A rule based system is built up, that takes into consider- For example , Forest": 
ation all the results of the previous steps and some of the IF me LE forest && 
DLM layers. Due to redundant but independently derived forest-pan ss forest && 4 
information contradictions may occur so that the final forest-DLM zm forest 
decision may not be unique and several possibilities for a 
class assignment to a pixel could exist. The rules also try THEN forest certain 1 
to assign reliability or certainty codes. In case of 
consistency, i.e. no contradiction occurs by analysing all ELSE 
GIS layers, the final class is given the code , most 
reliable*. In some contradictory cases, in particular when IF maxlikel ER forest && chat 
DLM data are not consistent to current remote sensing forest: pan m forest && I", 
; forest-DLM z forest hii 
interpretation the final class is taken from the most recent 
data. There is the possibility that something has changed THEN forest ‘certain 2 
or data in the DLM are not correct or the image DLM may not be up to date, 
interpretation was wrong. In any case a reliability code possibly update necessary 
d like ,,not reliable, to be verified" is assigned. 
| The rule system has a hierarchical structure beginning ELSE 
with the conditional check of the most probable result of IF maxlikei " forest && Fig. 
the maximum likelihood classification. If a pixel does not forest-pan Ete forest && 
satisfy any of the rules, also the second probability of the forest-DLM == forest 
maximum likelihood will be involved. The reliability Fig. 
code eventually depends on the level of hierarchy where THEN forest_certain_3 classi 
the final class assignment takes place and varies from should be verified landu 
certain 1 (for the most reliable assignment) down to ax class 
certain 4. The following examples ‘show an overview ELSE landu 
how rules nay be composed for the class ,Forest" and IF | liked = forest m 
Settlement" (for the layer abbreviations see table 1). forest-pan Ex forest && s E 
Let us explain here a few problems that might occur. The forest-DLM == forest | 
map derived DLM forest layer is geometrically very HEN oreste ais à | 5.1TI 
accurate. Small clear cuts or narrow bands through the T P 
® should be verified 
forest where roads or track are located are mapped. Those In aj 
topographic elements might be too small to be detected in n. 
the satellite images. In that case the image classification For example , Settlement: masc 
said clearly that forest exists although it does not in IF maxlikel us settlement && ere 
reality. Though it is unlikely that roads have changed it is texture-pan Ez settlement && 
rather likely that clear cuts have changed. Many changes focal9x9 of == yes 
of forests appear along their boundary. Detected situ_DLM 
differences may be caused by generalisation effects in the 
DLM, or originate in wrong forest classification (eg. THEN settlement_certain_1 
effects of shadows), but there might also have been an . 
actual change. Many of those problems cannot be ELSE 
completely solved by an automatic procedure. The 
algorithm can just give a hint what could have happened IF maxlike2 == settlement && 
and ask the user to check before accepting the program's texture-pan == settlement && 
class assignment. focal9x9 of = yes 
situ_DLM 
The next example should demonstrate how different the 
rules are for the various classes . If the settlement class is THEN settlement. certain, 3 
to be processed, the texture layer and DLM situation 
layer are involved, again together with the maximum ELSE 
likelihood classification. Since the pixel size of the DLM 
layer is very small and settlements are represented as IF maxlikel * settlement && 
individual building blocks rather than as one contiguous T n E e Le 
filled polygon the conditional check employs a focal crue pa is | && 
window filter that tells us whether a settlement pixels has situ_DLM yes | 
been hit or not. — | 
THEN settlement_certain_4 
Fig.2: 
* 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998