International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
of the ruptures of the coastal dunes, nouakchotian arca
(composed of gypsum and shells) witch [facilitates the
resurgence of the salted sheet of water. See the profile shown in
Figure 1 (Elouard, P., Faure, H., 1972).
2
5
2
P
=
@ SPOT Image
Copyright
1998 p
SCHEMATIC PROFILE OF THE AREA OF NOUAKCHOTT TO 16' 50'N |
..Littoral —. lagoon of marine continental
dune - “the Aftout — *— nouakchotian — +— fixed dune
area
44 beach lives dune
5
A m
ü Ea at^
Ted
HH Risk Mood by marine Incursion
Zz Stranding Risk
8 Flood Risk by water resurgence
— Limit of the validation zane
Administrative limit
Area of Nouakchott, image spotd 06-11-1998
Geographical projection, datum WGSS4
Figure 1. Context and factors of risk on Nouakchott
the most significant landscape unit, concerned by the risk, is the
urban zone classified in three categories, dense, less dense and
dispersed.
For the risk of stranding, the landscape unit factors of this risk
are in the East of the city, it acts of the old dune (red dunes
composed of organic materials of the quaternary period), the
ergs (large wind dunes) and of the dunes sharp.
2. THE SELF ORGANISING MAP OF KOHONEN
In the Self Organising Map of Kohonen model (SOM)
(Kohonen, T., 1984) the map is a square grid of 10*10 cells.
The SOM model allows to partition the whole of the spectra of
reflectances in 100 homogeneous and compact subsets (Fritzke,
B.. 1994) and (Badran, F., Thiria, S., Yagoub, M., 2001).
Each subset is associated to one cell of the map and is
characterized by a prototype spectrum which we call referent.
This model carries out the property of the conservation of
504
topology, Two close cells on the map are associated to two
close referents in reflectances space. The determination of the
referents and the partitions associated with cells is calculated by
the algorithm of Kohonen (Anouar, F., Badran, F., Thiria, S.,
1997).
The application of this algorithm requires initially the
preparation of the training database formed of representative
spectra of the studied problem. The algorithm functions in
several iterations, with each iteration, we presents successively
the spectra of the base of training which make it possible to
adapt the referents and the subsets of the associated partition.
The final map is obtained after several iterations. This algorithm
is known as "with not supervised training", because no a priori
knowledge, on the base of reflectance, is provided to the
algorithm. Thus, the quality of the map obtained depends
especially on the quality of the base of training and its
representative of the varieties of the spectra coming from the
studied problem.
At the end of the training, the algorithm of Kohonen associates
each cell of the grid a referring spectrum and a function of
assignment making it possible to assign any spectrum of
reflectance to the cell having the referent nearest within the
meaning of the Euclidean distance. The spectrum referent of the
cell will be representative of all the spectra of the pixels which
arc affected for him. By labelling each cell of the map, by a
given landscape unit, we obtain a model allowing to recognize
the unit of landscape in any pixels of the image.
Indeed, the assignment function attributes the reflectance
spectrum (in a given pixel of a map cell) to corresponding
referent. This referent will attribute in its turn, the pixel, the unit
of landscape which is affected for him. The labelling of the cells
of the map could be done by an expert (a physician expert) who
will analyze the spectra of reflectances allotted to the cell. This
step is facilitated by the possibilities of visualization of the data
which this algorithm allows.
For a given cell we can visualize the referent spectrum and the
subset spectra of the base of training which are affected for him.
In addition, the 100 cells of the map (10*10) gather the whole
of the data in 100 subsets which can them also be gathered in
relatively homogeneous sets the Figure.2 represents the whole
of the referents spectra assigned to each respective cell and their
classes. These cells are organized in similar groups.
3. METHODOLOGY
The followed methodology in this article was applied for the
classification of the colour of the ocean in LODYC laboratory
(Niang, A., Gross, L, Badran, F, 2003). We adapted it to
our problems on images with high resolution.
To apply the Kohonen Self Organising Map algorithm (SOM)
to the Spot images covering the area of Nouakchott, we took,
the sections of images covering the area of study (600*1240
pixels); this section is shown in Figure |.
Remember that the objective is to characterize reflectances of
the units of landscape, described above, from the 3 dates of the
same season on the area. Methodology is defined in the
following stages.
3.1 The satellite images
The used images in this work on the area of Nouakchott are spot
images made up of three bands of which the wavelengths are:
B1: 0.5 to 0.59 uum corresponding to the green band.
B2: 0.61 to 0.68 um corresponding to the red band.
Inter
B3:
The
11-0
11-1
11-0
We |
cons
part:
expe
appl
32
We
600
The
banc
num
accc
Spo
that
Whe
Afte
spec
base
SON
con:
on i
date
a ba
The
Case
lanc
ider
| of
vari
trair
3.3
Afte
app!
pre
con:
of t]
2).
repr
rule
refe
Org:
In a
this
sam