International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
Segmentation of image and extraction of the homogeneous
areas as forest types can be develop the classification results.
Hence, the main aim of this study was to compare of the pixel-
based classification method with object oriented classification
techniques on ETM- data for forest type Mapping in mixed
hardwood forests in the north of Iran. Among the object-
oriented techniques, recognition of which method can
accurately extract forest types on ETM+ data was another aim
in this study.
Beside these, in the some studies it was resulted that adding
artificial processed Bands facilitate improving the result
classification. The Ratio transforms are often used in image
processing to reduce radiometric effects of slope, sunlight
angles or seasonal variability (Eva Ivits & Barbara Koch, 2001).
The first three components contain more information contrary to
each band individually. The brightness and greenness axes of
the Tasseled Cap calculation can be useful in topographic
variation and to differentiate between closed forest canopy
conditions (Cohen and Spies, 1992). Then, in the study it was
also investigated that these bands can be improve the results in
the both classification methods.
2 Object oriented method
In general, the object-oriented approach and the image analysis
process can be divided into the two principal workflow steps,
segmentation and classification.
2.1 Segmentation
In a principal definition, “objects” or “local pixel group”
information is basis for the object oriented classification
approaches. Segmentation principally means the grouping of
picture elements by certain criteria of homogeneity
(Hildebrandt, 1996). In the object-oriented approach, an image
is subdivided into homogenous objects by segmentation
analysis. Therefore image will be heterogeneous and contrast of
objects will increase (see figure 1).
Segmentation can be done in multiple resolutions in different
levels of objects by criteria such as scale, color and form.
1B
Figure 1: small windows of study area 453 (RGB) before
segmentation (1A) and after segmentation (1B) show contrast of
objects in a forest area.
2.2 object oriented classification techniques
The meaningful primitive objects, which obtained by
segmentation, can be classified through two methods: Sample-
based classification by nearest neighbor classifier and
Rule-based classification by membership function technique.
The experiences were showed that when several different
feature order objects into classes, the nearest neighbor method
should be used and when only few discrete features can separate
classes from each other, use of membership function is an
optimal choice (Ivits. E. & Koch. B. 2000).
The nearest neighbor classifier, as a supervised classification
method needs training area in a multidimensional feature space.
It would be useful when user has no knowledge to describe
feature spaces. In the nearest neighbor method or the sample-
based method, the primitive objects are classified through
similarity to training units or segments for each class. The rest
of objects in the image are belonged to their nearest sample in
each class. It usually uses spectral information of bands as
feature space for description of classes that are to be classified.
After segmentation, in addition to spectral attributes, the objects
will have extra information such as shape, texture, context
attributes and topological relations between neighborhoods and
other objects. This information can be used for exact extraction
of each class in classification.
In the Membership Function (Rule-based) Method, segments
are classified by membership functions, which are based on
fuzzy sets of object features. Fuzzy logic is a mathematical
approach to quantify uncertain statements (willhauck, 2000). To
be aware of relevant information which can correctly classify
classes play important role in this method as a knowledge-based
system. In this method, the interpreters can define thresholds to
be belonging objects to each class by suitable attribute trough
fuzzy sets.
3. METHODOLOGY
3.1 Study area
The study area is located at educational and research forest,
Faculty of Natural Resources of Tehran University in the north
of Iran between 51?33'12"E and 51?39'56" E longitude and
36°32°08” N and 36°36°45 5” N latitude. The whole forest has
been subdivided to seven Districts. The study has been
performed on three districts (Patom, Namkhaneh and Gorazbon
respectively), are about 3000 hectares (see figure 2). Altitude
ranges from 50 m to 1350 m (sea level). Because of different
aspects and range of altitude, a Variety of forest types have been
established.
&
"esae
em
E. 57)
.. Legend
Parcels
|
« |I|/\/Districtions
3 | B] Study area
[
|
X
{ ;
A bs
\
t 1 Scale
tl 1:80000
Figure 2: location of study area in the research forest of Tehran
University (Kheyrodkenar) in the north of Iran.
3.2 Data
In order to investigate ETM-- data potential for forest types
mapping, a small window on 164-35 Scene from 2nd August
2000 was selected. Except for thermal bands, all multi-spectral
and panchromatic data were used for this study.
[n addition, some ancillary data extracted DEM such as aspect
and elevation maps were resized to spatial resolution of satellite
1106
pr
Tei
rej
ref
Th
to
trui
trat
met