The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1224
high-resolution spatial image through adaptive wavelet package
fusion algorithm to obtain high-resolution colour image. This
method can avoid the limits brought by using single feature for
fusion and make fusion information more abundant.
2.1 Band selection model by the optimal index principle
The greater the image’s standard deviation the more
information it will have, while the smaller the correlative
coefficient between image bands, the more independent they
will be and the less redundant information they will have. So
here the optimal index model (OIF) can be adopted.
£№ = ¿$,/¿1/1,1 (1)
1=1 1=1
Where S,• =standard deviation of the i image band
Ry ^correlative coefficient between i and j image
band.
The numerator is the sum of the mean square error of three
image bands, and the denominator is the sum of the correlative
coefficient between them. It is clear that the smaller the
correlation, the higher sparsely (information) these bands will
have. The 3 image bands of the whole image data that give the
highest OIF value are selected as the optimal image bands. If
we are interested in some special regions, we can define them
and get the optimal bands for them correspondingly.
2.2 Wavelet Package Algorithm
Wavelet package is composed of a few wavelet functions.
Assume that there is a wavelet function cluster [w n {x),n e N]
and these functions have the following relationship (Zhu,2000).
w 2n( x )=T, h k w n ( 2x ~ k ) (2)
k
W 2„ + lM =Yj^k W n{^ X ~ k ) (3)
k
When n=0, > v 0 (jc)=<p(jc)and W| (x) = v'W’ where <p( x ) is scale
function, y/{x) is wavelet function, [h k } and are
coefficients of low-pass and high-pass QMF filter and
correspond to scale function and wavelet function respectively.
Then |w H (x),x e N) is called wavelet package determined by
w 0 (.x)=tp(x). According to equation (2) and (3), w n (x)can be
reconstructed by w 2 „(x) and w n (x):
w ni 2x - 0= Y,{Pl-2k W 2n(* - *)+ ?/-2* W 2„ + l(* “ *)} (4)
k
Where ({/?*}, }) is QMF filter corresponding to {{h k }, {g*}).
So the decomposing and reconstructing algorithm can be
described as follows:
Assume that A * f{x) is approximation of wavelet package
w„ on scale 2 1 , which is
A n A x ) = X S A -Wn&x-k) (5)
k
Where S J nk = 2 2 f f(x)w n il J x-k)dx
’ J-oo
Then according to equation (1),(2), the decomposing algorithm
is
=2X-A
m
^2n.+l,I ~ XI ^m-21^n,m
And according to equation (3), the reconstructing algorithm is
Sn,k ~ y ^ J Pk-2l^2n,l + ' S y. C lk-2l^ J 2n+U ( 7 )
/ /
2.3 Wavelet package fusion algorithm for hyperspectral
data based on optimal index principle
To process hyperspectral data with wavelet package fusion
algorithm, optimal bands have to be selected first by optimal
index algorithm to obtain synthesized low-resolution color
image, and based on which, registered panchromatic image is
obtained. Then the color image and registered image are
decomposed by wavelet package algorithm. On each
decomposing layer, every image part, no matter high frequency
part or low frequency part, of the previous layer is analyzed
iteratively and comprehensively. After analyzing, sub-images
that are decomposed from the original two images are fused
through related fusing strategy, and the final fusion image is
reconstructed through inverse transformation.
Usually, the more the wavelet package decomposing layer, the
more details the fusion result will have. However, this is
achieved at the expense of increasing computing amount. With
the increasing of analyzing layer, the computing amount
increases quite rapidly and information will be more and more
heavily lost at the top layer. Thus wavelet package
decomposing layers are sometimes chosen between 3 and 5 and
more layers are not advisable. Here, wavelet package fusion
algorithm for hyperspectral data based on optimal index
principle can be summarized as follows:
(1) Select optimal bands of hyperspectral data based on
optimal index and construct synthesized low-resolution
color image.
(2) Initialize the wavelet package decomposing layer, the
decomposing coefficients and so on.