge-based systems in remote 
sensing analysis utilize evidential 
reasoning approach through 
Dempster-Shafer theory of evidence 
(for example: Lee et al, 1987; 
Srinivasan, 1990; Kontoes et al., 1993; 
Peddle, 1995). However, in this 
framework no evidential values were 
attached to rules and as a result no 
separation was made between the = 
knowledge base and the inference 
mechanism and no confidence 
calculations were made to evaluate each > s : t s 
accepted recognition conclusion. Since Figure 3a. Evidential knowledge-based iterative 
there were more than one rule that their classification mechanism 
set of conditions were fulfilled in a 
single pixel and lead to different m 
land-cover, the rules were manually and Oefendes | 
iteratively put in changing order till best 
recognition was achieved. Figures 3a,b 
describe evidential versus definite 
knowledge-based iterative classification 
mechanisms. Primarily, definitive 
approach was applied in order to 
examine the potential of multi-source 
integration approach for crop YES 
recognition in Mediterranean region. 
Progressive process will include 
evaluation of attached evidential values : : ; : 
for each rule, and the application of an Figure 3b. Definite knowledge-based iterative 
inference engine separated from the classification mechanism 
knowledge base. 
  
   
     
  
   
  
   
  
Knowledge base 
(Evidential rules) 
Inference engine 
        
     
  
Change of 
evaluation 
values 
   
     
  
Good 
Recognition 
quality 
Change of 
rules 
  
  
  
      
    
  
     
  
Change of 
rule 
sequence 
  
   
   
     
  
Good LE NO 
Recognition 
quality 
Change of 
rules 
2.1.3 Examples of “split-and-merge” criteria extraction 
Winter crops 
Most winter crops are hardly irrigated as their development considerably depends on rainfall amounts and temporal 
distribution. As a result their phenologies are similar and achieving high differentiation between them is not simple. 
However, it seems that NDVI temporal values reflect real minor modifications between different winter crops’ 
phenologies that are not recognized in unsupervised classification. In order to trace these changes, statistical analysis of 
mixed clusters was applied with ground reference samples and maximum and minimum values were found for 
sub-clusters. Nevertheless, only modifications that reflect real differences between crops were then formulated into 
splitting rules. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 283