Evaluation of forest and nonforest classification capability
of ILU image with dirrerent kinds of pixel size and coherence
Generation methodology
Classification result based on ground
truth data derived from TM
Classification accuracy result before and
after post-processing based on ground truth
data derived from TM are showed in table 5.
It showed that almost the same change
tendency has been observed with the
evaluation result based on Land Use Map.
That is to say, before post-processing, the
accuracy increases with pixel size no matter
what kinds of coherence methodology
although there is a little difference in the increment step; After post-processing.the
accuracy decreased with pixel size. In regard
to the effect caused by difference coherence
methodology on classification accuracy, we
also got consistent result with table 4: the
classification accuracy of new coherence
was lower then that of old coherence for all
the three kinds of pixel size.
Table 5 Classification Accuracy assessment result based
on ground truth data derived from TM (%)
| Pixel Size(m) | 50 | 75 | 100 |
| Coherence
method. | OC | NC | OC | NC | OC | NC |
| Before post-processing | 76.310 | 75.120 | 75.691 | 74.661 | 75.243 | 73.598
|
| After post-processing | 72.986 | 71.738 | 73.187 | 71.928 | 73.555 | 72.094 |
Note: OC stands for old coherence; NC stands for new
coherence
Synthetical Analysis
In order to get reasonable accuracy
assessment result, two kinds of ground truth
data have been used for this experiment.
Among them, the land use map covers the
whole Zengcheng County, so the evaluation
based on it can be considered as one in larger
region. Although much more attention has
been paid to control the boundary and type
errors during the digitizing and GIS
processing from hardcopy of land use map, it
was five year older then the currently used
SAR data. So we think it is necessary to use
one forest map derived from TM image as
ground truth data. Because the Landsat TM
image, which was acquired almost at the
same time with the SAR data, has been
carefully selected processed, so we have
much more confidence on the authenticity of
this TM derived forest and non-forest map
then the digital Land Use Map. Since that
the same kind of conclusion has been
achieved from the two kinds of ground truth
data, we think the result we got in this
experiment is reasonable and truthfully.
The accuracy assessment result before and
after class speckle sieving is inversive in
view of the classification effect caused by
different pixel size. Before class speckle
sieving, classification accuracy increases
with the pixel size; After class speckle
sieving, it shows the smaller the pixel size is
the higher classification accuracy will be got.
But in both situations, the accuracy change
from one pixel size to another is small. One
possible explanation maybe that the low
resolution ILU image has high multi-look
number, the image speckle on it must be
smoothed more seriously than small pixel
size, so the class speckle maybe depressed in
some way, as a result, the final classification
accuracy will be a little higher.
It's impossible and also not applicable to
use the raw classification result without class
speckle sieving for forest mapping. So we
think it is much more appropriate to use
post-processed classification result to map
forest and non-forest. Although the accuracy
increment after post-processing is limited, it
is better to use high resolution (low pixel
size in some way) for forest and non-forest
mapping, specially when we need to produce
one large scale land use map.
According to the specification for making
photoplan of remote sensing
[4]
( see table 8),
if SAR image is to be used for making
photoplan in the scale of 1:250 000, the
ground resolution should not be less then
75m. So 50 m resolution ILU image can be
used to generate map with scale to be equal
or lower then 1:250 000; 75m and 100 m
resolution ILU image can be used to
generate map with scale to be equal or lower
then 1:500 000. In consideration of this
specification, the 50m-pixel-size ILU image
is the only choice to generate 1:250 000
forest and non-forest map.
Not only for assessment of three kinds of
pixel size but also for that before and after
speckle sieving we got the same conclusion
that the classification accuracy based on ILU
image with new coherence methodology is a
little lower then that with ordinary coherence
methodology without topography correction.
That is to say, there is no need to apply this
kind of topography correction as used in this
experiment to the IQL System if the ILU
image is planed to be used only for forest
and non-forest mapping.
Conclusion
There is not so much difference for ILU
image with different pixel size to classify
forest from non-forest. ILU image with high
resolution such as 50m is preferred for forest
and non-forest mapping if this kind of data is
available. Otherwise image with low
resolution such as 100m and 75m can also be
used without too much accuracy loss. But if
the National Stand: Specification for Making
Photoplan of Remote Sensing should be in
conformity to, only the 50m ILU image can
be used for 1:250 000 forest mapping; The
75m and 100m ILU image can be used for
1:500 000 and 1:1 000 000 forest mapping.
There is no need to apply this kind of
topography correction as used in this
experiment to the IQL System if the
produced ILU image is only used for forest
and non-forest mapping.
Acknowledgement
All ERS SAR Tandem data have been
processed by IQL System at ESA ESRIN.
Many thanks to ESRIN staff who help us to
obtain and process all ILU images used for
this evaluation.
Reference
-
Urs vegmüller, Charles L. Werner, 1995,
SAR Interferometric signatures of forest,
IEEE Trans. Geosci. Remote Sensing.
Vol., 33, no. 5, pp. 1153~1161
- Urs vegmüller, C. L. Werner, D. Nüesch,
and M. Borgeaud, 1995, Land-surface
analysis using ERS-1 SAR interferometry,
ESA Bulletin, No. 81, pp. 30-37.
- B. Rosich, Li Zengyuan, “Forest
Mapping in China with ERS SAR:
Evaluation of Project feasibility and
Future Perspectives”, Rev. 1.1, July 1998.
- National standard of People’s Republic of
China: Specification for making
photoplan of remote sensing, GB 15968-
1995, published in 1995.12.29.
