The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B7. Beijing 2008 
1370 
University(BULC website),GLCC(GLCC website), were used 
mainly to decide candidates of training data. 
3.2.3 Thematic data: Some thematic data such as population 
density data, World mangrove atlas, snow data derived from 
MODIS, and shoreline vector data were used for individual land 
cover class mapping. Population density data used is LandScan 
Dataset which comprises a worldwide population database 
compiled on a 30” by 30” latitude/ longitude grid (Population 
website). World mangrove atlas is initially published by a hard 
copy (Spaldings 1997), and updating digital version is available 
through internet (UNEP-WCMC Interactive Maps website). A 
Global Self-consistent, Hierarchical, High-resolution Shoreline 
Database (GSHHS) is continuously updating vector shoreline 
data available through internet (GSHHS website). 
3.3 Regional land cover maps 
Approximately two hundred existing regional maps were used 
as reference to check intermediate classification result and to 
find suitable locations of training data. Table 1 shows only a 
part of them as examples. 4 
The Atlas of Canada 
National Land Cover Database 2001 (USA) 
Forests and Protected Areas of Brazil 
Forests and Protected Areas of Argentina 
Corine land cover 2000 (CLC2000) 100 m - version 
8/2005 
Vegetation Atlas of China(2001) 
Forest cover map of Insular Southeast Asia 
Agriculture and Forest Land in China,Mongolia and the 
States of the FSU 
Afghanistan Land Cover 2002 
Land Cover / Land Use of South Africa 
2001/02 Land Use of Australia Version 3 (non- 
agricultural and interpretive agricultural land use) 
Australia's Major Vegetation Groups 
Circumpolar Arctic Vegetation Map 
Global Distribution of Current Forests 
Table 1 Examples of regional maps used as reference 
4. LEGEND 
Twenty land cover types including water were defined using 
Land Cover Classification System versions (LCCS2) 
developed by FAO(LCCS website). In the actual mapping 
process, initially defined land cover legend was slightly 
modified through the collection of training data. In other words, 
definition of legend and collection of training data are 
interactive process. Table 2 shows the finally defined land 
cover legend for GLCNMO. 
code 
GLCNMO land cover class 
1 
Broadleaf Evergreen Forest 
2 
Broadleaf Deciduous Forest 
3 
Needleleaf Evergreen Forest 
4 
Needleleaf Deciduous Forest 
5 
Mixed Forest 
6 
Tree Open 
7 
Shrub 
8 
Herbaceous 
9 
Herbaceous with Sparse Tree / Shrub 
10 
Sparse Vegetation 
11 
Cropland 
12 
Paddy field 
13 
Cropland /Other Vegetation Mosaic 
14 
Mangrove 
15 
Wetland 
16 
Bare area, consolidated (gravel, rock) 
17 
Bare area, unconsolidated (sand) 
18 
Urban 
19 
Snow / Ice 
20 
Water Bodies 
Table 2. Land cover legend for GLCNMO 
5. TRAINING DATA 
Collection of training data is the most important part of land 
cover mapping. The following steps were performed in this 
study for fourteen land cover classes: 1 Broadleaf Evergreen 
Forest, 2 Broadleaf Deciduous Forest, 3 Needleleaf Evergreen 
Forest, 4 Needleleaf Deciduous Forest, 5 Mixed Forest, 7 Shrub, 
8 Herbaceous, 9 Herbaceous with Sparse Tree / Shrub, 10 
Sparse Vegetation, 11 Cropland, 12 Paddy field, 13 Cropland 
/Other Vegetation Mosaic, 16 Bare area, consolidated (gravel, 
rock), and 17 Bare area, unconsolidated (sand). 
(1) GLC2000 land cover map was used to select candidates of 
training regions 
(2) Landsat data observed around 2000 were used as a 
background image when training sites are selected by shape file. 
(3) MODIS NDVI seasonal patterns are drawn for each training 
polygon. If different training polygons of the same land cover 
class, for example cropland, have different NDVI patterns, 
these training data are assigned as different subclasses, for 
example cropland A and cropland B. 
(4) Trial of classification by supervised method 
(5) Compare intermediate classified result with regional land 
cover maps. It was also checked by National Mapping 
Organizations participating in Global Mapping project. 
(6) If no good agreement with regional land cover maps, 
unsuitable training data are deleted with the reference of 
Landsat, NDVI pattern, and Google Earth/Virtual Earth. 
Necessary additional training data are selected with reference 
of regional land cover maps and Google Earth/Virtual Earth. 
(7) Modification of training data are repeated until the 
classified result has no serious difference from regional land 
cover maps. 
Final numbers of training polygons are 824 for 105 subclass in 
Eurasia, 492 for 47 subclass in North America, 136 for 30 
subclass in South America, 124 for 39 subclass in Africa, and 
31 for 16 subclass in Oceania. Globally 1497 training polygons 
were collected. 
6. CLASSIFICATION 
6.1 Main classification 
Subclasses of fourteen land cover classes of Table 2 were 
classified by supervised method of MODIS data for each 
continent. Maximum likelihood method by ENVI software was 
applied for MODIS seven bands and NDVI of eight 16-day