THE FUSION OF SPOT PANCHROMATIC AND TM MULITSPECTRAL
IMAGE BASED ON MULTI-BAND BIORTHOGONAL WAVELET
LIANG Shouzhena, b, *, LI Lanyonga
a
School of Geomatics, Liaoning Technical university , Fuxin - [email protected]
b
Yantai Institute of Coastal Zone Research for Sustainable Development, Yantai - Liangshzh0816@[Link]
Comission VII,WG VII/6
KEY WORDS: Fusion, Multi-band,Biorthogonal ,Wavelet,TM,SPOT
ABSTRACT:
Remote sensing image fusion aims at integrating the information conveyed by data acquired with different spatial and spectral
resolution from satellite and aerial platforms. Image fusion is a very important issue in remote sensing and mapping applications
such as features extraction, image classification because the fused image can provide more comprehensive information than single
image. At present there are various methods developed to fuse or merge images from different sources, acquired at different time or
having different characteristics, particularly low-resolution multi-spectral and high-resolution panchromatic images. Some fusion
techniques or methods, such as PCA(principle component analysis), IHS (Intensity-Hue-Saturation) and Brovey transform ,can
provide superior visual high resolution multi-spectral images but ignore the requirement of high quality synthesis of spectral
information .In other words, there are some distortion of spectral characteristics in comparison with original multi-spectral
[Link] paper develop a nesw method which integrate multi-band wavelet and biorthogonal wavelt to merge remote sensing
images. Multi-band wavelet transform can deal with arbitrary integer ratio of images resolution, while biorthogonal wavelet
transform can reduce the ambiguity on the edge of image because it is symmetric and smooth. A example of fusion based on three-
band and biorthogonal wavelet transform is experimented with SPOT panchromatic image and Landsat TM image and the fused
image is compared with fused images based on orthogonal wavelet ,PCA,HIS and Brovey transform respectively in this paper. Some
parameters including mean gradient, combination entropy, standard deviation and correlation coefficient are adopted to appraise the
fused images.
1. INTRUDUCTION
However, the fusion method based on well-known two-band
Image fusion refers to a process that extracts redundant and wavelet, which have been successful for the fusion of SPOT
complementary information from a set of input images and panchromatic and other multispectral images, is not very
fuses it into a single and more complete image. The fused efficient for the fusion of images whose ratio of spatial
image should have more useful information content. The fusion resolutions is not 2n (n = 1, 2, 3, ...), e.g., for fusing a 10-m
of the two images can take place at the signal, pixel, or feature resolution
level. Presently many image fusion methods have been
proposed for combining remote sensing data .A detailed review panchromatic SPOT image and with 30-m resolution
on this issue was given by Pohl and Van Genderen (Pohl et multispectral TM images .Moreover, the two-band orthogonal
al,1998).The well-known methods are, for example, the HIS wavelet lacks symmetry ,which makes the color of fused image
(Intensity, Hue, Saturation) (Edwards et al,1994), PCA not been smoothly integrated into the spatial [Link] it
(Principal components Analysis) (Zhou,1998; Chavez,1989), produces some distortion of spectral characteristics in the
Brovey Transform (Gillespie et al,1987) .These methods are original multispectral [Link] biorthogonal wavelet is
conventional methods which have been studied widely. But symmetric and multi-band wavelet can effectively fuse diffent
these fusion methods have some deficiencies. At least in ratio of spatial resolution .So these problems can be solved if
comparison with the ideal output of the fusion,these fusion integrating multi-band wavelet with biorthogonal wavelet to
techniques can often produce poor [Link] most obvious fuse remotely sensed data. So in this paper, we adopt the new
problem is that it leads to some distortion of spectral fusion technique to merge SPOT panchromatic images with
characteristics. Recently developments in wavelet analysis multispectral TM [Link] technique can effectively fuse
provide a potential solution to these drawbacks. For example, differet or arbitrary ratio of spatial resolution such as 30m-
Nunez developed an approach to fuse a high-resolution resolution TM images and 10m-resolution SPOT panchromatic
panchromatic image with a low-resolution multi-spectral image images ,and in this approach the color distortion can be reduced
based on wavelet decomposition (Nunez et al,1999). Ranchin to a certain extent.
and Wald designed the ARSIS concept for fusing high spatial
and spectral resolution images based on the multi-resolution The structure of this paper is as follows. The first section is
analysis of two-band wavelet transformation (Ranchin et introduction .In this section mainly introduce the study and
al,2000). development on image fusion and the structure of the whole
* Corresponding author :Liang Shouzhen, E-mail:,liangshzh0816@[Link]
1179
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
paper. The second section discusses the theoretical basis and coefficients HL (variations along the columns), vertical
transformation characteristics of multi-band and biorthgonal coefficients LH (variations along the rows), diagonal
wavelet. Then the fusion method based on multi-band and coefficients LL (variations along the diagonals) (Gonzalez and
biorthgonal wavelet is implemented to fuse 10-m SPOT Woods, 2001). The three high frequency image is called detail
panchromatic and 30-m multispectral TM images Next, the image, which contain information of local details.
experimental result is compared with previous methods
developed for image fusion, such as IHS ,PCA and two-band When a = nj and b= njk,the resulting equation is
wavelet
1 (4)
2. BIORTHOGONAL AND MULTI-BAND WAVELET ψ j ,k (x) = ψ (n - j x - k )
j
n
2.1 Mltiresolution Analysis 1
WTf ( j, k ) = f ( x),Ψ j,k (t ) = ∫f (x)ψ (n - k )dx (5)
* -x j
j
2
Wavelets are functions in L Rdetermined from a basic wavelet n R
function by dilations and [Link] are used for
representing the local frequency content of functions. The basic
wavelet should be well localized in general, and the wavelet We denoted it as n-band wavelet. After n-band discrete wavelet
should have zero mean( Daubechics, I.,1992.). The basic transform, an image yields n2 images: one low-pass image and
method to construct a wavelet is Multiresolution Analysis.A n2-1 high-pass The high frequency image is called detail image,
Multiresolution Analysis (MRA) is defined by a sequence of which contain information of local details and low frequency is
closed subspaces (Vj ) j∈Z ,which approximates L2R, and a [Link].
function Φ ∈L2 R s an orthonormal basis for V0.
{0 }⋅ ⋅⋅ ⊂ V −1 ⊂ V 0 ⊂ V 1 ⊂ ⋅ ⋅ L 2 ( R )
{Φ 0 ,n ;Φ 0 ,n ( x ) = Φ ( x − n ), n ∈ Z }
Where Φ (x ) is a scaling function
2.2 Multi-band wavelet
Generally, a wavelet family is described in terms of its mother
wavelet, denoted as ψ(x). A daughter wavelet ψa,b(x) is defined
Figure1 two-band wavelet and three-band wavelet transform
by the equation
2.3 Biorthogonal wavelet
1 (1) Generally a function f(x) can be decomposed as a superposition
ψ a ,b ( x ) = ψ (( x - b ) / a )
a of the orthogonal basisψj,k(x), But to biorthogonal wavelet, a
function f(x) can be written (7)
where a, b∈R and a≠0;
a is called the scaling or dilation factor and b is called the
f ( x ) = WT f , ϕ j ,k ( x) = ∫ 〈 f ( x ),ϕ ( x ) 〉ϕ ( x ) dx (6)
j, k j ,k
translation factor. f ( x ) = WT f , ϕ j , k ( x ) = ∫ 〈 f ( x ),ϕ * j, k ( x )〉ϕ j , k ( x ) dx (7)
A common choice is a=2j and b=2jk, where j and k are integers.
The resulting equation is
Where 〈ϕ j ,k ( x), ϕ * j ' ,l ( x)〉 = δ jj 'δ kl (8)
1 (2)
ψ j ,k (x) = ψ (2 - j x - k )
2 j When applying the biorthogonal wavelet to decompose function
1 or images, the function ϕ * j ,k ( x ) is used, while the
WTf ( j, k ) = f(t),Ψ j,k (t ) = ∫f (t)ψ (2 x - k )dx (3)
* -j
2j R function ϕ j ,k ( x ) is applied to reconstrut function f(x) The
biorthogonal wavelet systems generalize the classical
orthogonal wavelet systems. They are more flexible and
This equation(2) is two-band orthogonal wavelet. Accordingly generally easy to design. One of the main reasons to choose
the wavelet transform of function f(x) is (3) biorthogonal wavelets over the orthogonal ones is symmetric.
Symmetric wavelets and scaling functions are possible in the
After 2-band discrete wavelet transform, an image yields four framework of biorthogonal wavelets. However, the
images: one low-pass image and three high-pass images. orthogonality no longer holds in biorthogonal wavelet systems.
Namely, approximation coefficients (labeled LL), horizontal However,that is the near orthogonal system.
1180
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
g=
1
( N x − 1)( N y − 1)
∑ (ΔI 2
x )
+ ΔI y2 / 2 (10)
3. IMAGE FUSION AND QULITY APPRAISAL
3.1 Image fusion
Where g is mean gradient
Wavelet-based fusion technique integrates the high-frequency
components of the higher resolution data with the low-
frequency components of the lower resolution data (or the lower N x , N y is the number of colums and lines respectively
resolution data) in a Multi-Resolution Analysis (MRA).
Currently wavelet-based image fusion methods used are mostly
Δ I x2 , Δ I y2 is the square sum of difference of adjacent pixel
based on two computation algorithms: the Mallat algorithm
(Mallat,1989; Ranchin et al,2000) and the à trous algorithm
(Aiazzi et al,2002; Shensa,1992).In this paper the Mallat 4. EXAMPLE AND FUSED IMAGE ANALYSIS
algorithm is used. The fusion process has been well described
(Yocky,1996; Aiazzi et al,2002). The first step is to co-register As an experimental study, the multi-band biorthogonal
two images precisely. It is generally preferable to register the wavelets-based image fusion method was applied to the10m
lower resolution image to the higher resolution image. In other resolution Spot panchromatic image and three 30m resolution
words the high resolution image is used as reference image. Landsat multispectral [Link] the ratio of spatial
However if the lower resolution image has georeference that is resolution is three,we adopt three-band biorthogonal wavelet to
to be retained, it may be desirable to use it as the reference merge two [Link] ,this paper also fuses two images
image. the process includes two sides:locality transform and based on PCA,IHS and Brovey in order to appraise the fused
resampling. Secondly it is to calculate the ratio of spatial image based on three-band biorthogonal wavelet.
resolution between different resolution images. If the result is
integer w,the value would be adopted as dimension of muti- From the fused images Figure2 ,it should be noted that both the
band wavelet transform .The high-resolution image would be spatial and spectral resolution have been enchanced. The fused
decomposed based on w-band biorthogonal wavelet transform. images contain the spatial information of high-resolution Spot
Accordingly if the w is not integer ,the Lease common multiple panchromatic and multi-spectral information of TM
of the resolution of two images must be [Link] [Link] image based on three-band biorthogonal wavelet
interpolation of high-resolution image is carried on based on has a better visual rusult in comparison with other three fused
lease common multiple Following this ,the decomposition of [Link] this paper appraises the fused image
new high-resolution is processed based on multi-band quantitively by four parameters :standard deviation ,mean
biorthogonal wavelet transform.A low frequency image and w2 gradient , combination and correlation. The result is Table1
-1 high frequency images are [Link] the low frequency
image of the higher resolution image is replaced by the low
resolution image . Finally, an inverse wavelet transform is
applied. The result is an image which merges the high-
resolution image and the low-resolution image.
3.2 Quality analysis and appraisal
In this paper several parameters including combination
entropy ,correlation coefficient ,mean gradient and standard
deviation , as used in other studies such as Wald et al.(1997),
Sun et al.(1998), and Li et al.(1998),are adopted to value
quality of the fusion image. Standard deviation and correlation
coefficient are parameter we are familiar with, so combination
entropy and mean gradient are only presented in this
[Link] to Shannon information theory，the larger the
entropy of image is , the richer the information and the better
quality of a image is .
n Figure2 the fused images based on,3-band biorthogonal
H ( x )= - ∑Pi log 2 Pi (9) wavelet , PCA, IHS and , Brovey
i= 0
In general the standard deviation and combination entropy can
reflect the information of image and deviation from original
where H ( x ) = the entropy of image image to a certain degree. The fused image based on three-band
i = the grey value of pixel biorthogonal wavelet transform has the most entropy among all
. n = the number of pixel of image fused [Link] this technique fuses more information than
other fusing [Link] addition, the standard deviation of
the biorthogonal wavelets based image fusion is more close to
the standard deviation of original images than these of other
Pi =the probability of i methods,which indicates this method has less distortion of
spectral characteristics .
1181
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
The correlation coefficient reflects the redundancy degree of which has the same resolution with low-resolution [Link]
images. According the table1,we know the original image the spectral information can be reseaved to the greatest extent
contains a plenty of redundant information and the correlation and the distoration of spectral characters can be resolved.
coefficient of fused image based on three-band biorthogonal Further we can draw a conclusion that n-band biorthogonal
wavelet is less than that of original image and other fused image wavelet has some advantages when the ratio of spatial
except of fused image based on PCA, which indicates that this resolution is n. So the multi-band biorthogonal wavelet can play
method has less redundant information . a impotant role in image fusion.
The mean gradient reflects the contrast between the details
variation of pattern on the image and the clarity of the image. 6. REFERENCES
Of all images including original TM image ,the fused image
based on three-band biorthogonal wavelet transform has the Aiazzi, B., L. Alparone, S. Baroni, A. Garzelli,2002. Context-
most mean gradient. Driven Fusion of Spatial and Spectral Resolution Images Based
on Oversampled Multiresolution Analysis, IEEE Transaction
According to previous statistics and analysis, we can draw a on Geoscience and Remote Sensing, 40(10), pp2300-2312
conclusion that the fusion method based on three-band
biorthogonal wavelet is efficient for merging low resolution Chavez P. S. and Kwarteng A. Y.,[Link] spectral
multi-spectral TM image and high resolution Spot panchromatic contrast in Landsat Thematic Mapper image data using
[Link] has many advantages in comparison with other selective principle component analysis, Photogramm. Eng.
methods when the ratio of spatial resolution of original image is Remote Sens., 55(3), pp.339–348
not [Link] other fusion methods have advantage in some
ways. Daubechics, I.,1992. Ten Lectures on wavelets, CBMSNSF
Series in Applied Mathematics, 61, SIAM,
Philadelphi,Pennsylvania, pp. 125-146
Method MG SD CC CE Edwards. K. and Davis .P. A.,[Link] use of Intensity-Hue-
Saturation transformation for producing color shaded-relief
34.07 46.72 0.296 images, Photogramm. [Link] Sens, 60(11), pp. 1369–1374
35.73 52.38 0.389
PCA 7.668 Gillespie .A. R., Kahle .A. B and Walker .R. E.,1987 .Color
26.95 53.30 -0.25
enhancement of highly correlated images—II. Channel ratio
31.77 53.14 0.647 and ‘chromaticity’ transformation techniques. Remote Sens.
32.12 46.45 0.774 Environ., 22, pp. 343–365
IHS 7.661
27.52 52.83 0.617
Mallat. S,1989. Theory for multi-resolution signal: The wavelet
33.39 52.81 0.776 representation, IEEE Trans. Pattern Anal. Mach. Intell., vol. 11,
34.31 46.07 0.661 no. 7, pp.674–693
Brovey 7.683
27.68 52.99 0.463
Nunez, J., X. Otazu, O. Fors, [Link], V, Pala, and [Link],
3-band 39.82 56.33 0.578 [Link]-based image fusion with addtive wavelet
Biorthogonal 44.63 52.93 0.523 7.786 decomposition, IEEE Transactions on Geoscience and Remote
Wavelet 33.95 58.83 0.00718 Sensing, 37(3):1204-1211
Original 35.89 52.89 0.587
7.70 .Ranchin, T. and Wald, L.,2000. Fusion of High Spatial and
Image 36.81 49.40 0.533 Spectral Resolution images: The ARSIS Concept and Its
(TM) 29.99 55.14 0.03 Implementation . Photogrammetric Engineering and
RemoteSensing, 66, pp.49-61.
Table1 the comparison of fused image of different methods
Shensa. M. J.,[Link] discrete wavelet transform: Wedding
the à Trous and Mallat algorithms, IEEE Trans. Signal Process.,
40(10), pp. 2464–2482
5. CONCLUSIONS
Wald,L., T. Ranchin, and M. Mangolini,1997. Fusion of
satellite images of different spatial resolution: Assessing the
In this paper, a new fusion method based on three-band
quality of resulting images, Photo-grammetric Engineering and
biorthogonal wavelet was presented to merge the multi-spectral
Remote Sensing, 63(6),pp.691-699.
TM and high resolution Spot panchromatic image. Several
parameter including mean gradient, standard deviation,
Yocky, D.A.,[Link] Wavelet Decomposition
correlation coefficient and combination entropy were adopted
Image Merger of Landsat Thematic Mapper and SPOT
to appraise the fused products. At last the result of experiment
Panchromatic Data, Photogrammetric Engineering and Remote
proves that the fusion method based on three-band biorthogonal
Sensing, 62(3),pp. 295-303.
wavelet is more efficient to fuse the TM 30m-resolution image
Zhou.J., Civco. D. L., and Silander .J. A.,1998.A wavelet
and 10m-resolution Spot panchomatic image than other fusion
transform method to merge Landsat TM and SPOT
methods. The reason is that biorthogonal wavelet is symmetric
panchromatic data,” Int. J. Remote Sens., 19(4), pp. 743–757.
and three-band wavelet can produce the low frequency image
1182
