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
The two global landcover data layers were clipped based on the
Chinese administrative boundary in a scale of 1:1,000,000, and
were transformed to the same projection as the reference data.
The reference data was then rasterized with 1-km resolution
and finally these 3 data sets were transformed according to the
crossover of them (tablel). After that, these data were
overlapped and the corresponding results were analyzed by GIS
tools.
2.3 Comparison and evaluation methods
2.3.1 Class area consistency: Class validation checks were
made to see whether the wetlands class characters (e.g. area) of
the global land cover data sets was in accord with that of the
reference data. We calculate the area consistency coefficients
between the wetland-related landcover types from the two
global data sets and the reference data using the following
equation:
) (1)
* 100
Where CC is the area consistency coefficient; Ki is the area of
the number i categories wetlands in the calculated global land
cover data set; and Ni is the area of the corresponding
categories in reference data. The bigger the consistency
coefficients, the better the consistency between the calculated
data we have, and vice versa.
e [i EE
2.3.2 Spatial consistency: The confusion matrix, namely
error matrix, is a standard format of the precision validation
(Zhao, 2003). Generally, the confusion matrix of classified
remote data was calculated to get the Kappa coefficients to
validate spatial agreement between the classification data and
the reference data (Li ef al, 2009). The confusion matrix of
wetlands between the MOD12Q1/GLC2000 and the reference
data were built, and the kappa coefficients were calculated.
36 Others
Another approach to validate the global data sets is the
pixel-to-pixel comparison between different data sets. We
adopt this approach to assess the spatial precision of the two
global land cover products. Specifically, the reference
rasterized data and the global landcover product were firstly
overlapped spatially. The pixels with the same value between
the global landcover data set and the reference data was
retained, while pixels with different values were labeled as
areas of disagreement with new value. So there are nine classes
in combination in the results. Because the non-wetland part
belongs to no data in results, only eight combinations were
summarized. These results describe the spatial location
consistency between the landcover data and the reference data.
Afterwards, the following equations (Wu et al, 2009) were
used to calculate the spatial consistency:
A
O= * 100% (2)
A+B
Where O is the spatial consistency coefficient; 4 is the pixel
account of agreement class, e.g. peatlands/peatlands,
waters/waters; B is the pixel account of disagreement, e.g.
waters/peatlands, non-wetlands/peatlands and so on.
3 Results and discussion
3.1 Class area accuracy
Table 2 shows that the wetland area of these two global data
sets are both less than the area of the reference data as a whole.
However, the GLC2000 data has higher consistency than
MOD12Q1 data with the reference data in area consistency
(77.27%>56.85%). At the same time the area consistency of
wetland water is much better than that of peatlands in both the
global landcover data sets.
Table 2 the results of the class accuracy comparison
Product Wetland 3 Wetland Peatlands ; Peatlands Total 2 Overall
waters(km^) waters(%) wetland(km^) _wetland(%) area(km’) (%)
GLC2000 196065 84.29 60272 37.13 256337 7727
MOD12Q1 171886 98.56 16712 10.30 188598 56.85
Reference data 169444 162315 331759
Table 3 the confusion matrix between two global landcover data sets and the reference data
Reference Peatland Wetland Map User Kappa Overall
data water Accuracy (%) Accuracy (%) coefficient accuracy
unclassified 91.90 59.81
GLC2000 Peatland 3.32 3.78 332 46.92 0.0886 19.81%
Wetland water 4.78 36.41 36.41 88.33
unclassified 95.12 61.62
MOD12Q1 Peatland 0.10 0.55 0.10 14.83 0.0913 18.91%
Wetland water 4.79 37.83 37.83 88.71
The higher consistency of wetland water implies that automatic
extraction of this wetland type can meet requirements in most
circumstances. The inconsistency occurring in the peatlands is
in accord with the results of existed research, which means that
there are still a great many uncertainties in the extraction of
peatlands by automatic computer classification.
3.2 Spatial consistency
3.2.1 The confusion matrixes: As a whole, wetlands in both
the data sets have very poor spatial agreement with the
