2008 International Conference on Computer Science and Software Engineering

A Remote Sensing Image Fusion Algorithm Based on Nonnegative Ordinal Independent Component Analysis by Using Lagrange Algorithm
Zhongni Wang, Xianchuan Yu*, Wang YuDai Sha
College of Information Science and Technology Beijing Normal University Beijing, 100875, China {Chuan.yu@ieee.org, bameinini@163.com}
AbstractData fusion on remote sensing is one of important problems in current image processing. The key of a successful image fusion is to find an effective and practical image fusion algorithm. To eliminate high-order image data redundancy for two different remote sensing images which is nonnegative, a new approach using the nonnegative ordinal independent component analysis(ICA) based on Lagrange algorithm for remote image fusion between panchromatic and multi-spectral images is proposed. Firstly, the multi-spectral image and the panchromatic image are registered with the error in a pixel. Then the independent components, obtained by nonnegative ICA transform, are done factor analysis to determine the sequence of independent components successfully. Finally, the fused image is obtained by applying image fusion rules. Visual and statistical analyses prove that the concept of fusion method based on nonnegative ordinal ICA is promising, and it does significantly improve the fusion quality with higher signal-to-noise ratio compared to conventional IHS and wavelet fusion techniques. Keywords-Image fusion; Independent Component Analysis; Factor Analysis; Remote sensing image

In this paper, to overcome these limitations, we propose a new image fusion algorithm for Landsat ETM+ and CBERS images. The new fusion method is Nonnegative ordinal Independent Component Analysis based on Lagrange Algorithm LNO-ICA . Experimental comparison with conventional fusion methods shows that the LNO-ICA fusion method outperforms these existing approaches II. NONNEGATIVE INDEPENDENT COMPONENT ANALYSIS

I.

INTRODUCTION

With the development of modern remote sensing technology, the spatial and spectral resolution of remote sensors has been greatly improved which leads to the diversity and complexity of data sources. The data fusion is an important tool for improving the data quality in remote sensing and can effectively integrate the images from different sources into one image. The image fusion techniques have become a hot problem in recent years [1-5]. Because of the importance of image fusion techniques, many image fusion algorithms have been proposed. Traditional image fusion methods in the remote sensing contain the intensity-hue-saturation (IHS) transform, principal component analysis (PCA), discrete wavelet transform (DWT) and etc. [6]. Although these methods improve the quality of the fusion result, there are still some limitations. For example, these fusion algorithms do not eliminate redundancy between different data, or just eliminate the low-order data redundancy, and do not consider the higher-order statistical properties of the signal [7].

The traditional fusion methods including IHS, PCA and DWT are all based on spatial or frequency domain which does not consider redundancy between the different signals. Although the PCA method could eliminate the low-order redundancy, it does not take the high-order redundancy into consider. According to the statistical theory, the most important information of image signal is always included in the statistical characteristics of high-order [8]. Independent component analysis which is recently developed from blind source separation is a novel high-order statistic signal processing method and it tries to transform an observed multidimensional vector into components that are statistically as independent from each other as possible [9-13]. ICA is very useful and gradually become a hot problem in signal processing. Recently ICA is widely applied in biomedicine signal processing, sound signal separation, communication, error diagnose, feature extraction, financial time sequence analysis, data mining, image processing and etc. A. Independent component analysis ICA is signal processing technique whose goal is to express a set of random variables as linear combinations of statistically independent component variables. The estimation of the data model of independent component analysis is usually performed by formulating an objective function and then minimizing or maximizing it. Therefore, the properties of the ICA method depend on both of the objective function and the optimization algorithm. Assume that there is an M-dimensional zero mean

= ( s1 , s2 ,..., sM )T , whose components are mutually independent. The vector s (t ) corresponds to n independent scalar valued source signal si . We can write the multivariate
vector s

978-0-7695-3336-0/08 $25.00 2008 IEEE DOI 10.1109/CSSE.2008.1380

610

p.d.f. of the vector as the product of marginal independent distributions.

We can use the Lagrange multipliers method to solve the problem, we assume the ij is Lagrange multipliers corresponding to the condition of u = w x 0 , denoting by = [ij ] ,
T

p(s) =
A data vector point t , such that

p i ( si )

(1)

x = ( x1 , x2 ,..., x )T is observed vector at

N

L = kurt(u) + tr(u) = kurt(wT x) + tr(wT x) (7)

Then doing partial derivative of L within Kuhn-Tucker conditions, the issues could be solving by following equations,

x(t ) = As (t )

(2)

Where, A is an N*M scalar matrix which is called mixing matrix. Sometimes we need the columns of matrix A, if we denote them by a j the model can also be written as,

L T 3 = 4E[x(w x) ] 3w + x = 0 w wT x = 0

(8) (9)

x = ai si
i =1

(3)

As the equation is non-homogeneous, there is no general solution for the equation. The both sides of equation (10) are multiplied by w , and then the iteration is as follows,

The goal of ICA is to find a linear transformation W of the correlative signals x that makes the outputs as independent as possible:

4 E[ x( wT x)3 ] 3

(10)

u (t ) = W x (t ) = W A s (t )

(4)

Where, u is an estimate of the sources. Intuitively speaking, the key to estimating the ICA model is non-Gaussianity. It is mainly because the non-Gaussian means independent. The estimation of data model of independent component analysis usually chooses a suitable objection function and then minimizes or maximizes it. It means that an ICA method depends on objection function and optimization algorithm.

C. Factor Analysis In the ICA model, it is easy to see that we cannot determine the order of the independent components.
The reason is that, again both equation.

s and A being unknown in

(1 1)

x = As =

(b
i

1
i

a i )( s i bi )

B. Nonnegative independent component analysis

As the ICA methods can not ensure that the independent components are nonnegative, but the remote image data is nonnegative. Several authors have introduced some algorithms for nonnegative ICA [14], yet these algorithms are all based on the assumption that the sources are well-grounded except for independence and nonnegative. We propose a new remote sensing image fusion algorithm based on nonnegative independent component analysis by using Lagrange arithmetic even in the case that the sources are not well-grounded. We choose the Kurtosis

We can freely change the order of the terms in the sum and call any of the independent components the first one. Formally, a permutation matrix P and its inverse can be substituted in the model to give x = AP -1 Ps . The elements of Ps are the original independent variables s j , but in another order. The matrix
AP -1 is just a new unknown mixing matrix, to be solved by the ICA algorithms.

kurt (u ) as the objection function.

kurt(u) = kurt(wT x) = E[(wT x)4 ] 3{E[(wT x)2 ]}2 (5)

With the constraint of u is as follows,

= wT x 0 , the objection function

(6)

However, the ambiguity of the independent components sequence has not been concerned. This is mainly because the ICA method was applied to different kinds of blind source separation problem. The uncertainty of the independent components sequence did not have an impact on the solution of the issue in several fields. However, when dealing with mineral resources prediction, the sequence of independent components (ICs) is quite important to explain the results. Therefore, the ordinal ICA algorithm uses factor analysis methods to obtain unique ICs. The basic idea of factor analysis is group by the correlation of variable, so that the variable in the same group has higher correlation, but in different group with lower correlation. Each group represents a basic structure, are called factor.

Max(kurt (u)) = Max(kurt ( wT x)) T u = w x 0

611

Assume a random vector x with p dimension, the mean of it is , covariance is . The factor analysis model is as (12).
X

= + Lf +

(1 2 )

Where

f1 , f 2 ,... f m is mutual factors and 1 , 2 ,... p is

specific factors or error, L = (lij ) pm , The elements lij of matrix L are called factor loadings. To establish factor model, first we should estimate factor loadings matrix and special variance. The common methods of parameter estimation has been used in such techniques as PCA method, the main factor method, the maximum likelihood method and etc.. Then, look for a reasonable interpretation of the factors. To reduce subjectivity of the interpretation, factor rotation is introduced. Therefore, the results are then easier to interpret. We use the varimax as the factor rotations technique. Therefore, to revert the corresponding original signals, we do factor analysis and see the mixed and independent components as new observed data. The two components which contribute most to the same common factor can be seemed as the corresponding components. III. REMOTE SENSING IMAGE FUSION BASED ON LNO-ICA

(b) Figure 1. Multi-spectral image and panchromatic image (a) original CBERS multi-spectral image with bands 1, 2 and 3; (b) Original ETM+ panchromatic image

matrix V stands for a picture. The figure 2 is the images with three different bands corresponding to the figure1 (a).

respectively corresponding to the three bands of multi-spectral image, P stands for the panchromatic image. So each column of

To eliminate high-order image data redundancy for two different Remote sensing images, we introduce LNO-ICA for remote image fusion between Landsat ETM+ panchromatic and CBERS multi-spectral images. The images after LNO-ICA as a new tool for image fusion both could improve spatial resolution and preserve spectral characteristics. The study area is located in the urban area of Zhuhai, China. Figure1(a) shows a CBERS multi-spectra image with bands 4, 3 and 2, acquired in 2001.Figure 1(b) shows the ETM+ panchromatic image, acquired in 2001.A scene of 256 256 pixels in size was selected for our experiments, and it includes some ground covers, such as the area for agriculture, forest, grass, man-made infrastructure and etc. Before the image fusion, the multi-spectral images were co-registered to the corresponding panchromatic images. In our experiment, as the resolution of images is 256 256. We can see the multi-spectra image with bands 4, 3 and 2and panchromatic image as a random matrix of (256 256) 4, so that the observed signal is V = [ R, G, B, P]T , here, R, G, B

(a) Band 4 (b) Band 3 (c) Band 2 Figure2. the 3 bands of Multi-spectral image

The detailed steps of this integrated fusion method are as follows, first, registering the multi-spectral image and the panchromatic image with the error in a pixel, and denoting by x = [ R, G , B, P]T . Then we obtain the independent components by nonnegative ICA, with it denoting by s = [ 1 , 2 , 3 , 4 ]T .therefore, we have a new matrix owned 8 vectors. The former four vectors are observed signals x = [ R, G , B, P]T and the latter four vectors are independent components s = [ 1 , 2 , 3 , 4 ]T . As the factor analysis can eliminate indeterminacy of ICA results, we do factor analysis to get 4 common factors. Next, find an effective fusion rule for fusion. The detailed fusion rule is as follows:

1 =(1 + 4 ) / 2 2 =( 2 + 4 ) / 2 3 = (3 + 4 ) / 2
1 2 3 4
obtained.
(a)

(1 3 ) (1 4 ) (1 5 )
independent of components signals observed

Where,

we

assume are the

the

estimate

x = [ R, G , B, P]T

sequentially. At last, the fusion image is

612

IV.

EXPERIMENTS AND EVALUATION

Before the image fusion, the multi-spectral images were coregistered to the corresponding panchromatic images. Experimental results compared with those of conventional fusion methods including IHS [15], PCA [16] and DWT [17] are shown in figure 3.

fusion between Landsat ETM+ panchromatic and CBERS multi-spectral images. At the same time, we use factor analysis to determine the sequence of ICA results. After applying the effective fusion rule, we can get more exact results. Experiments show that compared with those of conventional methods, LNO-ICA fusion method produces better fusion result. VI. ACKNOWLEDGEMENTS

This paper is supported by the National Natural Science Foundation of China (No. 60602035, 40372129), Natural Science Foundation of Beijing (No. 4062020) and program for New Century Excellent Talents in UniversityNCET-06-0131 .Thanks for the data supplier Wang Shan in Beijing Normal University.
(a) I HS (b) DWT [1]

VII. REFERENCES
Mitianoudis N, Stathaki T. Pixel-based and region-based image fusion schemes using ICA bases [J].Information Fusion, 2007, 8(2),pp.131 142. Pohl C, Van Genderen J L. Multi-sensor image fusion in remote sensing: concept, methods and applications [J]. International Journal of Remote Sensing, 1998, 19(5),pp.823854. Cardinali A, Nason G P. A Statistical Multiscale Approach to Image Segmentation and Fusion[C]. Philadelphia: 8th International Conference on Information Fusion, 2005,pp.475482. Huang xin,Zhang liang pei,Li ping xiang.Classification of High Spatial Resolution Remotely Sensed Imagery Based Upon Fusion of Multiscale Features and SVM [J].Journal of Remote Sensing, 2007, 01,pp.48-54. Zhang Yun, Hong Gang. An IHS and wavelet integrated approach to improve pan-sharpening visual quality of natural Color IKONOS and QuickBird images [J].Information Fusion, 2005, 6,pp.225234. Yang Rongling, Ehlers M., Usery E.L., et al. FFT-enhanced IHS transforms method for fusing high-resolution satellite images [J]. Journal of Photogrammetry and Remote Sensing, 2007,pp. 381392. Nikolaos Mitianoudis, Tania Stathaki. Pixel-based and region-based image fusion schemes using ICA bases [J]. Information Fusion, 2007, 8(2),pp.131142. Hyvrinen A, Hoyer P O, Inki M. Topographic Independent Component Analysis [J], Neural Computation, 2001, 13,pp.15271558. M. Chen, J. H. Xuan, D. R. Li, Q. Q. Qin, Y.H. Jia, Image Fusion Algorithm Based on Independent Component Analysis, OptoElectronic Engineering, vol.34,pp.82-87, 2007. Mitianoudis N., Stathaki T. Adaptive Image Fusion Using Ica Bases[C]. Toulouse, Acoustics, Speech and Signal Processing, 2006. Hyvrinen A, Oja E. Independent component analysisAlgorithm and applications [J].Neural Networks, 2000, 13(4),pp.411430. Jutten C, Herauh J. Blind separation of sourcesPart I:An adaptive algorithm based on neuromimatic architecture [J].Signal Processing 1991, 24(1),pp.110. Comon P. Independent component analysis, a new concept? [J], Signal Processing, 1994, 36,pp.287314. Plumbley M D.Condition for nonnegative independent component analysis [J], IEEE Signal Processing Letter, 2002, 9,pp.177~180. Chen Huaixin. A Multi-resolution Image Fusion Based on Principle Component Analysis[C]. Chengdu: Fourth International Conference on Image and Graphics, 2007. Choi M A. New Intensity-Hue-Saturation Fusion Approach to Image Fusion with a Tradeoff Parameter[J]. IEEE Transactions on Geoscience and Remote Sensing, 2006, 44(6),pp.16721682. Zhou J, Civco D L, Silande J A. A wavelet transform method to merge Landsat TM and SPOT panchromatic data [J], International Journal of Remote Sensing, 1998, 19 (4),pp.743757.

[2]

[3]

[4] (c) PCA (d) LNO-ICA Figure 3. the fused results: (a) I HS; (b) DWT; (c) PCA; (d) LNO-ICA

To assess the quality of the fused images, this experimental use the signal-to-noise ratio (PSNR) as statistical parameter to evaluate the effect fusion result. The quantitative results of different fusion method are shown in table 1.
TABLE I.
THE CONTRAST OF PSNR

[5]

[6]

[7]

PSNR IHS DWT PCA LNO-ICA R 14.5890 16.4525 11.3327 20.9196 G 13.8819 15.3591 11.3741 19.9235 B 13.8548 15.0530 11.4700 20.0556 The PSNR is used to measure the difference between images. By analyzing and comparing the all fusion methods from statistical parameters (table 1) and visual measurements (Fig. 3), we can draw conclusions that, the LNO-ICA fusion method produce better result with higher PSNR. From visual measurements aspect (Fig. 3), the same conclusion can be drawn. In other words, the images after LNO-ICA as a new tool for image fusion both could improve the quality of fusion image. V. CONCLUDING REMARKS

Ba nd

[8] [9]

[10] [11] [12]

[13] [14] [15]

ICA is a novel method for finding underlying components from multivariate statistical data and gradually become a hot problem in signal processing. To eliminate high-order image data redundancy for two different remote sensing images, we propose a new method that is LNO-ICA for remote image

[16]

[17]

613