27
Results
y
n
Combination
MSD(Level2)
SD
EN
CE
MI
UI
WA-WBA+
NG+
AMO+
RBV
LPT
61.2172
7.5279
1.9536
2.5195
0.6933
Morph
61.7087
6.8094
2.4341
5.7131
0.7192
DWT
60.2194
7.6022
1.7723
2.2389
0.6911
SiDWT
61.5087
7.1942
2.4854
5.1215
0.7235
Combination
MSD(Level3)
SD
EN
CE
MI
UI
WA-WBA+
NG+
AMO+
RBV
LPT
60.4544
7.5507
1.9186
2.2094
0.6927
Morph
61.7250
6.8123
2.4279
5.7029
0.7191
DWT
59.3618
7.5947
1.8653
1.9139
0.6785
SiDWT
61.5075
7.1966
2.4801
5.1154
0.7235
Combination
MSD(Level4)
SD
EN
CE
MI
UI
WA-WBA+
NG+
AMO+
RBV
LPT
59.7312
7.5428
1.8338
2.0376
0.6958
Morph
61.7268
6.8127
2.4276
5.7014
0.7191
DWT
58.9473
7.5800
1.8069
1.7212
0.6732
SiDWT
61.5045
7.1970
2.4766
5.1124
0.7235
Trough the experiments we mainly focused on the LPT,
Morph, DWT, SiDWT fusion algorithms. From the figure 3 we
can find use the adaptive multi-objective optimization in the
combing coefficients period as figure 2 demonstrates has much
better quality than other method. From the origin image (Figure
1), we find we can not differentiate water region and road
region easily in SAR image which has low resolution and
speckle noise. And the same time the SPOT image has high
space resolution, but the problem is that the same ground object
in the image has the different spectrum and brightness (e.g.
water region).To get more correct information from the two
origins (e.g. Land cover classification and enhance the
resolution of the SAR image), the efficient fusion scheme is
needed. Through the fusion methods tests we can definitely
distinguish waster region and road from fused image use
method (b) than method (a), and the crucial period is adaptive
multi-objective optimization (AMO) in the coefficients
combing to search the Pareto optimal weights as shown in Fig.
3(b). Therefore, the approach to image fusion that uses AMO to
search the adaptive fusion weights is optimal. This approach
can save up the optic features of the images and consider the
huge differences with SAR image, overcome the limitations of
simple combing parameters, and obtain the optimal fusion
performances.
Meanwhile, table 2 shows the evaluation indices of the fused
images from different multi-scale decomposition levels. From
table 2, the MI quality metric matches evaluations more closely
than do the other measures. Prefer the SiDWT the first and
MORPH the second adopted fusion methods and the MI
measure indicates highest values for these two fusion schemes.
For poor fusion approaches like the LPT pyramids and so on,
MI assigns them much lower values, matching evaluations. For
the DWT fusion schemes, MI shows a low value which is quite
lower than SiDWT close to the evaluation ratings from the
observers. CE and UI measures match the trends but not
remarkable in the cases.
As one can expect, it seems that standard quality metrics like
SD, to extract features from the fused image itself show no
apparent correlation to compare fusion method and subjective
evaluations. EN is used to measure global information content
in the fused image. The EN metric produces comparable quality
values for the DWT methods than others. This caused by the
sensitivity of the EN metric to noise and other dramatic
fluctuations in the image as perceptual. The entropy can not
discriminate between useful information and noise.
Furthermore, from the table 2 we can see the higher the
decomposition level is, the MI metric decrease and other
metrics have had no significant difference. These must caused
by the quality of the SAR image which had some speckle noise
and other special characteristics. As levels raised, more details
get in consider and computer, this may cause the introduction of
noise.
The above comments are based upon the average values
over some different region RS images. There may be some
exceptions for some individuals. For most cases, the majority of
information in the fused images is transferred from high
resolution optical images, and the information got from the
SAR is auxiliary but maybe crucial for the information
extraction and other applications.
5. CONCLUSIONS
In the paper, through subjective and objective performance
evaluations for the various fusion schemes applied in fusing
SAR with optical images for the information extraction. The
method using AMO to optimize the weight parameters of
combining is feasible and relatively effective, which can get the
Pareto optimal fusion result. Multi-objective optimization for
the fusion parameters can avoid the limitations of too heavy
dependence on the experience and so simple algorithm such as
choose-max or weighted average methods. The experimental
results show that man usually prefers the SiDWT or MORPH
fusion approaches for the task. The tests of objective quality
measures show that the Mutual Information (MI) better matches
the purpose than do the others.
One aspect that we would like to explore in the future is as
follows.
1) Using prior information. Our research employed little
extra information about the source images other than the pixel
values. If prior knowledge is available, we can attempt to use
such information to improve the performance. Using prior
information in the fusion is an important issue that should be
studied.
2) Research on the reduction of the metrics.There are so
many quality metrics need studied in the world and which
metrics combine is the most valuable and efficient.
Sophisticated and efficient combination of quality measures
should be a topic of future study.
3) Multi-spectral or hyper-spectral images fusion with SAR
for applications. The bands are Bi, B2, and B3 or more and the
