; monodates
petween the
. be with a
S to test a
:
2
ad land.
sked
sing).
he zones of
cus on the
(Fig.7). The
ler to get an
y continuous
uds" is then
'OT and thc
sification
This method proceeds by iterations in order to create 35 classes,
this number was set has in prior to take sufficiently into account
the radiométric variability of the zone.
We then proceed to a thematic calibration of this classified image.
To leave some zones of practice of which the work of soil is
known and of the general knowledge of the region, we assigned
the 35 initial classes a thematic content.
5. POLL PLAN
Once the classification achieved, we setup a polling plan. For
this, we localize our segments on the satellital image as well as on
the topographic map and proceed with ground data collection
contained in these segments as well as of their surfaces. In our
case, the area units is some segments (700* 700 m) in order to
definc them, a grid is superimposed on the image. Every clement
of the grid corresponds to a block of 100 segments, the sample is
thus in pulling K numbers at random between 1 and 100 (Fig.9).
¥ 22304: 5:090 gogo
C «4 C07 CONAN =
—
Fig. 9: Block representation (100 segments).
This method allows us to get a systematic sample aligned of K
repetitions. Remote Sensing with some images of the year,
represents an important asset for «frame» the results of ground
obtained through the summary some segments. This means that
the agricultural statistics are always finally achieved on the
ground, and that remote sensing is only a support in order to give
back these results of ground more «reliable» in the course of time.
The whole theoretical segments, which cover the totality of the
territory serves a basis of poll for the pull of a sample, through
which could take up statistics of surface.
In the case of a sample no stratified, or with a stratification found
on some steady criterias in the time, the basis of poll does not
degrade from a year to another.
6. GEOREFERENCING OF DATA TO OBTAIN
AGRICULTURAL STATISTICS
Because of the diversity of the basis information used, errors and
imprecisions of all nature can take place at various levels of the
gcoreferenced system.
+ The quality of content of the achieved stratification depends
on, one hand, on the nomenclature and, on the other hand, on the
identification of the land cover on some images straightened
geometrically. The geographical precision of the stratification
depends mainly on some constraints fixed by the sampling plan.
= The radiometric quality of the images is essential in order to
reach good performances in classification. The geographical
precision of images results of course of the spatial resolution of
sensor and of the precision of the geometric correction.
= The structure of the investigated segments is bipolar, with a
vectorial information of drawing some parcels included in the
segment and a list of referenced attributes to the parcels.
6.1 Minimisation of the superimposing errors:
It concerns the adequacy between the position of the segments
established from investigation documents and their coordinates in
the straightened satellital data. The geographical superimposing
of segments in the images allow to reduce errors, which will have
a detenoration of the final quality of the regression.
Image 10: Superimposing ground segments data on the image.
The approach adopted for the inventory of cultures by remote
sensing passes by a certain number of stages of processing and
control. Indeed, it includes four levels of information, which are
the stratification, the satellital data, the segments as well as the
ground statistics. The segments are superimposed on the image
by using of the geometric correction in order to limit errors du to
localization of drawing of reference marks. The crossing of the
stratification with the satellital data allow us to get new layers, on
which we choose some segments.
Some classifications are achieved, thereafter we achieve the
matrix of confusion which is a stage of control. The results of the
classification are thereafter compared to the ground statistics by a
regression. Finally, an indicator is calculated in order to measure
the efficiency of remote sensing.
6.2 Regression estimator:
In order to determine a given surface by remote sensing, we
determine a law of a whole "classification by remote sensing" to
"ground truth." This law could be determined by the least squares
line. In remote sensing, we satisfy ourselves by estimating this
law by a certain number of ground truth and their associate
classification (following figure).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 297
I
IN
|
