International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
2. METHODOLOGY
2.1 Data Analyse
Three types sensor multispectrum images are discussed in this
paper for compression experiment. Table 1 gives the
multispectrum images information.
Table 1 Test Images
of the same objects. Although the pixels of the same object
have some difference, there are presenting the same object and
have the same geometric feature such as the edge and structure,
which mean the strong correlation. And these band redundancy
is determined by the resolution of speactrum. High spectrum
resolution has higher spectrum redundancy .
Image Name | Sensor pipe Resolution Area Image Size(Pixel)
TM Landsat 5* 20m Suburb of Wuhan 512x512
Imaging
E Spectrum 16 100m Qinian Mountainous Area | 480x480
Instrument
TDF MODIS 34 200m South China 1354x4670
*TM has 7 bands, while the correlation between band 1 and 2
is too high and the resolution of band 6 is not same to others
band, we omit band 2 and 6.In fact only 5 bands is in use and
we number them T1, T2, T3, T4, TS in turn.
Table 2 Multispectrum Images Correlation Analyse Results
Table 2 is the spectrum correlation among the 5 band Landsat
images, 34 bands MODIS image and 16 bands imaging
spectrum sensor images.
Band Reference ; ;
Sensor Type Number [mae Correlation Coefficient Average
LandSat 0.855/0.983 ;
TM S 3 /0.950/0.974 0.941
Imaging 0.939/0.940/0.942/0.943/0.944/
Spectrum 16 16 0.945/0.947/0.948/0.949/0.95/ 0.946
Instrument 0.95/0.95/0.95/0.95/0.95
0.962/0.956/0.960/0.960/0.965/0.970/0.975/0.980/
0.985/0.990/0.995/1.000/0.995/0.990/0.990/0.992/
MODIS 34 is 0.995/0.999/0.998/0.996/0.992/0.989/0.986/0.985/ 0:951
0.983/0.981/0.9790.977/0.974/0.974/0.975/0.976/0.975
2.2 Correlation Analyse of Multispectrum Images
The compression technique of multispectrum images is an
urgent problem for the remote sensing image storage, image
database and transmission. All the compression technique is
realized through remove data redundancy from the Shannon
entropy. Different kinds of images have the different
redundancy characteristic. The multispectrum have two kind of
redundancy: spatial redundancy and spectrum redundancy .
The spatial redundancy presents the correlation of the neignbor
pixels in some specific band which is simular to single band
remote sensing image and the compression can be realized
through general compression algorithms such as JPEG2000.
The spectrum redundancy including the statistic redundancy
and structure redundancy . Thus, the multispectrum images
compression focus on both the pixel correlation between
neighbour pixels and edge structure of some object in different
band images.
There are various compression methods for spatial
decorrelation(Jain, 1991), relatively few studies for spectral
decorrelation across bands have been presented. Saghri, 1995
use Karhunen-Loeve transform (KLT), which is theoretically
the optimum method to spectrally decorrelate the image data.
Comparing to the common single band remote sensing image,
the character of the spatial correlation among the multispectrum
images is the mode localization. The multispectrum images are
a group of images to the same region in different bands. Each
band has an image while the imaging objects are the same.
HIRIS has 192 bands thus there are corresponding 192 images
As table 2 shown, the band image correlation of the same
region is near 95%, which means there are strong spectrum
correlation among the multispectrum remote sensing and these
spectrum correlation redundancy will be higher as the band
increases. And the compression algorithm proposed in this
paper mainly focus on this spectrum redundancy .
2.3 Multi-band Wavelet Analyse
Wavelet transform is always used in the image compression.
After wavelet transformation, the image is divided into different
presentation. Thus, the different quantification policy is design
to the wavelet coefficient to obtain compression. While, the 2D
wavelet are indeed well-suited to present smooth and textured
region of images, the 2D wavelet description of edges are high
inefficient in some specific scale. The multi-band wavelet can
express edge or object geometric feature more accurately. This
can be used for edge reserved compression to obtain high
reconstruct image quality. In this paper, the 3D wavelet
transformation is adopt as the multi-band wavelet analyse to
obtain much precise edge. Following is the 3D wavelet
transformation analyse.
(1) Multiple scale analyse
2
The multiple scale analyse of spatial domain E (R) refers to
pt ;
the spatial list which fulfil the following
conditions:
56
Internat
Q
Thus, 4
followi
AN
nez
where
[7
(2) Co
As the
propert
filterin;
Thus, tl
equatio
(3) Im
As the
wavele
Taken
of M-1
a
JH,
