i| 2004
ite that
'aterial
pectral
'raphic
ion of
nilarity
present
it has
se two
and, it
tly the
ce CEM
j) other
ded by
onding
. better
CEM
0.96
0.84
f each
vel. So
sure IS
1easure
in this
space,
ond to
ue.
ond to
ues.
ted as
pes of
and all
we are
ore we
W:
ts il.
ci ib
ct if.
ns only
and R1
| target
ASI-32
curacy
5, these
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
e. f.
a - rs
x X 4 KE €
«MT c5 yam e.
&,
bue $us i
“ * er App P 394r V
ee À
d.
"d * 1 ya 6
** Ÿ & € à
8 vr t o. o.
2 a4
i iif. So b
e. i
Figure 4. The similarity and decision images for the analysis of
the CASI-48 image: (a) and (b) for the SVM, (c) and
(d) for the MSAM, and (e) and (f) for the CEM.
CASI-48 CASE
R1 R2 R3 R1 R2 R3
Overall 1.998 | 096 | 095 | 094 | 2006 097
Accuracy
Overall | ges |:087- | 083°] 079 | 0.89] 03s
Kappa
Table 3. Accuracy Parameters of Fusion images.
MSAM
le oe
pO n as
i ; E
e —— ———9
|
pos
|
JA----4-——--—»SVM
el d
] e- s á bus
X
CEM
Figure 6 The 3-D space of decision fusion.
83
Figure 5. The similarity and decision images for the analysis of
the CASI-32 image: (a) and (b) for the SVM, (c) and
(d) for the MSAM, and (e) and (f) for the CEM.
8. CONCLUSION
From the primary evaluation, i.e. from the initial techniques, it
can be conclude that these techniques may be useful for some
applications. For example, if the goal of material mapping is to
extract the boundary of building, the results of CEM are more
reliable for both quality and quantity evaluations. However, in
both image sets, the results are a more or less similar. But at this
level of evaluation, we can also find that spectral similarity
measure can be useful, especially when the used measure can
compensate the linear effects due to the geometry of the scene:
in that case, the result of similarity measures such as the
MSAM should be considered for target detection. In other
hand, the size of pixels is a sensible parameter in the context of
this study. In urban area, the sizes of manmade objects like
residential or non-residential buildings are various. In our case,
related to this subject, we can say that the analyses of CASI-32
with two meters of resolution provide better results. Moreover,
the high spectral resolution is a capability of hyperspectral
imagery but the different between 32 and 48 bands is not
considerable. Also, we can say that the spectral resolution could
be of importance when the materials are nearly similar, for
example different type of roof materials, vegetations, etc. In
addition, it must be added that these techniques are useful for
pure pixel material mapping and that for resolving the problem
of mixing like it might occur at the border of the buildings,
another kind of modelling have to be considered.
