International Archives of the Photogrammetry, Remote Se
nsing and Spatial Information Sciences, Vol XXX V, Part B7. Istanbul 2004
4
data. These data were imported as thematic layers in further 4 RESULTS
n classification process.
> 4.1 Pixel-based classification
e 3.3 Ground Truth
m In the pixel-based classification method, the suitäble image
h For accuracy assessment of classification results and analyses €.g. Tasseled Cap calculation, Principal Component
st comparison of the pixel-based and the object oriented analysis and some suitable ratioing transformations were
n classification methods, a sample ground truth map of forest type applied to create the new artificial bands.
S Was used. It had already been generated for another project in Based on ground truth Map some training area was selected for
this study area through a systematic sampling network. [t each type. The best bands were selected based on spectral
[S contained 193 plots that each plot was ] hectare in size. In properties of training area by seperabilty measures such as
d addition to plantation arca, Six types was recognized by Bhattacharya distance index (table 2). The maximum likelihood
d dominant species frequency of 100 thick trees: classifier has been chosen as a suitable classifier and was
n applied two times on different data set. First, ETM-- bands
Frequency percent of species | (multispectral and panchromatic bands) only accomplished
s Pure Fagus (hecch) | 290 *6 Fagus Orientalis ^
n
The best
artificial bands
[Mixed Fagus | 50.90% Fagus Oras
[Pure Carpins 01 >90% Corpus Beteieg —
Mixed Alnus (alder)
Peal, Pca2, Pca3,
Brightness, Greenness, Pcal, Pca3,
d : ECT: Ei Ratio(NIR-G), Brightness,
y Mixed Hardwood | Other species, under 307% — Ratio(NIR/G ) GATES.
h ; se <n : : Ratio(NIR/R+G), Ratio(NIR /G),
Table 1: recognizing of forest types by fre uency percent of 100 : ;
high va species e TERN E 2002) Katio(NIR-MIR/NIR+MIR), Ratio :
: Ratio(NIR -R/ NIR *R) (NIR/R+G)
To avoid a breaking on the borders of samples due to difference classification. Second, the best artificial processing bands were
between magnetic and 3cographic north (4° in north of Iran and added to ETM+ bands and classification was fepeated,
g geograp (
A) the ground truth map was rasterized to 10 meters Table 2: The ETM+. processing and Post processing bands
tL selected by seperabilty measures
th
id 4.2 Object oriented classification approaches
1s
n In this study, first aim Was to test the suitable object oriented
n classification methods for forest type extraction. So, the nearest
le neighbor classifier, membership functions and integration of
nt both classifiers were examined individually:
in
IS os 4.2.1 The Nearest Neighbour (Sample-based) Method
AN
3 " The multiresolution Segmentation was individually instructed
M oar on georefrenced data by suitable scale parameters and
a homogeneity criterion. The class hierarchy was made with
p MA forest types and non-forest classes (shadow and road).
The experiences in the pixel-based classification showed that
use of only one or few bands could not well extract forest types.
Therefore, the feature Space by use of all mean bands was
defined for description of each class. Some suitable segments
ns Figure 3. sampling grourid truth map of types in the study arca were selected as Samples for each class and then classification
a e was done on segments.
34 pre-processing of data 4.2.2 Membership Function (Rule-based) Method
The first step before the image analysis, it is necessary to per- de ; i
processing of images for classification. Since, the raw data are In according to ground truth information, main three Dominant
received as orbit-oriented images and were registered to any types (mixed Fagus, mixed Carpinus and mixed hardwoods) are
— references; The ETM+ bands were geo-referenced in two steps almost covered in Whole study area. Therefore, a few feature
n in PCI software. First, the panchromatic image was geo- Spaces were selected to separate them, Descriptions were
referenced by a digital 1:25000 maps and ground control points, determined through sample Information by comparison ‚of
Then other images have been geo-referenced by using “ image overlap rates of mean objects in each band as well as by optical
to image” technique. All images in corresponding to ground interpretation. According to results of the Primitive Study, the 2,
truth map were resized to 10 meters resolution by second order 4, 7 and pca3 bands could separate main types rather than other
es transformation and re-sampled by cubic convolution resampling bands. J dub € ; ;
ist method. The membership function Was used in three multiresulotion
ral segmentations. In the first super level, the segmented objects
were generally classified to main types and non-forest classes in
ct class hierarchy by use of mean 2,4 and 7 bands and similarity
ite
expressions. In the middle level, to accomplish the refinement
1107
