multi-
for the
values.
ion or
'opped
uld be
nd the
: cloud
on and
ection,
stically
IS data
'82
VIODIS
AVHRR
«
by TM
onal to
.data by
ustering
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
(Pan et al., 2003; USGS - NASA Distributed Active Archive
Center, 2004a), maximum likelihood classification (Liu et
al, 2003), decision tree classification (Townshend et al.,
1999), and hybrid algorithm of decision tree and neural
network (Friedl et al, 2002). Running et al. proposed a
simple new logic for classifying global vegetation based on
the structural characteristics of vegetation (Running et al.,
1995). Since the classification result in this study is the
input parameter for the hydrological model, the critical
feature of the classification method is to control the
classification result using simple parameters.
2.2 Data used
MODIS products are being used by scientists from a variety
of disciplines, including oceanography, biology, and
atmospheric science. Most users will obtain data products
by ordering them through an ordering system (Conboy). The
product named "Surface Reflectance 8-Day L3 Global 250m"
(abbreviated as MODO09QI) is used in this study. MODIS
250 m Surface Reflectance is a two-band product computed
from the MODIS Level 1B land bands | and 2 (centered at
648 nm and 858 nm, respectively). The product is an
estimate of the surface spectral reflectance for each band as it
would be measured at ground level if there were no
atmospheric scattering or absorption (USGS - NASA
Distributed Active Archive Center, 2004b). The 45 periods
of MOD09QI products in the whole of 2002 are re-projected
to the Equirectangular (latitude/longitude) projection by
means of MODIS Reprojection Tool (USGS - NASA
Distributed Active Archive Center, 2004c). The spatial
coverage is from 50 degrees north to 20 degrees north in
latitude, and 90 degrees east to 150 degrees east in
longitude, and the resolution is 7.5 arc seconds. The
following multi-temporal metrics are derived from the time
series:
1. Annual maximum NDVI (NDVI ann max)
2. Annual minimum NDVI (NDVI ann min)
3. Annual amplitude of NDVI (NDVI ann amp)
4. Annual minimum band 1 reflectance (Refl ann min)
5. Annual minimum band 2 reflectance (Ref2 ann min)
6. April minimum NDVI (NDVI apr min)
7. June maximum NDVI (NDVI jun max)
8. August minimum NDVI (NDVI aug min)
Annual maximum/minimum of NDVI/reflectance (1,2,4, and
5) are the *quasi" maximum/minimum value, that is, the
second maximum/minimum value in the 22 periods from
April 23 to October 8 is selected for the purpose of avoiding
the snow and erroneous data. Annual amplitude of NDVI is
derived from NDVI ann max - NDVI ann min. April
minimum, June maximum, August minimum NDVI is pure
maximum or minimum value in the period from Apr. 7 to
May 9, Jun. 2 to Jul. 4, and Aug. 5 to Sep. 6, respectively.
2.3 Algorithm
The supervised decision tree classification method is
selected in this study for the reason of the easiness in
controlling of the classification result. The decision tree
classification algorithms has significant potential for land
cover mapping problems (Friedl et al. 1997), and its
performance is acceptably good in comparison with that of
837
other classifiers, except with high-dimensional data (Pal et
al., 2003). The scheme of the classification method in this
study is shown in Figure 2. The eight threshold values are
arranged in the each steps of the decision tree, and the result
of the land cover classification is tuned by these threshold
values. Ref2 ann min is applied to extract water area since
the spectral reflectance of water in near-infrared wavelength
band is much lower than that of land surface.
NDVI ann max represent the most active status of the
vegetation in the year, hence it is applied to categorize the
non-vegetation and less-vegetation from other vegetated
land. NDVI ann min is applied to the vegetated area for the
discrimination between evergreen and deciduous vegetation.
Refl ann min is also applied to the vegetated area to
discriminate tree type and grass type vegetation based on
the rough assumption that the tree type land surface is
generally “darker” in visible wavelength as against grass
type land. The time series of NDVI apr min, NDVI jun max,
and NDVI aug min are jointly used to extract double
cropped agricultural area especially in the downstream of
Yellow River basin. The first cropping season is from middle
of February to end of May, and second season is from July to
middle of November in this region. This phenological
characteristic is unique compared with other natural
grasslands or single cropped agricultural fields. Therefore, if
NDVI apr min is greater than NDVI jun max and
NDVI jun max is less than NDVI aug min, that pixel is
categorized as double cropped agricultural field.
Réf2 ann min.
+ True
Water >< Threshold : False
d Ce Threshold 1: 0.03
NDVI ann max Threshold 2: 0.11
Non-vege >
Threshold 3: 0.375
Threshold 4: 0.375
Threshold 5: 0.029
Threshold 6: 0.3
Threshold 7: 0.5
Threshold 8: 0.026
| <Threshold2”
NDVI ann max
Less-vege . « Threshold3 7
NDVI ann in in
<
^w. Threshold4^
YN, AP mb
| Evergreen | | Deciduous |
Réfl ann min Réf] ann min,
^x Threshold8^ "«Threshold5^
Forest Grassland Forest | Grass type |
A
-”
NDVI anfi_amp » Thresholdé NN
NDVI_ann_max > Threshold? ae time Senes
Y M
Agri. (double)
Grassland Agri. (single)
Figure 2. Flow of the land cover classification
