'anbul 2004
d using all
lecisions or
irces into a
with respect
we combine
ive analysis
c
on map
nation
ziven in the
> evaluation
> concept of
easure was
; in order to
elaxation of
A formal
set function
follow the
BcZ and
ly.
(5)
ced the so
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
A-Fuzzy Measure: Let Z = Ei Zube the set of available
change detectors.
For each change detector z;to be combined, we associate a
fuzzy measure g, (z ;) indicating its performance in the class À.
For a given pixel, let h, (z;) be the objective evidence of the
change detector z;for the class &. The set of change detectors is
then rearranged such that the following relation holds:
h, (zı )= iE (z,)2 0.
We obtain an change
detectors A = iz. so that A,= z, and A; = A4; ( Uz;.
ascending — sequence of
The fuzzy measures of the obtained change detectors are
constructed as
- g (4) e.) Ag, (4,4) (es)
For each class, A is determined by solving an n-1 degree
equation (Cho and Kim, 1995; Cho, 1995):
n
[[b+2g,(e;)]=1+2 (6)
i=l
Fuzzy Integral: for a given class k the Sugeno fuzzy integral is
computed as
n :
Is(k)- [hog= Max|Min(^, (z; ). 2, (4))] (7)
In the present case 7 is equal to 2. Also, the computation of the
fuzzy integral would only require the knowledge of the
importance of each source expressed by the fuzzy measure (Cho
and Kim, 1995). These quantities can be computed by several
ways. In this study, for each change detector the fuzzy measure
is defined as being the fuzzy accuracy per land cover class
computed on a validation set.
3. EXPERIMENTAL RESULTS
The study area is a portion of a coastal region located in the
north of Algeria (Algiers), for which two SPOT images were
selected to test the validity of change detector combination. The
first image has been captured in May 1989, while the second
image was captured in June 1991 (Figure 2). Due to the weak
rainfall in this period, the study site has undergone important
changes. Therefore, we were interested in those changes from
water, construction, and vegetation to nakedly soil. However,
the satellite data depict other changes caused by the presence of
clouds in the second image. To avoid all surprising effect of this
factor, an additional class ‘X = Clouds’ was taken into account.
X denotes whatever land cover class. Hence, the selected land
cover categories are listed in table 1.
711
3.1. Quantitative evaluation
Unlike the hard classification techniques, the fuzzy set theory
provides several measures for accuracy assessment beyond the
standard error matrix. A number of approaches are available as
the fuzzy distance measure and the fuzzy entropy. In this paper,
we use the fuzzy accuracy (FA) per land cover class as well as
the fuzzy overall accuracy (FOA), for performance evaluation
(Bärdossy and Samaniego, 2002).
(b)
Figure 2. Coastal region of Algiers
(a : image of 1989, b : image of 1991)
Class labels Description (1989 c» 1991)
1 Water = water
2 Vegetation = vegetation
3 Construction = construction
4 Soil = soil
5 Construction = soil
6 Vegetation = soil
7 Water = soil
8 X = clouds
Table 1. Classes of interest