International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
rectification accuracy (measured by root-mean-square error)
based on a 2™ order transformation model was 1.7 pixel unit.
Since the study area is quite flat with average elevation of 600
m, there is no need for topographic normalization.
The purpose of the classification is to create meaningful
information class value from image by providing pixels of
similar spectral characteristics. The supervised classification
approach that requires priori information about land cover was
used to categorize the pixels in a group. The GPS data and
visual interpretation was used as the priori knowledge in this
study.
Instead of a specific land-cover classification system, a new one
was prepared and applied in the area in terms of predominant
crops. The scheme firstly categorized into three main classes as
(1) agricultural area, (2) forest and pasture land, and (3) non-
agricultural areas. These major classes are then divided in sub-
classes (Table 1), The pistachio areas are included in orchards
sub-classes of agricultural areas.
2.2.1 Collection of spectral signatures:
A signature is a range of brightness values of pixels for each
band of digital images. These values have statistical meaning
(mean, standard deviation etc.) so that each pixel is assigned to
a class based on how well it matched with this signature.
Collection of signatures can be done by selecting training pixel
samples that are to be representative of particular class.
Training samples for the study area were collected using
ERDAS-IMAGINE's area of interest (AOI) tools. Training
samples were tried to be selected from spectrally pure pixels
representing target class. Training sites were determined with
visual inspection and by overlying the locations of GPS data
onto image. Approximately ten to fifteen training sites for each
land-cover class were chosen by considering a good spatial
distribution. As a result, only the signatures left in signature
files were going to be assigned to the specific classes after
running classification algorithm. These classes are accepted as
real land cover of study area and used for the estimating of
pistachio areas. Creation of signature files by means of training
samples is performed separately for each county. So there were
nine different signature files ready for supervised classification
process.
2.2.2 Classification Process: A variety of classifications
algorithm methods have been used in remote sensing. One of
these methods is maximum-likelihood classification algorithm
which assumes that the data are normally distributed and based
on the probability that a pixel belongs to a particular class
(Jensen, 1996).
In this project, the maximum likelihood classifier is used
because it is still accepted as the most accurate time intensive
classification algorithm compared to the others (Weber and
Dunno, 2001). Maximum likelihood classifier algorithm takes
into account class variability by using the covariance matrix for
each signature (ERDAS, 1999). This algorithm was applied to
previously subseted images of each county for which signature
files had been created. Some photo interpretations were also
used to increase the accuracy of digital classification.
To determine the accuracy of the classification process, an
accuracy assessment test was conducted for all classified
images. Classification results are usually summarized as
matrices which presents classification errors. These errors
includes omission and commission errors. À common error,
appears when a pixel wrongly assigned to a class of classified
image, is measured by user accuracy (Zhuang, 1995). Kappa
statistics is also important parameter for
classification errors. It express “proportionate reduction in error
generated by a classification process compared with the error of
a completely random classification” (ERDAS, 1999).
3. Results and discussion
Area statistics for main crops in both province and county basis
were produced from classified images. Total agricultural land
was determined as 309213 ha in the province. Field crops
constituted 148477 ha area, while orchards and vineyards
covered 160735 ha (Table 1).
Table 1. Disrtribution of agricultural lands by county.
evaluation of
Arable land Fruit crops
Total (%) in (%) in
aeri. total total
land (ha) Área agric. Area agric.
(ha) area (ha) area
S.Kamil 47926 12873 27 35053 73
S.Bey 49837 20833 42 29005 58
Araban 21028 16107 77 4921 23
Islahiye 15046 11905 79 3141 2]
Kargam 28480 10923 38 17557 62
Nizip 55985 12989 25 42996 77
Nurdagi 22835 21096 92 1739 8
Oguzeli 48118 32443 67 15675 33
Yavuzeli 19956 9308 47 10648 53
TOTAL 309213 148477 48 160735 52
Wheat, barley, pulses and chickpea cover 125101 ha (84 % of
field crop areas). Rest of the field crop area is covered by
cotton, maize, vegetables and other crops. Pistachio areas
constitute 58% of the total orchards in the province. Vineyards
and olive groves together cover 35% of the area. The rest of the
area is other fruits. Accuracy assesment test results were given
in Table 2 for each county. Overall accuracy changes between
74.5% and 85.6%.
Table 2. Overal accuracy for each county.
Overall Overall
County Accuracy County Accuracy (%)
(%)
SEHITKAMIL 82,10 NIZIP 77,10
SAHINBEY 80,86 NURDAGI 85:59
ARABAN 77,44 OGUZELI 75,98
ISLAHIYE 74,71 YAVUZELI 81,10
KARGAMIS 79,60 Average 79,38
With this project remote sensing technology has been used to
support area statistics for important crops. Since there was only
county boundaries for project area as a base map, analysis were
carried out on county basis. Cadastral maps that could supply
more details were not available in digital format. Moreover,
high resolution satellite images would be more effective to gel
more accurate statistics and maps.
Similar crop calendars of different crops (cotton, second crop
maize, pepper) caused some classifying errors in the study area.
Interna.
On the
and v
backgre
irrigatic
affects
differen
differen
groves
as man
training
increase
was als
Wrong]
manual
4. Conc
Agricul
use infc
use ma]
of them
land bu
effectiv
Remote
date lan
Is getti
pioneeri
effective
This st
LANDS
current
and accı
Applied
paramet
able to c
area. Th
ture car
study s|
could be
