EE ER 
one used by OPIT (*) is the result of a process involving repeated 
use of computation of multidensity and compactness parameters. The 
following procedures have been done: 
l- Choice of regions of reference (i.E. training fields) 
to compute: 
* the multidensity vector of reference for 
this kind of region 
* The most appropriate size of window d 
2- Reclassification of the per point classified image 
using the multidensity technique and regional signa- 
tures obtained in l. 
3- Deletion of all the transitional areas which are 
created by the window used in step 2. These areas are 
often linear and can be recognized by a compactness 
analysis. 
4- Expansion of the remaning parts. 
5- Definition of compact forest using the per point 
classified image ( see figure 2 for this work on a 
part of this image). 
6- Selection of the linear water bodies (i.e. rivers and 
swamps, lakes,...). 
7- Combination of 4-5-6 (5 and 6 being more imporatnt 
than 4). 
The result of this work is shown in figures 5-1 and 5-2.The 
step 1 follows the per point classification which is not reported 
here. Phases 3 and 4 are necessary due to the use of the window 
which is often large (11 by 11 is the smallest size of window of 
all these experiments). On this test site we were working on soil 
mapping in agricultural land where forest was also spread out. With 
step 5 we only keep the main forest where it was of no use to 
publish the map obtained by multidensity. In the step 7 the mask 
of this compact forest overlays the unuseful results of multidensity 
COMPARISON WITH MANUAL INTERPRETATION 
  
  
In the experiment done on the Chiang Mai region in Thailand, 
we have not compared a manual interpretation and an unsupervised 
numerical process, but attempted to analyse how these numerical 
analyses can help the photointerpreter. So one defines training 
fields, firstly in each regign delineated by the interpreter to 
check the obtained boundaries with those of the interpretation, 
and secondly in two slightly different parts of one region to 
know the extent of the variability within particular kind of 
landscape that the interpreter may distinguish but cannot map 
(see fig 6). 
Although not quantified, the results (fig 7) have been esti- 
mated to be very similar for the first point and helpful for the 
second. A detailed description of this experiment can be found 
in +2); 
RELATIONS BETWEEN SCALE AND SIZE OF WINDOW 
  
  
A previous study (6) conducted with about 100 soil and 
geological maps of different countries (Europa, Africa, US) and 
ofdifferent scales (between 1/5.000 and 1/3.000.000) has shown 
that there was a fairly constant mean width of unit drawn on maps. 
(The estimated relation was of Log, _1/W = -.4 -.14Log, Scale 
for Europa and US, where W is the méan width of units Por a map; 
see also (3)) 
Using an average of .9cm encountered with physiographic maps 
(*) OPIT: Operation Pilote Interministerielle de Teledetection. 
  
198 
(the 
estimr 
proce 
(the 
exper 
1/400, 
1/300, 
1/200, 
1/100, 
mean 
it ce 
ion i 
Other 
great 
windc