1052
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B7. Beijing 2008
rely on the difference of the calculational mehtod of the
mathematical model parameters. Thus calulating model
parameters corresponding to the fusion method is the main task
when implementating the method. Compared to Wang’s work
(Wang,2005), the paper mainly concentrates on two aspects.
First, the paper presents a generalized model for remotely
sensed data pixel-level fusion, which has a wide range of
applicability. The various commonly used remote sensing data
fusion algorithms can be deduced to the generalized model.
Second, the implementation technique base on the generalized
model only calculates the model parameters impacting the last
fusion results and discards the processing steps not affecting the
fusion results, saving computational time.
3. THE GENERALIZED MODEL
According to the imaging mechanism and the ideal
pan-sharpening results of multispectral image, the presented
generalized model is formulated by,
(1)
( ' ,7) : Spatial and textural details extracted from the
panchromatic band by a certain calculation.
a
Oc,i,j). The coefficients modulating
into
xs
L
(k.ij)
The presented model expressed by equation (1) can clearly
describe the mathematical relationships among the original
multispectral image, the spatial details extracted from the
high-resolution panchromatic image, and the adopted fusion
strategy. In another word, the spatial and textural features
extracted from the panchromatic band are imported into the
multispectral image in terms of the fusion coefficients and the
fusion result is the image whose features are enhanced by the
panchromatic image. The fusion operations are fulfiled pixel by
pixel, band by band after calculation of
and } but
in the course of calculation of and , not only the
pixel value of location (i,j) of the lower resolution multispectral
k th band and the panchromatic band but also the whole
statistical information and neighbor pixels of loacation (i,j) are
used.
the methods calculating parameters
C
O’D include:
g
1) the linear combination method: obtaining the < ~ x ’ y) after
subtracting multispectral bands’ linear combination from the
panchromatic band, such as IHS, PCA, RVS, Brovey,
Block-regression (Zhang and Yang,2006);
2) filter method and multi-scale analysis method: obtaining the
(x ' y) after subtracting its filtered or multi-level decomposition
results from the panchromatic band, such as SFIM, LCM, A
trous (N'u~nez,1999), GLP (Aiazzi,2002), ARSIS method
(Ranchin,2003).
oc
( k,x,y ) is determined by following factors: the panchromatic
and multispectral relative spectral response, spectral range of
the panchromatic and multispectral bands, the GIFOV(Ground
projected Instantaneous Field Of View) of panchromatic and
multispectral bands, the landscape properties and land cover
classes, radiometric calibration method of different sensors, the
temporal properties, the correlation between the panchromatic
and multispectral bands, the average value, variance and other
statictical characteristics of the multispectral and panchromatic
bands.
cc
The methods calculating parameters include:
1) Constant Value, such as IHS, PCA, RVS, A trous, LCM;
2) Spectral Distortion Minimum: such as SFIM, Brovey, A
trous, Block-regression;
3) Context-based Decision (CBD), such as GLP, ARSIS.
The model formulated by equation (1) is more comprehensive
and applicable than Wang’s model. The fusion coefficients of
wang’s model are limited to the cases of constant value and
spectral distortion minimum, and can not describe the fusion
coefficients for LCM, ARSIS and GLP fusion algorithms. For
the method extracting spatial and textural details, Wang’s
model include the filter and linear combination mehtods while
the generalized mode proposed in this paper supports the
additional methods used in LCM and GLP fusion algorithms. In
a word, the generalized model can characterize most of
commonly used remote sensing data fusion algorithms
including not only the IHS, PCA, A trous, Brovey, HPF
algorithms but also RVS, GLP, LCM, ARSIS, wavelet
decomposition plus PCA transform, wavelet decomposition
plus IHS transform, and the authors proposed Block-regression.
4. DEDUCTION FOR COMMONLY USED FUSION
ALGORITHMS
In this section, three categories of fusion algorithms mentioned
in section 1 are deduced to the generalized model, i.e. the
proposed equation (1), through the mathematical transformation.
Through the deduction, the conclusion can be drawn that
different fusion technique rely on the difference of the
S< Cl (k .
calculation of parameters {l,J> and ( ’ l,J> .
4.1 Component Substitution Fusion Technique
The typical algorithms applying component substitution fusion
technique include IHS, PCA, LCM and RVS fusion algorithms.
To illustrate the deduction for this technique, following is the
transformation steps taking PCA fusion algorithm as an
example.
XS
The lower resolution multispectral band k is resampled to
have the same size as the higher resolution panchromatic band
P an after those bands are co-registrated: XSk ~ rs P^ xs k) ? and
after the resampling the implementatiom steps for PCA fusion
algorithm are as follows (Shettigara, 1992):
1 ) Calculating the correlation matrix of the n lower resolution
multispectral bands, ^ is equal to 4;