International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. [stanbul 2004
Band #1 Band #3 Band #5 Band #7
Bol 177 7177 P5E.0 A BU ANA
NISI: Pe] a. ini
edeMeeMeMeeAMb poe keen o [p UN xu dj DEC ws
j a PA : EN ; NG
don Tue Was CIN
Ber NE ne t] dd foot naan tte er mines Na dar oa
f Mr nm E FM C RU y 0
1 1 1 1 i Fun ' ¢ LA ‘ £s '
da |: A : 5s acu 1100 Le ee an ry i zn s
d d Dd S 2.5 d A 3 T d
Band 42 Band #4 Band #6
BEL T 10 1 d 256.0 ae 1 256.0 oiu "s x
ude—iebeeeheeeeld E--34412€g——1—3———dl [- ET
Cw 3 0 | on
ux ILI ul |. EN led ln A PU UM TO MIS
NN rae ET
A cur Pate
ga | 2 3N 1. 1 31 Ting ace ha | OK | Cancel
4 ; NN GL hs
Figure 3. Brightness mean value of rural area
Land cover classification system made by ministry of
environment in Korea is used in this study. It has 7 classes
including rural area, forest area, grass area, agriculture area,
wetland area, barren area, water area. Rural area includes
residential area, commercial area, traffic and public facilities.
Forest area includes broad-leaved tree area, needle leaf tree area
and mixed area. Grass area includes green tract of land,
graveyard and hillock. Agriculture area includes rice field, farm
and arable land. Wetland area includes swamp, salt field and
tidal flat. Water area includes river, lake and sea (Park, 2001).
Brightness value in pixel may be variable according to
acquisition time. Difference of brightness value according to
time in forest and grass area may be larger than that of other
areas. So, feature extracting from images is constructed as
database bimonthly (January, March, May, July, September,
November). User should use feature information close to
acquisition time of database for classification. In feature
database, segmentation objects which are generated using two-
neighbour centroid linkage region growing method (Hong,
1991) have feature information. Segmentation objects include
feature information selectively and constructed as database
(Table 2). Figure 3 shows brightness mean value of rural area in
March for example among feature information. Feature
database is constructing in nowadays.
Brightness ‚AnsscHed eap Band ratio
transformation
bl mean brightness mean | b2/b1 mean
bl std. brightness std. b2/bl std.
b2 mean greenness mean b3/b2 mean
b2 std. greenness std. b3/b2 std.
b3 mean wetness mean b4/b2 mean
b3 std. wetness std. b4/b2 std.
b4 mean haze mean b4/b3 mean
b4 std. haze std. b4/b3 std.
b5 mean b5/b4 mean
b5 std. b5/b4 std.
b6 mean b7/b5 mean
b6 std. b7/b5 std.
b7 mean
b7 std.
Table 2. Feature lists for classification training data in database
566
3.3 Processing and result
Satellite imagery information management center (SIMC) of
Korea archives past Landsat images and receives Landsat-7
ETM+ images. Using Landsat image database and reference
database, we are constructing feature database for land cover
classification as mentioned above. When Landsat images are
classified, feature database will help users to operate few steps
for land cover classification as shown figure 4. First,
segmentation should be processed using several parameters
(Figure 5). The segmentation process needs for minimum
parameters considering users don’t have profound knowledge.
Level for combing means threshold of region growing. Scale
means minimum size of segment. Level for merging means
threshold of neighbour segment for merging. Figure 6 shows
segment of Landsat image in this study. Images are classified
using feature database after image segmentation.
| Reference DB |
Landsat image DB
Segmentation
| Feature extraction & training
Landsat
Feature DB image
| Classification |
|
Land-cover
classification
map
Figure 4. Brief flowchart in this study
Internat
TCI Ue
That is
So, use
Figure
Figure
produce
classific
classific
did bar
Accurac
of time
accurac
method
users.
