2007 asprs

DISCRETE WAVELET TRANSFORM EFFECT ON HYPERSPECTRAL IMAGE
CLASSIFICATION PERFORMANCE USING LOSSLESS COMPRESSION
ALGORITHMS WITH COLOR & QUALITY SCALABILITY UNDER JPEG2000
David Negrete, Vikram Jayaram, Bryan Usevitch
Signal Processing & Communication Research Group
The University of Texas at El Paso
500 W. University Ave., El Paso, TX 79968-0523
{Negrete, Jayaram, Usevitch}@ece.utep.edu

ABSTRACT
In this paper classification performance of Hyperspectral images using two different data decorrelating transforms is
investigated. The tradeoff between the discrete wavelet transform (DWT) and the principal component analysis
(PCA) is measured using different compression algorithms found in the JPEG2000 standard. These cross-component
transforms can be evaluated in order to obtain a comparison of how well each transform can decorrelate data,
reorder data optimally, speed of implementation, and effect on mineral classification performance. In addition, two
scalability options defined by JPEG2000 were used: color and quality. Compression algorithms were carried out for
the DWT and the PCA using the same set of parameters; the same Hyperspectral image, the same classification
algorithm—spectral angle mapper (SAM), and the same type of compression: lossless compression. The purpose of
this comparison is to be able to determine how much partial decompression is needed in order to achieve a 90% hit
rate when classifying several minerals. This evaluation is focused primarily on two types of Mica called Aluminumrich and Aluminum-poor Mica. Our past research using PCA has shown extremely good overall classification
performance. However, this transform has the disadvantage of higher computational complexity. This prompted us
to try out an alternate lower complexity transform such as the DWT. The classification results were found poor as
compared to PCA. The complex nature of our classification does not allow us come to a complete judgment of this
finding. Thus, other data sets involving low complex nature of classification shall be dealt in our future work. The
simulation results conclude that the DWT is seven times faster than PCA and allows parallel processing huge data
sets.

INTRODUCTION
Hyperspectral images are large data sets which are used by geologists to detect and classify minerals. These
images are obtained by airborne or satellite-borne satellites by a method called remote sensing. Every mineral or
plant reflects and/or emits a unique amount of electromagnetic energy distributed among many frequency bands.
Remote sensing works by measuring this energy and then using special software to decode this “spectral signature”
and assign it to its proper source, be it a plant or mineral. The Hyperion (Hyperspectral Imager) provides a high
resolution Hyperspectral image capable of resolving 220 spectral bands (0.4 - 2.5 µm) with a 30 meters pixel
resolution. The data used in these experiments is the Mt. Fitton scene in southern Australia. The layers or bands of
this Hyperion data are highly correlated to each other in the spectral direction. Decorrelation of these bands allows a
more compact representation of an image by packing the energy into fewer numbers of bands. Part 2 of the
JPEG2000 standard recommends the use of either the PCA or the DWT for decorrelation. Although PCA is optimal
transform when it comes to efficiently packing band energies in fewer dimensions nevertheless its computationally
inefficient nature leads us to employ DWT for decorrelating the bands. The DWT can be implemented by either a
convolution or a lifting scheme. We have used the lifting scheme implementation of DWT due to its lesser
computations and memory requirements. In our experiments the 5/3 Daubechies filter is used as we are performing
lossless compression.
The Hyperion instrument can image a 7.5 km by 100 km land area per image and provide detailed spectral
mapping across all 220 channels with high radiometric accuracy. The Hyperion instrument, in conjunction with
advanced land imager (ALI) instrument which provides accurate radiometric calibration on orbit using precisely
controlled amount of solar irradiance, forms the standalone camera system on the Earth Observatory-1 (EO-1)
satellite [7]. The analysis of Hyperspectral Imagery by means of such spectrometers (Hyperion) exploits the fact that
each material radiates different amount of electromagnetic energy throughout all the spectra. This unique
ASPRS 2007 Annual Conference
Tampa, Florida May 7-11, 2007

characteristic of the material is commonly known as the spectral signature which can be read using such airborne or
space borne-based detectors. The data used for experiments was collected by the Hyperion over the semiarid Mount
Fitton area which is located in the Northern Flinders Ranges of South Australia. Centered at −29° 55´ S and 139°
25´ E, about 700 kilo-meters NNW of Adelaide. This Hyperion image consists of 194 atmospherically corrected
contiguous spectral bands. Each band has dimension of 6702 X 256 X 2 bytes, which is equal to 634.857 MB in
memory utilized by the whole data cube. The most obvious problem with using data cubes this big is that storage
becomes a problem when managing multiple files. In addition, moving these files is not easy due to their size.
Therefore, an appropriate algorithm is needed to compress files that are larger than usual. On the other hand, the
compression algorithm should minimize the difference, or error, between original and decompressed images when
using lossy algorithms.

Figure 1. Hyperspectral remote sensing.

JPEG2000 FRAMEWORK
The recent JPEG2000 standard has many advantages compared to earlier compression standards. Such
advantages include superior low bit rate performance, bit rate scalability and progressive transmission by quality or
resolution. The coder and decoder for JPEG2000 use a wavelet based compression method which implements the
embedded block coding with optimized truncation (EBCOT) [5] scheme. The use of wavelets to decompose a
signal, unlike the old standard JPEG that relied on the discrete cosine transform (DCT) to approximate a signal, has
proved to have numerous advantages. For example, wavelets provide localized support which is more useful to
approximate high frequency components in an image. In addition, with appropriate coding algorithms, such as
embedded zero-tree wavelet (EZW) by Shapiro [6] or Set Partitioning in Hierarchical Trees (SPIHT) by Said and
Pearlman [9], it has been shown that the wavelet transform yields better compression than the cosine transform [9].
These algorithms are useful since they provide an ordering of data according to its importance. Specific details
about the DWT and the ordering of data will be explained later.
Quality scalability is achieved by dividing the wavelet transformed image into code-blocks. After each codeblock is encoded, a post-processing operation determines where each code-block’s embedded stream should be
truncated in order to achieve a predefined bit-rate or distortion bound for the entire data. This bit-stream
rescheduling module is referred to as Tier 2. It establishes a multi-layered representation of the final bit-stream,
guaranteeing an optimal performance at several bit-rates or resolutions. Figure 2 illustrates the general flowchart of
the JPEG2000 architecture. The Tier 2 component optimizes the truncation process, and tries to reach the desired
bit-rate while minimizing the introduced distortion or mean squared error (MSE), utilizing Lagrangian rate
allocation principles. This following procedure is known as post compression rate-distortion (PCRD) optimization
[4] and the basic principles behind it are extensively discussed in [5].
The new image coding standard JPEG2000 does not only offer better compression performance than the
previous one, but it also offers more options when decompressing an image. The four types of scalability presented
are: spatial, resolution, color, and quality scalability. The first two take advantage of the two-dimensional wavelet
transform to decompress a portion of the image or reduce its dimensions respectively. Specific details about these
types of scalability can be found in [5]. However, the performance comparison between the PCA and the oneASPRS 2007 Annual Conference

dimensional DWT used to decorrelate the image components do not take advantage of these two scalability options.
Therefore, is not pursued in the scope of this paper.

Figure 2. JPEG2000 architectural flowchart.

The other key advantage of the JPEG2000 standard in HSI compression is that due to its ordering of data,
smaller files can be obtained from a compressed file by truncating the bit stream where needed. In other words, a
compressed file using 4 bits per pixel of an image could be truncated in half to obtain a file using 2 bits per pixel.
Due to these advantages and the availability of color and quality scaling, which will be explained later, the
JPEG2000 standard was chosen as the method to compress the Hyperspectral images. The JPEG2000 standard has
been established and its software implementations have started to become available. However, a three-dimensional
transform has not been determined nor specified by the standard for multi-dimensional electro-optical images, which
are those that have more than three color components. Therefore, it is now required to use a transform that will
remove redundancy in the spectral direction. HSI compression has relied mainly on the use of the PCA transform to
transform images in the spectral direction [1]. Its main advantage is that it not only allows the most decorrelation
possible among spectral bands, but also the best energy compaction, which in turn results in better mineral
classification. On the other hand, a main disadvantage of the PCA is that the covariance matrix, which is used to
decorrelate among frequency bands, has to be calculated, and therefore is data-dependent. In addition, its
implementation is computationally expensive and does not help to decrease the number of time and computations
needed to classify materials. On the contrary, decompression times can be increased by several times.

USE OF DECORRELATING TRANSFORMS
Color and quality scalability are options that can be applied to HSI or its 3-D transformed version. Before
explaining color and quality scalability, an important property of the PCA and DWT needs to be considered. After
transformation of the data cube in the spectral direction, an interesting result is obtained. Figure 3 shows a hypothetical
distribution of energy after the ten components of an image cube undergoes transformation.
It can be seen that most of the energy in the data cube is compacted into the first few bands. In addition, each
band has less energy than the ones before. Therefore, data is compacted and ordered by the transform. Color
scalability takes advantage of this order by decreasing the number of color component bands that are used to
decompress an image. In other words, the image decoder only decompresses a predefined fewer number of bands,
assuming the rest to be zero, and performs the inverse transform to approximate the original image.
One of the advantages of color scalability is that it can reduce the amount of time spent on decompressing an
image since only a few components are used. Quality scalability relies on the energy compacting property of the
transform. Instead of using the first bands of a transformed data cube like color scalability, quality scalability assigns
larger amounts of the bit budget to those bands that are believed to be more important. This means that by
distributing the bit budget accordingly, a measure of error can be optimally decreased.


Figure 3. Energy plot of an image with ten components after transformation.

COMPRESSION PERFORMANCE
Since JPEG2000 supports compression of images with more than three color components, but does not define
the that transform in the color component direction, the most important aspect that has to be addressed is that of
decorrelation among the bands, which will result in higher compression ratios. The cross-component transform had
to be written and implemented separately from current JPEG2000 compression software in order to use off the shelf
software. In addition, once a Hyperspectral image has been transformed, it is divided into its color components.
Each file contains a single color component band, which results in a total of 194 files that can then be compressed
independently. Lossless compression algorithms were used for all experiments. The method is carried backwards
when decompressing—component bands are first merged and then inverse transformed to obtain the original image.
The PCA is defined as the transform that achieves the most decorrelation possible among bands. Since
compression is directly related to decorrelation, the PCA transform is also the one that achieves most compression.
Therefore, the first item to be compared is compression performance. Previous results show that when using the
PCA transform and lossless compression, the resulting 194 files had a combined size of 271,022,749 bytes, or
258.467 MB. The compressed file is only 40.71% of the original file. However, since the decoder requires the
transformation matrix used for the PCA, two more files have to be included. The first file contains the mean vector
and its size is 776 bytes. In addition, the second file of size 304,842 bytes contains the transformation matrix. This
results in a total file of size 271,328,367 bytes, equivalent to reducing the original file to 40.76% of its size. The
next step was to repeat the procedure, with the only change being the use of the DWT to decorrelate component
bands. The compression algorithm was followed exactly. Now, the 194 files had a combined size of 298,548,385
bytes, or 284.718 MB. This yields a compression percentage to about 44.84% of the original size. This means that
the DWT—a transform that is significantly faster, is not data-dependent and does not require additional files for
decompression—has a compression performance of about 4% less than that of the PCA.

COLOR SCALABILITY USING JPEG2000
Cross-component transforms, such as the DWT and PCA, have the objective of compacting energy into as few
component bands as possible. By using only the most important bands of an image, some compression can be
achieved. This is what is referred to as color scalability. Using a cross-component transform makes it possible to
distinguish which bands are more important in an image.
Just like the PCA, the DWT tries to compact as much energy into the first bands as possible. However, the
DWT does not achieve as good data ordering as the PCA. In other words, after using the DWT, the component
bands do not necessarily decrease in energy. Using the first component bands to decompress an image might result
in a reduction of mineral classification performance due to the fact that these bands do not contain the most
information. Therefore, two methods were followed to decompress an image using color scalability. The first
involves using the order of the component bands after implementing the Discrete Wavelet Transform. This means,
that the first bands are always used. The second method requires using another parameter to measure the importance
of the component bands. This results in having the bands reordered. Using the order of component bands after
applying the DWT is required for this method. The algorithm followed is shown in Figure 5.


Figure 4. Comparison of energy distributions among bands after PCA and DWT transforms.
Applying the cross-component transform is the first step required. As described previously, this transform is
implemented separately of current JPEG2000 implementations. Once this is done, each component band is saved in
a new file, therefore creating 194 files, which are compressed independently. This is where color scalability takes
place: only a few bands are decompressed. Let the number of component bands used be N. Obviously, N can only
be an integer less than or equal to 194. These bands are then merged, and the rest of the component bands are
assumed to be zero. The inverse transform is performed, which results in an approximation of the original HSI.
Finally, mineral classification is performed and results are taken. Several compression rates were obtained by using
a set of few numbers of component bands. This means that this method was followed when using 7, 13, 25, 49, and
97 component bands to approximate the original image. One of the advantages of color scalability is the
considerable reduction of time required to decompress an image, since only a few bands are decompressed. In
addition, this method is also easily implemented since additional files are not required. For instance, a
transformation matrix, such as the one used for the PCA, is not needed. However, it is expected that this method
will result in slightly worse classification performance because of the energy compaction properties of the DWT.

Encoding Stage

Hyperspectral
Image

Hyperspectral
Image
approximation

Forward
Transform

Split
Color Bands

Compress 194
Color Bands

Inverse
Transform

Merge
Color Bands

Decompress N
Color Bands

Decoding Stage
Mineral
Classification
Figure 5. Color scalability algorithm.

Alternative Method: Ordered by Variance
This method requires the use of a parameter to measure component band importance. One important
assumption is made here: the variance of a component band is directly related to the amount of information it
contains. Therefore, the variance of each component band was used to measure its importance. In other words,
bands with higher variances are more important and are decompressed first than bands with smaller variances. The
procedure followed for this method is almost the same as the first one. The only differences include having to
calculate the variance of each component band and creating a file with the order of the bands according to their
variances.
The variance of each band is calculated once the Hyperspectral image has been split into 194 files. Their use is
to distinguish which component bands contain more information. Therefore, a file with the order of the variances is
created. In other words, the number of the component band with highest variance is saved first, followed by the
number of the second highest variance band, and so on. This file is not strictly required by the decoder, but it is
created in order to avoid having to recalculate variances each time a file needs to be decompressed. The file is used
to know the order in which component bands need to be decompressed. When the required number of component
bands has been reached, the procedure described previously is followed exactly. Although this method is expected to
yield better classification performance, many of the advantages for the previous method are still present. For
instance, decompression times are also reduced. The only additional computations come from the calculation of the
variance for each component band, but this is done only one time for each image. In addition, the overhead created
by this method is very small, almost negligible.

QUALITY SCALABILITY USING JPEG2000
Quality scalability works by finding optimal bit allocations among different sources of information, such as
component bands. There exist numerous techniques to find this optimal bit allocation. The objective of this
allocation is to minimize a measure of error, subject to the total bit rate constraint. MSE is commonly used as the
error parameter because of many mathematical advantages that is has, such as being a continuous function that can
be differentiated when needed. In this manner, a technique called Lagrange multipliers is used to obtain this bit
allocation among the component bands, which is detailed in [5] and [2]. Lagrange multipliers are used because a set
of equations is subject to a constraint, which in this case is the available bit budget. An advantage of using Lagrange
multipliers is the simplicity involved. The set of equations that solve the problem are usually simple, and their
calculation generally involves few computations. Let R be the available bit rate that will be distributed among the N

component bands. The constraint in this case will be

R=

1
N

N

∑R
k =1

g (x ) = 0 = R −

and

k

1
N

N

∑R

k

k =1

.

Here, Rk is the bit rate assigned to each component band. Using MSE to measure error, the distortion introduced in
the kth band, σ rk is directly related to the quantizer error, and it is defined as
2

2
σ rk = ε 2 2 −2 R σ k2 ,
k

where ε2 is a constant dependent on the input distribution and quantizer, which will later cancel out. This means that
this technique is appropriate for any type of distribution or quantizer. The input variance of each component band is
denoted by σ k . The function to be minimized is the sum of these errors given by
2

N

N

k =1

k =1

2
f ( x ) = σ r2 = ∑ σ rk = ∑ ε 2 2−2 Rk σ k2 .

The method of Lagrange multipliers indicate that such problem can be solved by

∇f ( x ) + λ ∇ g ( x ) = 0
∂
∂Rk

⎛ N 2 −2 Rk 2
⎡
1
⎜ ∑ ε 2 σ k + λ ⎢R −
⎜
N
⎣
⎝ k =1

N

∑R
k =1

K

⎤⎞
⎥⎟ = 0 .
⎟
⎦⎠

Solving this equation for Rk will eventually lead to the next result we arrive at

σ k2
1
Rk = R + log 2
,
1
2
⎡ N
⎤ N
2
⎢∏ (σ j )⎥
⎣ j =1
⎦
which displays the relationship between the variance of each component band and the geometric mean of the
variances to the fraction of the bit budget assigned to it. In other words, just like when using component scalability,
component bands with higher variances are believed to be more important than others. The difference now is that
component bands might not be decompressed completely; instead, truncated bit streams defined by the JPEG2000
standard will be decoded. These shorter bit streams will always be close to being optimal for lossy compression
when using multiple layers during encoding.

Using Log of Variances
The method of Lagrange multipliers reaches a very simple set of equations as a result. Each component band’s
bit rate will be the sum of the desired target bit rate and half of the log2 of the ratio of the input variance to the
geometric mean of all the variances. The latter part of the equation is what gives name to the result of using
Lagrange multipliers: log of variances. Implementing these equations is relatively straightforward and does not
require much processing time. The only computationally-intensive task that has to be performed before solving
these equations is calculating the variance of each component band. The exact same procedure was followed when
using color scalability, with the difference that the method of Lagrange multipliers will require the use of the actual
values of each variance. In this manner, the variance of each component band is calculated independently and then
saved for later use. This creates an additional file, which is very small in size. In fact, using a 32-bit coefficient to
represent each variance will be enough. This results in an additional file of size 4 x 194 = 776 bytes, which is less
than 1 KB, thus creating relatively no overhead.
The actual procedure followed is somewhat similar to the one used when doing color scalability. Figure 6
shows the algorithm and presents some differences. The actual bit rate allocation has to be calculated each time a
new compression rate is desired. A problem of using Lagrange multipliers is that the equations to be solved are not

forced to result in non-negative numbers. In other words, the optimal bit allocation might require some component
bands to have negative bit rates, which is obviously not possible. Bands not being decompressed will have a bit rate
of zero, which will result in an increase of the target rate. Therefore, an iterative process is followed to prevent
going over the target rate. Bands with higher variances are assigned their bit rate according to the equations given
by the method of Lagrange multipliers. Bands with negative bit rates are assigned a zero bit rate instead. The new
bit rate mean (calculated in bits per pixel per band) is calculated. If the target rate has been reached the process
stops, if not, the next component band is assigned and the process is repeated. Once the process is stopped,
component bands are decompressed and merged together. The inverse transform is performed and the image is
classified. This algorithm was repeated for several bit rates: 0.25, 0.5, 1.0, 1.5, and 2 bits per pixel per band.
The purpose of using quality scalability is to minimize distortion introduced when quantizing or compressing
coefficients. Therefore, this approach might result in slightly better mineral classification performance than using
color scalability. On the other hand, one of its disadvantages is that optimal results will not be necessarily obtained
when using Lagrange multipliers. However, its simple implementation represents an alternative to avoiding
excessive computations.

RESULTS
One of the objectives of this paper is not only to measure the compression performance of the DWT, but also to
evaluate how much information is lost when using lossy algorithms. In this manner, the Hyperspectral images were
classified and the results were used to compare the DWT to PCA—which achieves most image compression
possible and results in better classification performance.
Mineral classification was focused on two similar materials: Aluminum-Rich and Aluminum-Poor mica.
Software used to classify minerals use data of a Hyperspectral image an assign a mineral to each of its pixels. This
is achieved by assigning each pixel vector of length 194 to one of the spectral signatures used for classification. The
problem is that Al-Rich and Al-Poor mica have almost identical spectral signatures. Therefore, the classification of
both types of mica represents a good parameter to measure classification performance. In addition, other minerals
have a similar spectral signature, which increases the complexity of the problem. Figure 6 show the algorithm
followed to classify minerals and compare them.
There exist numerous methods to classify Hyperspectral images. However, in order maintain consistency with
our past results and the classifier’s popularity, the usage of SAM with PCA was recommended. The SAM classifier
requires only one parameter, which is the threshold (in degrees.) This parameter was kept at 5° for all cases. When
an image has been classified, it has to be compared to a file containing correct mineral classification. This file is
called ground truth and was obtained by classifying the original image. Ground truth was compared to each
classification of compressed images, which results in three different possibilities when comparing two pixels. Table
6.1 displays the events possible and their meaning when comparing images.
Table 1. Hit, Miss, and False Alarm definitions
Event

Ground Truth

Compressed Image

Definition

Hit

Mica

Mica

Mica is classified correctly

Miss

Mica

Other

Mica was not detected

False Alarm

Other

Mica

Mica was erroneously found

It is obvious that hit rates and miss rates will always add up to 100%. This is why only hit and false alarm rates
will be shown and described in this chapter. The purpose of each algorithm used was to obtain a 90% hit rate and
have a false alarm rate as low as possible.

Classification Performance using Color Scalability (First Method)
Previous results showed that using the PCA as the cross-component transform produced hit rates over 90%
when using only 30 of the 194 bands. Since the DWT does not achieve as good of an ordering of data as the PCA, it
is expected that classification performance will be somewhat degraded. Figure 6 shows the comparison between the
PCA and DWT when classifying Al-Rich Mica.


Al-Rich mica

Al-Rich mica

100

80

90

70

80

60

60

PCA
DWT

50
40

False Alarm Rate

Hit Rate

70

50

PCA
DWT

40
30

30

20
20

10

10
0

0
0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

Number of bands

50

60

70

80

90

100

Number of bands

Figure 6. PCA and DWT hit rate and false alarm plot using Color Scalability (1st Method)—Al-Poor Mica.
Classification performance were found different for both minerals. Hit rate increases slowly, while false alarm
rate will also increase sometimes. The only difference worth noting is that the 90% hit rate mark was achieved
when using 97 component bands. The actual hit rate was 93.42% having a false alarm rate of 10.01%. The
difference between the PCA and DWT performances was somewhat expected due to the data ordering achieved by
both transforms. If the assumption made that input variance is directly related to the amount of information
contained in a component band, then the alternative method will achieve better classification performance.

Classification Performance using Color Scalability (Alternative Method)
Sorting the component bands is expected to yield better classification performance. Since the PCA transform is
ideal in the sense that all the component bands will always be ordered optimally, then sorting these bands will only
decrease classification performance. Therefore, this method had not to be performed for the PCA transform, and the
results are equal to those of the previous section. Therefore, only the comparison between this method and the
previous will be described.
It can be seen that this method starts with worse classification performance. On the other hand, this method
begins to reach the same performance when using about 37 component bands, and actually performs better after 49
component bands. Figure 7 also shows a problem of using the DWT: using wavelets has a tendency to increase
false alarm rates at specific numbers of component bands. For example, when using 49 bands the false alarm rate
for both cases is over 39%, but using the alternative method yields a hit rate of 88%, which is almost 17% more than
the first method. The plot in Figure 7 also shows that 90% hit rate can be achieved when using about 57 component
bands—a significant improvement over the first method, which did not even reach this rate when using 97 bands.
The performance comparison when classifying Al-Poor Mica is somewhat similar.
Once again, the classification performance is deteriorated but starts increasing very rapidly. In fact, this method
starts doing better than the first when using 25 component bands, with the only exception that false alarm rate is still
high at this point. On the other hand, using 49 component bands decreases false alarm rate to 6.36%, which is very
close to the value of 3.92% when using the PCA. In addition, this method achieves 87.87%, only about 10% less,
hit rate than the PCA. Of course, this loss has its advantages: a faster and data-independent transform. In other
words, decompressing the same file will take less time when using the DWT. Moreover, the exact same transform
will work for all compressed files, which is not the case for the PCA.


Al-Rich mica

Al-Rich mica
80

100
90

70

80
60

H it Rate

60

First Method
Alternative Method

50
40

False A larm Rate

70
50

First Method
Alternative Method

40
30

30
20
20
10

10

0

0
0

10

20

30

40

50

60

70

80

90

0

100

10

20

30

40

50

60

70

80

90

100

Number of bands

Number of bands

Figure 7. DWT hit rate and False Alarm plot using Color Scalability (Both Methods)—Al-Rich Mica.
Al-Poor mica

Al-Poor mica
80

100
90

70

80

60

H it Rate

60

First Method
Alternative Method

50
40

False A larm Rate

70

50

First Method
Alternative Method

40
30

30

20
20

10

10

0

0
0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

Number of bands

Number of bands

Figure 8. DWT false alarm rate plot using Color Scalability (Both Methods)—Al-Poor Mica.

Classification Performance using Quality Scalability
Classification results were not available for this type of scalability; therefore, the procedure was followed when
using the PCA and the Discrete Wavelet Transform. Figure 9 shows the comparison between the two transforms
when classifying Al-Rich Mica, at several bit rates, in bits per pixel per band (bpppb.)
The comparison demonstrates that the performance of the DWT is usually about 15% less than the PCA. In
addition, the DWT reaches 90% hit rate at about 1.8 bpppb. Similarly, the false alarm rate has similar behavior to
that of the PCA when using above 1.25 bpppb, which is less than 10%. Therefore, a bit rate of about 1.8 bpppb
might be enough to have a classification performance as good as that of the PCA. This bit rate yields a compression
ratio of 1.8 bpppb to the original 16 bpppb, which is equivalent to 11.25%. The classification algorithm was
repeated for Al-Poor Mica. Figure 10 illustrates the performance of the DWT and PCA.
Figure 10 also shows that better results are achieved when classifying Al-Poor Mica. For example, the hit rate
when using the DWT starts at about 20% less, but decreases constantly. In fact, the 90% hit rate mark is achieved
under 1.25 bpppb, approximately at about 1.125 bpppb. Moreover, the false alarm when using the DWT is always
very close to that obtained when using the PCA. Bit rates above 0.6 bpppb have false alarm rates smaller than 10%,
which is very close to PCA results. A comparable performance between the DWT and the PCA can be obtained
when using 1.125 bpppb, which is equivalent to a compression ratio of 7.03%. Quality scalability has the advantage

that it is capable of achieving better classification performance than color scalability. This is because quality
scalability has the objective of minimizing a measure of distortion, which in this case is the MSE. In addition, the
implementation of the method of Lagrange multipliers does not represent a computationally expensive task, yet
good results are obtained.

Al-Rich mica

Al-Rich mica

100

80

90

70

80
60

H it Rate

60

PCA
DWT

50
40

False A larm Rate

70
50

PCA
DWT

40
30

30
20
20
10

10
0

0
0

0.25

0.5

0.75

1

1.25

1.5

1.75

2

0

0.25

0.5

bits per pixel per band

0.75

1

1.25

1.5

1.75

2


Figure 9. PCA and DWT false alarm rate plot using Quality Scalability—Al-Rich Mica.

Al-Poor mica

Al-Poor mica

100

80

90

70

80
60

H it Rate

60

PCA
DWT

50
40

False A larm Rate

70
50

PCA
DWT

40
30

30
20
20
10

10
0

0
0

0.25

0.5

0.75

1

1.25

1.5

1.75

2

0

0.25

0.5


0.75

1

1.25

1.5

1.75

2


Figure 10. PCA and DWT false alarm rate plot using Quality Scalability—Al-Poor Mica.

Computational Efficiency
Previous sections detail the comparison in classification performance between the PCA and DWT. The PCA
being the theoretically ideal transform is expected to have better results for both classification performance and
compression. However, its biggest disadvantage is its computationally expensive implementation. Therefore, the
DWT is a means of achieving faster decompression times while sacrificing classification performance. As seen
previously, the use of the DWT decreases performance slightly. This section will now compare the time it takes for
both transforms to be implemented. Only the actual implementation times required to transform an image were
measured, since classification will require the exact amount of time for both transforms. Table 2 shows the results
obtained for each transform.


Table 2. Implementation times for PCA and DWT
Transform
Time (in seconds)

PCA

DWT

2837

378

As seen in the table, the DWT is about 7.5 times faster, which is a significant reduction in the time spent each
time an image is decompressed. Furthermore, this transform does not require any overhead, such as the
transformation matrix, which eliminates the need to calculate additional information each time an image is
transformed. Therefore, the DWT is a very efficient and fast transform that does not require additional files to
transform images.

CONCLUSIONS
The discrete wavelet transform represents a method to obtain reasonable classification performance results
while maintaining a low number of computations. The results show that the DWT is capable of reaching 90% hit
rate when using 50 component bands and color scalability. This number is less than double of that required by the
PCA, which are 30 component bands. Similarly, when using quality scalability, the bit rate has to double or triple to
achieve the 90% hit rate. However, the advantage of using the DWT is that it is faster–about 7 times faster than the
PCA. Although PCA has a superior classification performance, these results demonstrate that the DWT could be
used as an alternative in less complex classification problems. One other main advantage of DWT, that is allows
dividing a huge data set in to parts and pre-processing each of these independent of the other. This enables parallel
processing which cannot be done with PCA.

REFERENCES
Vikram Jayaram, Bryan Usevitch and Olga Kosheleva, “Detection from Hyperspectral images compressed using
rate distortion and optimization techniques under JPEG2000 Part 2,” in Proc. IEEE 11th DSP Workshop &
3rd SPE Workshop (cdrom), August 2004.
Max Garces and Bryan Usevitch, “Classification performance of reduced color band Hyperspectral imagery
compressed domain under JPEG2000 Part 2,” in Proc. IEEE 11th DSP Workshop & 3rd SPE Workshop
(cdrom), August 2004.
M. D. Pal, C. M. Brislawn, and S. P. Brumby, “Feature Extraction from Hyperspectral Images Compressed using the
JPEG-2000 Standard,” in Proceedings of Fifth IEEE Southwest Symposium on Image Analysis and
Interpretation (SSIAI’02), Santa Fe, NM, pp. 168–172, April 2002.
Silpa Attluri, Vikram Jayaram and Bryan Usevitch, “Optimized Rate allocation of Hyperspectral Images in
compressed domain under JPEG2000 part 2” in Proc. IEEE 2006 Southwest Symposium on Image Analysis
& Interpretation (SSIAI), Denver, Colorado, March 2006.
D. Taubman and M. Marcellin, JPEG2000 Image Compression Fundamentals, Standards and Practice. Boston,
Dordrecht, London: Kluwer, 2002.
Bryan Usevitch, “A Tutorial on Modern Lossy Wavelet Image Compression: Foundations of JPEG2000”, IEEE
Signal Processing Magazine, Vol. 18, Issue 5, September 2001.
T. J. Cudahy, R. Hewson, J. F. Huntington, M. A. Quigley and P. S. Barry, “The performance of the Satellite borne
Hyperion Hyperspectral VNIR–SWIR imaging system for mineral mapping at Mount Fitton, South
Australia,” in Proceedings of IEEE International Geosciences and Remote Sensing Symposium IGARSS01,
Sydney, July 2001.
Shaw Gary and Manolakis Dimitris, “Signal Processing for Hyperspectral Image Exploitation”, IEEE Signal
Processing Magazine, Vol. 19, Issue 1, January 2002, ISSN 1053-5888.
Said Amir and William A. Pearlman, “Signal Processing for Hyperspectral Image Exploitation”, IEEE Signal
Processing Magazine, Vol. 19, issue 1, January 2002, ISSN 1053-5888.


2007 asprs

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to 2007 asprs (20)

More from Pioneer Natural Resources (15)

Recently uploaded (20)

2007 asprs