You are on page 1of 6

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 35, NO.

5, SEPTEMBER 1997 1171

Preserving Radiometric Resolution in Remotely


Sensed Data with Lossy Compression
James C. Tilton, Senior Member, IEEE, and Mareboyana Manohar, Member, IEEE

Abstract— Recent advances in imaging technology make it offs in which certain scientific returns are sacrificed for the
possible to obtain remotely sensed imagery data of the Earth sake of others. In this paper we model the radiometric version
at high spatial, spectral, and radiometric resolutions. The rate at of this form of lossy compression by truncating a specified
which the data is collected from these satellites can far exceed
the channel capacity of the data downlink. Reducing the data number of lower order bits followed by lossless compression
rate to within the channel capacity can often require painful of the remaining higher order bits. We call this approach
trade-offs in which certain scientific returns are sacrificed for Truncation followed by Lossless Compression (TLLC). Using
the sake of others. We focus here on the case where radiometric the TLLC approach, the data rate can be set to within the
resolution is sacrificed by dropping a specified number of lower channel capacity by selecting the appropriate number of lower
order bits (LOB) from each data pixel. To limit the number
LOB’s dropped, we also compress the remaining bits using order bits dropped. We have found that this method produces
lossless compression. We call this approach “truncation followed reasonable rate distortion performance for compression ratios
by lossless compression” or TLLC. We then demonstrate the less than 5 or 6. However, for larger compression ratios, the
suboptimality of this TLLC approach by comparing it with the distortion continues to increase exponentially as the amount
direct application of a more effective lossy compression technique of truncation increases.
based on the JPEG algorithm. This comparison demonstrates
that, for a given channel rate, the method based on JPEG lossy Much better rate distortion behavior can be obtained by
compression better preserves radiometric resolution than does using other lossy compression approaches. For the lossy
TLLC. compression approaches we have studied, the rate distortion
Index Terms—Data acquisition, data compression, image cod- performance is either linear or sublinear. We report here our
ing, image communication, satellite applications. results using a hybrid method based on the JPEG (Joint
Photographic Experts Group/Discrete Cosine Transform) [1]
lossy compression algorithm. These results are similar or
I. INTRODUCTION somewhat better than results we have obtained with other

T HE imaging sensors onboard satellites are capable of


scanning the Earth at very high spatial, spectral and
radiometric resolutions. Downlink channel capacity is often
lossy compression approaches (e.g., Vector Quantization [2]
and the more recently developed Model-based Vector Quan-
tization [3]). For a given data rate, this improved distortion
a major limiting factor for the resolution at which the data is behavior over TLLC can be looked upon as a preservation of
collected. Image compression techniques can be used to reduce radiometric resolution, or equivalently as a gain in radiometric
the data rate from the imaging sensor to within the downlink resolution over TLLC.
channel capacity. Ideally, decompression of the downlinked In this paper we first describe the TLLC approach in more
data should result in fully lossless recovery of the image detail, and give a summary description of the JPEG algorithm
data as sensed onboard the satellite. However, the amount and how we have combined it with lossless compression of the
of compression possible from lossless techniques is bounded quantized residual in order to improve performance in terms
by the entropy of the source. This entropy bound limits the of maximum absolute error. We then derive a measure of gain
amount of compression that can be obtained to the range of 2–3 in radiometric resolution of a particular lossy compression
for most NASA image data sources. This is usually insufficient approach over TLLC. Finally we demonstrate the gain in
to reduce the sensor data rate to within the channel capacity. radiometric resolution provided by the JPEG algorithm over
Large amounts of compression can, instead, be obtained the TLLC approach with imagery data from three remote
using lossy compression techniques. In fact, a crude form of sensing instruments: the Landsat Thematic Mapper (TM), the
lossy compression is most often used in these cases: the tempo- Advanced Solid-state Array Spectroradiometer (ASAS), and
ral, spatial, spectral, and/or radiometric resolutions are limited the Advanced Very High Resolution Radiometer (AVHRR). Of
to produce a data rate that can be handled by the channel these, TM imagery data is at 8-bit resolution, while imagery
capacity. Establishing these limits often requires painful trade- data from the other two are at 10-bit resolution.

Manuscript received March 7, 1995; revised March 12, 1996. This work II. TRUNCATION FOLLOWED BY
was supported by RTOP funding from the Office of Advanced Concepts and LOSSLESS COMPRESSION (TLLC)
Technology, NASA Headquarters, Washington, DC.
J. C. Tilton is with NASA Goddard Space Flight Center, Greenbelt, MD Truncation followed by Lossless Compression (TLLC) is a
20771 USA (e-mail: tilton @istbsun.gsfc.nasa.gov). compression approach that can be used to advantage under
M. Manohar is with the Department of Computer Science, Bowie State
University, Bowie, MD 20715 USA. certain special conditions (e.g., very noisy data, very low
Publisher Item Identifier S 0196-2892(97)05518-6. available computational power and very limited channel ca-
0196–2892/97$10.00  1997 IEEE

Authorized licensed use limited to: M Th Koelen. Downloaded on June 12, 2009 at 11:39 from IEEE Xplore. Restrictions apply.
1172 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 35, NO. 5, SEPTEMBER 1997

pacity) [4]. However, TLLC is not normally used as a general coefficient quantization and lossless compression [1]. The
compression approach. As noted in the introduction, we use original image is first partitioned into nonoverlapping 8
it here as a model for the design practice of setting the 8 pixel blocks. Each block is then independently transformed
radiometric resolution to a lower value than sensor technology using the DCT. The DCT coefficients are then quantized using
would allow, so as to keep the data rate produced by the sensor a quantization table that is designed using the HVS (Human
within the limits of channel capacity for bringing the data from Visual System) contrast sensitivity function (quantization ta-
the sensor to Earth. We are not advocating TLLC as a general bles can be tailored to particular applications, see [9]). The
compression approach. first coefficient of DCT transformation is designated the DC
Let the radiometric resolution of the image data collected coefficient and is proportional to average brightness of the
at the instrument be bits/pixel and the channel capacity block. The quantized DC coefficients from each block are
be bits/pixel ( ). The TLLC approach reduces concatenated and are compressed using DPCM (Differential
the bit rate from to no more than by dropping a Pulse Code Modulation) using 1-D causal prediction. The
number of lower order bits, , and losslessly compressing the remaining 63 quantized coefficients from each block, called
result. Here is chosen such that the lossless compression the “AC coefficients,” are zig-zag scanned to convert the 2-
of remaining – bits results in an output bit rate of no D array into a 1-D array, and are then losslessly compressed
more than bits/pixel. The lossless compression approach using a Huffman table that is transmitted to the decoder as a
that consistently performed best in the cases we tested uti- part of the header information.
lizes the coding model for lossless encoding specified in the We used the sixth public release of the Independent JPEG
JPEG still image compression standard [1] combined with the Group’s free JPEG software, which we obtained from UUNET
Witten–Neal–Cleary version of arithmetic coding [8]. (The archive at ftp.uu.net:graphics/jpeg/jpegsrc.v6.tar.gz. The re-
implementation we used can be obtained from the anonymous lease date was August 2, 1995. Spectral correlations are
ftp location dftnic.gsfc.nasa.gov:software/unix/crushv3/*.) In not easy to exploit in JPEG, as there are no standards for
some cases prediction mode “1” (value of the previous pixel decorrelating the bands of multispectral image data (JPEG
in the same line) produced the best results in JPEG lossless does however, allow red, green and blue decorrelations by
coding, and in other cases prediction mode “7” (average of converting them to luminance and chrominance components.
the values of the previous pixel in the same line and the See [1, pp. 18–20, p. 503]. Therefore, we compressed each
previous pixel in the same column) produced the best results. band of the multispectral images independently in our tests.
The prediction mode used for each result is noted in the When we applied the JPEG algorithm directly to NASA
Experimental Results section. imagery data, we found that the maximum absolute error ob-
The image data is reconstructed following TLLC by loss- served was often larger than acceptable, even while the mean
lessly decompressing the data, scaling it up by a factor of ,
squared error was acceptable. We have devised a simple hybrid
and adding a bias term of [or ].
approach to limit the maximum absolute error, while achieving
For completeness we comment here on certain compression
nearly the same overall compression ratio. A side effect is a
approaches that are similar to, or appear to be similar to,
slightly improved mean squared error. For convenience, we
our TLLC model. We have already noted that, under certain
call this approach JPEG/hybrid.
special conditions, TLLC is used for lossy compression. In this
The JPEG/hybrid approach is as follows.
regard, the algorithm of Rice and Plaunt [4] employs the TLLC
approach as an option. Other lossy compression approaches 1) Apply the JPEG algorithm to the image data using
may at first glance appear similar to our TLLC approach, but quality factor .
are quite different on closer examination. The TLLC approach 2) Decompress the data and subtract the resulting image
splits each data point into higher order bits (HOB) and lower from the original image data.
order bits (LOB), losslessly compresses the HOB and discards 3) Divide the difference image by a quantization factor
the LOB. The approach described by Lo and Krasner [5] splits ( ).
each data point into HOB and LOB, losslessly compresses 4) Losslessly compress the quantized difference image
the HOB and keeps the LOB in its uncompressed state. The using the lossless JPEG compression algorithm used
approach describe by Zhang, Loew, and Pickholtz [6] splits in TLLC.
each data point into HOB and LOB, losslessly compresses the The coded data consists of the lossy JPEG compressed data
HOB and lossily compresses the LOB (with “conventional” file plus the lossless JPEG compressed quantized difference
block transform coding). Finally, the approach described by image (we include the quantization factor, , in the name of
Bassiouni, Tzannes, and Tzannes [7] splits each data point this file).
into HOB and LOB, losslessly compresses the HOB and To decompress the data, decompress the quantized differ-
lossily compresses the LOB (using an approach similar to ence image, multiply it by the quantization factor ( ) and add
the JPEG/DCT algorithm). None of these later three totally the result to the decompressed lossy JPEG image.
discards the LOB as does TLLC. We could have used any of a number of approaches to
losslessly compress the quantized difference images. We tried
III. JPEG/HYBRID “gzip” (version 1.2.4, release date August 18, 1993, from
The JPEG lossy compression algorithm consists of three ftp.uu.net) and the standard UNIX “compress” programs.
successive stages: Discrete Cosine Transformation (DCT), Both produced lower compression ratios than the lossless

Authorized licensed use limited to: M Th Koelen. Downloaded on June 12, 2009 at 11:39 from IEEE Xplore. Restrictions apply.
TILTON AND MANOHAR: PRESERVING RADIOMETRIC RESOLUTION 1173

JPEG approach we employed. However, since the quantized TABLE I


difference images consists mainly of zero values, we suspect (a) CR VERSUS MSE AND MAE ON TM DATA, TLLC RESULTS (AVERAGE OF
BANDS 1–7). (b) CR VERSUS MSE AND MAE ON TM DATA, JPEG/HYBRID
that a simple run-length encoding approach may perform well, RESULTS (AVERAGE OF BANDS 1–7). (c) GAIN IN RADIOMETRIC RESOLUTION BY
but we haven’t tested this approach. USING JPEG/HYBRID RATHER THAN TLLC ON TM DATA. THE JPEG/HYBRID
MSE IS ESTIMATED BY LINEAR INTERPOLATION FROM QUALITY FACTORS ( ) Q
THAT GIVE CR CLOSEST TO THE TLLC RESULT. 1b
IS CALCULATED FROM (5).
IV. CALCULATING RADIOMETRIC RESOLUTION GAIN (d) GAIN IN RADIOMETRIC RESOLUTION BY USING JPEG/HYBRID RATHER
THAN TLLC ON TM DATA. THE JPEG/HYBRID CR IS ESTIMATED
In the TLLC approach, the radiometric resolution of the BY LINEAR INTERPOLATION FROM QUALITY FACTORS ( ) THAT GIVE Q
input image is explicitly reduced by bits by the truncation MSE CLOSEST TO THE TLLC RESULT. 1b
IS LINEARLY INTERPOLATED
process. We show here that the mean squared error (MSE) FROM THE TLLC “BITS DROPPED” RESULTS WITH THE CLOSEST CR
distortion resulting from the truncation varies exponentially
with , the loss in radiometric resolution.
The mean squared error is defined formally as

MSE (1)

where and are the th pixels from the original


and reconstructed images, respectively, and is number of
(a) (b)
pixels in the image.
For TLLC with bit truncation, the pixel error after
reconstruction may be one of the integers
. We assume throughout
this section that a bias term is employed in reconstruc-
tion [the final result is the same if bias term is
employed]. Assuming a uniform distribution of these pixel (c)
error values, the expected mean squared error is given by

MSE (2)

(d)
(3)

We have observed that the uniform distribution assumption amount of compression. We use this approach as one method
holds best for lower values of . Equation (3) can be derived for calculating the effective radiometric resolution gain
from (2) using the Euler–Maclaurin summation formula [10] of the JPEG/hybrid method with respect to TLLC in the
with . From (3), we can obtain in terms of MSE by following section.
solving the quadratic equation in and taking giving We must note here that one would not expect that the
MSE (4) JPEG/hybrid method would produce error values with a uni-
form error distribution. In addition, for higher values of ,
Equation (4) can be used to compute the loss of radiometric the TLLC method also often produces error values that do
resolution due to the mean squared error distortion from lossy not follow a uniform error distribution. However, (5) is still
compression. We can thus compare performance of lossy useful in these cases as a convenient index for improvement
compression techniques in terms of radiometric efficiency. in MSE, since still does correlate with MSE. However,
Assuming two lossy compression approaches (method 1 and cannot be literally interpreted as the number of bits of gain in
method 2) give the same amount of compression, let the radiometric resolution in nonuniform error cases.
MSE distortions from the two lossy methods be and
, respectively. Let and be the corresponding loss of V. EXPERIMENTAL RESULTS
radiometric resolution computed from (4). Now if , Three different multispectral image data sets are used in our
there is gain in radiometric resolution, , by using method experiments. The first data set consists of a 2048-by-2048 pixel
2 instead method 1 which is given by subimage of a seven-band Landsat Thematic Mapper (TM)
scene collected in 1991 (path/row 46/28) from over the Gifford
Ponchot National Forest in the state of Washington in the
(5) United States of America. The radiometric (pixel) resolution
of this data is 8 bits.
Using (5) lossy compression techniques can be compared Table I(a) shows the compression ratio (CR), mean square
in terms effective radiometric gain by using one with lesser error (MSE) and maximum absolute error (MAE) results from
distortion than the other compression technique for a given the TLLC approach for the TM data set (prediction mode

Authorized licensed use limited to: M Th Koelen. Downloaded on June 12, 2009 at 11:39 from IEEE Xplore. Restrictions apply.
1174 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 35, NO. 5, SEPTEMBER 1997

Fig. 1. The histograms of the error residual for each band for the Fig. 3. The histograms of the error residual for each band for the
3-bits-dropped case on the TM data. These histograms show that the closer Q = 77; qt = 16 case on the TM data. These histograms show that
the MSE error is to the value calculated via (3), viz. 5.50, the closer the the error distribution is more like a Gaussian or Laplacian distribution rather
histogram appears to be that of a uniform distribution. than a uniform distribution.

Fig. 2. The histograms of the error residual for each band for the Fig. 4. The histograms of the error residual for each band for the
4-bits-dropped case on the TM data. These histograms show that the closer Q = 55; qt = 16 case on the TM data. These histograms show that
the MSE error is to the value calculated via (3), viz. 21.5, the closer the the error distribution is more like a Gaussian or Laplacian distribution rather
histogram appears to be that of a uniform distribution. than a uniform distribution.

“1” was used). We also list the MSE calculated through (3). TABLE II
OFFSET AND RESULTING MSE MEASUREMENTS FOR THE CLIPPED
We have found that the amount of agreement between the AND TRUNCATED AVHRR DATA. THE AVERAGE MSE IS 25.79
measured MSE and calculated MSE is a good indicator of
how well the uniform distribution assumption holds. As a
demonstration of this relationship, Figs. 1 and 2 show the
histograms of the error residual for each band for the 3 and 4
bits dropped case, respectively.
The CR, MSE, and MAE results for various quality factor
( ) values are given in Table I(b) for our JPEG/hybrid method
higher than 8 bits. Statistical analysis of this data showed
for lossy compression (prediction mode “1” was also used in
that the main concentration of dynamic range is in a 9 bit
the compression of quantized residuals in JPEG/hybrid). The
range, so we offset each band by a certain amount and then
gain in radiometric resolution by using JPEG/hybrid rather
clipped it to the 9 bit range (0–511). Finally, we truncated
that TLLC, , is given in Table I(c) and (d). In Table
the resulting image pixels to 8-bits prior to JPEG/hybrid
I(c), is calculated from (5). In Table I(d), is linearly
compression by dividing by two. We choose the offsets to
interpolated from the TLLC “bits dropped” results with the
closest compression ratio. We see from Table I(c) and (d) that approximately minimize the MSE after reconstructing the
ranges from 0.6 to 2.4, with the higher values coming clipped data to the original range. The offsets and resulting
from the higher “bits dropped” cases. MSE are given in Table II. After we ran our JPEG/hybrid lossy
Figs. 3 and 4 show the histograms of the error residual for compression approach on the offset, clipped and truncated
each band for the case and the , data, we computed the MSE by decompressing the data,
case, respectively. These histograms look more like multiplying by two, adding back the offset, and comparing
Gaussian or Laplacian distributions than uniform distributions. it with the original data. This process reduced the value of our
This being the case, the in Table I(c) cannot be literally test somewhat, since we could not then compare JPEG/hybrid
interpreted as the number of bits gain in radiometric resolution. compression with TLLC for 1 or 2 bits dropped for this data.
However, the values in Table I(c) are very close to the values Table III(a)–(d) gives our results for the AVHRR data set in
in Table I(d), which were obtained by interpolation rather than a manner similar to that used to present our results for the TM
(5), indicating that a loose interpretation of the values data in Table I(a)–(d). (Prediction mode “7” was used in TLLC
in Table I(c) as the number of bits of radiometric gain is and the compression of quantized residuals in JPEG/hybrid.)
reasonable. We note in Table III(a) that the calculated MSE is within
The second data set is 2048 column by 1460 row portion 5% of the measured MSE for less than 6 bits dropped. The
of a five-band Local Area Coverage (LAC) data set from difference between the calculated and measured MSE goes up
the Advanced Very High Resolution Radiometer (AVHRR) substantially for 6 and more bits dropped, indicating that our
instrument, collected on June 14, 1991 from over Indonesia. uniform distribution assumption is violated in those cases (as
The pixel resolution of this data is 10 bits (stored as 16 corroborated by an inspection of the error image histograms).
bits/pixel). We see from Table III(c) and (d) that ranges from 0.1
We ran into a problem in performing the JPEG/hybrid to 1.7, with the higher values coming from the higher “bits
compression on this data set since the baseline JPEG im- dropped” cases. We note that our values calculated from (5)
plementation we employed does not handle pixel resolutions in Table III(c) for 6 and more bits dropped are suspect because

Authorized licensed use limited to: M Th Koelen. Downloaded on June 12, 2009 at 11:39 from IEEE Xplore. Restrictions apply.
TILTON AND MANOHAR: PRESERVING RADIOMETRIC RESOLUTION 1175

TABLE III TABLE IV


(a) CR VERSUS MSE AND MAE ON AVHRR DATA, TLLC RESULTS (AVERAGE OFFSET AND RESULTING MSE MEASUREMENTS FOR THE
OF BANDS 1–5). (b) CR VERSUS MSE AND MAE ON AVHRR DATA, CLIPPED ASAS DATA. THE AVERAGE MSE IS 0.466959
JPEG/HYBRID RESULTS (AVERAGE OF BANDS 1–5). (c) GAIN IN RADIOMETRIC
RESOLUTION BY USING JPEG/HYBRID RATHER THAN TLLC ON AVHRR DATA.
THE JPEG/HYBRID MSE IS ESTIMATED BY LINEAR INTERPOLATION FROM
Q
QUALITY FACTORS ( ) THAT GIVE CR CLOSEST TO THE TLLC RESULT. IS 1b
CALCULATED FROM (5). (d) GAIN IN RADIOMETRIC RESOLUTION BY USING
JPEG/HYBRID RATHER THAN TLLC ON AVHRR DATA. THE JPEG/HYBRID CR
IS ESTIMATED BY LINEAR INTERPOLATION FROM QUALITY FACTORS ( ) THAT Q
GIVE MSE CLOSEST TO THE TLLC RESULT. 1b
IS LINEARLY INTERPOLATED
FROM THE TLLC “BITS DROPPED” RESULTS WITH THE CLOSEST CR
TABLE V
(a) CR VERSUS MSE AND MAE ON ASAS DATA, TLLC RESULTS (AVERAGE
OF THE 6 SELECTED BANDS). (b) CR VERSUS MSE ON ASAS DATA,
JPEG/HYBRID RESULTS (AVERAGE OF THE 6 SELECTED BANDS). (c) GAIN IN
RADIOMETRIC RESOLUTION BY USING JPEG/HYBRID RATHER THAN TLLC ON
ASAS DATA. THE JPEG MSE IS ESTIMATED BY LINEAR INTERPOLATION FROM
Q
QUALITY FACTORS ( ) THAT GIVE CR CLOSEST TO THE TLLC RESULT. 1b
IS
CALCULATED FROM (5). (d) GAIN IN RADIOMETRIC RESOLUTION BY USING
JPEG/HYBRID RATHER THAN TLLC ON ASAS DATA. THE JPEG/HYBRID CR
Q
IS ESTIMATED BY LINEAR INTERPOLATION FROM QUALITY FACTORS ( ) THAT
GIVE MSE CLOSEST TO THE TLLC RESULT. 1b
IS LINEARLY INTERPOLATED
(a) (b) FROM THE TLLC “BITS DROPPED” RESULTS WITH THE CLOSEST CR

(c)

(a) (b)

(d)

of the noted violation of the uniform distribution assumption.


However, the results from Table III(d), which are not affected (c)
by this violation, show that the calculated results in Table III(c)
are reasonable despite the violation.
The third data set is six bands from a 510 column by 520
row portion of a 29-band data set from the Advanced Solid-
state Array Spectroradiometer (ASAS) instrument. ASAS is
reported to have 12-bit radiometric resolution, but the selected (d)
data set only has a 10-bit pixel resolution. This data set was
collected on August 4, 1989 over the Konza Prairie Research
Natural Area near Manhattan, KS. It is designated 92161621 approach on the offset and clipped data, we computed the
in vol. 4 of the FIFE CD-ROM series [11]. The six selected MSE by decompressing the data, adding back the offset, and
bands are bands 5, 9, 13, 17, 21, and 25 from the nadir view comparing with the original data.
(designated 921 161 621.406, .410, .414, .418, .422, and .425). Table V(a)–(d) gives our results for the ASAS data set in a
We had the same problem with this data set as we had with manner similar to that used to present our results for the TM
the AVHRR data set in performing the JPEG/hybrid compres- and AVHRR data. (Prediction mode “7” was used in TLLC
sion on this data set since the baseline JPEG implementation and the compression of quantized residuals in JPEG/hybrid.)
we employed does not handle pixel resolutions higher than 8 We note in Table V(a) that the calculated MSE is within
bits. However, we were able to get around this problem more 3% of the measured MSE for less than 4 bits dropped. The
satisfactorily with the ASAS data set. Statistical analysis of difference between the calculated and measured MSE goes up
this data showed that the main concentration of dynamic range substantially for 4 and more bits dropped, indicating that our
was in an 8-bit range, so we were able to offset each band by uniform distribution assumption is violated in those cases (as
a certain amount and then clip (not truncate) to the 8-bit range corroborated by an inspection of the error image histograms).
(0–255). We choose the offsets to approximately minimize the We see from Table V(c) and (d) that ranges from 0.5
MSE after clipping. The offsets and resulting MSE are given to 1.1, with the higher values coming from the higher “bits
in Table IV. After we ran our JPEG/hybrid lossy compression dropped” cases. We note that our values calculated from

Authorized licensed use limited to: M Th Koelen. Downloaded on June 12, 2009 at 11:39 from IEEE Xplore. Restrictions apply.
1176 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 35, NO. 5, SEPTEMBER 1997

(5) in Table V(c) for four and more bits dropped are suspect [10] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions
because of the noted violation of the uniform distribution with Formulas, Graphs, and Mathematical Tables. Washington, DC:
U.S. Gov’t Printing Office, Nov. 1970, p. 16.
assumption. However, the results from Table V(d), which are [11] D. R. Landis, D. E. Strebel, J. A. Newcomer, and B. W. Meeson,
not affected by this violation, show that the calculated results “Archiving the FIFE data on CD-ROM,” in Proc. 1992 Int. Geoscience
in Table V(c) are reasonable despite the violation. and Remote Sensing Symp., Houston, TX, May 26–29, 1992, pp. 65–67.

VI. CONCLUSIONS
James C. Tilton (S’79–M’81–SM’94) received the
We have demonstrated on three types of NASA imagery B.A. degree in electrical engineering, environmen-
data collected from spaceborne platforms that the JPEG/hybrid tal science and engineering, and anthropology, the
M.E.E. (electrical engineering) degree from Rice
lossy compression is more effective at preserving the radiomet- University, Houston, TX, the M. S. degree in optical
ric resolution of the data than the commonly used truncation sciences from the University of Arizona, Tucson,
approach. The truncation approach (which we call TLLC) is a and the Ph.D. degree in electrical engineering from
Purdue University, West Lafayette, IN, in 1976,
model for the engineering practice of reducing the radiometric 1976, 1978, and 1981, respectively.
resolution of a sensor so that the data rate from the sensor He is currently a Computer Engineer with the
falls below that of the channel capacity of the downlink. Our Applied Information Sciences Branch (AISB) of
the Earth and Space Data Computing Division, NASA Goddard Space
tests show that JPEG/hybrid lossy compression effectively Flight Center, Greenbelt, MD. He previously worked for Computer Sciences
preserves from about 0.5 bit to over 2 bits in radiometric Corporation, from 1982 to 1983, and Science Applications Research, from
resolution versus truncation followed by lossless compression, 1983 to 1985, on contracts with NASA Goddard. As a member of AISB, he
helps define future requirements for image analysis information extraction and
depending on the data characteristics and the amount of data compression in support of NASA programs. He is also responsible for
compression selected. We hope that this demonstration will designing and developing computer software tools for space and Earth science
inspire satellite sensor designers to explore lossy compression image analysis and compression algorithms, and encouraging the use of these
computer tools through close interactions with space and Earth scientists.
as an alternative method of reducing satellite sensor data Dr. Tilton is a member of the IEEE Geoscience and Remote Sensing,
output rates to match the available data downlink channel Computer, and Signal Processing Societies, as well as the Phi Beta Kappa,
capacity. Tau Beta Pi, and Sigma Xi honoraries. From 1992 through 1996, he served
as a member of the IEEE Geoscience and Remote Sensing Society Adminis-
trative Committee. He currently serves as an Associate Editor for the IEEE
REFERENCES TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING.

[1] W. B. Pennebaker and J. L. Mitchell, JPEG Still Image Data Compres-


sion Standard. New York: Van Nostrand Reinhold, 1993.
[2] A. Gersho and R. M. Gray, Vector Quantization and Signal Compression.
Mareboyana Manohar (M’89) received the B.E.
Boston, MA: Kluwer, 1991.
[3] M. Manohar and J. C. Tilton, “Model-based VQ for image data degree in electrical engineering and the M.Tech.
archival, retrieval, and distribution,” in Proc. Conf. on Visual Information degree in electronics instrumentation from Regional
Processing IV, SPIE Proc./SPIE AeroSense, Orlando, FL, Apr. 17–18, Engineering College, Warangal, India, in 1971 and
1995, vol. 2488, pp. 190–198. 1973, respectively, and the Ph.D. degree in com-
[4] R. F. Rice and J. R. Plaunt, “Adaptive variable-length coding for effi- puter science from Indian Institute of Bangalore,
cient compression of spacecraft television data,” IEEE Trans. Commun. India, in 1984.
Technol., vol. COM-19, pp. 889–897, 1971. He is currently an Assistant Professor in the
[5] S.-C. B. Lo, B. Krasner, and S. K. Mun, “Noise impact of error-free Department of Computer Science, Bowie State Uni-
image compression,” IEEE Trans. Med. Imag., vol. 9, pp. 202–206, versity, Bowie, MD. He previously worked for
1990. ISRO Satellite Center, Bangalore, from 1974 to
[6] Y.-Q. Zhang, M. H. Loew, and R. L. Pickholtz, “A combined-transform 1984. In 1985, he was an Assistant Professor in the Department of Computer
coding (CTC) scheme for medical images,” IEEE Trans. Medical Imag., Science and Engineering, Indian Institute of Technology, Madras. He was a
vol. 11, pp. 196–202, 1992. recipient of National Research Council Fellowship award and was associated
[7] M. A. Bassiouni, N. S. Tzannes, and M. C. Tzannes, “High-fidelity with Space Data and Computing Division of the Goddard Space Flight
integrated lossless/lossy compression and reconstruction of image,” Opt. Center, Greenbelt, MD, from 1986 to 1989, where he was involved in
Eng., vol. 32, no. 8, pp. 1848–1853, 1993. development of parallel algorithms for image processing on Massively Parallel
[8] I. H. Witten, R. M. Neal, and J. G. T. Cleary, “Arithmetic coding for Processor (MPP). From 1989 to 1992, he worked for the Universities Space
data compression,” Commun. ACM, vol. 30, pp. 520–540, 1987. Research Association (USRA), Greenbelt, as Senior Research Associate,
[9] V. Ratnakar and M. Livny, “Performance of customized DCT quan- where he was involved in image compression research. He worked for Hughes
tization table on scientific data,” in Proc. 1994 Science Information STX, Greenbelt, during 1992 as Principal Scientist where he continued his
Management and Data Compression Workshop, Greenbelt, MD, Sept. research in image compression research. His current interests include image
26–27, 1994, pp. 1–8. registration, image compression, neural nets, and parallel processing.

Authorized licensed use limited to: M Th Koelen. Downloaded on June 12, 2009 at 11:39 from IEEE Xplore. Restrictions apply.

You might also like