International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
The two fuzzy change detectors used independent training data
but they were evaluated on the same data set comparatively to
the combiner. The derived FA and FOA measures from the
different approaches are reported in table 2. As can be seen, the
fuzzy integral outperformed the individual change detectors
although the category water — soil has performed somewhat
better with the simultaneous analysis based change detector.
This empirical finding is due to the fact that the difference of
performance between the two change detectors is important. In
such a case the fuzzy integral produces an accuracy lower than
that of the most precise change detector. In other classes the
fuzzy integral gives the best fuzzy accuracy rates yielding to a
significant improvement of the FOA rate. In fact, this
combination rule tends to increase the overall fuzzy accuracy by
equalizing the fuzzy accuracies in individual classes.
Class CA (94) SA (94) FI (96)
1 98.38 89.06 100
2 67.54 73.00 87.30
3 78.52 75.52 90.74
4 83.91 75.94 97.20
5 61.84 82.61 86.76
6 77.21 80.90 88.84
7 17.65 49.69 46.08
8 75.90 66.78 87.39
FOA 70.56 74.43 86.56
Table 2. Fuzzy accuracy values obtained for the individual
change detectors against the combination process
3.2. Visual inspection
The visual inspection of the resulting change detection map
indicates how the considered system generalizes. Since we are
using fuzzy systems, the output is hardened by affecting each
pixel to the land cover class which corresponds to the maximum
fuzzy membership value. Thus, in the final change detection
map, each land cover class takes a particular color. Figure 3
shows the resulting maps obtained for the two change detectors
as well as the combined system. In this figure only the three
change classes (Classes whose labels are 5, 6, 7) are depicted.
As can be seen, the two change detectors produce considerable
misclassification rates. According to figure 3.(a), the
comparative analysis neglects an important change surface in
the river and so detects badly the class water — soil. Moreover,
it presents a considerable amount of omission in the class
construction — soil. Instead, the simultaneous analysis based
change detector presents important overestimation rates in the
classes construction — soil and water — soil. In fact, errors in
the class 5 are related to the clouds which were not selected as
belonging to the class X — clouds, while errors in the class 7
are due to changes in vegetal areas which have not been
considered in the training set. As shown in figure 3.(c), the
fuzzy integral performs better, and can increase the detection
accuracy in the different change classes while reducing the false
alarm rate. This means that the fuzzy integral combines the two
change detectors by extracting the correct decision from both so
that it derives the best final decision.
Figure 3. Change detection maps (a : Comparative analysis,
b : Simultaneous analysis, c : Fuzzy integral)
BI Construction = soil E Water © soil
= Vegetation = soil
|
|
Intern
In thi
was ii
accur:
comb
chang
and t
combi
used
square
classe
demoi
fuzzy
increa
alarm:
chang
given
gives
accura
smalle
the cle
combi
Bärdo:
classif
on Ge
Carlot
remote
surveil
pp. 18
Cho, |
networ
Transc
Cho,
networ
Internc
375.
Decr, |
imager
Geogr
Adelai:
Lambii
multite
change
Remote
Liu, Q
Using I
cyberne
Rosin,
on Geo
Verikas
Gelzini
compar
Science