Istanbul 2004
ociated (o two
ination of the
s calculated by
F. Thiria, S.,
initially the
representative
1 functions in
ts successively
| it possible to
>d partition.
This algorithm
use no a priori
ovided to the
ained depends
uning and its
ming from the
)nen associates
a function of
/ spectrum of
'est within the
referent of the
ie pixels which
the map, by a
1g to recognize
he reflectance
corresponding
> pixel, the unit
ing of the cells
an expert) who
o the cell. This
tion of the data
ectrum and the
fected for him.
ither the whole
be gathered in
ents the whole
e cell and their
98.
applied for the
YC laboratory
e adapted it to
'orithm (SOM)
chott, we took,
idy (600* 1240
reflectances of
> 3 dates of the
lefined in the
akchott are spot
elengths are:
nd.
d.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
B3: 0.79 to 0.89 pun corresponding to the near infra-red band.
The dates of the images are:
11-03-1989 11:44:49
11-12-1995 11:45:42
11-06-1998 11:41:06
We took the same section for the 3 available dates. Each section
consists of 300 lines * 600 columns * 3 bands. They are the
parts of the images on which we have validated data by the
experts and digitized on a SIG. Once the method validated, we
apply it to the totality of the zone of study.
3.2 Constituting of the base of training
We chose the reorganization of the matrix of image made up of
600 lines * 300 columns * 3 bands in 5040 lines and 3 columns.
The lines correspond to pixels and the columns to the channels.
We thus obtain for each pixel its three numerical values in each
band of reflectance. From this new matrix, we transformed the
numerical values for each channel into reflectance value
according to the formula given by the guide of user of
Spotimage society. By supposing that surface is lambertian and
that the reflection is isotropic:
p-Cn*zríiGn *cos9* Es... (1)
Where i- index of the channel, 1 to 3
Gn = calibration gain of channel i, value given on the
header file of the images.
0 = solar zenith angle.
Es = Solar irradiance equivalent to channel i given by
the Spot site, depends on the HRV and the generation
of Spot.
After the transformation of the images matrices into values of
spectral reflectances, we proceeded to the constitution of the
base of training which is used as a basis of application for the
SOM algorithm. This base is consisted of the three dates, by
considering that 3 successive lines can present same information
on an image, we chose then to take a line from three of each
date of image and then to combine their matrix. Thus, we obtain
a base made up of 3 dates of our area of study.
The objective of this stage is to constitute a base where all the
cases of figure of reflectances representing of the units of
landscape are present. The base of training is thus made up; it is
identified as multi dates. Standardization, between -1 value and
| of the values of reflectances, is applied in order to reduce the
variations of the values between the dates within the base of
training.
3.3 Application of the SOM algorithm
After the constitution of the multi dates base of training, we
applied the SOM algorithm, whose principle is explained
previously. We thus, obtained the Kohonen map which
constitutes the spectra of calculated reflectances representative
of the referents of different cells the base of training (see Figure
2). We observe on the map, that the spectra of referents
representing the subsets are localised by complying with the
rule of conservation of topology and similarity: the close
referents present similar spectra. Thus the referents are
organised in the map, from the higher reflectances to the lower.
In addition to the conservation of topology, the characteristic of
this map is the possibility to visualize the pixels spectra of the
same subset and their geographical localisation.
505
carte des référents de la base d'apprentissage
bi ti t4 [i fi d i
z pn 1 od
E red z = © sh 11110
|I: 1 eti ur I DN
zl E +} es
an ri "m ac an = A, qc es
ji i — FA sr
Tan + Dé, great tig po BEST ^e 1 mA 1
Fi cya UT ST!
BR | ponen BE. tr 1 a
t tij." SES !
Fi 1 F1 |i i
i | SNS
nly gs on az ERE BT EE ee
3n KA c Fi | IS i
ET | Nds
ER A e ER EAD
3.3 | 1-4 i i
| — — i i bd i
ciiciies NUI
d m Re 4 FE 22 By | pls 7 za
Fi FA | ii |
aile SEIEN
za mE ga Ch mm. Joe ze, ies
1 ICT { 7 T 1
T 13 | I
=} 1 — A r4 Ets DU EE | i
ani. as, pe, 4 aa queo. TEA m DO An.
[o d = 11 1E | | ]
~ 11 { ; [INI | N
93 94 95 SB 97 98 99 100
Figure 2. Spectra of the referents classified on the Kohonen
map
3.4 The application of the HAC classification
In order to reveal these similar groups, we applv an Ascending
hierarchical clustering (HAC) to the Kohonen map. While
supposing at the beginning that we have 100 subsets, each one
being represented by a cell. Each iteration merges two subsets
in only one and thus reduces the number of subset of one unit.
In this manner, the number of subset falls during iterations to
arrive at only one subset when the algorithm arrives at the end.
It is possible to stop this algorithm with a given iteration and to
thus recover a definite number of subsets.
The application of the HAC requires the use of a criterion of
similarity between the sets. Several criteria are defined in the
literature; we chose that of Ward for this application. This stage
permits to gather the subsets (cells) having similar reflectances,
therefore to gather all local variability of the landscape.
We stopped the HAC to 15 classes. 15 is the optimum estimated
number fixed by the experts on the study area.
We note that each class gathers close cells on the Kohonen map
(thus similar subclasses). Each class on the map is represented
by a colour (Figure.2), it can be also visualised geographically
on the original image.
The visualization of the 15 classes gathering the 100 calculated
subsets, are labelled to make correspond one or more classes to
a landscape unit on the ground. These units present the elements
factors of risk. The class of dunes is an example compared to
the urban zone.
3.5 The multi dates application
After having identified the classes which correspond to the units
of landscape, the following phase consists of the assignment of
the results of the base of training classification on each date of
image. This phase makes it possible to locate each unit of
landscape on each date of images and to visualize the evolution
of each class on the three dates available.
The classification is realised without calculation but by simple
assignment of the 15 classes on each date. Thus we see the
evolution of the two risks on the various dates (stranding and
^
flood) compared to the urban areas of Nouakchott. Figure 3
shows the various stages of the methodology of work.
