IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002
3.4 Used methods of extracting thematic information from
Satellite Images
Updating the database with remotely sensed information re-
quires methods of extracting the desired thematic information
out of the raw data.
As part of this work different techniques are used for identify-
ing the different land-use-classes. For the identification of water
bodies, which are important for the hydrological analysis of the
project area, different classification methods have been tested.
The most suitable method was an algorithm for a decision-tree
(DT) classification. It's based on spectral values and additional
information of DEM and slope.
It's possible to differentiate the spectral class of the water
bodies from the land-use classes “vegetation”, “road and settle-
ment” and “non-irrigated field”. This can be done using the
spectral response of the classes. But as Figure 5 shows, it’s not
clearly possible to separate the water bodies from the shadow
areas. A reliable classification depends on the separation of the
water bodies from the other classes. To achieve this more infor-
mation is needed, because the spectral response is not enough
for the distinction of the classes for sure from the SPOT XS or
from the SPOT XI scene. As additional information, used to
separate the classes, a DEM was used. Analysing the slope of
the possible water bodies was a great help separating the
classes, because water bodies are directly related to flat slope,
while shadow areas are related to more steep slope
200 —-——+—————
180 non-irrigated
160 road and field
140 settlement FA
120 À
z I |
Z 100 | [ |
A
60 ) J | |
vegetation N
40 water 0
bodies shadow
20
0 È
Band 1 Band 2 Band 3 SWIR
Figure 5. The spectral response range for typical object classes
in SPOT2 XS. rsp. SPOT4 XI multispectral data
(Feng et al 2002).
This decision-tree (DT) algorithm was then compared to more
traditional methods like the supervised maximum-likelihood
classification (MLC) and the unsupervised ISODATA method
(Lillesand, et al., 2000).
The DT algorithm showed a very high accuracy in comparison
to the also tested MLC and ISODATA classifications. This is
not that surprising, because additional information was used.
The use of elevation data together with the sensor data was
found to be able to improve land cover discrimination also by
using the MLC algorithm (Haala, et al., 1999).
The extraction of the vegetation-¢lasses was performed in two
steps. At first the vegetation information was extracted, using a
threshold method within the multispectral scene. Secondly,
non-vegetation pixels were masked out and classification of the
combined image from SPOT XS (10.08.1998) and the NDVI
was performed on the vegetation pixels only.
For the land use/land cover classification it's important to un-
derstand the spectral response characteristics of each land use/
land cover type. The irrigated fields are distributed mainly in
valley, the broadleaf, pine and shrub are concentrated in the
hills. The prevention tree belts are extensively distributed over
the whole county.
Two methods, supervised classification (MLC) and unsuper-
vised classification, were performed in the classification. The
classification results. showed that rice and some of the
woodland cannot be distinguished by their spectral response
alone. More spatial parameters including the terrain factors
were used to improve the classification results. This method
was very effective to finally distinguish between the rice fields
and the woodlands.
3.5 Improving the classification with data fusion
To further improve the classification results the high-resolution
SPIN-2 data from 1995 with 1,5m resolution was fused with the
mid-resolution SPOT XI data, taken 1999.
The fused data didn't fulfil all the expectations. Neither the re-
sult of the supervised nor of the unsupervised classification did
improve with the fused data. The data fusion therefore did sig-
nificantly improve the visual interpretability of the data. Digi-
tising linear features and updating thematic maps with the
newly fused data was easier then using just the coloured SPOT
XI data or just the greyscaled SPIN-2 data. The same result was
achieved by using CBERS-1 data and SPIN-2 data. There's
more than one reason for these results. SPIN-2 images are ana-
logue images made with the Russian KVR-1000 camera. The
used SPIN-2 data from the Nanjing area, though offering a high
spatial resolution, also suffer from a low signal-to-noise ratio.
The images were taken at different times, making the fusion
more problematic and less useful. Finally it's important to
remember, that an higher spatial resolution doesn't guarantee a
better classification result.
3.6 Land Cover Map
The integrated land use map was digitised using thematic maps
and other relevant data. Afterwards it was edited and updated
by data processed from satellite images. Linear elements
including roads and rivers, and polygon elements such as
550
