Results and Discussions - Hyperspectral Image Fusion - page 152

Table 10.2 Performance measures for various techniques for visualization of the moffett 2 data
Fusion technique Variance Entropy Avg gradient Relative Fusion Fusion

bias b

σ

2

H g ¯ factor FF symmetry FS

Bilateral filtering technique 476.86 6.02 4.65 0.22 1.62 0.24

Bayesian technique 487.65 5.68 6.10 0.33 1.81 0.30

Variational technique 467.94 5.95 4.75 0.17 1.79 0.32

Optimization technique 593.04 6.12 5.52 0.39 1.62 0.32

Three band selection 481.41 5.59 4.86 0.45 1.01 0.66

Piecewise linear function 421.35 5.93 4.35 0.20 1.64 0.29

Color matching function 403.44 5.94 4.34 0.22 1.64 0.26

(Bold font denotes the best performance)

It can be inferred that the selection of three bands for display may lack in information
content, but it may not bear any relationship with the sharpness or the contrast in
the image. It may select the bands possessing high amounts of visual contrast and
sharpness. The values of relative bias b are low for all the techniques except the band
selection and the optimization-based one. The selection of bands deals with only a
small fraction of the input hyperspectral data, and thus, it does not perform well in
terms of the participatory performance measures. The optimization-based solution
explicitly deviates from the relative bias and the fusion symmetry (FS) in order to
generate an output that is intended to maximize certain output characteristics. The
variational technique too, iteratively modifies the fused image for well-exposedness.
Though this indicates a good participation from the constituent input bands (in the
form of a higher fusion factor), a uniformity in their participation is not guaranteed,
as indicated by a comparatively higher value of fusion symmetry. Therefore, the
values of participatory measures in the case of these two techniques do not match
well with the qualitative description seen in Fig. 10.1 .
It may also be noted that the fusion factor (FF) that reflects the amount of mutual
information between the fused image and the input spectral bands is quite high for
all techniques except the band selection technique.
We now provide the results of fusion over the second dataset—the moffett 3 data
in Fig. 10.2 . These data depict the region in surroundings of the first dataset (see the
latitude-longitude information in Table 10.1 ). Therefore, the variety in the features,
and thus, the nature of corresponding fusion results are quite similar to those of
the moffett 2 data. The only notable difference lies in the colors of the results of
the band selection technique and the CMF technique. We restate that the coloring
schemes employed for all the results are purely pseudo-coloring schemes, and hence
the change in the image due to the color or hue change should not be construed
as a measure of its quality. However, the change of colors in the result of band
selection given in Fig. 10.2 e is due to different choice of bands as compared to the
ones in case of the moffett 2 data. The saturation in the results of the technique using
the PLF is still persistent in Fig. 10.2 f. While the results of the bilateral filtering-
based fusion shown in Fig. 10.2 a are visually appealing, some regions in the right