International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004
In this study, a method based on logical channel approach is
presented. Limitations introduced by logical channel method are
relieved through adjustment of training set so as to provide
additional sensitivity to ancillary data. Method is applied on a
selected rural land to extract land cover information.
Data and Study Area
A region that shows variety both in morphology and land cover is
selected near Ankara in central Anatolia. Morphological structure
is uneven dominated by volcanic mountainous terrain with
dissected stream valleys; besides there also exist flat regions.
Superior land cover classes in the region belong to the typical
continental environment of central Anatolia. Native vegetation is
mainly rangelands (Anderson et al. 1976), they are composed of
common steppe vegetation species in central Anatolian regions
where typical continental climate is prevalent. The native shrubs
and brushes of the study area are steppe species of maximum 2-
meter height, distributed densely in the terrain. Herbaceous
rangelands of the study area are poorly vegetated lands with
herbaceous plants of maximum 20-30 cm height. In some areas,
herbaceous rangelands are mixed with the native shrubs and
brushes of the study area. Moving through north, particular areas
are dominated with trees composed mainly of coniferous and
partially of deciduous tree species. Apart from the natural land
cover, particular land use classes are present in the study area;
those are primarily agricultural, residential, industrial and
transportational.
A subscene of Landsat 7 ETM of May 2000 including bands
1.2.3,4,5,7 is the primary source of the data analysis. Ancillary
data is composed of Digital Terrain Model (DTM), slope and
aspect of the study region.
A group of data is set apart from the classification and used for
obtaining ground truth information only. Those data are; IRS
panchromatic image with 5 m resolution, forest map form General
Directorate of Forest, digital land cover and land use map from
General Directorate of Rural Affairs, aerial photograph stereo
pairs and field observation data.
Preliminary Data Processing
1/25000-scaled topographical map served as a basis to
georeference all available data. Remotely sensed data used in the
study are free of systematic errors but they have unsystematic
errors due to alterations in altitude and attitude. Geometric
correction is made via Ground Control Points (GCPs) obtained
from topographical map. This procedure is followed by image
rectification.
A 30x30 meter DTM was produced from 10-meter interval
contours digitized from 1/25000-scaled topographical map.
Consequently, derivatives of DTM; slope and aspect map with
30x30 meter cell size were generated.
Classes
Classification level denotes the level of thematic detail for
classification. Since the level of classification is dependent on the
sensor system and image spatial resolution, the level of
classification for the study was set taking the image's information
capability into account. Primary data source for the study; Landsat
7 ETM with 30x30 meters resolution is appropriate for performing
a first level classification (Jensen, 1996).
54
The land cover and land use categories in the study area are
composed of five Level I classes which are;
e Urban and Built-up Land
e Agricultural Land
e Range Land
e Forest
e Water Bodies
From the five Level I classes in the study area, two were excluded.
These classes are Urban or built-up land and water bodies.
Whereas land cover information can be directly interpreted by
means of spectral characteristics of an image, additional
information sources are needed to reinforce the image data in
order to identify whether the area mentioned is an area associated
with human activities (Lillesand and Kiefer, 1994). The data is
usually a thematic map or information regarding the type of use of
a specific area or construction and often becomes more critical
than the spectral data. Since the remotely sensed imagery is the
primary data source for this study, surpass of an ancillary data is
unacceptable. Other reason for excluding built-up land class is
related to artificial human effect. Human factor when exceeded a
trade-off between required development area and present suitable
area, is often challenging. Land use associated with human
activities can be practiced anywhere even unusual, regardless of
the topographical restrictions, but dependent on other parameters
instead. The reason why water bodies were excluded from the
analysis is; clear water bodies with distinct and unambiguous
spectral signatures are the most easily extracted information class
within a multispectral image, hence there is no need to support
classification of such water bodies with additional information.
As a consequence, land cover classes remained are; (1)
Agriculture, (2) Range Land and (3) Forest. At this point,
rangelands in the study area were reevaluated, because; rangelands
of the area obviously consist of two contextually different
categories, which are herbaceous rangeland and shrub rangeland.
A subdivision for rangeland category is made although it may
violate Level I of generally acknowledged classification schemes
(Anderson et al., 1976; CORINE, 1993). As a consequence of this
subdivision; ultimate list of land cover ended up with four classes;
e Agriculture
e Rangeland-shrub (Range-shrub)
e Rangeland-herbaceous (Range-herb)
e Forest
METHOD
Image classification for this study aims to convert spectral data
into four land cover classes. A conventional supervised
classification algorithm; maximum likelihood is selected.
Maximum likelihood classifier clusters pixels into information
classes by means of training data based on probability distribution
models for the cases of interest. (Favela and Torres, 1998).
Maximum Likelihood classifier is the most commonly used
supervised method and is supposed to provide better results
compared to the other supervised methods (Foody et al., 1992;
Maselli et al., 1995).
This study attempts to integrate topographical information into
conventional supervised classification through particular
adjustment on the training data. A five-phased methodological
Ini
fr
PREPROCKESSINC
BASIC
ANAL YVYSES
ACCURACY