ogra- 
cut 
ap- 
nd 
of a 
ial 
diff- 
paper 
eneity 
This 
most 
both 
natural 
oothed 
echni- 
ppress 
es 
the 
one 
sent 
GC) 
uted 
et 
f the 
ne 
for 
sses 
den- 
for 
Pis+ To 
ity 
on of 
cape 
an 
give a complete characterisation of the study area, and thus form 
the set of regional signatures. Each signature defines the nature 
of a component region. The regions of reference delineated by 
polygons from which the reference Vector is calculated must be 
well known areas. 
.4). Vector Decision Rule 
In the C dimensional space of Pi, one decides if one point 
belongs to.one region or another, according to the minimum distance 
rule. The computed distance is$"| xi-vi| : 
5) Choice of the Window Size ' 
The choice of the best window size is made by testing diffe- 
rent sized windows, and choosing the one which best reflects the 
heterogeneity of the region of reference. 
Using the same test on different reference regions, one can 
estimate the compatibility of a given size for each landscape. 
6) Regional Structure - Heterogeneity/Homogeneity 
The aim of the multidensity process is to give a synthetic 
map. The process is successful if the result is at least less 
noisy than the initial classified image. - 
To determine this point, one uses a compactness index. It is 
computed from connexity relations (see for example 1). There is 
a connexity relation if two pixels are neighbours. For instance, 
on a rectangular Landsat grid, one pixel is adjacent by its sides 
to 4 pixels, and by its sides and angles to 8 pixels. These confi- 
gurations are called 4-connexity and 8-connexity. We have conside- 
red only the 4-connexity case. In testing connexity relations two 
parameters have been computed in a scanning window of the same: 
size as above. They are: 
  
  
  
a) N , the number of pixels belonging to the same 
class as the central pixel of the window, 
b) E , the number of "external" connexity relations 
(i.e. when two adjacent pixels have two different labels and one 
of them has the same label as the central pixel of the window). 
The ratio N/E can then be built. N/E is related to the shape 
of the set of pixels of the same class as the central pixel of 
the window: for example, if the N/E ratic is high, this means 
that this class is very compactly located inside the window. The 
interpretation of the ratio N/E is the same as this of the 
surface/perimeter ratio. 
Shapes can also be grouped by dividing the two dimensional 
graph N,E. The use of Nand E to compare images before and after 
multidensity is explained in Figure 1. These parameters permit us 
to distinguish regions by structure and may be a complement to 
miltidensity analysis (see Figure 2). 
  
GEOLOGICAL AND SOIL APPLICATIONS 
  
The ICAR techhique has been applied mainly in soil and 
geological mapping in Africa (Upper-Volta, Mali), Asia (Thailand), 
and France (Vesoul, Tours-Saumur, Caen). Some of these results 
are shown in the paper and the appendices. 
In Figure 3 the different types of rocks identified in Vesoul 
area are displayed.  Schists have been identified because they were 
covered by a mixture of crops; clays were covered by both crops 
and meadows, and dolomites mainly by dry meadows and crops. The 
area covered by this analysis was of about 500 sq km (20km by 25km) 
In Figure 4 the table shows the relation between the regions 
identified by ICAR, and the series previously mapped in the Chiang 
Mai valley (Thailand). 
The example of figure 5, chosen on the same test site as the 
197