Band Selection: Revisited - Hyperspectral Image Fusion - page 120

Fig. 8.3 Plots of PSNR of the fused images for various values of κ on the Hyperion and the AVIRIS
data using the bilateral filtering-based fusion technique over the subsets selected using the output-
based scheme. The resultant image from fusion of the entire dataset using the same technique is
used as a reference in each case

the reference image. Similar to the input-based band selection, the values of PSNR
drop very gradually when 65, this scheme selected 71

bands of the urban hyperspectral image which is nearly

κ is increased. For κ = 0 .

1

/ 3 -rd the entire dataset,
however, the fused image provides a PSNR above 30 dB which indicates a good
visual match between the two images. For the same value of , 110 bands of the

moffett 2 data were selected, yielding the corresponding fused image with a PSNR of
nearly 43 dB. A PSNR above 40 dB can be observed in the resultant image generated
from fusion of only 12 bands when the threshold parameter

κ

is set to 0.70 in the
case of the moffett 2 hyperspectral image. This result compares very well with that
for the input-based band selection scheme shown in Fig. 4.5 . Therefore, both the
schemes of band selection provide a subset of bands which is capable of producing
comparable results with only a fraction of the data. Similar characteristic plots have
been observed for other pixel based fusion techniques also.
Although visual comparison is a subjective measure, it is often important, espe-
cially when the primary purpose of fusion is the visualization of scene contents by a
human analyst. We provide some of the output images resulting from fusion of sub-
sets of the moffett 2 hyperspectral image where the bands have been selected using
the output-based scheme. The fusion results of the output-based band selection are
provided in Fig. 8.4 . The images in Figs. 8.4 a-d are obtained from the subsets of
the moffett 2 data where the cardinality varies with relation to the variation in the
threshold parameter

κ

. Figure 8.4 a shows the result of fusion of the entire dataset,
and thus it acts as the reference for the evaluation purpose. Figure 8.4 b is the result
of fusion of 142 bands by setting

κ

40,
the band selection scheme does not discard any of the input bands, and hence the
fused images obtained for the values of

κ to be 0.50. As for the values of 0 ≤ κ ≤ 0 .

κ less than 0.40, are essentially the same