FR1.L09 - PREDICTIVE QUANTIZATION OF DECHIRPED SPOTLIGHT-MODE SAR RAW DATA IN TRANSFORM DOMAIN
1. Predictive Quantization of
Dechirped Spotlight-Mode SAR Raw Data in
Transform Domain
Takeshi Ikuma, Mort Naraghi-Pour*
Department of Electrical and Computer Engineering
Louisiana State University
Baton Rouge, LA
Thomas Lewis
Air Force Research Laboratory
Dayton, OH
2. 2
Presentation Outline
Circular spotlight-mode SAR, Motivation
Previous work
Autoregressive modeling of IDFT transformed SAR data
Predictive encoding
Block predictive quantization: scalar, vector
Predictive trellis coded quantization
Numerical results
Conclusions
3. 3
Circular Spotlight-Mode SAR
We are interested in circular spotlight-mode SAR
Radar periodically emits a linear FM chirp pulse and receives,
dechirps, and samples the reflected pulses
A large volume of data is generated that must be downlinked
for processing and archiving
Downlink channel has limited bandwidth z
Need on-board compression
q: azimuth angle
of SAR RAW* data q
* Not SAR Image Compression
y
x
4. 4
Previous Work
Block Adaptive Quantization (BAQ)
Simple scalar quantizer, adapted to the signal power
Implemented in exiting systems
NASA Magellan Mission
NASA Shuttle Imaging Radar Mission C
More Effective Method?
Samples of both I and Q channels of SAR raw data are
largely uncorrelated
However, SAR image exhibits some correlation
Transformed data may exhibit some correlation
5. 5
Previous Work, cont’d
Paper Method Pre-Proc. Quantize Post-Proc.
r
Kwon (1989) BAQ Normalization SQ
Arnold (1988) CCT VQ
Franceschetti SC-SAR 1-bit SQ
(1991)
Benz (1995) FFT-BAQ Normalization & SQ w/bit
2-D FFT allocation
Bolle (1997) R & AZ comp DCT SQ Huffman
Owens (1999) TCVQ Trellis coding VQ
Baxter (1999) Gabor/TCQ Gabor trans. VQ Huffman
Trellis coding
Poggi (2000) Range VQ
compression
Magli (2003) NPAQ LPC SQ Arithmetic
6. 6
Spotlight CSAR Data
CSAR data samples are uncorrelated
Zero-mean Gaussian distributed
Signal power varies slowly over time
Example: AFRL Gotcha data set (about 42,000 returns from full 360°)
Magnitude of Raw Data Formed SAR Image (CBP, 512 returns)
7. 7
Transformed Data
If there are strong reflectors in the scene, range-wise IDFT of
CSAR data exhibits correlation along azimuth.
Isotropic reflectors appear as sinusoidal traces in the
transformed data. Anisotropic reflectors appear as partial
sinusoidal traces.
IDFT of SAR data
High magnitude sinusoidal trace from metallic cylinder object
8. 8
Transformed Data, cont’d
Develop block adaptive AR model for IDFT data across returns
(azimuth) for each fixed IDFT bin
AR model can capture strong reflectors and homogeneous field
Example: AR(1) Model of Gotcha Data
Blocks with higher signal power AR poles close to the unit circle
Companion Paper:
T. Ikuma, M. Naraghi-Pour and T.
Lewis,
“Autoregressive Modeling of
Dechirped Spotlight-Mode Raw
SAR Data in Transform Domain,”
Poster presentation today.
9. 9
Block Predictive Coding
Using the AR modeling, we develop predictive coding techniques
for compression of SAR data
AR Estimator: Burg’s method
Predictive Encoder:
Predictive quantization Encoder
Scalar: TD-BPQ (DPCM)
Vector: TD-BPVQ
Predictive Trellis Coded Quantization
Decoder
10. 10
Transform Domain Block Predictive Quantization
All signals are complex-valued
ikMv i ,q
Q(x) : 2 identical scalar quantizers
rkMb i ,q Q(x) ˆ
ekMb i ,q
for I and Q channels
Designed for zero-mean
2 2
k ,q Gaussian input with variance k ,q
Predictor states initialization
rkMb i ,q First block : BAQ encoded.
A(z)
ˆ
rkMb i ,q
Subsequent blocks: Last L
a k ,q coded samples of previous
block
DPCM Encoder
L: Predictor Order
11. 11
TD-Block Predictive VQ
There is some correlation between neighboring IDFT bins
Code multiple (Nb) IDFT bins together to take advantage of
this correlation Predictive VQ
Model a block of data as a vector AR process
Treat each IDFT bin as a separate channel. Use generalized AR
estimators for vector process. Each AR coefficient is now a matrix
Innovation process comprises independent circular complex Gaussian
processes with zero mean and different variances
Vector quantizer codebook
Basic codebook designed with LBG algorithm for circular complex
Gaussian training samples with zero mean and unit variance
For each data block, basic codebook is transformed according to
estimated covariance matrix given by vector AR estimator
12. 12
Predictive Trellis Coded Quantization
We have also applied predictive trellis-coded quantization
(PTCQ) for coding of IDFT data
Two design considerations: Trellis and codebook
Amplitude Modulation Trellises
Exhibit reasonable resistance against error propagation
Codebook Design:
Based on 32QAM symbol constellation
Scaled according to variance estimation from AR analysis
(Can be optimized by training it with LBG)
Viterbi algorithm is used for encoding
14. 14
Numerical Results
Numerical Results are obtained for AFRL Gotcha dataset
Performance measure: Average SNR of formed SAR
images
119 images are formed each from 352 returns (roughly 3° azimuth)
using convolution back-projection algorithm
Bit Rate: Fixed to 2 bits per real sample
TD-BPQ, TD-BPVQ, TD-BPTCQ, & BAQ are compared in
terms of
Average SNR
Per-Image SNR
as prediction order and block size are varied
15. 15
TD-BPQ: SNR vs. Predictor Order
M = 256
SNR improves by 2.5 dB by introducing prediction
(from L = 0 to L = 1)
16. 16
TD-BPQ: SNR vs. Block Size
Prediction order L = 4
Reasons for SNR loss:
Mb small – poor AR estimates
Mb large – data non-stationary
17. 17
Performance Results
L = 4, M = 256, TD-BPVQ: 2 IDFT bins together
1 dB 1.5 dB
BAQ
TD-BPQ
5 dB TD-BPVQ
TD-BPTCQ
SNR for each of the three schemes experiences fluctuations across
images due to the anisotropic nature of the scene
18. 18
Formed SAR Image Comparison
BAQ TD-BPVQ
9.7 dB 14.4dB
Original
TD-BPQ TD-BPTCQ
13.5 dB 14.9 dB
19. 19
Conclusions
Significant correlation is observed in IDFT of dechirped
CSAR data
Three predictive encoding algorithms are applied to
transformed data:
TD-BPQ: Scalar DPCM coding in IDFT domain
TD-BPVQ: Vector predictive coding in IDFT domain
TD-PTCQ: Predictive Trellis Coded Quantization
The predictive quantization can provide up to 6 dB
improvement in average SNR
Any Questions?