Evaluation of forest and nonforest classification capability
of ILU image with dirrerent kinds of pixel size and coherence
Generation methodology
Geometric Correction Reference
Image
One scene of ESR-2 SAR PRI product
image (1997.07.12) has been acquired and
used for rice mapping in this County. DEM
image of this County have been used with
the satellite track parameters and imagery
equations to generate the orth-rectified SAR
image. This orth-rectified SAR image would
be used as reference image to geo-reference
all ILU images.
Evaluation Methodology
The six ILU images had to be geo-referenced
to the orth-rectified SAR image
(image to image) firstly in order to fully use
the ground truth data for signature collecting
and accuracy assessment. Then the
signatures of several typical kinds of land
use type were collected based on the ground
truth data. After that, one classification
methodology would be applied to do the
classification. The classification result would
be recoded to forest and non-forest and
compared with the forest and non- forest
ground truth data pixel by pixel to determine
which kinds of ILU image can produce best
forest and non-forest classification result.
The main methodology was to be explained
simply as the follows.
Classification Methodology
Once a set of reliable signatures has been
created, the nest step is to perform a
classification of the data. Each pixel is
analyzed independently. The measurement
vector for each pixel is compared to each
signature, according to a decision rule, or
algorithm. Pixels that pass the criteria that
are established by the decision rule are then
assigned to the class for that signature.
There are two decision rules used here.
The first decision rule is parallelepiped
decision rule, the data file value of the
candidate pixel is compared to upper and
lower limits. These limits can be either:
- The minimum and maximum data file
value of each band in the signature;
- The mean of each band, plus and minus a
number of standard deviations; and
- Any limits that you specify, based on your
knowledge of the data and signatures. This
knowledge may come from analysis of
signature.
The first limits definition is applied to all
ILU image’s classification here.
The second decision rule is maximum
likelihood. This algorithm assumes that the
histograms of the bands of data have normal
distributions. If this is not the case, better
results may be achieved using the
parallelepiped decision rule. For signatures
such as water body and urban, some
histograms of them are really not normal
distribution. So if we perform a first-pass
parallelepiped classification before
maximum likelihood classification, better
result may be obtained. The developed
classification flow diagram is showed in Fig
5.
Fig. 5 Classification Flow Diagram
Post-processing of classification result
The classification result has several kinds
of land use types: forest, crop, urban, water
body and layover area. Forest and layover
area is combined to form the final forest type
because almost all layover area fills into the
area of forest area in land use map. The other
types such as crop, urban, water body are
combined to form the final non-forest type.
In this forest and non-forest map, there are
lots of class speckles that should be removed
through some algorithm before the
classification result can be used for mapping.
This classification post-possessing can be
called class speckle sieve. The minimum
group of connected pixels that will be sieved
as class speckle may be called sieve distance,
which should be decided according to the
class speckle density, pixel size and ground
truth. For small pixel size, there may be
much more class speckles then big pixel size,
the sieve distance can be set a little larger.
One class speckle filled area can be chosen
as sieve effect assessment window when
compared with corresponding ground truth
area with the sieve distance to change
increasingly. The sieve distance that can
effectively get rid of class speckles in the
assessment window will be used for the
whole studied region.
Classification Accuracy Assessment
Method
There are two kinds of land use types both
in the classification result and the ground
truth data. The two images will be compared
pixel by pixel, the pixel number of forest
classified as forest (FtoF), non-forest as non-forest(
NFtoNF), forest as non-forest (FtoNF)
and non-forest as forest (NFtoF) will be
calculated. The accuracy in percentage can
be computed as:
Accuracy(%)=(FtoF+NFtoNF)/(FtoF+NFto
NF+FtoNF+NFtoF)*100
Both the classification accuracy before
and after classification result post-processing
will be shown and analyzed.
Result and Analysis
Classification result based on the ground truth data of Land Use Map
Classification accuracy result before post-processing
based on ground truth data of
land use map is showed in table 4. From
pixel size 50m to 100m, the accuracy
increases for both old coherence and new
coherence methodology. But the increase
step is very small: less then 1.00%. In the
view of the effect caused by difference
coherence methodology on classification
accuracy, we can find that the result of new
coherence is lower then that of old coherence
by about 1.60% for all the three kinds of
pixel size.
From the classification accuracy result
after post-processing we can find that from
pixel size 50m to 100m, the accuracy
decreases for both old coherence and new
coherence methodology with one step less
then 0.30%. The result of new coherence
was still lower then that of old coherence for
all the three kinds of pixel size, and the mean
difference is 1.48%.
Table 4 Classification Accuracy assessment result based on
ground truth data of Land use map (%)
| Pixel Size(m) |
50m | 75m |
100m |
| Coherence |
OC | NC |
OC | NC |
OC | NC |
| Before post-processing |
73.344 | 71.747 | 73.824
| 72.226 | 74.709 | 73.045 |
| After post-processing | 76.891
| 75.504 | 76.646
| 75.189 | 76.485
| 74.899 |
Note: OC stands for old coherence; NC stands for new coherence
