lid 
Inc 
se 
gs 
ior 
are 
es 
jet- 
ior 
er 
"he 
pic 
ASE 
r oi 
na 
Fol 
ate 
Sec 
tra 
atoi 
ass 
ac- 
ven 
; in 
ate 
the 
| is 
ars- 
ter, 
axt- 
ass 
are 
of) 
ate 
The knowledge base of the system consists of two 
main parts : 
| - facts about the status of the world model -> GIS. - 
procedures for feature extraction / parameter estim- 
ation. - procedures for predicting future states. 
Il - rules for the selection of procedures for feature 
extraction. - rules for updating the model, given 
data(source). 
Efficiency of hypothesis evaluation is needed because 
with higher resolution (including digitized photos) and 
an increasing number of sources the volume of data 
increases more rapidly that the volume of information. 
Efficiency can be achieved through the top down, 
backward chaining of rule by gathering the statistics of 
the degree of change of probability from prior to 
posterior as a function of {class , feature -> source}. If 
for a certain class a certain feature does not signifi- 
cantly change the likelihood for that class then there is 
no need to evaluate P(Class|feature) for that combin- 
ation in future. This is also the case if for an object the 
P(Class) is so low that it is very unlikely to lead to a 
significant P(Class|f). 
P.M. : Bayes , P(CI|f)xP(f)-P(fÍ|CI)xP(CI) , P(f[|CI) would 
have to be very high to compensate for a low P(CI). 
5 Concluding remarks . 
Most of the present publications on the subject of 
using multisource data is at a pre scientific level of 
picture processing. One of the more favoured painting 
receipes is to play with the IHS transform. Another 
favourite is to throw data of different sources and 
hence incompatible physical dimension together into a 
principal components analysis. This disregards not 
only the incompatibility of the physical units but also 
the restriction of linear transforms to additive vector 
models. 
Discussion with experts in visual image interpretation 
who have looked at e.g. SPOT+SAR -> HSI pictures 
does not provide more representable knowledge than 
can be derived from physical modelling. The useful 
knowledge of interpretation experts is in the field of 
context dependent prior probabilities related to comp- 
lex spatial relations or to complex processes involving 
901 
human activies such as destroying the environment. 
Their expertise is best used in defining sensible 
hypotheses about object's states and about processes 
and the relation between priors and context. 
Progress in computer assisted image analysis is most 
rapid in those areas where models can be defined for 
the relationship between object class, model parameters 
and data(source). Examples are model based analysis 
of buildings [Schutte,1992] and plants in digitized aerial 
photos, the use of vegetation indices and (DEM) model 
based analysis of SAR radar [Woodhouse, 1990]. 
Backward chaining of classification and parameter 
estimation rules allows efficient handling of missing 
data, and the omission of data which is not relevant to 
the evaluation of a current hypothesis. 
The GIS which is used to contain the world model must 
have the possibility to store the relevant P(class) 
vectors as these are required for a multi source 
updation of the model of the world. 
The combination of Bayes and Markov relations can be 
used to estimate states of the system as a function of 
time. 
The above formulated meta rules have resulted in a 
research agenda at ITC aimed at model based image 
analysis, in cooperation with the University of Twente. 
Research into a GIS with likelihoods is executed in 
cooperation with the Rijks Universiteit of Utrecht (the 
Camotius project). 
The definition of the relationship {class , parameter, 
data(source)} is central in the treatment of data from 
various sources. The knowledge base with rules for the 
relation Class -> image processing procedure, is under 
construction in a PhD project [Fang,1992]. Rules for 
Class , parameters, data(source) -> feature extraction 
‚will have to be added (in the problem analysis part).