The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1371
composite periods in 2003 for every 1.5 months. Classified
result of subclass level were merged into fourteen classes.
6.2 Partial mapping
Six land cover classes, which are Water, Urban, Tree open,
Permanent snow/ice, Mangrove, and Wetland, were extracted
independently as follows. Furthermore Antarctica and Regions
over 80 degree north latitude were mapped separately due to
unavailability of the MODIS data in these regions.
6.2.1 Shoreline and inland water: Unsupervised classification
and visual labelling are the main method to extract water area.
MODIS seven bands of three periods in 2003 were used for
cluster analysis. Three 16-day periods in a year were selected
from less snow seasons. Labelling was done with the reference
of Landsat TM/ETM+, GLC2000, and global shoreline GSHHS
data (GSHHS website).
There are mainly two reasons for errors of extracted water areas.
One is that shallow water areas are labelled as land. The other
case is that terrain shadow areas are assigned as water. These
error parts were corrected manually.
6.2.2 Urban: Urban areas were extracted using population
density data (Population website), MODIS/NDVI data, and
DMSP/OLS data. Common areas after threshold processing for
these three data were assigned as urban area (Alimujiang 2007).
The main input data is population density data. MODIS/NDVI
was used to exclude large green area such as a park in
populated area. DMSP/OLS was used to exclude villages in
developing countries with large population.
latitude are added to complete global coverage. Figure 1 shows
the produced GLCNMO. At the time of writing this paper,
validation of the result has not yet completed.
8. CONSIDARATIONS
Through this mapping project, the following things were
learned.
- Trial of land cover mapping by unsupervised classification
had problem in labelling process for clusters with small areas.
To cope with this problem, unsupervised with labelling method
can be applied only partially. This means difficult area for
labelling must be classified by other method, i.e., supervised
method. For this reason, unsupervised method was not used as a
main classification method in this project.
- Many times, more than ten times, of classification trials are
required in the modification process of training data. Therefore
one trial of classification process for one continent must be
completed within a limited time, for example a few days.
Maximum likelihood method was selected by this condition.
- Maximum Likelihood method used in this project has a
condition that number of training pixels of one class must be
more than the number of input variables. For some class such as
cropland, it is difficult to collect large homogeneous training
site. Therefore in order to reduce the minimum number of
training pixels, number of input variables were reduced by
using only eight 16-day periods instead of using all 23 periods.
6.2.3 Tree open: Global percent tree cover data were
produced by the authors using the same MODIS data used in
this study (Rakhmatuloh 2007, ISCGM website). 15% to 40% of
tree cover areas of this product were assigned as “Tree Open.”
6.2.4 Permanent snow/ice:
Permanent snow/ice areas were mainly mapped by threshold
processing of 5-km global percent snow cover data,
MODIS/Terra Snow Cover Global 0.05Deg CMG (MOD10C2)
data set(MOD10C2 website). Some part was mapped by
unsupervised and labelling processing of 1-km MODIS data.
6.2.5 Mangrove: Mangrove were extracted by visual
interpretation of Landsat ETM+ images and by digitizing World
Atlas of Mangrove.
6.2.6 Wetland: The common wetland areas in GLC2000 and
IGBP-DISCover were used as wetlands in GLCNMO.
6.2.7 Antarctica: Shoreline of Antarctica was copied from
Global Map Antarctica which is originally coastlines of VMAP
Level 0 (NGA).
6.2.8 Regions over 80 degree north latitude: Regions over
80 degree north latitude were mapped using GLC2000 because
of no availability of the used MODIS data.
7. MAPPING BY INTEGRATION
The GLCNMO was produced by integration and resampling.
Six independently extracted classes were overlaid on classified
result with fourteen classes by maximum likelihood method for
five continents and two ocean regions. Then approximate 1-km
pixels were resampled to exact 30 arc second pixels. And, five
continental and two island regions land cover data were
mosaicked and Antarctica and regions over 80 degree north
- Without existing regional land cover maps and Google
Earth/Virtual Earth, this mapping project has not completed.
Accumulation of existing regional maps gives reliable
knowledge of global land cover to the map producer, and they
are useful to check the intermediate classified result.
Combination of existing regional maps and Google
Earth/Virtual Earth are effective to select training data.
- Through the water area mapping, we found errors in even the
most reliable existing water area data, GSHHS. Clustering of
MODIS data of selected seasons are found to be effective for
the extraction of water area. However shallow water and terrain
shadow area must be manually corrected.
- Urban, mangrove, and wetland are difficult to be classified by
MODIS data. They need to be extracted by other methods.
Urban was successfully mapped in this project. However global
extraction of mangrove and wetland must be studied in the
future.
- Though cropland was classified by supervised method in this
study, it is a difficult class to be classified due to its
heterogeneity even in the same region. Satellite data of Landsat
or ASTER level resolution are more suitable for cropland
mapping, however such better resolution satellite data has
usually lack of phenological information due to long
observation interval. Cropland mapping is one of future subjects
in global land cover mapping.
9. CONCLUDING REMARKS
As one of the activities of Global Mapping project by more than
100 National Mapping Organizations, a new global land cover