ınbul 2004
[area are
excluded.
er bodies.
preted by
additional
ze data in
associated
he data is
e of use of
ye critical
sery is the
ary data is
id. class is
'xceeded a
nt suitable
th human
rardless of
parameters
| from the
ambiguous
ation class
to support
mation.
are: (1)
this point,
rangelands
y different
rangeland.
igh it may
n schemes
nce of this
ur classes;
pectral data
supervised
| selected.
information
distribution
res, 1998).
nonly used
tter results
et al., 1992;
mation into
particular
hodological
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004
framework was proposed for developing a procedure for the
integration (Figure 1).
GROUND TRUTH
PRIMARY DATA ANCILLARY DATA RS Pancomatie
LANDSAT bands Topographical parameters Land Es map
— d Aerial Photo
Field Observation
geometric correction of
minor data
geometric correction of
topographical map
: I.
i {oem of |
;| Digital Terrain Mode! (DTM) ||
i |derivation of| derivation o |
i ASPECT SLOPE |
| 1
geometric correction
of band 1.2,3,4,5,7
1st Level Classification
Classes: Agriculture, Forest, Range-shrub, Range-herb
Classification method: Maximum Likelihood
selection of class
spectral signatures
Ti spectral
PREPROCESSING
BASIC
DEFINITIONS
Pi
No
«a correlation Correlation
w
a correlation
_
el] e DTM and SLOPE remained
«| c En
E selection of class
7. topographical signatures
< Ti topographical
» selection of T1 selection of T2
| CLASSIFICATION | CLASSIFICATION |
us ! Input bands: Input bands: |
12,34,5,7, |
a | DTM SLOPE | |
|
|
- mea ©
Jy |
zu
Ol =
zi test with
SIG Ground truth data
con
à
«x
Figure 1: General Framework of the Study
First phase: Procedures involved basically involves
understanding class spectral characteristics. A certain time was
devoted to understanding visual components of land cover classes
in the study area making use of particular band combinations and
other reference data. Training samples were selected for all classes
overall the image, ensuring that they are good representatives of
each information class. Selected training set was tested both for
seperability and representativity, if not satisfied with the results; it
was modified and tested again. This procedure continued since a
balance between sample size and sample error was supplied. A
Training Set Dendogram is used to obtain the results of a
hierarchical analysis of the class signatures in graphic form
(Figure 2). The spectral seperability of signatures were tested by
"Transformed Divergence".
Separability
9.00 9.50 0.99 1.43 1.98
[]11 (36.63x)
2 2 (24,423)
EN SET TES
M 4 4 « 6.74%)
Training Class
Figure 2: Seperability of Initial Training set by means of Transverse
Divergence measurement (1) Agriculture, (2) Range-shrub, (3) Range-
herb, (4) Forest
Class spectral signatures compose the initial training set for thc
multispectral image data. However, this set is not used for training
the classification procedure, rather it serves as prior information
for the later redefinition of training data.
Second phase: Quantification of the relationship between land
cover classes and topographical parameters; elevation, slope and
aspect is involved. Dependening on the significant relationships,
ancillary topographical data that may contribute to improvement
of classification accuracy is determined.
Four land cover classes and the topographical parameters were
tested for correlation. Land cover data involves training samples
of land cover classes; agricultural land, range-shrub, range-herb.,
and forest. Those samples have been collected randomly from all
over the study area and are spectrally good representatives of their
associated classes, so, they formed an adequate test set.
Topographical data merely involves the pixel values spatially
corresponding to spectral training samples.
Point Biserial Analysis is performed for quantifying the
correlation between topographical parameters (interval scale) and
land cover classes (dichotomous scale). The correlation
coefficients obtained ranged between minimum of 0.02 to
maximum positive of 0.65, and maximum negative of 0.41 (Table
1); where 0 denotes there is no correlation, 1 is perfect correlation
and -1 is perfect negative correlation.
Topographic Land cover Correllation
Parameter class Coefficient
Elevation agriculture 0.62
Slope agriculture 0.65
Aspect agriculture 0.02
Elevation range-shrub -0.34
Slope range-shrub 0.48
Aspect range-shrub -0.11
Elevation range-herb. -0.41
Slope range-herb. 0.08
Aspect range-herb. 0.06
Elevation forest 0.1
Slope forest -0.5
Aspect forest 0.24
Table 1: Point Biserial Correlation coefficients for four land cover
classes and topographical data
The result of the point biserial correlation analysis indicated the
relation between specific land cover classes and the topographic
parameters. The significance test verified that correlation
coefficients greater than approximately 0.30 are significant.
Significance level, often called the p value is the probability that a
statistical result as extreme as the one observed would occur if the
null hypothesis were true.
As a consequence of the point biserial correlation analysis; aspect
parameter with very low correlation coefficient was incidentally
excluded from the remaining part of the study. Elevation and slope
data were quantified for use as ancillary input for classification.
Third phase: Ancillary topographical data; elevation and slope
were examined for topographical signatures selection. A
procedure similar to that performed in the first phase was carried
on. However, this time the aim 1s to define the representative sets