The numbered results of this image, show us that thc cercals in a
general manner was well represented with a surface of 33.95% of
the image, distributed in two classes; cereals good output
(5,00%) and cereals weak output (28.95%). Note that this
classification gave good results in spite of the error rate which is
of 1.46% (Array 4).
4.2 Estimation of the classification:
We calculate with the aid of the matrix of confusion. the
percentage of pixels well classified, as well as the statistics of
classification. We notice that the percentage of pixels well
classified for the class cereal good vield is of 73.74% (Array 6). a
number of pixels of this class arc matched with the classes cereal
weak yield and reforestation (Array 5). The other classes appear
well represented with some percentages of pixels well classified
variable between 79.31% and 91.75%.
répartition 1 2 3 4 5 6 7 8 Répartition
classes
Cereals G.Y ac e Rack cM, par ue us 99
Cereals W.Y 15,093 ^". uie 150 116
Courses Ey RN EE RR. 32
Fallow land IUS. WS COSS V orar PCS 79
Bare soil Ph ig 789 5 97
Forest idv cr c UN 110
Maquis HOP T-IR IPSOM 92
irrigualed zones - 5 - - - 2? 11 M9 102
UA
Result. classif. 100 115 45 80 97 110 92 108 747
Array 5: Confusion matrix of classification by maximum
Likelihood.
With the aid of this matrix we calculate some statistics of
classification [FOURNIER ct al.] (Array 6):
— « statistics »: inform us on the classes which were
overestimated as well as the onc that was under estimated by the
classification.
~ «% of well classified” »: give us the percentage of
pixcls well classified by class.
~ «thematic significance” »: allow us to have an idea
on the part of pixels assigned to the class in question with regard
to the result of the classification.
Statistics Statistics % well Thematic
Classes classified significance
Cercals GY 101.01% 73.74% 73.00%
Cercals WY 99.14% 79.31% 80.00%
Courses 86.54% 61.54% 71.11%
Fallow land 101.274 86.08% 85.0094
Bare soil 100,00% 91.75% 91.75%
Forest 100.0094 81.82% 81.82%
Maquis 100.00% > 90.22% 90.22%
Irripuated zones 105.8874 89.22% 84.26%
Array 6: Descended counts of the matrix of confusion.
The global percentage of pixels well classified is in the order of
81.71% this represents a good rate. The assessment of this result
allowed to note a certain number of errors and imprecisions
(confusions. overestimates and under evaluations). Many of these
Statistics = Marge of classification / Marge of. ground truth
Percent of well classified = Diagonal / Marge of ground truth
Thematic significance = Diagonal / Marge of result classification
problems arc du to the fact that the given satellitales monodates
data present a strong level of spectral confusion between the
different types of covering. A good precision will be with a
multitemporal approach. These unprecisions led us to test a
method of classification bv parcel (fig. 7).
Images SPOT XS
Y
Mask agglomération. bad land.
water
| Images SPOT masked |
Y
| mage segmentation |
1 à Non supervised
classification
ground
reality
patterns groupmg |
10 classes
[ land cover classes 3
Fig.7: Automatic classification (parcels processing).
With regard to the previous approach, we masked the zones of
uncovered soil and water surfaces so that to focus on the
following processing on the only zones of interest (Fig.7). The
following phase consists of segmenting image in order to get an
image in which the homogencous regions separated by continuous
contours. An automatic classification bv "dynamic clouds" is then
achieved taking into account. the. XS channels of SPOT and thc
segment image that wc has just created.
IMAGE 8: Resulting image obtamed by automatic classification
(parcels).
296 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
This metl
this numt
the radior
We then |
To leave
known ar
the 35 ini
Once the
this, we Ic
the topog
contained
case, the
define the
of the gri
thus in pu
This metl
repetition
represent:
obtained
the agric
ground, a
back thes:
The whol
territory s
which cou
In the cas
on some
degrade fi
6. (
Because ¢
imprecisi
georeferei
+ The q
on, one h:
identifica!
geometric
depends n
