1051
GENERALIZED MODEL FOR REMOTELY SENSED DATA PIXEL-LEVEL FUSION
Zhang Jixian, Yang Jinghui*, Li Haitao, Yan Qin
Chinese Academy of Surveying and Mapping, Beitaiping Road 16, Beijing 100039, P. R. China
*: Corresponding author. Email -jhyang@casm.ac.cn. Tel: +86-10-88217730
Commission VII, WG VII/6
KEYWORDS: Remotely Sensed Data, Fusion, Generalized Model, Implementation, PCA
ABSTRACT:
A generalized model characterizing most remotely sensed data pixel-level fusion techniques is very important for theoretical analysis
and applications. This paper focuses on the establishment of a generalized model for most data fusion methods, which is helpful to
quantitatively analyze and quickly implement different data fusion techniques. As an example, the PCA fusion method is selected to
demonstrate the availability of the generalized model through the generalized model based implementation.
1. NOMENCLATURE
XS k : the k th band of the lower resolution multispectral image;
P a the higher resolution panchromatic band;
P an : the degraded panchromatic band;
P anA : approximation coefficients after n level GLP
(Generalized Laplacian Pyramid) or a trous wavelet
decomposition;
w
P D : detail coefficients after n level GLP or a trous wavelet
decomposition;
xs L
k : the k th band of multispectral image resampled or
relatively processed to have same size as the panchromatic
band;
xS H .
k : the k th band of the higher resolution multispectral image
after fusion;
xsl xsl
(kjj) • the pixel value of location (i,j) of the band * ;
xs 11 . . . xs H
• the pixel value of location (i,j) of the band k ;
: spatial and textural details of location (i,j) extracted from
the panchromatic band;
CL 8 . xs L
. the fusion coefficients modulating ( ' ,7) into ( * ' ,7) . 2
2. INTRODUCTION
So far, many pixel-level fusion methods (Carper, 1990,
Shettigara, 1992, Hill, 1999, Liu,2000, Zhou, 1998,
Ranchin,2003) for remote sensing image have been presensed
where the multispectral image’s spatial details are enhanced by
adopting the higher resolution panchromatic image
corresponding to the lower resolution multispectral image.
Therefore, the main principle of remote sensing data fusion
focuses on the maximum enhancement of its spatial details on
the condition of minimizing distortion of multispectral image’s
spectral characteristics. When correlation between the
multispectral and panchromatic images is not high, it is often a
mutual contradiction between maintenance of spectral
characteristics and enhancement of spatial details. Thus the
choice of fusion algorithm is determined to emphasize spectral
features or spatial details according to a specific application.
Typical algorithms of remote sensing data fusion can be divided
into three general categories (Zhang and Yang,2006):
component substitution fusion technique (Chavez, 1991,
Carper, 1990, Shettigara, 1992, Hill, 1999), modulation-based
fusion technique (Chavez, 1991, Vrabel, 2000, Liu, 2000, Zhang
and Yang,2006) and multi-scale analysis based fusion technique
(Zhou, 1998, Ranchin,2003, N'u~nez,1999, Pradhan,2006,
Aiazzi,2002). The typical algorithms of component substitution
fusion technique are IHS transform fusion algorithm (Carper,
1990), PCA transform fusion algorithm (Shettigara, 1992), LCM
(Local Correlation Modeling) fusion algorithm (Hill, 1999) and
RVS (Regression Variable Substitute) fusion algorithm
(Shettigara, 1992); the fusion algorithms of the
modulation-based technique include Brovey transform fusion
algorithm (Vrabel,2000), SFIM (Smoothing Filter Based
Intensity Moulation) fusion algorithm (Liu,2000) and high pass
filter fusion algorithm (Chavez, 1991); the fusion algorithms
based on the multi-scale analysis mainly include wavelet
decomposition based fusion technique (Zhou, 1998,
Ranchin,2003, N'u'nez, 1999, Pradhan,2006) and Laplacian
pyramid decomposition based fusion technique (Aiazzi,2002).
When various fusion algorithms are studied, an issuse whether
these algorithms can be described by a generalized
mathematical model (Tu,2001, Wang,2005) is ignored. The
model can reflect the main features of the fusion process by a
simple mathematical formula. The establishment of a
generalized model will contribute to relatively theoretical
analysis and fusion algorithm design in the light of a specific
application. Also the model is beneficial to qualitative and
quantitative analysis of fusion technology from different aspects.
The most important aspect is that the establishment of a
generalized model will reveal that different fusion technique