The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 
Figure 5. work flow of the fusion 
3.2 fusion and sharpen 
Scattered single points or small targets disturb during 
classification even though they usually are comprised with 
special target,but in some cases it is not necessary to pay 
more attention, So a special designed Vector filter (Ma 
Yanhua,etc,2006) is designed to erase odds and ends points, 
which is named a vector filter based neighborhood 
clustering(NCVF). The filter is used on both kinds of the 
images and it makes the fusion easier. 
Polynomial registration algorithms were used to as a rough 
registration means to get the special relationship of the pixels in 
two imaggs, the hyperspectral image are not re-sampled. 
Because the two kind images are captured at the same time ,so 
the luminance in corresponding bands are high correlation, the 
registration between the super pixel and the high special 
resolution image pixel can be defined more accurately. 
Next step is to get the grads of both image .The hyperspectral 
image grads can get vector grade first and then translate it to 
scalar quantity, such as ||x||. 
Review every hyperspectral image pixel of their grads and the 
grade of corresponding high special image area to decide 
weather it’s a pure pixel, the criterion can be set according with 
the image. 
Produce the edge image of the high special resolution image to 
segment it as the map to be filled with spectral information or 
classifying information, edit it artificially if necessary. 
The pure pixel spectral fill to the edge image in their registered 
position, and diffusing with a balloon algorithm to cover the 
other pixel in the same segments. If there are more than one 
spectral in a single segment, the spectral should be calculate 
their mean. Spectral will leak out if there are un continuousness 
in the edge, so restrict has to set up to avoid spectra’s leaking 
out at small holes in the edge. Better image segment algorithms 
will be helpful. 
The unpure pixel can be get by pixel decomposing. Now that 
the pure pixels are filled in the image, the indétermination 
during pixel decomposing is reduced. Linear decomposing ruler 
is used in this article. 
The decomposing can be done step by step with spectra that 
already known in neighbor area of the pixel to be 
decomposed .First step decompose those composed with two 
types of material (super pixels that cover two segments in edge 
map) and one of then is known already, Then those composed 
with three types of material.Of cause, more step pixel 
decomposing is meaningless because of error. 
After decomposing, the spectral are filled in the edge map just 
as the pure pixel before-mentioned. 
3.3 experiments 
The method is not well developed yet ,original experiments is 
just finished .One experiment images are showed in figure 6, 
the hyperspectral image is of 100*40pixel, and RGB image is of 
328*235 (the images were acquired before the AMIRS 
finished ,so the resolution of the RGB image is lower 
relatively).The results is showed in figure 7 and figure 8. 
Figure 6. Hyperspectral image and 
high spatial resolution image