The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1126
Target region
coverage classification based on Aster image. (Han.2006)
It is the region about 12km toward east & west and about 11km
toward south & north among construction target region for
multi-functional administrative city (Seijong-si) of Korea.
(Ficture 2) It is total 72.996 in area including Yeongi-gun (3
myeon and 28 ri) and Gongju-si (2 myeon and 5 ri) in aspect of
administrative region.
ULtLat) S6.59Ü!ÊQa UL(LOft) 12?. I ?35361 2
UWLafi 36.4820849 UR(Lon) 127.44257221
5 ROI, vegetation (cultivation), road, water, forest, residence
class is set using maximum likelihood algorithm and 35 training
data sets were designated. Overall accuracy of 94.27% and
Kappa coefficient of 0.86 were acquired.
Figure 2. Testing area
3. EXPERIMENT
3.1 Image
Aster image used in pixel based classification is most high in
resolution and it is VNIR image providing 15m resolution in
visible range and near infrared and multi-purpose satellite
KOMPSAT2 image of Korea is used for classification of high
resolution image. The satellite was launched on Jul. 28th 2006
and MSC with capacity for panchromatic lm, multi-spectrum
4m and swath width of resolution 15km in sun synchronous
orbit of orbital altitude of 685km. Acquired image is *ID:
msc_071005015048_06336_10811273__PS image taken on Oct.
5th, 2007 and high resolution 1 m color image is acquired with
HPF image fusion method. Geometrical correction was process
of affine conversion selecting the 25 points from 1:1000
topographic map. Image re-sampling was completed with
nearest neighbor method. Also, ortho rectification was
processed suing the lm DEM data acquired from contour line.
3.2 Pixel Based Coverage Classification
1) Low Resolution Image Coverage Classification
Supervised classification was carried out for pixel based
Table 1 Pixel based classification accuracy
Overall Accuracy = (73364/77817) 94.27%
Kappa Coefficient = 0.8664
Ground Truth (Pixels)
Class
Road
Forest
Cultivatio
n
Water
Residence
Unclassified
0
0
0
0
0
Road
5437
9
14
54
30
Forest
1
29174
5
17
19
Cultivation
104
77
23114
983
300
Water
70
17
233
9367
59
Residence
211
253
208
804
6272
Total
5823
29530
23574
11225
6680
Figure 3 Land coverage map using Aster image
2) High Resolution Image Coverage Classification
To increase the accuracy of classification using high resolution
image, object based classification method was applied. Land
coverage map was drawn up as result of classification to
compare the accuracy of classification and it is produced as
raster data. Also, accuracy was compared by producing the land
used map with GIS technique using land category provided in
cadastral map and distribution rate for each class. Data
processing work flow chart is as figure 4. RXD detector was
used for extracting algorithm. RXD is the algorithm verified
that it is clearly effective in detecting the opaque and non
transparent object and dead pixel or line does not have
influence in detection though it is abnormally generated.
(Chang et al. 2002) (Reedl.S & X. Yu, 1990)
Work Flow
Image registration & Ortho
rectification ÎGCP & tm OEM)
Sample extraction
Ut 3738/9 100.216029.000
t,fl: 3/0879,100. 219029.800
Feature extraction algorithm
High resolution coverage
classification
Classifying land use Category [■
Grouping
spectral & Shape based
Segmentation (resolution)
Classification
Grouping
J
Themai
S da
¿c map
sses
7~~~~
Figure 4. Work flow of study