151
transition probability
land usage
previous year
exposition
>
Rule
northern slope
=> no wine
cf = 90%
wine
sensor
channels
>
>
>
>
>
>
>
land usage in
training areas
A
water
forest
wine
fruit
Figure 4: Combination of evidence from multiple processsing models
4. CONCLUSIONS AND FUTURE WORK
Human image interpreters apply rules that derive addi
tional evidence from ancillary data for identifying parti
cular classes in a spectral classification. An example of
such a rule is the following:
Wine often grows on southern slopes between 50 and
300 meters elevation.
A MYCIN-like approach combining this type of rule
with probabilities derived by statistical classification is
described in (Desachy, 1989). Rather than interprete
these rules on a pixel-by-pixel basis, we intend to trans
late such a rule to an algorithm that computes additional
evidence for land-use classes from the available geogra
phic data. That is, the rule will be compiled into a
dedicated processing model. Other sources of evidence
in land-use classification lead to new types of processing
models such as the transition probabilities between land-
use classes, as handled in the work of H. Middelkoop at
ITC Enschede (Janssen, 1990).
Future work will concern the combination of evidence
computed by multiple processing models including su
pervised spectral classification, rules on ancillary data,
and transition probabilités in parallel (figure 4). There
are a number of alternate methods of combining eviden
ce, such as MYCIN-type certainty factors (Buchanan,
1984), the Bayesian method (as used in Middelkoop,
1989), and the Dempster-Shafer approach (Gordon, 1985).
At present and in future stages of the RESEDA research,
we will be working on three levels in parallel, thus
conforming to the following general guideline of the
RESEDA project:
- On the application level, we are becoming familiar
with the user’s requirements, and we are trying to
verify our methodology by cooperating with an
environmental project. Currently, RESEDA is in
volved in the Integrated Rhine Project (IRP), which
is directed by the Environmental Ministery of the
Federal State of Baden-Württemberg. One scientist
from our team (a biologist) is permanently engaged
in this part of the project.