A Block-based Inter-band Lossless Hyperspectral

Image Compressor
Marko Slyz, Lei Zhang
1
Abstract: We propose a hyperspectral image compressor called BH which consid-
ers its input image as being partitioned into square blocks, each lying entirely within
a particular band, and compresses one such block at a time by using the following
steps: rst predict the block from the corresponding block in the previous band, then
select a predesigned code based on the prediction errors, and nally encode the pre-
dictor coecient and errors. Apart from giving good compression rates and being
fast, BH can provide random access to spatial locations in the image.
We hypothesize that BH works well because it accommodates the rapidly changing
image brightness that often occurs in hyperspectral images. We also propose an intra-
band compressor called LM which is worse than BH, but whose performance helps
explain BHs performance.
1 Introduction
Each location in a hyperspectral image records the intensity of light received at many
dierent frequencies, both visible and invisible. These many frequencies make iden-
tifying the imaged substances easier [1], but increase the data size.
Several lossless hyperspectral image compressors compress a pixel by rst predict-
ing it from pixels in other bands. In order to account for variations in image statistics,
the methods typically use more than one predictor. They can be roughly classied
according to where they use any given predictor:
on an entire band [2] is an example which also shows a principled way for choos-
ing the order in which to process bands.
on a subset of a band [3] uses each predictor on a row of pixels, [4] uses each
predictor on a set of square blocks, and [5] uses each predictor on an irregularly-
shaped (and usually not connected) region that lies within a single band.
on a single pixel [6] transmits side-information to indicate which of a small set
of predictors to use for any given pixel, while [7] and [8] design their predictors
anew for each pixel.
1
Marko Slyz was with the Department of Electrical and Computer Engineering, McMaster Uni-
versity, Hamilton, ON, Canada. He is now with Bitfone Corporation, Laguna Niguel, CA, USA. Lei
Zhang is with the Department of Electrical and Computer Engineering, McMaster University, Hamil-
ton, ON, Canada. (email: {mslyz,johnray}@mail.ece.mcmaster.ca) This work was supported in
part by CRESTech and PCI Geomatics.
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
Other approaches to hyperspectral image compression are vector quantization and
transform coding; [5, 2] briey surveys many of these. We believe that linear pre-
dictive coding works well enough to be worth studying on its own, and often times
more-complicated methods use linear predictive coding as one of their parts, e.g. [5].
One of our contributions is a compressor called BH which belongs to the subset-of-
a-band category. BH uses each predictor on a single square block, and predicts only
from the corresponding block in the adjacent previous band. At least the predictive
part of BH is basically a variant of [3] and [7], as Section 2 discusses.
BH has a useful combination of several good properties: it gets good compression
rates; requires only about twice the time for compressing or decompressing as merely
reading or writing and byte-order swapping the original image; is relatively easy
to implement; could be implemented to compress the dierent spatial regions in the
image independently so that random access to them is easy; and could be implemented
to require the buering of relatively small portions of the imagemaybe 4.3 Mb while
compressing, and 112 Kb while decompressing.
Another of our contributions is a hypothesisbased on an observation about
hyperspectral image statisticsfor why BH and similar methods work well. Finally,
we propose a compressor, called LM, that is in part based on this same observation
but that does only intra-band prediction. Overall LM is worse than BH, but LMs
performance helps explain BHs performance.
Well cover these subjects in the following order: First well describe how BH and
LM work, then the proposed physical model that helps explain why they work, and
nally, experimental results supporting our claims.
2 The BH Compression Algorithm
Let x be an arbitrary hyperspectral image, and let x
i,j,k
denote the pixel in the ith
line, jth column, and kth band. In this paper, wavelength increases with k.
Suppose that each of the images bands is partitioned into blocks which are usually
square and B pixels on a side, except that near the images edges the blocks might
be smaller. Suppose also that the borders of each bands blocks line up. Let x
i,j,k
be
a column vector containing the pixels in the block whose top left corner is at (i, j) in
band k. The exact order in which pixels are placed into x
i,j,k
doesnt matter except
that it must always be the same.
BH compresses one block at a time. It compresses the stack of blocks covering a
given spatial location all at once, in order of increasing k. When compressing a block
in the rst band, x
i,j,0
, the algorithm assumes that the previous block, call it x
i,j,1
,
is all zeros. To compress x
i,j,k
the algorithm
1. computes the number g that makes g times x
i,j,k1
as similar as possible in the
least-squares sense to x
i,j,k
, namely
g
def
=
x
T
i,j,k1
x
i,j,k
x
T
i,j,k1
x
i,j,k1
(1)
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
2. produces a quantized version of g, called g, with a uniform scalar quantizer that
has 1024 levels stretching from 0.1 to 3;
3. computes the error vector
e
i,j,k
def
= gx
i,j,k1
x
i,j,k
(2)
4. computes the mean absolute error,
M
def
= average{|e
i,j,k
|} (3)
5. applies to M a nonuniform scalar quantizer each of whose bins corresponds to
a dierent precomputed entropy code (Section 3 describes how to choose the
quantizer and codes);
6. transmits g, transmits the bin number from Step 5 and, using the code from
Step 5, transmits the elements of e
i,j,k
.
BH uses the same quantizer and set of entropy codes for every image, and doesnt
adapt them as it runs. To do decompression, BH simply receives g and the bin
number, then uses the code corresponding to that bin number to uncompress the
elements of e
i,j,k
, and nally sets x
i,j,k
equal to gx
i,j,k1
e
i,j,k
.
As mentioned, BH is particularly similar to [3], especially to [3]s SE-o1B predictor
which predicts each pixel from its corresponding pixel in the previous band plus an
ane term. The major dierence is that BH uses each predictor on a spatially
adjacent region of pixels which are likely to have similar characteristics, and not on
a very elongated region of pixels which arent. BH is also particularly similar to [7],
diering in that BH doesnt do a predictor design for every single pixel. Finally, many
video compressors are similar to BH in that they compress blocks in the current frame
(band) by referencing blocks in the previous one. Some dierences are that BH doesnt
compensate for motion since there shouldnt be any, and that BH rst multiplies the
reference block by a scalar before subtracting it from the current block. [9] examined
the use of standard video compressors for lossy hyperspectral compression.
Excluding Step 6, BH requires about eight arithmetic operations per pixel. Specif-
ically, on a per block basis, Step 1 requires about 4B
2
1 operations, Steps 3 and 4
each require about 2B
2
operations and, for our quantizer implementations, Steps 2
and 5 require about 7 and 25 operations respectively on average. This doesnt include
memory access, coding, and I/O times, which seem to be very signicant, but because
counting their operations is more dicult we instead refer the reader to Section 6s
experimental results.
BHs random access capability comes about because the compression of a stack
of blocks corresponding to a particular spatial region, {x
i,j,0
, . . . , x
i,j,1
}, where is
the number of bands in the image, doesnt depend on pixels outside this stack.
BHs small buering capability comes about because during decompression BH
needs to keep just the data corresponding to a single stack in memory. Compression
also works on only a single stack at a time, but at least for frequency-sequential storage
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
orders (like in [10]s images) reading a whole row of stacks at once is convenient. [7]
could also be made to use small buers by storing only the rows of the image that
are necessary for prediction, but their A&P methods would need to be redesigned.
3 The Training Process
The training process to nd BHs quantizer and set of entropy codes is as follows:
run the compression algorithms rst four steps on some training data; collect the
resulting values of M and e
i,j,k
; sort the values of M, partition them into some number
n of contiguous equinumerous groups, and record the maximum (or minimum) values
of M in every group as the quantizer thresholds; design a code for each groups
errors; and, nally, store the quantizer and codes with the rest of BHin the current
implementation theyre compiled into the executable. Section 6s experiments all use
a version of BH trained on the M and e
i,j,k
values from 18068 blocks, each 16 16
pixels in size except near the images edges, that were selected from random locations
in the Moet Field radiance image [10].
We pick n and the block size B by trial-and-error. This is straightforward because
the algorithms performance doesnt seem very sensitive to the choices for n and B
once they reach a particular range of values. Figure 1 shows an example of this for
block size. Here the overall compression rate doesnt change much for blocks of size
8 8 up to blocks of size 256 256. The values we use in Section 6s experiments are
B = 16 and n = 25.
0
1
2
3
4
5
6
7
8
4 16 64 256 1024
r
a
t
e

(
b
p
p
)
B
total
error
overhead
Figure 1: The total compression rate (top line) on the Cuprite image [10], the rate for
just the errors (middle line), and for just the overhead (bottom line). The quantizer
and entropy codes are kept xed to the Moet Field-trained ones described above.
[4]s use of an equiprobable quantizer is similar to BHs, while [11, 2.2] uses such
quantizers in a somewhat less related way. Other kinds of quantizers likely would
work as well.
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
The current BH implementation uses canonical minimal redundancy codes as de-
scribed in [12], using their Table-Lookup method with x = 8 for decompression. Code
lengths for this design essentially come from the Kraft-coder algorithm with tree ll-
ing that [13] describes, although the ordinary Human algorithm would have worked
just as well.
4 The LM Compression Algorithm
Let w
i,j,k
, the context, be a vector of pixels that are all in band k, and that are taken
from some xed set of locations near i, j. A simple compressor, call it LMi, predicts
x
i,j,k
from its context w
i,j,k
, using a predictor designed to be good at predicting the
previous few pixels at location i, j from the corresponding contexts in each of their
bands. LMi is basically a backward-adaptive version of [3], with the compression
proceeding rst in the frequency direction. A more sophisticated compressor, call it
LM, trains several predictors each using a dierent kind of context, and then selects
the predictor that worked best on the past few pixels.
A more detailed description of LM is as follows. Let u vary from 1 to some number
r. Let w(u)
i,j,k
be one of r dierent contexts, chosen so that contexts with the same
index u but from dierent bandse.g. w(u)
i,j,k
and w(u)
i,j,m
for m not equal to
ktake their pixels from the same spatial locations and in the same order. Also let
u index a set of positive integers {M
u
}. Then to code the pixel at x
i,j,k
, LM
1. for each u forms and solves the equation

w(u)
i,j,k1
.
.
.
w(u)
i,j,kMu

a
u

x
i,j,k1
.
.
.
x
i,j,kMu

(4)
for a
u
in the least-squares sense.
2. picks the a
u
that minimizes the error in predicting some number N of past
pixels, i.e. that minimizes

w(u)
i,j,k1
.
.
.
w(u)
i,j,kN

a
u

x
i,j,k1
.
.
.
x
i,j,kN

(5)
3. uses this a
u
to predict x
i,j,k
, and selects a Golomb code based on (5) as in [14].
A slightly more exible approach is to allow N to be dierent for this code
selection than it was for the a
u
selection.
The decompressor can do these same steps since all of them depend only on pixels
that it has already seen.
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
The motivation for LMi and LM is the idea that the nearby spectra, or maybe
some weighted average of them, are similar to the current one, and that the diculty
is picking which weighted average to use.
It might seem at rst that LM isnt any more powerful than LMi, since one could
choose to let LMis context vector contain the pixels in the union of all of LMs
contexts, and then least-squares would eectively pick a good subset. Unfortunately
a large context like that one would require more training data, but this data might
not be available if the image statistics change near i, j, k. [15]
5 A Hypothesis About Why BH Works
BH does well on a particular model which seems to t hyperspectral images at least
to some extent. The model says that the spectra near any arbitrary location i, j in
the image are each the product of a constant underlying radiance spectrum v, and a
scalar brightness factor c
i,j
. In other words,
x
i,j
c
i,j
v for (i, j) R (6)
where R is a relatively small region of the imaged area. A rationale is that the c
i,j
takes
into account rapidly varying brightness changes due to changes in terrain inclination
and possibly other factors [1, p.12], leaving the components of the spectrum that vary
slowlylike the substance being imaged, the atmospheric eects, and the spectral
properties of the illuminationin v.
Visual inspection of groups of spectra selected from randomly-chosen locations in
some of the [10] images helps support this model, though this is somewhat subjective.
Figure 2 shows a common situation where adjacent spectra from a region in the Jasper
image follow (6) in that they have essentially the same shape, and vary mainly in
brightness. Thats at least the main trend; a point against this, which is observable
in the color version of Figure 2, is that the spectra occasionally change their relative
positions, especially near sharp transitions. The other common situation is when
spectra taken from a small region change their separation with increasing wavelength,
still having peaks and troughs in the same places, but it seems that these changes
are usually gradual enough that the model is still approximately correct over small
intervals of bands.
The standard Spectral Angle Mapper (SAM) algorithm [16, p.12-13] is also es-
sentially based on the idea of dierent brightness factors multiplying an underlying
spectrum (SAM ignores these brightness factors), although SAM doesnt use the idea
that spatially nearby spectra are similar.
Now (6) is interesting because BH and similar methods can perfectly predict pixels
that follow it. To see this, let x
i,j,k
be a block of size B B at an arbitrary location
in the image; let g be the minimizer, as in (1), of gx
i,j,k1
x
i,j,k

2
; and let c be
a vector having B
2
elements that contains the brightness factors for each of the B
2
spectra in the block. If (6) holds for the block, with R containing the blocks B
2
locations, then
gx
i,j,k1
x
i,j,k

2
= gcv
k1
cv
k

2
(7)
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
0
500
1000
1500
2000
2500
3000
3500
4000
0 50 100 150 200
r
a
d
i
a
n
c
e
band
Figure 2: A set of 25 image spectra from near location 320,1020 in Jasper.
where v
k1
and v
k
are the k1st and kth elements of the underlying radiance spectrum
v. Now by choosing g to equal v
k
/v
k1
, (7) will be zero.
Instead of predicting across frequency bands like BH, LM predicts a spectrum from
an adaptively weighted combination of its neighbors. This ought to accommodate
(6), but it should also be able to accommodate more signicant dierences. One
disadvantage of LM is that, to the extent the model is true, BH has many data
samples (B B dierent pixel pairs) to do its estimation while LM typically has
fewer.
Finally, Figure 1 suggests, as does direct examination of the underlying spectra,
that R is not necessarily all that small. The spectra appear similar to some extent
for even quite large blocks, although for large blocks there are more cases of spectra
having signicant dierences that the model doesnt even approximately account for.
6 Experimental Results
Table 1 shows compression rates on the raw AVIRIS radiance images from [10]. The
image names are listed in the left-hand column, with the AVG\l.a. row denoting
the average excluding the Low Altitude image.
BG is a version of BH that uses Golomb coding, selecting the coding parameter
for each block based on the blocks errors as in [14]. LM (1/1) uses two rst-order
predictors for every pixelone predicting from the spectrum at i 1, j, and one from
the spectrum at i, j 1and trains these on the previous two image bands. LM (5/1)
also used two predictors, a fth order one and a rst order one, training them on the
previous 30 and 2 image bands respectively. We chose these particular parameters as
the best performing ones out of many candidates.
The remaining results are from the literature. [5] uses vector quantization to clas-
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
sify the image spectra, but then uses inter-band linear prediction. LPVQs predictor
[17] is based on vector quantization and isnt linear. Its somewhat similar to LM in
that it predicts within bands, and that the predictions in one band indirectly inuence
the predictions in others. SLSQ-OPT[8] uses linear prediction based both on pixels
within and outside the current band. The standardized JPEG-LSs results are from
[17].
[7] doesnt generally have results broken down by image, but does quote 4.95 bits
per pixel as their best rate on the set excluding Low Altitude.
image BH BG LM (1/1) LM (5/1) [5] LPVQ
SLSQ-
OPT
JPEG-LS
cuprite 5.12 5.15 5.49 5.28 4.68 5.11 4.94 7.66
jasper 5.23 5.25 5.51 5.42 4.62 5.67 4.95 8.38
low altitude 5.52 5.52 5.77 5.60 5.54 5.26 8.00
lunar lake 5.12 5.15 5.48 5.26 4.75 4.95 4.95 7.48
moet 5.26 5.26 5.48 5.35 4.62 5.44 4.98 8.04
AVG 5.25 5.27 5.55 5.38 5.34 5.02 7.91
AVG\l.a. 5.18 5.20 5.49 5.33 4.67 5.29 4.96 7.89
Table 1: Comparison of Compression Rates. All values are in bits per (16-bit) pixel.
Table 2 shows BHs compression and decompression speeds. It also has columns
for NULL, a trivial compressor that just writes (for compression) and reads (for
decompression) the image in its raw format, only doing conversion from big-endian
to little-endian order. We did these experiments on a commodity PC with a 2.53GHz
Pentium 4 with 753Mb of RAM, and an IDE hard drive with DMA turned on. In
implementing BH we did pay some attention to run-timefor example, by using our
own, simpler, I/O buering rather than the one that comes with getchar() and
putchar().
The compression speed on Low Altitude seems anomalous. This probably has to
do with memory paging, which is necessary for Low Altitude and not for the others
because the current BH implementation buers the entire image, and only in Low
Altitudes case wont this buer t into core memory.
image size BH NULL
Ct Dt Ct Dt
cuprite 579 58 51 29 22
jasper 678 68 64 34 29
low altitude 968 455 247 49 54
lunar lake 375 41 33 19 15
moet 533 51 48 27 23
Table 2: The size of each image in megabytes, and the times in seconds for compressing
(in the Ct columns) and decompressing (in the Dt columns).
The current, not optimized implementations of LM take about 1030 seconds to
compress Cuprite in the (5/1) case, and about 590 seconds in the (1/1) case. Decom-
pression times are similar.
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
[7]s best method (as quoted in Table 1) compresses at 1150 Kb/s, and decom-
presses at 2860 Kb/s on an 800 MHz Pentium-III. Multiplying these speeds by 3 to
account for the dierences in machine speed, which is optimistic, gives about 172 sec-
onds to compress Cuprite and about 70 seconds to decompress it. A dierent variant
in [7] that gets 5.10 bits/pixel would similarly require about 50 seconds to compress
Cuprite, and about 51 seconds to decompress it. [5] says that their method requires
1180 seconds, also on an 800 MHz PC, to compress an image. Although [5] doesnt
give a time for decompression, its likely a lot faster since theres no VQ design then.
The other papers dont provide specic run times. LPVQ would seem to have
slow compression since it requires multiple unconstrained VQ designs. But LPVQs
forward-adaptive nature likely makes decompression about as fast as BHs. SLSQ-
OPT seems like it would take a long time for both compression and decompression
since it needs to design an order-19 linear predictor for each pixel in either case.
7 Discussion
BHs average compression rate is about 11% worse than [5]s, the lowest-rate lossless
hyperspectral compressor that were aware of, but BH ought to be a lot faster. BH
also gets worse compression than SLSQ-OPT, but again ought to be much faster.
BHs compression rates are more similar to [7]s, especially to their faster methods.
Its not surprising that [7] compresses better, because [7] designs a dierent predictor
for every pixel, and not just for groups of them like BH; but for this same reason [7]
should be signicantly slower and it doesnt seem to be. Some possible reasons for
this are the overly-optimistic adjustment for clock speed dierences mentioned above,
that [7] codes some of the bands without prediction, and that I/O and memory access
is so signicanttaking up at least 1/3 to 1/2 of the time in BHs case according to
Table 2that [7]s extra few operations per pixel dont matter much.
The major dierence between BH and [5] is that [5] can accommodate spatial
dierences in image statistics on a ner scale, since it groups together similar spectra
before doing prediction, while BH just uses a xed partition into blocks. The rela-
tively small gap between BHs and [5]s performance seems to support the idea that
most of the ability to compress comes from being able to accommodate dierences in
brightness. LMs performance seems consistent with this as well since, as mentioned
in Section 5, LM can also accommodate spatial dierences in radiance spectra on a
ner scale than BH, but in fact LM does a little worse.
Our nal observation is that BG gets only about 1% worse compression than BH.
Implementing Golomb codes in a straightforward way is easier than implementing
canonical codes, but canonical codes can be somewhat faster.
Acknowledgements
Were grateful to Xiaolin Wu for guiding us and helping us to design the compression
algorithms, and to Kristi Markham, William Morris, and Richard Pollock for helpful
discussions.
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.
References
[1] R. B. Smith, Introduction to hyperspectral imaging.
http://www.microimages.com/getstart/hyprspec.htm, 2001.
[2] S. Tate, Band ordering in lossless compression of multispectral images, IEEE T. on
Computers, vol. 46, pp. 47783, April 1997.
[3] R. Roger and M. Cavenor, Lossless compression of AVIRIS images, IEEE T. on
Image Processing, vol. 5, pp. 71319, May 1996.
[4] B. Aiazzi, L. Alparone, and S. Bartoni, Near-lossless compression of 3-d optical
data, IEEE T. on Geoscience and Remote Sensing, vol. 39, pp. 25472557,
November 2001.
[5] J. Mielikainen and P. Toivanen, Clustered DPCM for the lossless compression of
hyperspectral images, IEEE T. on Geoscience and Remote Sensing, vol. 41,
pp. 29432946, December 2003.
[6] N. Memon, K. Sayood, and S. Magliveras, Lossless compression of multispectral
image data, IEEE T. on Geoscience and Remote Sensing, vol. 32, pp. 28289,
March 1994.
[7] J. Mielikainen, P. Toivanen, and A. Kaarna, Linear prediction in lossless compression
of hyperspectral images, Optical Engineering, vol. 42, pp. 10131017, April 2003.
[8] F. Rizzo, B. Carpentieri, G. Motta, and J. Storer, High performance compression of
hyperspectral imagery with reduced search complexity in the compressed domain, in
Data Compression Conference, pp. 47988, 2004.
[9] T. Wilkinson and V. Vaughn, Application of video-based coding to hyperspectral
imagery, in Hyperspectral Remote Sensing and Applications (S. Shen, ed.), vol. 2821,
pp. 4454, SPIE, 1996.
[10] JPL, AVIRIS free standard data products.
http://aviris.jpl.nasa.gov/html/aviris.freedata.html, 1997.
[11] M. Weinberger, G. Seroussi, and G. Sapiro, LOCO-I: A low complexity,
context-based, lossless image compression algorithm, in Data Compression
Conference, pp. 140149, 1996.
[12] A. Moat and A. Turpin, On the implementation of minimum redundancy prex
codes, IEEE T. on Communications, vol. 45, pp. 12001207, October 1997.
[13] M. Liddell and A. Moat, Length-restricted coding using modied probability
distributions, in Proceedings of the 24th Australasian Conference on Computer
Science, (Gold Coast, Queensland, Australia), pp. 117124, IEEE, 2001.
[14] M. Slyz and D. Neuho, Some simple parametric lossless image compressors, in
Proceedings of the IEEE International Conference on Image Processing, vol. I,
(Vancouver), pp. 1247, September 2000.
[15] H. Ye, G. Deng, and J. Devlin, Adaptive linear prediction for lossless coding of
greyscale images, in Proceedings of the IEEE International Conference on Image
Processing, (Vancouver), September 2000.
[16] F. Kruse, A. Lefko, J. Boardman, K. Heidebrecht, A. Shapiro, P. Barloon, and
A. Goetz, The spectral image processing system (SIPS)interactive visualization
and analysis of imaging spectrometer data, Remote Sensing of Environment, vol. 44,
pp. 145163, May-June 1993.
[17] G. Motta, F. Rizzo, and J. Storer, Compression of hyperspectral imagery, in Data
Compression Conference, pp. 33342, 2003.
Proceedings of the 2005 Data Compression Conference (DCC05)
1068-0314/05 $ 20.00 IEEE
Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on September 10, 2009 at 04:02 from IEEE Xplore. Restrictions apply.