nes
ine
nts
rful
lds.
ore
ain
age
out
ude
| to
and
put
nce
ost
nts
ion
| to
een
ues
ysis
ient
d to
| to
red
rity
ige,
ent
and
lant
the
ion
ere
ion
the
am?
C
C
=
300 m and 2500 m above sea level. The area is
characterized by a variety of landcover types, including;
forest areas, open areas and farmland which were suitable
for the purpose of this study.
WB.
su Ann AL M nu La VU 1.30 000 13 OO. 340 (0K
4 070 000 us
4 0635 000
4 060 000 4 060 000 m
N
0 1000 2000 km À
J nap
m
330 007 735 000 TJT O00 ài
|
F5 000 i FAT OV 325 000
Figure 1. Location of the study area
3. Methodology
Using raw bands of Landsat in the classification process is
a widely used way of extracting vegetation maps. But the
statistical similarities of vegetation spectral responses,
spatial resolution of data and presence of similar species
sometimes do not allow obtaining the desired results from
the original bands. As a first step of the study raw bands of
Landsat ETM belonging to Mediterranean region were
classified. The maximum likelihood method was used to
classify the image because, unlike the minimum distance
and the parallelepiped classifiers, this technique takes into
account both the spectral variability within and between
classes (Fahsi et al,, 2000 ).
The classification legend was determined by using the
available data such as forest management maps and
reconnaissance field survey results. While forming the
training data, this legend was taken into account and eight
different vegetation classes; Callabrian Pine, Black Pine,
Taurus Fir, Taurus Cedar, Farmland, Sparse vegetation
were discriminated.
To check the accuracy of the results, ground truth data set
with 26 reference point were determined using the 1/25
000 scaled forest management map of the region. When
the training set was applied on the classified image, an
overall accuracy of 62.96% was obtained, which is not
satisfactory for this kind of studies.
7 BiroundTruthi
Ground Truth Raster...
-
Figure 3. Error matrix of classification performed on raw
bands.
At this step a new method was implemented, in order to
increase the accuracy of the result. Suitable vegetation
indices and image components were produced by using
205
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
Principal Component Analysis (PCA) which is a technique
for removing or reducing the duplication or redundancy in
multispectral images and for compressing all of the
information that is contained in an original n-channel set
of multispectral images into less than n channels or, more
specifically, to their principal components (Ricotta ef al.,
1999).
In this study the main inputs of the feature components are
the indices. Two sets of indices were used; the first set
includes the vegetation indices which directly give the
spectral response of chlorophyll by using the ratio between
red and NIR bands. The second set was used to remove the
soil noise by changing slope value of red and NIR bands.
First set of indices are most commonly used remote
sensing tools for extracting green vegetation cover that
employ red and near infrared vegetation such as
Normalized Difference Vegetation Index (NDVI) (Drake
et al., 1999). In addition to NDVI, Global Vegetation
Index (GVI), Infrared Percentage Vegetation Index (IPVI),
Transformed Vegetation Index (TVI), and Tasseled Cap
Greenness Index were used. Equations of these indices are
given in Table 1.
Normalized NIR-red
Difference NDVI- ----------- X 255
Vegetation Index NIR+red
: GV I--0,2848* TM1-0,2435* TM2-
Global esee 0,5439*TM340,7243*
TM4-40,0840* TM5-0,1800* TM7
Greenness = -0.2848( TM 1)-
Greenniess 0.2435(TM2)-0.5436(TM3)+
0.7243(TM4)+0.0840(TM5)-
0.1800(TM7)
Transformed TVI=100 * [((NIR - red) / (NIR +
Vegetation Index red))/2)*0,5]
Infrared ]
Percentage IPVI= ---- (NDVI+1)
Vegetation Index 2
Soil Adjusted
Vegetation Index
#5 eio NIR-red
Sonal SAV] = IH)
L=0 for low NIR+red+L
vegetation cover
MSAVII = (( NIR-red) / (NIR + red
Modified Soil +L)x(1+L)
Adjusted
sata : L= 1-( 2* slope * NDVI * WDWI)
Vegetation Index 1 WDWI = NIR — slope * red
MR dS MSAVD - 1/2*(Q*(NIRHD)-
* + 2; (NIR- 1/2
Vegetation Index 2 (NIE D-S Rr)
Table 1. Indices used in this study
Principal components of these six indices were calculated
and the lists of image eigenvalue loadings for this
transformation on all vegetation indices are given in
Table 2. According to this table, correlations of PCI with
the indices are very high except IPVI. This means PCI has
a great amount of information of these 5 indices. To
include the spectral information of IPVI in the analysis,
PC2 is used because IPVI has a high loading value in this
component. 92.27 percent of the spectral information was
collected on the first two principal components; PCI and
PC2 are selected as feature components of vegetation
indices.
