International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
5. RESULTS AND DISCUSSION
5.1 Transferability test
Table 4 lists the average results of the accuracy assessment of
the four test sites in the two geographic regions coast and desert
(classification of four base classes water, built-up, non built-up,
and roads). The original classification results from the rule base
which was developed for the specific image. The transferred
classification is produced applying one rule base to the other
test sites without making any changes and the adapted one Is
with adjustments to the new test site. The original classification
accuracies with 89% for Coast and Desert are higher than for
the transferred and adapted classification results. The results
show that adaptation of the rule base increases the accuracy to
an acceptable level.
Table 4: Accuracy Assessment of original, transferred, and
adapted classification (four base classes)
Coast Desert
Origi- | Trans- | Adap- | Origi- | Trans- Adap
nal ferred ted nal ferred | -ted
Overall | 89% | 74% | 78% | 89% | 82% | 86%
Accuracy » it
Overall
0.83 0.57 1.0.65 |. 0.6 027 | 0,56
Kappa | ? ; > i
[n addition to the classification of the four base classes, a more
detailed classification was performed. Dependant on the image
content, the class water for example was divided into sea, lake,
and river and built-up into high- and low-density built-up. The
detailed classifications comprise up to eight feature classes.
Comparing the OA of the base and detailed classification, as
expected, it is clear that the base classifications achieve equal or
higher accuracies.
Analysing individual feature classes indicates similar accuracies
of the transferred and adapted classification for the class water.
Therefore it can be assumed that water is transferable.
Adaptation of the rule base results in an increase of the quality
for the classes built-up and non built-up. There is no clear trend
for the class roads. In the coast region, the adapted roads are
worse than the transferred and vice versa for the desert region.
The dependency of the image characteristics can be clarified
especially with roads: Broad, paved roads are more easy to
extract than. narrow, dirt tracks which are built from the
surrounding material.
E
d
f:
utr S ad
dea
= 4 d m NUR ds e 1 ES b Ts
Figure 3: left: IRS image Jijel, centre: transferred, right: adapted
classification result
Figure 3 shows a classification example of the transferability
test. On the left hand side is a subset from the IRS image of the
test site Jijel. In the center the classification result from the
transferred rule base from the test site Mostaganem is displayed
and to the right the adapted result. The transferred result
contains only one water class (dark blue). The sea and one lake
are correctly classified as water, most of the river is not
classified. After adapting the rule base both lakes (bright blue)
are correctly classified by taking into account context
information (neighborhood to other than water classes) and
most of the river is extracted. Classified built-up areas are
displayed in red, non built-up in green, and roads in yellow.
Table 5: Accuracy Assessment transfer of the rule base from
Mostaganem to Jijel (four base classes)
Mostagenam > Jijel
[Original | S
> l'ransferred | Adapted
Jijel
Overall Accuracy | 87 % 80 % 87 9^
Overall Kappa 0.8 0.69 0.8
The accuracy assessment results from the transfer of the rule
base from Mostaganem to Jijel (four base classes) are listed in
Table 5. The OA, as well as the OK of the transferred
classification, are lower than the original classification of Jijel
and the adapted result. By adapting the rule base an
improvement can be achieved which produces an OA and OK
equal to the original classification.
5.2 Base rule set
Similar to the transferability test (section 4.1), three accuracy
assessment steps were performed for each image. The first is the
accuracy of the transferred base rule set, the second is from the
adapted. rule set and the third is the accuracy from the
classification with added rules to the base rule set. Figure 4 is
an example of the calculated accuracies and it reports the OK
for the four test sites, in addition to the accuracy of the original,
individual classifications (four base classes) For example,
figure 4 shows that the OK for Jijel is 0.61 and it increases to
reach 0.65 and 0.7 after the adapting and addition phases,
respectively. The figure also shows that the average OK of the
transferred rule set is 0.67, 0.69 for the adapted rule sets and
0.71 for the classifications with the added rules. In spite of the
higher accuracies achieved after adding new rules, it still falls
below the average OK (0.83) for the original classifications.
The OA shows similar results (Figure 5). The average OA for
the transferred base rules set is 79% and it increases to 82%
after adding new rules. Again, this is less then the average
accuracy (89%) of the original classifications.
The average Kappa values of the coast classification results for
individual feature classes are presented in figure 6. Except for
the class water, the original Kappa values have the best
accuracy. For water the accuracies of the transferred, adapted,
and original classifications are nearly the same. 'The accuracies
of the transferred classifications of the classes non built-up and
built-up are considerable lower than the original Kappa values
but with an average of 0.7 and 0.55, respectively, they are
higher than those of roads. This linear feature class delivers the
worst results with a maximum Kappa value of 0.39 for the
transferred and added rules classification but the individual
classification are not much higher with a Kappa of 0.42. Roads
are quite difficult to extract because of their special geometry
(narrow, elongated) and the image resolution of Sm is not
sufficient to achieve better results with object-based
classification.
1062
Inte
And
ada]
imp
tran
enh
betv
rule
cert.
com
ima,
ada]
"Du Xx [2] e
Overall Kappa
eo
10
«n
Overall Accuracy
tn
—
Fig
Tran:
by 2
new
comp
requi
set
class
imag
