reference data (table 3). The peatlands’ precision of the 
GLC2000 is obviously higher than the one in MODI2QI 
(3.32%>0.10%). In both data sets, the wetland waters’ accuracy 
is also better than that of peatlands, which are close at 36.41% 
3.2.2 Pixel-to-pixel comparison: Table 4 is the result of 
consistent coefficients based on equation 2. From table 6 we 
can obviously conclude that the wetland water agreement of the 
two data sets is similar, in which the agreement proportion of 
the GLC2000 and MODI2QI is 36.88% and 40.94%, 
respectively. But the agreement of peatlands of both data sets is 
much lower, in which the GLC2000's coefficient is larger than 
MODI2QI's (3.21%>0.09%). In general, the comparison of 
these two data sets with reference data is not satisfied. 
Table 4 the results of the spatial agreement 
  
  
Data product Peatlands (7o) Wetland waters (%) 
GLC2000 321 36.88 
MODI2QI 0.09 40.94 
  
3.3 Discussions 
3.3.1 Errors distribution among the global landcover types 
In order to find out which landcover types are easily confused 
with wetlands, the global land cover data sets were overlapped 
spatially with reference data and the results were summarized 
into table 5. The landcover types that are not wetland-related in 
global data sets, but are wetland types in the reference wetland 
data, were considered as omissions. 
Regarding to Peatlands, the maximum omission of the 
GLC2000 data set occurred in Herbaceous Cover (closed-open) 
(13). And then Cultivated and managed areas (16), Bare Areas 
(19), Sparse herbaceous or sparse shrub cover (14). Moreover, 
the omission proportion with waters (20), closed deciduous 
broadleaved tree Cover (2) and deciduous needle-leaved tree 
Cover (5) are also great. 
The reason that omission occurred mostly in herbaceous cover 
(45.88%) is partly because of the different definition of wetland. 
The meadow was regarded as wetland in reference data sets. 
Whereas, the regularly flooded shrub and/or herbaceous cover 
is considered as the herbaceous cover in the GLC2000, in 
which the definition of the Herbaceous cover is “herbaceous 
cover, closed to open (>15%)”, therefore, the overlapping 
definition between the regularly flooded shrub and/or 
herbaceous cover and herbaceous cover exists. Due to the 
absence of a special peatlands class here in the GLC2000, we 
transformed the regularly flooded shrub and/or herbaceous 
cover (15) into the peatlands class according to their definition. 
Another reason is the unsupervised classification approaches 
used in the GLC2000 data set, which has low classification 
precision of grassland and marshes so that the producer 
precision of the Herbaceous Cover is only 49.8%, and the user 
precision is also just 40.5% (Herold et al., 2008) . In addition, 
There exists similar spectrum responses of the cropland and 
Herbaceous Cover. Another reason is the mixed pixel. As the 
scattered distribution of the cropland around the wetlands and 
the coarse (lkm) resolution of the GLC2000, the inevitable 
existence of the mixed pixel leads to confusion between the 
peatlands and cropland. Although cropland could represent a 
clear texture feature, the texture feature is not available on the 
images at 1km resolution, while it reflects very clearly on the 
TM images which the reference data is based on. A typical 
example is the flooded wetland in a river valley where lots of 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
   
and 37.83%, respectively. Both the user precision and the 
producer precision of the GLC2000 and MOD12Q1 show that 
the wetland waters’ precision is better than the peatlands’. This 
is same as that of the area comparison above. 
croplands and flooded wetlands merge indiscriminately with 
each other. 
According to the definition of the Intertidal zone/Shoal/Bay 
class in the reference data that is “the coastal beach with 
vegetation cover<30% and the bottom substrates that consist of 
rocks, gravel, mixed sand and stones or the mucky”, this 
wetland type could be partly considered as bare area in the 
GLC2000 data set. A similar situation may also occur in the 
wetland type of the Delta at estuary/alluvial/ sand island in the 
reference data set. In view of the overlapping definition of class 
between these two data sets, we extracted the classes of the 
flooded wetlands, the Intertidal zone/Shoal/Bay and the Delta 
at estuary/alluvial/ sand island from peatlands, merged them as 
new one class "flooded area", and compared them spatially 
with GLC2000. The results show that area of the peatlands' 
omission decrease from 13059 km! to 5193 km?. The omission 
proportions of the flooded area in the Herbaceous 
Cover(closed-open)(13), bare areas(19), croplands(16), Sparse 
herbaceous or sparse shrub cover (14), waters (20) and closed 
deciduous broadleaved Tree Cover (2) are 38.26%, 16.51%, 
14.21%, 10.35%, 9.98% and 1.95%, respectively. 
The mixing up of peatlands and water largely arises from the 
various acquisition dates of satellite images and the different 
spatial resolutions between these two data sets. This 
phenomenon is especially obvious in arid and semi-arid regions 
where the change of wetland water area would be twice or even 
more in a year. Compared to a relatively high spatial resolution 
of reference data source, the mixed pixel in the GLC2000 data 
sources also contribute to a large extent to the omission of 
peatlands. 
The confusion between the peatlands and forest cover classes 
(such as closed deciduous broadleaved Tree Cover (2) and 
deciduous needle-leaved Tree Cover (5)) is possibly relevant in 
the forest swamps in most areas of North-eastern China. 
Patches of forest swamps are distributed in the alpine areas 
forest belt of China, especially the coniferous tree cover and 
mixed forest tree cover in the cool temperate zone (Niu & Ma, 
1985). Forest swamp classes were included in the reference 
data and this may lead to the confusion between peatlands and 
forests, which can be validated in the following analysis of the 
omission's regional distribution. 
The landcover class omission of the MOD12Q1 data set is 
greatly similar with that of GLC2000, in which the main 
landcover types include: Grasslands (10), Croplands(12), Open 
Shrublands (7), Barren or Sparsely Vegetated (16), Water 
Bodies (0)and Mixed Forest(5). The reason is also same as 
those of GLC2000, though they are based on different images 
data. At the same time, the MODIS IGBP product may 
overestimate woody cover proportions (Pflugmacher et al, 
2007). This phenomenon can be well explained by the 
confusion of the peatlands and mixed forest. 
With regard to wetland water, In GLC2000, the landcover types 
that wetland water is mistakenly classified as include cropland 
(16), herbaceous cover (closed-open) (13), bare areas (19), 
evergreen needle-leaved tree cover (4), regularly flooded shrub 
and/or herbaceous cover (15) and snow and ice (21), 
respectively (table 5). Moreover, a similar situation occurs for 
the MOD12Q1 data set, in which the most confused landcover