Edge-Preserving Solution - Hyperspectral Image Fusion - page 52

A smaller value of the constant C in Eq. ( 3.5 ) results in boosting the finer details
whichmay result in over-sharpening. On the contrary, a too high value of C smoothens
out fusion results which makes it average-like . We have selected the value of C to
be equal to 50 in this experiment while the hyperspectral bands are a 16-bit data.
In the case of the RGB version of the fusion result, the three pre-final stage images
have been assigned to the red, green, and blue channels of the displaying device. The
RGB visualization of hyperspectral images inherently involves pseudo-coloring, and
hence, the coloring scheme does not bear any physical meaning.

3.7 Experimental Results

The experimental results of the presented techniques are typically provided in the
same chapter for most of the topics. However, the subsequent chapters of this mono-
graph present three more techniques for visualization of hyperspectral images. We
would like to provide a detailed comparison of these techniques together which, we
believe, would facilitate a quick and easy analysis of these fusion solutions to the
reader. Therefore, instead of providing separate results for each of the techniques, we
have provided the results of all the techniques presented in this monograph, as well
as some other existing hyperspectral image fusion techniques together in a specially
dedicated Chap. 10 at the end. Also, these results will also involve a quantitative
performance evaluation of these techniques using various objective performance
measures. We would like the readers to be familiar with these performance mea-
sures before we discuss merits and demerits of the fusion techniques on the basis of
these measures. In Chap. 9 , we explain in detail various performance measures that
quantify various aspects related to fusion of a large number of images. In the present
chapter, however, we shall illustrate in brief the experimental results for fusion over
just a couple of datasets.
We demonstrate the efficacy of the presented solution using a real world hyper-
spectral data provided by the Hyperion imaging sensor used in the EO-1 spacecraft
for the earth observation. The dataset consists of 242 bands (0.4-2.5 m) with a 30m

spatial resolution. The selected data set depicts the urban region of Palo Alto, CA.
The dimension of the Hyperion data cube is (512

µ

242). This dataset is a part
of third sub-block of the original dataset. We refer to this dataset as the urban dataset
throughout this monograph. We have employed a three-stage hierarchical scheme for
fusion over this dataset. The fused images in the first stage comprise a bandwidth of
approximately 110-120nm, which are generated by fusing 12 contiguous bands. The
second-stage fusion generated three images from18 resultant images of the first-stage
fusion. The first fused image as shown in Fig. 3.4 a results from fusing bands 1-74,
corresponding to the wavelengths from 350.505nm to 887.235nm. The second fused
image [Fig. 3.4 b] of the pre-final stage represents the fusion over bands 75-154. The
bandwidth of this fused image of the urban data scene is 807.531nm (from 887.240
to 1694.771nm). The last fused image at this stage has a bandwidth of 888.508nm,
which represents the fusion of bands 155-242 of the original hyperspectral image

× 256 ×