d'Agriculture de Haute-Saone), M. Bruneau (Centre d'Etude Geogra-
phique et Ecologique Tropical), M/s Kilian and Raunet (Institut
de Recherche Agronomique Tropical). ICAR processing has been ap-
plied to eight test sites: three in France, four in Africa and
one in Asia, analysing different problems in each place.
DEFINITION OF REGIONS
The smoothing processes often used to generalise the result of a
classification do not preserve or take into account the spatial
heterogeneity of the original image. Only textural analysis
registers spatial heterogeneity, its one drawback being the diff-
iculty of simultaneous work on several channels. Hence, this paper
tests a process which permits one to use the spatial heterogeneity
of the image in several channels.
1) Classification
The starting point of the work is a four channel image. This
is reduced to a single channel by classifying each pixel.In most
cases a supervised maximum likelihood classification is used.
The number of classes is between 7 and 13.
2) Heterogeneity/Homogeneity of Regions
Whichever classifier is used the result appears noisy, both
because of the way the data are recorded and because of the natural
heterogeneity of landscape. This natural heterogeneity is smoothed
by the scanner as a function of the pixel size. Processing techni-
ques such as smoothing, labelling relaxation or filtering suppress
local heterogeneity e.g. within fields, but present techniques
do not deal with regional heterogeneity e.g. field patterns.
In a visual analysis one attempts to define regions by the
similarity of patterns of units, hence, in digital analysis, one
defines statistical parameters to characterise these regions.
These parameters are the percent of pixels of each class present
computed relative to all the pixels of a scanning window.
If we take: C as the number of classes ( 1$€i or j&C )
Ni as the number of pixels of class i computed
at a given moment in the window,
then Pi, the percent of pixels of class i in the window, is:
Pi socNi/Sj Nj
The numerically computed parameter to define a region is a set
of Pi:
(Pl;P2,P3,P4;...Pi;..5.5Pc)
known as the regional signature.
Given the definition of Pi, the number and the nature of the
classes can be chosen in many different ways. For example, one
may decide to compute statistics of agricultural land cover for
each point relative to all points ( excluding all other classes
such as forest, urban area, water,....) This give a relative den-
sity of agricultural land, and the technique can be repeated for
forest, water, etc.. In addition multiple computations may be
done to produce the density of individual crops, e.g. wheat,
maize, sorghum. f
This process is called MULTIDENSITY.
3) Empirical Definition of Regional Types
Each point is now characterised by a multidensity vector of Pi. To
define regions from this numerical matrix, we need multidensity
vectors for reference. These are obtained from the computation of
Pi within given areas, each one representing a type of landscape
to be identify. This set of multidensity reference vectors can
196
LL
give c
the se
Of a.«
polygc
well }
.4)
belonc
rule.
5)
rent f£
heterc
estime
6)
map.
noisy
comput
a conr
on a 1
to 4r
gurati
red or
parame
size ec
class
(i.e.
of the
"
of che
the wi
that t
inter
surfac
C
graph
multic
to dis
miltic
GEOLOC
T 7
geoloc
and Fı
are sl
]
area e
covere
and me
area c
In .Fic
identi
Mai vi
r
dose ei (A —