China Window

1
China Window
Global Land Cover 2000
2

• Data preparation
• climatic stratification of China
• VGT datas preparation
• remove the cloud contamination
• synthesizing the geo-code image
• reference data
• Classification system
• the land cover classification system-LCCS
• stratums land cover classes from LCCS
• Classification process and result
• Accuracy assessment

3
Working process

• Phase 1climatic stratification and data
  preparation.
• Phase 2unsupervised classification and
  labeling.
• Phase 3accuracy assessment and result integrity.

4
Climatic stratification
Why should we performed the climatic
stratification ?
Vegetation growth and distribution are most
associate with climatic factors such as
temperature and moisture. The study area was
divided into sub-area based on temperature and
moisture condition, this is called climatic
stratification. This will contribute to ?
simplifying the study area into small and
environmentally homogeneous sub-region, which
will help to reduce the confusion in different
sub-regions. ? increasing the mapping accuracy .

5
the climatic stratification in China
According to the situation of the aridity
and the above ten accumulated temperature in
centigrade, at the same time considering the
region integrity, the study area was divided into
nine stratum.
Note T is the above ten accumulated temperature
in centigrade. Ar is the aridity
Ar0.16T/r, where r is the precipitation in
millimeter.
6

Result of Climatic stratification
7
Which data is used for classification ?
Original images
36 ten-days NDVI images
Environment data
8
Data preparation Removal cloud contamination
With the Harmonic Analysis of Time Series (HANTS)
technique , the cloud contamination were removed
for 36 NDVI images.
before removing cloud contamination
after removing cloud contamination
the first ten days in August
9
Comparison of HANTS processing
10
Synthesizing of the geo-code image

• Geo-code data was used as a single band for the
  classification.
• Annual average temperature and precipitation were
  interpolated into 1km resolution
• GScokriging method
• 313 meteorology stations
• DEM data was generated from 11 million
  topographical map.

11
(No Transcript)
12
AHP METHORD
Firstly, three climatic images were normalized to
eliminate the dimension. Secondly, to construct
the judging matrix. For each sub-region, the
matrix is given by the experts who have abundance
geography knowledge. Because of the subjectivity
of different experts, we use a coherence index to
test the accuracy. For the three factors, three
dimension matrix is constituted, and the
eigenvector and the maximum eigenvalue (?max)
can be gained. The eigenvector of the P matrixs
maximum eigenvalue that passed the consistency
test is the fraction of the climatic factors. The
consistency test is calculated with the formula
For every different dimension matrix, three
have a different RI, if CI/RI0.1, the value
given by the expert passed the test.
13
This time, we only selected three experts to
assign the matrix values. Following is the
max-eigenvalue and CR value of nine stratums.
The synthesizing image G(x,y) is calculated by
the following formula, G(x,y) 1T(x,y)
2R(x,y) 3E(x,y) The G(x,y) will be regarded
as a band to take part in the classification. The
three factors coefficient is showed in the
following table.
14
the fraction of each sub-region
15
Synthesized geo-code image
sj
im
nec
nc
cc
tb
ec
sc
16
Reference data
l         11,000,000 Vegetation map in China
l         11,000,000 land-use map of
China l         11,000,000 administration
map l         Beijing, China remote sensing index
map l         1km land-use data
17
Classification system
According to GVMUs requirement, the
classification is performed by using FAOs
classification systemLCCS, For the nine
sub-region, there are 21 classes.
18
Classification process and result
We regarded the geo-code image as a single band
and added to the NDVI image(which already had 36
bands) the sub-region was masked into forest land
and non-forest land. Unsupervised classification
is performed in each sub-region. For the NWC, IM
and SJ region, because of the poor vegetation,
the classification result with NDVI is not
satisfied. The original image of the last ten
days in August is used. For the city, the
unsupervised classification is impossible because
of the different density and chroma, we used
supervised classification to gain the main
cities. overlaying forest and non-forest into one
image. And then mosaic nine sub-region into whole
one.

19
(No Transcript)
20
Classs proportion
21
Class code from LCCS
22
Accuracy assessment
three sample sites Dingxi county in GANSU
province Yanchi county in NINGXIA province Daqing
city in HEILONGJIANG province The reference data
were land cover maps at a scale of 1100 000,
interpreted from TM data in 2000 with intensive
field works. These existing land cover maps
were rasterized into grid data with 100 meters.
And the land cover data from VGT S10 data was
converted into 100 meters resolution. There are
10 classes in the sample sites broadleaved
deciduous forest, slope grassland, plain
grassland, meadow, lake, swamp, farmland, bush,
desert grassland, desert.
23
The sample site distribute
24
Confusion Matrix of Classification in Yanchi
unit
25
Confusion Matrix of Classification in Dingxi
unit
26
Confusion Matrix of Classification in Daqing
unit
27
Meta data of Land cover data
28
Conclusion
The land cover mapping was a complex works. For
the area with poor vegetation, as north western
China, instead of using NDVI dataset, the
original image was used. This time, only one
dataset of last decade of August image, we
believe that multi-temporal images can be
better. The geographic factors is useful for the
classification of the vegetation distribution.
And the AHP method is tested to be reliable, it
will be used widely in the future. For the low
and medium scale land cover mapping, the
stratification was necessary. For all the
land-cover classes, water bodies is easy to
identify, and for the urban area, it is
impossible to identify with NDVI data. The
supervised classification of original image was
used for the urban area. Hants program greatly
improved the classification which removed the
cloud and generated consistent dataset for same
class.
29
Thanks !!!
30
(No Transcript)