1225
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B7. Beijing 2008
(3) Design low-pass and high-pass decomposing filter that
corresponds to the scale function p( x ) and wavelet
function y/{x) respectively and design low-pass and high-
pass reconstructing filter at the same time.
(4) Decompose panchromatic image and synthesized
hyperspectral color image on each layer by wavelet
package decomposing algorithm and get the decomposed
sub-images of both high and low frequency.
(5) Fuse the panchromatic and hyperspectral sub-images and,
in order to extend the application area of information
composition, we propose an adaptive wavelet package
fusion algorithm based on region features and the brief
process is:
(a) Assume the synthesized hyperspectral color image
after wavelet package decomposing is A( X, y) and
the panchromatic image is B(x,y), separate the
colors of image A(x,y) to get three sub-images
defined as Aj (jc,jp)( k = 1,2,3 and j is scale
coefficient) and then apply histogram equalization to
image B(x,y) and Aj(x,y).
(b) Open a M x N window (usually 3x3) for
image Aj ( X, and i?( X, jp), and in every image
window, compute the square error D i , energy E i
and information entropy S ( .
(c) Calculate the pixel’s weigh on image Aj (.XjT’) and
B(x,y)
W' = a* Ei+b* D' + c* Sy, (8)
Where a , b , c are weights of every feature and their
default value is 1.
(d) Get the fused sub-images by the following equation
Fi(x,y) = (A)(x,y)-W a +W,,-B k (x,y))KW'+W b ) (9)
Interpolate reconstructed image inversely by the above wavelet
package reconstructing filters and get the final fusion image.
3. HYPERSPECTRAL IMAGE FUSION EXPERIMENTS
3.1 Experimental data
The color image composition based on optimal bands and the
fusion of high-resolution panchromatic image and hyperspectral
image are two main parts of our experiment. The data used in
our experiment was received in Oct 2003. It is PHI data of
Shanghais Physics institution and is composed of 124 wave
bands ranging from 400 to 850nm. The flight altitude is 2163
meters and the latitude and longitude of the data is 31.18°-
31.20° and 121.46°-121.51° respectively. In the original data,
we selected the Yangpu bridge region (647x721 pixels) as our
research interest. Additionally, for our fusion experiment and to
obtain high-resolution color fusion image, we collected some
high-resolution panchromatic images in the same area.
3.2 Results and analysis
In order to test the validity of our methods, we conducted a few
experiments by the above PHI data. And the detailed fusion
process is shown as follows.
(1) Preprocess the hyperspectral data, and this step includes
the correction of radiation, atmosphere, and geometry and
so on [8].
(2) Preprocess the high-resolution images with geometry
correction, select the optimal hyperspectral bands by the
optimal index model and construct synthesized low-
resolution color image.
(3) Match hyperspectral and high-resolution images. This
pixel matching step is quite important for image fusion and
aims at eliminating the differences between images
obtained by different sensors on the aspects of resolution,
time, angle and confirming that every pixel of both images
corresponds to the same spatial position. During the
registering process, select some evenly distributed feature
points on both images, and then register the images by
these points through quadratic polynomial algorithm. After
registering, resample the images through bilinear
interpolation algorithm and output the images to end the
image registering process.
(4) Fuse the registered low-resolution color image and high-
resolution panchromatic image by wavelet package
algorithm described above.
Practically, we can constraint the selection of the red, green and
blue band of the synthesized image according to the wavelength
range and the original images’ center to reduce the blindness
and amount of computing. In our experiments, the red band was
confined to the 43-46 th bands, the green band was confined to
the 22-43 rd bands and the blue band was confined to the 1-21 st
bands. The result of optimal bands was the 55 th , 33 rd and 10 th
for the red, green and blue band respectively and the
corresponding optimal index was 56.881. Figure 1 is
synthesized low-resolution color image by optimal bands. It can
be seen that the synthesized image greatly keeps the spectral
characteristics of the original information and can provide
favorable conditions for image understanding. Figure 2 and 3
are the final fused images of the Yangpu bridge region, and it is
clear that the fused images’ spatial resolution and clearness are
greatly enhanced: automobiles on the bridge have more clear
forms; marks in the garden’s center are more prominent; and
the overall color information is satisfying.