Fractional Focusing and the Chirp Scaling Algorithm

With Real Synthetic Aperture Radar Data

by
Judith A. Northrop
A Thesis Presented in Partial Fulllment
of the Requirements for the Degree
Master of Science
Approved April 2011 by the
Graduate Supervisory Committee:
Antonia Papandreou-Suppappola, Chair
Andreas Spanias
Cihan Tepedelenlioglu
ARIZONA STATE UNIVERSITY
May 2011
ABSTRACT
For synthetic aperture radar (SAR) image formation processing, the chirp scaling al-
gorithm (CSA) has gained considerable attention mainly because of its excellent target
focusing ability, optimized processing steps, and ease of implementation. In particular,
unlike the range Doppler and range migration algorithms, the CSA is easy to implement
since it does not require interpolation, and it can be used on both stripmap and spotlight
SAR systems. Another transform that can be used to enhance the processing of SAR
image formation is the fractional Fourier transform (FRFT). This transform has been re-
cently introduced to the signal processing community, and it has shown many promising
applications in the realm of SAR signal processing, specically because of its close asso-
ciation to the Wigner distribution and ambiguity function.
The objective of this work is to improve the application of the FRFT in order to
enhance the implementation of the CSA for SAR processing. This will be achieved by
processing real phase-history data from the RADARSAT-1 satellite, a multi-mode SAR
platform operating in the C-band, providing imagery with resolution between 8 and 100
meters at incidence angles of 10 through 59 degrees. The phase-history data will be
processed into imagery using the conventional chirp scaling algorithm. The results will
then be compared using a new implementation of the CSA based on the use of the FRFT,
combined with traditional SAR focusing techniques, to enhance the algorithms focusing
ability, thereby increasing the peak-to-sidelobe ratio of the focused targets. The FRFT
can also be used to provide focusing enhancements at extended ranges.
ii
To my family, friends and colleagues,
I could not have completed this degree
without your never ending support.
iii
ACKNOWLEDGEMENTS
Without hesitation, I sincerely want to express my deep gratitude to Professor Antonia
Papandreou-Suppappola for encouraging my studies and research. This thesis would not
be possible without her detailed analysis, recommendations and efforts on my behalf. She
has also remained my strongest advocate in representing my needs to the ASU Graduate
College. Professor Papandreou-Suppappolas enthusiasm in the eld of signal processing
and devotion to teaching and her students has truly been inspiring.
I would also like to thank Professors Andreas Spanias and Cihan Tepedelenlioglu
for their interest in my research, and agreeing to serve as members of my supervisory
committee.
A very special thank you to Cynthia Cantrell, for her patience and constant encour-
agement while supporting me through school, jobs, relocation and all of my endeavours
no matter how trying.
Heartfelt thanks are due to Toni Pardi for her friendship and constant encouragement
throughout the past year, not only for support while implementing the fractional Fourier
transform, but for her detailed editorial guidance.
I must thank my proof readers, Janette Cantrell, Forest Brown and Ken Ward for
having interest in my research and being willing to lend a hand.
I must recognize Jim Sagmiller at Northrop Grumman for initially inspiring me to
study SAR signal processing, and my former colleagues at Lockheed Martin in Goodyear,
Arizona, for countless hours in collaboration studying radar processing concepts.
Finally, a special thank you to Ian Cumming and Frank Wong for writing digital
processing of Synthetic Aperture Radar Data; and Walter Carrara, Ron Goodman, and
Ronald Majewski for writing Spotlight Synthetic Aperture Radar: Signal Processing Al-
gorithms. These two textbooks have provided me with countless hours of study and in-
valuable reference for my interest in SAR signal processing.
iv
TABLE OF CONTENTS
Page
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Research, Motivation and Objectives . . . . . . . . . . . . . . . . . . . . 1
1.2 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Chirp Scaling Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Chirp Scaling Algorithm Description . . . . . . . . . . . . . . . . . . . . 6
2.2 Example of Chirp Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Constant Shift (bulk RCMC) . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Shift Varying Linearly with Range (differential RCMC) . . . . . . 13
2.3 Chirp Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Range Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Azimuth Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6 Range Variant Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Fractional Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1 Denition and Properties of the Fractional Fourier Transform . . . . . . . 21
3.2 Relation of FRFT with Other Transforms . . . . . . . . . . . . . . . . . . 24
3.3 The Discrete FRFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.1 FRFT Implementation in MATLAB . . . . . . . . . . . . . . . . . 28
4 Linear Frequency-Modulated Chirp Detection . . . . . . . . . . . . . . . . . . 29
4.1 Fractional Autocorrelation Function . . . . . . . . . . . . . . . . . . . . 29
4.2 LFM Detection Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2.1 GLRT Detector . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2.2 Fractional Autocorrelation Function Detector . . . . . . . . . . . 33
4.2.3 RADARSAT LFM detection accuracy . . . . . . . . . . . . . . . 36
4.2.4 RADARSAT FRFT FM Rate Detection Resolution . . . . . . . . 39
v
CHAPTER Page
5 Fractional Focusing and the Chirp Scaling Algorithm . . . . . . . . . . . . . . 43
5.1 Fractional Chirp Scaling Algorithm (FrCSA) . . . . . . . . . . . . . . . . 43
5.2 Fractional Focusing with Contrast Maximization . . . . . . . . . . . . . 44
5.3 Fractional Focusing and the CSA . . . . . . . . . . . . . . . . . . . . . . 49
6 Application of Enhanced FRFT Focusing Technique Using RADARSAT Data . 52
6.1 Processing Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.1.1 FRFT Algorithm Implementation . . . . . . . . . . . . . . . . . . 53
6.2 RADARSAT-1 Scene Phase History Data . . . . . . . . . . . . . . . . . 54
6.3 FM Rate Estimation Using Fractional Autocorrelation and Contrast Max-
imization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.4 Fractional Auto Focus for Chirp Scaling Algorithm . . . . . . . . . . . . 66
6.4.1 Fast-time FRFT chirp optimization . . . . . . . . . . . . . . . . . 66
6.4.2 Azimuth FRFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.4.3 Chirp scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.4.4 Range FRFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.4.5 Range matched lter . . . . . . . . . . . . . . . . . . . . . . . . 67
6.4.6 Range IFRFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.4.7 Azimuth matched lter . . . . . . . . . . . . . . . . . . . . . . . 68
6.4.8 Azimuth IFRFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.4.9 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
7.1 LFM Detection Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 74
7.2 Fractional Focusing and the CSA Implementation . . . . . . . . . . . . . 75
7.2.1 Contrast Maximization . . . . . . . . . . . . . . . . . . . . . . . 75
7.2.2 Enhanced Fractional Focusing (EFrF) using the CSA . . . . . . . 75
7.3 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.3.1 LFM detection using the Discrete Chirp-Fourier Transform . . . . 76
vi
CHAPTER Page
7.3.2 Multi-component LFM Rate Resolution . . . . . . . . . . . . . . 76
7.3.3 LFM detection and instantaneous frequency . . . . . . . . . . . . 77
7.3.4 Matched ltering and time-bandwidth product . . . . . . . . . . . 77
7.3.5 Time-frequency methods for moving target and inverse synthetic
aperture radar (ISAR) processing . . . . . . . . . . . . . . . . . . 79
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
vii
LIST OF FIGURES
Figure Page
2.1 Range migration of a single target in 2-D processor memory, with each cell
corresponding to an in-phase quadrature or I & Q complex sample format. . . 8
2.2 Flow diagram of the chirp scaling algorithm. Here, Rg corresponds to range
and Az corresponds to azimuth. . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Satellite collection geometry, Figure taken from [1]. . . . . . . . . . . . . . . 10
2.4 Range and azimuth reference to image data array. Each cell (corresponding
to an in-phase quadrature or I & Q complex sample format) represents the
range resolution
r
and azimuth resolution
a
. . . . . . . . . . . . . . . . . . 11
2.5 Time-frequency chirp scaling (constant shift). . . . . . . . . . . . . . . . . . 13
2.6 Chirp scaling effect on the range position of three targets with the same FM
rate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.7 RADARSAT-1 range Doppler (Vancouver Airport BC). . . . . . . . . . . . . 18
2.8 Range cell migration (Vancouver Airport BC). . . . . . . . . . . . . . . . . . 18
3.1 LFM projection in the TF plane. . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 LFM projection in the rotated TF plane using the FRFT. . . . . . . . . . . . . 22
3.3 Fractional TF plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 Two different views of the FRFT of a rectangular signal. . . . . . . . . . . . 26
4.1 Plot of FrACF detector statistic L(K) in (4.17) for different = tan
1
(K),
where = /2. The true LFM rate is K
a
= 1808 and the estimated one is
K
/
a
= 1807.9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Range FM rate estimation (N = 1349, K
r
= 721.4 GHz/s (see Table 4.1)). . . 37
4.3 Range standard deviation FM rate estimation. . . . . . . . . . . . . . . . . . 37
4.4 Range bin FM rate estimation (N = 705, K
a
= 1808 Hz/s (see Table 4.2)). . . 38
4.5 Azimuth standard deviation azimuth FM rate estimation. . . . . . . . . . . . 38
4.6 Azimuth FrACF ROC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.7 Multi-component TF plane (normalized axis). . . . . . . . . . . . . . . . . . 41
viii
Figure Page
4.8 Minimum detectable range resolution. . . . . . . . . . . . . . . . . . . . . . 41
4.9 Multi-component detection, N = 4096. . . . . . . . . . . . . . . . . . . . . . 42
5.1 Block diagram of FrCSA algorithm. Here R denotes range, and Az denotes
azimuth. Figure taken from [2]. . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2 Wigner distribution representations of matched lter and range-bin signals. . 45
5.3 (WD) or pulse after matched ltering: using classical FT (upper plot) and
optimized with FRFT optimization (lower plot). . . . . . . . . . . . . . . . . 46
5.4 Comparison of target response for range-bin 969 using classical FT (F
1
r
)
(blue-line) and optimized FRFT (F

r
) (red-line). . . . . . . . . . . . . . . . . 47
5.5 MATLAB mesh plot of 2-D contrast maximization search; sweeping K
a
and . 48
5.6 MATLAB scaled image of 2-D search, highlighting maximum contrast at:
Ka = 1804.5 and = 1.0002. . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.7 Block diagram providing an overview of fractional focusing with CSA. . . . . 50
6.1 Plot of image clutter for [MNR[ < [CNR[, showing the relationship of total
noise
N
with the total noise in the NRA,
N
NRA
. Asimilar graphic is provided
in [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.2 RADARSAT-1 phase-history data (Vancouver Airport BC). . . . . . . . . . . 59
6.3 Range FM rate detection using FrACF. . . . . . . . . . . . . . . . . . . . . . 60
6.4 Range SNR for RADARSAT-1 scene (Vancouver Airport BC). . . . . . . . . 60
6.5 RADARSAT-1 range Doppler (Vancouver Airport BC). . . . . . . . . . . . . 61
6.6 Azimuth FM rate detection using the FrACF. . . . . . . . . . . . . . . . . . . 62
6.7 Azimuth CNR for RADARSAT-1 scene. . . . . . . . . . . . . . . . . . . . . 62
6.8 Azimuth FRFT optimized parameters using contrast maximization. . . . . . . 63
6.9 Clutter-to-noise (CNR) ratio for each pulse from a single look complex (SLC)
image (compressed in both range and azimuth). . . . . . . . . . . . . . . . . 64
6.10 Pulse FRFT optimization values obtained using the FrACF. . . . . . . . . . . 65
ix
Figure Page
6.11 Pulse FRFT optimization values using FrACF; for =7e
5
, K
r
=2.8e
8
using Equation (4.19). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.12 Range matched lter phase and magnitude for pulse 2048. . . . . . . . . . . 68
6.13 RADARSAT-1 scene (Vancouver, BC). . . . . . . . . . . . . . . . . . . . . . 70
6.14 Full resolution chip of the RADARSAT-1 scene. . . . . . . . . . . . . . . . . 71
6.15 MATLAB point target analysis at target location (2424,2415). Target upsam-
pled by a factor on 16 and displayed with MATLAB routine imagesc. . . . . . 72
6.16 Point target fast-time comparison (azimuth pulse 2424). . . . . . . . . . . . . 73
7.1 Chirp: time-bandwidth product for 1 1. . . . . . . . . . . . . . . . . 78
x
LIST OF TABLES
Table Page
4.1 Range matched lter parameters . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Azimuth matched lter parameters . . . . . . . . . . . . . . . . . . . . . . . 35
6.1 RADARSAT-1 Fine-beam:2 parameters . . . . . . . . . . . . . . . . . . . . 55
6.2 RADARSAT-1 Vancouver scene parameters . . . . . . . . . . . . . . . . . . 55
6.3 RADARSAT-1 Vancouver scene contrast comparisons . . . . . . . . . . . . . 71
xi
CHAPTER 1
Introduction
1.1 Research, Motivation and Objectives
Synthetic aperture radar (SAR) is a microwave imaging technique capable of pro-
viding high-resolution imagery from data collected by a relatively small antenna [4]. The
SAR takes advantage of the forward motion of the platform while transmitting and receiv-
ing short waveforms or pulses to form the equivalent of a long antenna. It is a coherent
system, in that it retains both the phase and magnitude of the received echos which are
synthesized in the signal processor to produce high resolution imagery [5]. Compared to
passive electro-optical systems, SAR is an active system capable of collecting imagery in
day/night and cloud cover conditions.
SARwas initially developed by Carl Wiley at Goodyear Aircraft Company in Goodyear,
Arizona, in 1951 and was originally called Doppler Radar Beam Sharpening. With the
onset of Moores law and the digital processor, this technology has continued to evolve
from the early optical processors, where phase-history was recorded on lm and pro-
cessed on an optical bench, to todays high speed digital processors. For strip-mapping
SAR, where the antenna is xed with respect to the platform, high resolution radar images
are formed on digital processors using two-dimensional (2-D) correlation algorithms. For
spotlight SAR, where the antenna is steered to continuously illuminate the same spot,
image formation relies on the principle of computer tomography [4]. Another imaging
mode is inverse synthetic aperture radar (ISAR), where the antenna or radar beam is sta-
tionary and the movement of the target is used to create the synthetic aperture. Based on
the imaging modes, different image formation algorithms are used. Strip-mapping SAR
uses range Doppler, range migration and chirp scaling algorithms for image formation
[1], whereas spotlight SAR uses the polar format, range migration and chirp scaling al-
1
gorithms [6]. The image formation methods used by ISAR combine the range Doppler
algorithm (RDA) along with time-frequency methods such as linear frequency-modulated
(LFM) chirp detection approaches [7, 8]. Note that SAR signals containing moving tar-
gets (including ISAR) are multi-component LFM signals and, in order to separate and
focus this energy, efcient LFM detection methods are necessary [8].
For high quality imagery, autofocus methods are necessary to remove quadratic and
higher order phase errors caused by uncompensated motion between the antenna and the
scene being imaged. Other phase error sources are algorithm/hardware limitations and
propagation effects [6]. To correct for these phase errors (which can cause image de-
focus), autofocusing techniques are used to estimate the phase error and correct it using
a phase cancellation lter. As with the image formation algorithms, different autofocus
methods are more applicable to the imaging modes. For strip-mapping, map-drift and
contrast maximization methods are used [1], and for spotlight, phase gradient autofocus
(PGA), map-drift and contrast maximization methods are commonly used [6, 9].
The chirp scaling algorithm (CSA) for SAR image formation was rst introduced in
1994 [10] and has had wide spread acceptance, mainly due to its good focusing capabil-
ities and ease of implementation. The CSA provided a means of performing range cell
migration correction (RCMC) using fast Fourier transform (FFT) complex multiplication
methods. This approach is inherently phase preserving and can be implemented more
efciently than the interpolation method used by the conventional RDA, where accuracy
is limited to the size of the interpolation kernel and depends on range varying coef-
cients [1, 6]. Another feature of the CSA is that it provides secondary range compression
(SRC) in its 2-D match ltering process. The SRC is an efcient method for handling the
range-azimuth coupling, which becomes increasingly necessary for large-swath, large-
beamwidth and high squint angle applications.
Recently, the fractional Fourier transform (FRFT) has been used to enhance the fo-
2
cusing capabilities of the CSA [2]. The FRFT is a subset of the Fourier transform (FT)
that was introduced in the eld of optics by Namias in the 1980s [11] and then to the
signal processing community in 1994 [2, 12]. Whereas the FT forms orthogonal coor-
dinates used to dene the time-frequency (TF) plane, the FRFT generalizes the FT by
providing a continuum of fractional domains between the conventional TF domain [13].
The FRFT achieves this by rotating the TF plane in the counter clockwise direction with
the use of a rotational parameter , which is periodic with period 2. Since the FRFT is
a generalization of the FT, it is also a linear, unitary transform.
Having unitary and Hermitian operators associated with the FRFT [14, 15] allows for
a denition of the fractional autocorrelation function [13]. The fractional autocorrelation
function at arbitrary angles corresponds to radial slices of the ambiguity function (AF)
passing through the origin at the same arbitrary angles, reducing a 2-D AF search to a
1-D fractional autocorrelation function search [16, 17]. Using a fast discrete FRFT ap-
proximation with the fractional autocorrelation function provides an efcient detection
method, especially for LFM signals, in low signal-to-noise ratio (SNR) environments.
To characterize the accuracy of LFM detection using the FRFT in varying SNR environ-
ments, the Cramr-Rao lower bound (CRLB) for LFM mean squared error is often used
as a threshold for comparing the fractional autocorrelation function results.
The integration of the CSA with the FRFT is called the fractional chirp scaling al-
gorithm (FrCSA) [2]. Specically, the FRFT enhances the CSA processing using a local
optimization procedure (LOP), and the resulting FrCSA provides signicant improve-
ments to the SNR and sidelobe reduction ratio (SLRR) when compared to classical FFT
implementations.
One measure of SAR image quality is the quadratic phase error (QPE) [1, 6], which, if
not corrected, will contribute to elevated sidelobes and a broadening of a targets impulse
response width (IRW). A mismatch of frequency-modulation (FM) chirp rates between
3
the received signal and the matched lter will contribute to QPE, thereby decreasing im-
age quality. One cause for the mismatch is the uncompensated motion between the radar
antenna and the target. In addition to using the FRFT to enhance the focusing capabilities
of the CSA, recent works have indicated that the FRFT has desirable properties that can
be used to develop new methods for correcting this mismatch [2, 8, 18]. To use the FRFT
to enhance the CSA, we will leverage the FrCSA concept of using the LOP to produce
an optimal response by rst detecting the FM chirp rate mismatch using the fractional
autocorrelation function. The FM rate mismatch will be corrected in the range direction
on a pulse-by-pulse basis in the unprocessed phase-history data prior to CSA processing.
Since this occurs prior to image formation, it can be considered an update to the motion
compensation to correct for line-of-sight (LOS) translational motion. This update causes
changes in the distance between the target and sensor. For corrections in the azimuth
direction (QPE caused by estimation error of the effective radar velocity), we will use an
LOP to maximize a targets contrast, thereby providing improvements in SNR and peak-
to-sidelobe ratio (PLSR). Having the ability to maximize contrast using the FRFT is a
result of the extra degree of freedom provided by the use of the transform [2].
1.2 Thesis Organization
The rest of this thesis is organized as follows. Chapter 2 provides the CSA signal
model and is the basis for improvements provided by the application of fractional opti-
mization techniques. The signal model is derived in Ian Cummings and Frank Wongs
book, digital processing of Synthetic Aperture Radar Data [1]; it is presented in this
section and will be cited throughout this thesis. Another book by Walter Carrara, Ron
Goodman and Ronald Majewski Spotlight Synthetic Aperture Radar: Signal Processing
Algorithms [6] provides a derivation of the CSA and is also referenced in this section.
Chapter 3 provides an overview of the fractional Fourier transform by providing the
4
FRFT kernel and listing some of the key properties, specically the relationship with the
FT and the FRFT additivity property. The relationship of the FRFT with the Wigner dis-
tribution, and thus the ambiguity function, is also presented. This relationship is needed
to demonstrate how the fractional autocorrelation function is used in LFM detection. The
limitation of not having a fast FRFT algorithm comparable to the FFT is noted along with
a method for computing the continuous FRFT with complexity O(Nlog
2
N), where N is
the signal length.
For the FRFT to be a useful tool in SAR signal processing, a transform rotational
parameter is necessary to tune the transform for an optimal response. By concentrating on
the mismatch between the CSA matched lter FM rates and the FM rates of the returned
SAR signal, LFM detection can be used to obtain the FRFT optimal parameters. Chapter
4 examines LFM detection accuracy by comparing the FRFT detection performance with
the CRLB for LFM signals.
Chapter 5 presents the proposed focusing technique which is based on enhancing the
CSA using the FRFT with optimizations obtained using a local optimization procedure.
This fractional focusing method is a rst attempt at implementing the FrCSA as provided
in [2], with modications necessitated by using optimized FRFT rotational parameters
for a RADARSAT-1 scene.
In Chapter 6, we use a local optimization procedure to detect the FM rate mismatch
between the signal and the matched lter using the fractional autocorrelation function.
The mismatch in FM rates are used to optimize the FRFT pulse compression response
for a RDRASAT-1 scene, by providing the optimizations to the enhanced focusing CSA
as provided in Chapter 5. Comparisons of the CSA and enhanced CSA are provided by
evaluation of contrast and point target analysis.
5
CHAPTER 2
Chirp Scaling Algorithm
2.1 Chirp Scaling Algorithm Description
The chirp scaling algorithm (CSA), also known as the differential range deramp algo-
rithm, was presented in [1, 6, 10] as a renement of the classical range Doppler algorithm
(RDA) to provide very high quality synthetic aperture radar (SAR) imagery. The CSA
has received widespread acceptance due to its efcient implementation and the ability to
perform secondary range compression (SRC), which can limit the focusing accuracy of
higher squint and wide-aperture systems using the RDA. This improvement is a benet
of the phase-history data being available in the 2-D frequency domain where SRC can
be made azimuth frequency dependent. While the RDA algorithm uses interpolation to
implement range cell migration correction (RCMC), the CSA uses (a) frequency shifting
to correct for the constant migration component, and (b) chirp scaling to correct for the
linear migration component. The net result of the scaling and shifting operations, also
referred to as differential and bulk RCMC, respectively, is that a targets energy is aligned
to the correct range position.
As the scaling operation is better matched to chirp-encoded signals, a requirement for
the CSA is that the signal or phase-history data must be chirped in the range direction. If
the collection system employs a de-ramp or de-chirp on receive, as many do, the chirp-
encoding will need to be reapplied prior to the scaling operation. Also, the CSA works
well for both stripmap and spotlight SAR collection systems [6].
We consider a signal, such as a linear frequency-modulated (LFM) chirp, that is given
by
s(t) = exp( j2(K
r
/2)t
2
) (2.1)
with frequency-modulation (FM) rate K
r
Hz
2
, that is transmitted by a SAR. The signal is
6
reected off a point reference target and is migrated at different azimuth locations due
to constant changes in range throughout the synthetic aperture. The time taken for the
signal to travel to the target and return is given by = 2R/c, where R is the range of the
target from the radar and c = 310
8
m/s is the speed of wave propagation in light. This
scenario is depicted in Figure 2.1. The received baseband signal after the range migration
is given by [1],
s
r
(t, ) = A
0
w
r
_
t
2R

c
_
w
a
_

c
c
_
e
j4 f
0
R

/c
s
_
t
2R

c
_
, (2.2)
where
R

=
_
R
2
0
+V
2
r

2
_
1/2
, (2.3)
is the azimuth time, A
0
is an arbitrary complex constant,
c
is the beam center offset
time, w
r
(t) is the rectangular range envelope function, w
a
() is the sinc-squared azimuth
envelope function, f
0
is the radar center frequency, R

is the slant range at azimuth time

, R
0
is the slant range of closest approach, and V
r
is the effective radar velocity.
To correct for this range migration and shift the signal to its proper location, a fre-
quency modulation is applied to a chirp-encoded signal to achieve a shift or scaling of
the signal [19]. Note that the maximum shift or scale change that can be implemented
by the frequency modulation cannot be too large in order to avoid any problems with the
associated change in the signals center frequency and bandwidth. This restriction is mit-
igated by applying RCMC in two steps so that only the difference in range cell migration
(RCM) at different ranges is corrected in the chirp scaling operation and the bulk RCM is
completed in the 2-D frequency domain along with SRC [1]. The
0
point in Figure 2.1 is
also referred to as the broadside range position, which is the position where the azimuth
frequency f

= f

c
. Since the energy of a target should align along a constant range, the
range to a target in this pulse will be the range for which the target energy of all pulses
are corrected to. In other words, no RCMC will be applied to the pulse corresponding to

0
.
7
The CSA consists of four fast Fourier transforms (FFTs) and three phase multiplies.
Its derivation can be provided using Equation (2.2) for the received baseband signal and
the range equation in (2.3). Note that the signal is transformed into the 2-D frequency
domain with the use of 1-D FFTs and the principle of stationary phase (POSP) which is
used to approximate the Fourier transform (FT) integrals.
Figure 2.1: Range migration of a single target in 2-D processor memory, with each cell
corresponding to an in-phase quadrature or I & Q complex sample format.
The CSA steps are presented in [1], and Figure 2.2 is a ow diagram of the algorithm.
Given phase-history data which is in the range-time/azimuth-time domain, the CSA steps
are as follows.
1. The azimuth FFT is rst computed in order to transform the received data, s
r
(t, ),
into the range Doppler domain.
2. Chirp scaling is applied, using a phase multiply to equalize the range migration of
all targets.
8
3. A range FFT is used to transform the data to the 2-D frequency domain.
4. A phase multiply is performed with a reference function, which applies range com-
pression, SRC, and bulk RCMC in the same operation.
5. A range inverse FFT (IFFT) is performed to transform the data back to the range
Doppler domain.
6. A phase multiply is performed to apply azimuth compression with a time varying
match lter. A phase correction is also required as a result of the chirp scaling in
step 2, which can be incorporated into the same phase multiply.
7. The nal azimuth IFFT is computed to transform the compressed data to the SAR
image domain.
Figure 2.2: Flow diagram of the chirp scaling algorithm. Here, Rg corresponds to range
and Az corresponds to azimuth.
Figure 2.3 shows the collection geometry for an orbiting platform. The CSA is sim-
ilarly applicable to airborne platforms. The slant range variable R is actually the instan-
taneous slant range R

that continuously changes as the beam footprint moves across the

9
target at V
r
. As shown in (2.3), a targets slant range will be hyperbolic due to the constant
change in range from the broadside position of the antenna.
Figure 2.3: Satellite collection geometry, Figure taken from [1].
Figure 2.4 provides a denition of the image data array and demonstrates how the
azimuth and range directions are referenced. The conventions in Figures 2.1 and 2.4
will be used throughout this study, in that range will always be increasing from left to
right and azimuth will always be increasing from bottom to top. The term fast-time will
indicate the range direction and range-bin will be used to indicate a column vector at a
constant range. The term slow-time will indicate the azimuth direction and pulse number
will indicate a row vector of constant azimuth.
During the imaging operation or collect, pulses are received at the pulse repetition
frequency (PRF) and sampled at the range sampling rate to form the 2-D data array suit-
able for image formation processing. In Figure 2.4, the variables
r
and
a
refer to the
range and azimuth resolutions, respectively. For a radar signal, range is the linear dis-
tance along the line of sight from the radar to target. For stripmap systems with relatively
10
small squint, range is approximately perpendicular to the radar ight path and its resolu-
tion is given with
r
= c/(2BW), where c is the speed of wave propagation in light and
BW is the transmitted bandwidth. For two objects to be resolved,in the range direction,
they must be separated by a distance greater than the resolution
r
. Azimuth is the linear
distance in the direction parallel to the radar ight path. It can also be considered as the
along-track direction in an image, since it is the relative along-track position of an ob-
ject within the antennas eld-of-view, following the radars line of ight. For stripmap
systems with relatively small squint, the azimuth direction is perpendicular to the range
direction, and the resolution of an image in the azimuth direction is normally quoted as

a
= L
a
/2, where L
a
is the length of the antenna along the azimuth direction. For two ob-
jects to be resolved,in the azimuth direction, they must be separated by a distance greater
than the resolution
a
.
Figure 2.4: Range and azimuth reference to image data array. Each cell (corresponding to
an in-phase quadrature or I & Q complex sample format) represents the range resolution

r
and azimuth resolution
a
.
11
2.2 Example of Chirp Scaling
As noted, the chirp scaling concept was initially explained in [19] and the example
discussed in this section for SAR range processing was provided in [1]. Since the term
scaling is customarily used to indicate time or axis scaling, we wanted to provide this
example to clarify its use in the CSA and also show how the residual phase term is a
product of the scaling operation. In particular, by chirp scaling, we only mean multiply-
ing an LFM chirp by an exponential linear phase term to cause a frequency shift in the
LFM chirp; then we show that this is equivalent to a time shift of the LFM chirp.
2.2.1 Constant Shift (bulk RCMC)
In order to demonstrate the CSA with a constant shift, we multiply the LFM chirp
with chirp rate K
r
, and shifted in time t =t
a
,
s
0
(t) =exp
_
j2(K
r
/2)(t t
a
)
2
_
(2.4)
with the following complex exponential linear phase signal (also called the scaling func-
tion)
s
p
(t) =expj2K
r
(t t
a
)t (2.5)
to obtain the scaling operation
s
1
(t) =s
0
(t)s
p
(t) = exp
_
j(K
r
/2)(t t
a
)
2
_
exp j2(t t
a
)t (2.6)
This multiplication actually results in frequency shifting the LFM chirp s
0
(t) by f
sc
=
K
r
t. This multiplication is also referred to as chirp scaling. By rearranging the terms in
Equation (2.6), we can also express it as,
s
1
(t) =exp
_
j(K
r
/2)(t t
a
+t)
2
_
exp
_
jK
r
t
2
_
. (2.7)
12
Equation (2.7) shows that the frequency-shifted (or scaled) LFM chirp by f
sc
= K
r
t
is equivalent to time-shifting the LFM by t
sc
= t plus an undesired residual phase that
can be removed during azimuth matched ltering [1].
Figure 2.5: Time-frequency chirp scaling (constant shift).
Figure 2.5 is a plot of the time-frequency plane showing how the scaling or frequency
shift f
sc
= K
r
t causes the t shift from the original center t. This constant shift in fre-
quency is the bulk RCMC in the CSA.
2.2.2 Shift Varying Linearly with Range (differential RCMC)
Figure 2.6 shows three equally spaced targets in the time-frequency (TF) plane having
the same FM rate. The scaling function (dashed line) is itself an LFM signal and for this
example, K
r
is positive causing a compression of the range scale. This compression
results in the targets shifting closer to the center reference point. In the derivation of the
CSA, all targets are shifted in relation to a common reference location in range (R
re f
)
which is a line where no time shift is applied. In Figure 2.6, R
re f
will correspond to
the center target where t = 0. For this example, we will use the following LFM scaling
13
function,
s
p
(t) =exp
_
j2(K
r
/2)t
2
_
. (2.8)
Figure 2.6: Chirp scaling effect on the range position of three targets with the same FM
rate.
When we multiply s
0
(t) with t
a
= 0 in (2.4) with the LFM scaling function in (2.8), the
resulting scaling transformation is a change in the FM rate of the LFM chirp from K
r
to
K
r
(1+). Specically, the new LFM chirp signal is given by,
s
sc
(t) =s
0
(t)s
p
(t) = exp( jK
r
t
2
)exp( jK
r
t
2
) = exp( jK
r
(1+)t
2
) (2.9)
which changes the frequency of each target by f
sc
= K
r
t. In other words, the shift in
time or range is provided by t and is proportional to the distance from R
re f
or t =0. The
frequency shift f required to move the target from t
b
to t
/
b
can be obtained from,
f t
b
=K
r
(t
/
b
t
b
) =K
r
t
/
b
. (2.10)
This relationship can be shown in the time-frequency plane using the Wigner distribution
of the original LFM chirp and the transformed LFM chirp. Specically, it can be shown
that [20]
W
s
sc
(t, f ) =W
s
0
(t, f +K
r
t) (2.11)
14
where
W
s
0
(t, f ) =
_

s
0
_
t +

2
_
s

0
_
t

2
_
exp(j2 f )d (2.12)
is the Wigner distribution of s
0
(t).
2.3 Chirp Scaling
In the azimuth-time, range-time domain, a targets range curvature will depend on
both the azimuth and range position. In the azimuth-frequency range-time domain, also
referred to as the range Doppler domain, the range curvature will not depend on the
targets azimuth location since all target energy will align along a constant range [6]. For
this reason, chirp scaling is applied in this domain using a 2-D or azimuth dependent
range phase multiplication.
The rst step for the CSA is to compute the FT of the received signal into the range
Doppler domain. Since a closed form expression for the transform is not available, an
approximation will be used by applying the POSP [1, 21]. Specically we take the FT of
s
r
(t, ) in Equation (2.2) from the azimuth time to the Doppler f

to obtain
S
rd
(t, f

) =Aw
r
_
t
2R
0
cD( f

re f
,V
r
re f
)
_
W
a
( f

c
)exp
_
j
4 f
0
D( f

,V
r
)
c
_
exp
_
jK
m
c
2
_
1
D( f

,V
r
re f
)
D( f

re f
,V
r
re f
)
_
_
R
0
D( f

,V
r
)

R
re f
D( f

,V
r
)
_
2
_
(2.13)
where A is a complex constant, W
a
( f

) is the Doppler spectrum envelope or FFT of

w
a
() in (2.2), D( f

,V
r
) is the migration parameter in the range Doppler domain for the
effective radar velocity V
r
and is given by,
D( f

,V
r
) =

1
c
2
f
2

4V
2
r
f
2
0
, (2.14)
15
and K
m
is the modied range FM rate
K
m
=
K
r
1K
r
cR
0
f
2

/(2V
2
r
f
3
0
D
3
( f

,V
r
))
. (2.15)
For LFM chirps, the azimuth dependent scaling function is given by
s
sc
(t, f

) =exp
_
jK
m
_
D( f

re f
,V
r
re f
)
D( f

,V
r
re f
)
1
_
t
2
_
(2.16)
assuming no shift was applied to the target at R
re f
.
With the S
rd
(t, f

) and s
sc
(t, f

) dened as in (2.13) and (2.16) respectively, the azimuth

dependent chirp scaling is achieved with
S
1
(t, f

) =s
sc
(t, f

)S
rd
(t, f

). (2.17)
2.4 Range Compression
After the chirp scaling phase multiplication, the phase-history data is transformed to
the azimuth-frequency, range-frequency domain by taking the FT of S
1
(t, f

) in the range
direction, from t to f
t
, to obtain S
2
( f
t
, f

). The data is now in the 2-D frequency domain

and the POSP is used again to evaluate the integral providing S
2
( f
t
, f

)
S
2
( f
t
, f

) = A
1
W
r
( f
t
)W
a
( f

c
)
exp
_
j
4R
0
f
0
D( f

,V
r
)
c
_
exp
_
j
D( f

,V
r
)
K
m
D( f

re f
,V
r
)
f
2
t
_
exp
_
j
4R
0
cK
m
D( f

re f
,V
r
re f
)
f
t
_
exp
_
j
4
c
_
1
D( f

,V
r
re f
)

1
D( f

re f
,V
r
re f
)
_
R
re f
f
t
_
exp
_
j
4K
m
c
2
_
1
D( f

,V
r
re f
)
D( f

re f
,V
r
re f
)
_
_
R
0
D( f

,V
r
)

R
re f
D( f

,V
r
)
_
2
_
, (2.18)
where W
r
( f
t
) is the range spectrum envelope or FFT of w
r
(t) in (2.13). The second
phase multiplication in the CSA is the range matched lter, which is a 2-D lter with
dependencies on both azimuth and range and is applied to all range-frequencies with a
16
dependence on f

. The matched lter applies the bulk RCMC and SRC by removing the
second and fourth exponential terms of (2.18) resulting in S
3
( f
t
, f

).
S
3
( f
t
, f

) = A
1
W
r
( f
t
)W
a
( f

c
)
exp
_
j
4R
0
f
0
D( f

,V
r
)
c
_
exp
_
j
4R
0
cD( f

re f
,V
r
re f
)
f
t
_
exp
_
j
4K
m
c
2
_
1
D( f

,V
r
re f
)
D( f

re f
,V
r
re f
)
_
_
R
0
D( f

,V
r
)

R
re f
D( f

,V
r
)
_
2
_
(2.19)
The last step in range compression is to compute the 1-D inverse FT of S
3
( f
t
, f

) from f
t
to t, returning it to the range Doppler domain
S
4
(t, f

) = A
2

r
_
t
2R
0
cD( f

re f
,V
r
re f
_
W
a
( f

c
)exp
_
j
4R
0
f
0
D( f

,V
r
)
c
_
exp
_
j
4K
m
c
2
_
1
D( f

,V
r
re f
)
D( f

re f
,V
r
re f
)
_
_
R
0
D( f

,V
r
)

R
re f
D( f

,V
r
)
_
2
_
(2.20)
where the range envelope
r
(t) is the IFFT of W
r
( f
t
) in (2.19). The signal S
4
(t, f

) is
now compressed in range with the range curvature corrected for all targets in the scene.
Figure 2.7 shows the range compressed data in the range Doppler domain. For this scene,
the amount of RCM is small and would only be noticeable in the full resolution scene
data.
17
Figure 2.7: RADARSAT-1 range Doppler (Vancouver Airport BC).
Figure 2.8 plots the range cell migration throughout the processed aperture at the range
reference location R
re f
. As the plot indicates, the total RCM is 137 m and if we use
r
=
c/(2BW) with the bandwidth BW=30.11 MHz, then the range resolution is
r
=4.97 m
indicating the total RCM is 27 range-bins or pixels, which means the complete RCM
of 27 pixels is spread across the 4096 samples of the azimuth array. This is a plot of the
fourth exponential term in (2.18), thus it is the bulk RCMC.
Figure 2.8: Range cell migration (Vancouver Airport BC).
18
2.5 Azimuth Compression
The third phase multiplication is the azimuth matched lter which is the conjugate of
the rst and fth exponential terms of (2.18). The rst term is the azimuth modulation
and the fth term the residual phase that is created in the rst phase multiply (scaling
function). After applying the range dependent azimuth matched lter, azimuth compres-
sion is completed with a 1-D inverse FT. As a result, the scene data is now compressed in
both range and azimuth, providing a single look complex (SLC) image
S
5
(t, ) = A
4

r
_
t
2R
0
cD( f

re f
,V
r
re f
_

a
(
c
)exp j(t, ) (2.21)
where
a
() is the IFFT of W
a
( f

) in (2.20). There are different signal processing en-

hancements that are traditionally used in SAR processing but since we will be comparing
performance with the FRFT, these enhancements will not be implemented but discussed
in the conclusions.
2.6 Range Variant Scaling
The pervious section made an assumption that the transmitted pulse is an LFM chirp
and both V
r
and K
m
are range invariant. In general, this is not the case in that V
r
and
K
m
will vary in range providing higher order terms in the scaling function s
sc
(t, f

). To
account for these higher order terms, which are normally very small, we need a more
general expression for the scaling function,
s
sc
(t, f

) =exp
_
j2
_
t
0
K
m
q
t
(u, f

)du
_
(2.22)
where
q
t
(u, f

) =
_
D( f

re f
,V
r
)
D( f

,V
r
)
1
_
t +
2R
re f
c
_
D( f

re f
,V
r
)
D( f

re f
,V
r
re f
)D( f

,V
r
)

1
D( f

,V
r
)
_
. (2.23)
19
To evaluate (2.22), the terms K
m
q
t
(u, f

) can be expanded as a power series around

t. By assuming the frequency of the scaling function is locally linear, the function can be
approximated as,
s
sc
(t, f

) =exp
_
j
_
g
0
t +g
1
t
2
_
(2.24)
where the coefcients g
0
and g
1
vary with t and f

. With this approximation, the residual

phase (last term in (2.18)) is,

res
(t, f

) =
4
c
2
K
m
g
1
K
m
+g
1
_
R
0
D( f

,V
r
)

R
re f
D( f

,V
r
re f
)
+
cg
0
4g
1
_
2

g
2
0
4g
1
. (2.25)
20
CHAPTER 3
Fractional Fourier Transform
3.1 Denition and Properties of the Fractional Fourier Transform
The fractional Fourier transform (FRFT) of order , 2 2, is a generalization
of the classical Fourier transform (FT), and it can be interpreted as a rotation by an angle
=/2 in the time-frequency (TF) plane. It can also be considered as a decomposition
of the signal in terms of linear frequency-modulated (LFM) chirps [13, 22]. As a one-
parameter subclass of the class of linear canonical transforms, most of the properties of
the FRFT are special cases of the general properties of linear canonical transforms [23].
In Figures 3.1 and 3.2, we compare the FT and FRFT of an LFM chirp in the time-
frequency (TF) plane. We see that the bandwidth of the LFM chirp signal in the FT
domain is spread along the frequency axis as shown in Figure 3.1. If the TF plane is
rotated using the FRFT with angle , the energy distribution of the chirp is concentrated
in the rotated domain, as shown in Figure 3.2 [18]. This rotation of the TF plane makes
the FRFT an attractive transform to the signal processing community, especially for SAR
signal processing and LFM detection. Specically, the (t, f ) coordinate system is rotated
over an angle =/2 counter clockwise, and the rotated variables are given by (u

, v

)
in the TF plane in Figure 3.2. The variables can be obtained using the following rotation
matrix or by applying the rotation operator R

:
_

_
u

_
=
_

_
cos sin
sin cos
_

_
_

_
t
f
_

_
(3.1)
= R

(t, f )
21
The rotated axes are obtained as
u

=t cos + f sin (3.2)

v

=t sin + f cos (3.3)

and are more clearly demonstrated in Figure 3.3.
Figure 3.1: LFM projection in the TF plane.
Figure 3.2: LFM projection in the rotated TF plane using the FRFT.
22
Figure 3.3: Fractional TF plane.
The continuous one-dimensional (1-D) FRFT is dened by means of the transforma-
tion kernel [12],
K(t, u

) =
_

_
_
1 j cot e
j2
t
2
+u
2

2
cot j2u

t csc
, if ,= n, n N
(t u

), if = 2n, n N (3.4)
(t +u

), if = 2n+, n N
e
j2tu

, if = (/2)n, n N
For a continuous-time signal x(t), the FRFT is thus dened as
X(u

) = (F

x)(u

) =
_

x(t)K(t, u

)dt (3.5)
=
_

_
_
1 j cot e
j2
u
2

2
cot
_

x(t)e
j2
t
2
2
cot
e
j2u

t csc
dt, if ,= n, n N
X(u

) = x(t), if = 2n, n N
X(u

) = x(t), if = 2n+, n N
_

x(t)e
j2tu

dt = X(u

) = X( f ), if = (/2)n, n N
23
Given that F is the FT operator such that (Fx)( f ) =
_

x(t)e
j2t f
dt and F

is the
FRFT operator in Equation (3.5), then the FRFT possesses the following properties;
1. Zero rotation: F
0
= I where (Ix)(t) = x(t) is the identity operator
2. Consistency with the Fourier transform: F
/2
= F
3. Additivity of rotations: F

= F
+
4. 2 rotation: F
2
= I
5. Inverse FRFT: (F

r
)
1
= F
()
where (F

)
1
F

= I.
In addition, the FRFT kernel satises the following useful properties:
1. K(t, u

) = K(u

, t)
2. K(t, u

) = K

(t, u

), where * indicates complex conjugation.

3. K(t, u

) = K(t, u

)
4.
_

K(t, u

)K

(t, u

)dt = (u

), where (u

) is the Dirac delta function

5.
_

K(t, u

)K(u

, z

)du

= K(t, z
+
)
Additional FRFT properties can be found in [12] and [23].
3.2 Relation of FRFT with Other Transforms
Consider a signal x(t) and its FRFT (F

x)(u

). The Wigner distribution (WD)

W
x
(t, f ) of a signal x(t) was dened in Equation (2.12) in Chapter 2 as
W
x
(t, f ) =
_

x
_
t +

2
_
x

_
t

2
_
exp(j2t f )d.
The following relationship can be shown between the WD of the FRFT of a signal
and the rotated, over an angle , of the WD of the same signal [23, 24]. Specically,
24
W
x

(t, f ) = R

W
x
(t, f ) = W
x
(t cos 2 f sin, t sin +2 f cos), where the rota-
tion operator R

is dened in Equation (3.1).

Since several Cohens class of time-frequency distributions are obtained by convolv-
ing a kernel characterizing the distribution with the Wigner distribution, an important
relationship exists between the FRFT and those TFRs [24].
3.3 The Discrete FRFT
There are several challenges to the calculation of the FRFT; to date, there is no fast
method similar to that of the FFT, however approximations have been proposed for both
the discrete FRFT and continuous time transforms [12, 23, 24, 25, 26].
We need to introduce a new fractional operator F

r
that emphasizes the order of the
FRFT rather than the angle . It is thus related to the previous denition with = /2
or
F
/2
r
= F

or F
1
r
= F
/2
= F. (3.6)
The new operator F

r
has the same properties as F

and will be used as the operator for

the discrete approximation to the FRFT.
Figure 3.4 provides two different views of an FRFT of a rectangular signal with N =
64 samples, which sweeps the FRFT order 0 1. As mentioned in Section 3.1, the
FRFT is periodic in 2, and it is also symmetric about /2 with strong oscillations as
approaches 0 or . In Figure 3.4 we notice strong oscillations for < 0.5 and the
expected sinc pattern for = 1, which is in agreement with the FT property F
1
r
= F.
25
Figure 3.4: Two different views of the FRFT of a rectangular signal.
A denition of the discrete FRFT (DFRFT) is provided in [27] which generalizes
the denition of the discrete FT (DFT) in the same manner as the continuous fractional
FRFT generalizes the continuous FT. This denition of the DFRFT is based on a set of
eigenvectors of the DFT matrix, which consist of the discrete counterparts from the set
of Hermite-Gaussian functions. It satises the following desirable properties: (a) the
DFRFT is a unitary transform; (b) it preserves the additivity property F
b
F
a
= F
a+b
; (c)
it reduces to the DFT when the order is unity; and (d) it approximates to the continuous
FRFT.
A fast O(Nlog
2
N) approximation for computing the continuous FRFT is provided
in [25] which maps the N samples of the original signal to N samples of the FRFT. The
FRFT of an arbitrary signal x(t), with angle , is dened in Equation (3.5). In particular,
for angles that are not multiples of , the computation of the FRFT corresponds to the
following steps: (a) the signal is rst multiplied by an LFM chirp; (b) the FT of the signal
product is computed; (c) perform some scaling; (d) multiply by an LFM chirp.
Two methods for computing the approximation are provided in [25]. One method
26
which does not include the use of Fresnel integrals is,
(F

r
x)
_
m

2x
_
=
A

2x
e
i2()(
m
/2x)
2
N

n=N
e
i2(
(mn)
/2x)
2
e
i2()(
n
/2x)
2
x
_
n
2x
_
(3.7)
where = /2, = cot , = csc and
A

=
exp(jsgn(sin)/4+ j/2)
[ sin[
1/2
Since the summation in (3.7) is recognized as a convolution, the algorithm can be com-
puted with an FFT providing the FRFT with complexity O(Nlog
2
N).
The result of the rst chirp multiplication with x will cause a shearing in the TF plane,
making the bandwidth or time-bandwidth product twice as large. Since this is twice
the original support, x will need to be interpolated or up-sampled containing at least 2N
samples [26]. It will be assumed that this interpolation was done before x is used in
(3.7). For more detailed discussions of shearing in the TF plane see Appendix A of [25]
with an additional derivation in [23]. Also, the convolution of x with a chirp will cause
a similar shearing; see [20] for an additional discussion on the effect of time-frequency
distributions.
As noted and displayed in Figure 3.4, the FRFT exhibits strong oscillations as ap-
proaches 0 or 2. To adequately sample the FRFT in these regions would drive complexity,
so (3.7) is restricted to 0.5 [[ 1.5. There are different routines using the fractional
additivity property to work around this constraint of limiting [0.5, 1.5]. For [0, 2],
the method provided in [25] is,
1. If 0.5 1.5 or 2.5 3.5, evaluate the integral directly
2. If 0.5 0.5 or 1.5 2.5 then F

= FF
1
27
3.3.1 FRFT Implementation in MATLAB
There are two MATLAB programs available for computing (3.7), which are freely
available on the web. The rst is fracF and can be obtained at the website for the book
[23]. The second is fracft and is distributed as a routine within the MATLAB time-
frequency toolbox DiscreteTFDs, developed by J. ONeill. A complete analysis of the
two routines is provided in [26] and will not be covered here, however the difference in
their implementations will drive a decision to use fracF for the implementation in Chapter
6.
The main difference is that fracF operates on an even number of samples and uses
FFTs for upsampling in (3.7). The fracft routine operates on an odd number of samples
and uses sinc interpolator for upsampling. The sinc interpolator is dened as x(t) =

n=
x
n
sinc
_

T
(t nT)
_
where T is the sampling period and sinc(x) = sin(x)/x, for
x ,=0 and sinc(0) =1 [28]. It is noted in [26] that fracft provides slightly improved results
due to the symmetric kernel, however the odd number of samples will be problematic for
optimized FFT sizes needed in implementation.
28
CHAPTER 4
Linear Frequency-Modulated Chirp Detection
4.1 Fractional Autocorrelation Function
As the fractional Fourier transform (FRFT) is a linear canonical transformation [15], it
can be associated with a fractional autocorrelation function (FrACF). That is an important
function, for example, in estimating the rate of linear frequency-modulated (LFM) chirps
using one-dimensional (1-D) searches, especially at low signal-to-noise ratios (SNRs)
[15, 16, 17]. Using a fast discrete FRFT approximation, the 1-D searches can replace the
more computationally intensive 2-D ambiguity function (AF) searches since the FrACF
at an arbitrary angle corresponds to a radial slice of the AF passing through its origin at
the same arbitrary angle [15].
Consider a continuous-time signal s(t) and the unitary and Hermitian fractional oper-
ation (R

s)(t) presented in [15] as

(R

s)(t) = s(t cos)e

j2t sin
e
j
2
cos sin
.
Using the fractional operation, the FrACF, FA

(), of s(t) is simply the correlation be-

tween the signal s(t) and its fractionally operated signal in the time domain. Specically,
the FrACF is given by
2
,
FA

() =s, (R

s) = e
j
2
cos sin
_
s(t)s

(t cos)e
j2t sin
dt. (4.1)
Two special cases of the FrACF are obtained for = 0 and = /2, and are given by
FA
0
() =
_
s(t)s

(t )dt (4.2)
FA
/2
() =
_
[s(t)[
2
e
j2t
dt =
_
S( f )S

( f )d f (4.3)
where S( f ) is the conventional Fourier transform (FT) of s(t). The FrACF can also be
expressed in terms of the AF, AF
s
(t, ), of the signal s(t) or its FRFT, (F

r
s)(t) = S(u

)
2
Unless otherwise indicated, all integration limits are from -innity to innity.
29
in Equation (3.5). These two relationships are given as [16],
FA

() = AF
s
( cos, sin) (4.4)
=
_
S(u

)S

(u

)du

(4.5)
where the AF of s(t) is dened as
AF
s
(, ) =
_
s
_
t +

2
_
s

_
t

2
_
e
j2t
dt. (4.6)
The relationship between the FrACF and the AF in Equation (4.4) shows that the
FrACF at angle is equivalent to radial (of value ) slices of the AF, evaluated along
the orientation . This relationship is similar to the one that the time-correlation in (4.2)
shares with AF
s
(t, 0), which is a slice of the AF, evaluated along the horizontal axis.
Equation (4.4) will be exploited in the next section, for detecting LFM chirps.
Assuming a signal duration of N samples, an efcient method for calculating the
FrACF is given in [16], using a fast O(Nlog
2
N) approximation of the FRFT [25]. The
method makes use of another approach to compute the FrACF by taking the inverse FT of
the squared magnitude of the FRFT of the signal s(t) with the orthogonal angle parameter

= +/2. Specically, we can re-write (4.1) as

FA

() =
_
[S(u

)[
2
e
j2

. (4.7)
The efcient FrACF implementation follows from the discrete-time approximation of
Equation (4.7) [15]. Specically, with M different angles
m
, m = 1, . . . , M, the FrACF
can be effectively implemented by rst uniformly sampling the signal s(t) using the sam-
pling period T
s
to obtain the samples s[n] = s(nT
s
), n = 1, . . . , N; then the FRFT can be
calculated at orthogonal angles
m
+/2 for the sampled signal using the discrete FRFT
in [25]; the squared-magnitude s[n] of the FRFT is computed, and then its inverse FT is
obtained to yield the FrACF of the signal.
30
4.2 LFM Detection Accuracy
Amotivation for LFM estimation in SARsignal processing is to reduce the frequency-
modulation (FM) mismatch between the expected and received waveforms due to non-
linear motion between the radar and target [8, 18, 29]. In doing so, the fractional opti-
mization operators are used to tune the FRFT to produce an optimal response in pulse
compression, providing an increased SNR and improved peak sidelobe ratio (PSLR) per-
formance. To characterize the FM estimation accuracy, the Cramr-Rao lower bound
(CRLB) is used as the threshold for estimation performance [30].
Consider a received signal of the form,
s(t) =Aexp[ j(Kt
2
+ f
0
t)] +w(t), T
d
/2 <t < T
d
/2 (4.8)
where T
d
is the signal duration, A is the amplitude, K is the FM rate, f
0
is the initial
frequency and w(t) is zero mean complex additive white Gaussian noise (AWGN). The
mean squared error (MSE) of the FM rate estimation is given by [17],
MSE =
1
B
B

l=b
(K
b
K)
2
(4.9)
where B is the number of Monte Carlo simulation trials, K
b
is the estimated FM rate at
the bth simulation and K is the true FM rate. The CRLB for an unbiased FM rate MSE
estimator is given by [17],

2
K
=
90f
4
s
2
2
N(N
2
1)(N
2
4)SNR
(4.10)
where f
s
is the sampling frequency, N is the number of samples and the SNR is A
2
/
2
,
where
2
is the variance of the real and imaginary parts of the complex noise (note, the
real and imaginary parts have equal variances). As indicated in [17], the matched l-
ter (MF) detector and generalized likelihood ratio test (GLRT) detector provide better
estimation performance, especially in low SNR environments. For our simulations, we
31
compare the FrACF based detector with the GLRT detector [31].
4.2.1 GLRT Detector
The GLRT statistic for the received signal in Equation (4.8) rst computes the max-
imum likelihood estimate (MLE) of the unknown FM rate K. As the MLE cannot be
computed in closed form [31], it is given by
K
MF
= arg max
K
C
_
[S( f )
_
e
j(Kt
2
+f
0
t)
e
j2t f
dt[e
j2t f
d f . (4.11)
The typical denition of a matched lter is a correlator that will maximize the cor-
relation SNR. For the statistic provided in (4.11), instead of maximizing the correlation
SNR, we will maximize the correlation contrast, where contrast is dened as the ratio of
the standard deviation over the mean of the pixel magnitudes [9]. For SAR processing,
contrast is a common image quality metric normally calculated over a spatial region in an
image with pixels contained in a 2-D array encompassing the spatial region. For the test
statistic provided in (4.11), contrast will be calculated using the 1-D signal S(f).
C =
E([I[
2
)
[E([I[)]
2
(4.12)
where I is the pixel magnitude and E() is statistical expectation.
For the maximization procedure in (4.11) to estimate the FM rate, we will use the
golden search algorithm to reduce the computational burden of sweeping from the lower
to upper bounds [32]. The golden search algorithm will always nd a minimum or maxi-
mum value, however it can converge at local extrema. To help reduce the number of iter-
ations and avoid converging on local extrema, it helps to constrain bounds of the search
parameters to known limits. This can be done by rst using a course resolution search,
choosing its maximum value, and then using it as input to the golden search. Note that
we also use the same search algorithm for the FrACF optimization.
32
4.2.2 Fractional Autocorrelation Function Detector
For the LFM signal in (4.8), if both the FM rate K and the initial frequency f
0
are
unknown, then the optimal GLRT detector can be obtained by integrating all possible
lines in the Wigner distribution time-frequency plane [33]. If f
0
is not a parameter of
interest or is known, then a different detector was proposed in [16], that is based on the
AF of the received signal. Specically, the detector uses the fact that all LFM chirp
signals have a linear support region that crosses the origin of the AF, thus minimizing the
region of search for the estimated FM rate. The detector then becomes the integration
of the magnitude of the AF along lines that only go through the origin in the AF plane.
Specically, the estimated FM rate is,
K
AF
= arg max
K
_
[AF
s
(t, Kt)[dt (4.13)
which is the maximization of all line integrals of the AF along lines whose slopes corre-
spond to any possible values of the FM rate K.
In order to simplify the computation of the 2-D search in the AF plane in (4.13), the
relation between the AF and FrACF in Equation (4.1) is used. Specically, from (4.13),
the line in the AF plane is given by = Kt and has slope K. If we assume that this radial
line makes an angle going through the origin, then a different equation can be written
for the integral of the AF in (4.13), that makes this angle = tan
1
(K). This integral is
given by,
_ _
[AF
s
(t, )[( cos t sin)dt d (4.14)
and it can be further simplied to,
_
[AF
s
(t cos, t sin)[dt. (4.15)
33
Using (4.4), the integral can also be written in terms of the FrACF of the signal s(t) as,
L(K) =
_
[FA
tan
1
(K)
()[d. (4.16)
As a result, the detector will maximize this FrACF equation, over all possible values of
K,
K
FrACF
=arg max
K
L(K) = arg max
K
_
[FA
tan
1
(K)
()[d. (4.17)
Figure 4.1: Plot of FrACF detector statistic L(K) in (4.17) for different = tan
1
(K),
where =/2. The true LFM rate is K
a
= 1808 and the estimated one is K
/
a
= 1807.9.
Figure 4.1 is a plot of the FrACF estimate provided by (4.17) with the lower and upper
FRFT order constrained between 1.125 and 1.143.
Since we are interested in FM rate estimation, we will be using (4.8) with A = 1
and f
0
= 0. The range matched lter parameters are provided in Table 4.1 and azimuth
matched lter in Table 4.2.
34
Table 4.1: Range matched lter parameters
Parameter Symbol Value Units
Sampling rate f
s
32.317 MHz
Range FM rate K
r
721.35 GHz/s
Pulse duration T
r
41.74 s
Number of range samples N
r
= f
s
T
r
1349 samples
Table 4.2: Azimuth matched lter parameters
Parameter Symbol Value Units
Approximate FM rate K
a
1800 Hz/s
Pulse repetition frequency (PRF) F
a
1256.98 Hz
Pulse duration T
a
0.56 s
Number of range samples N
a
= F
a
T
a
705 samples
The relationship between the FRFT rotational parameter and the FM rate of a dis-
cretely sampled signal is given in [34]. For an LFM chirp e
j2(zt
2
+bt+c)
, with instanta-
neous frequency 2zt +b, the optimal FRFT order parameter is,

opt
=
2

opt
=
2

tan
1
_
f
/t
2z
_
(4.18)
where (a) f = f
s
/N is the frequency resolution; (b) t is the time resolution; (c) f
s
is the
sampling frequency; and (d) N is the number of samples. Thus we can re-write (4.18) as

opt
=
2

opt
=
2

tan
1
_
f
2
s /N
2z
_
(4.19)
The FM rate statistic provided by (4.17) and the GLRT statistic provided by (4.11), were
generated using 200 Monte-Carlo simulations. For the comparison results shown next,
the FM rate estimation MSE is plotted in dB using 10log
10
_
1
MSE
_
with signal SNR
spanning -5 to 20 dB.
35
4.2.3 RADARSAT LFM detection accuracy
In this section, simulations are performed with RADARSAT-1 LFM parameters pro-
vided in tables 4.4 and 4.2 to characterize the FrACF detector performance in varying
SNR. We use the results of this chapter to require a minimum SNR for LFM detection
using real RADARSAT-1 data in Chapter 6. Figure 4.2 is a plot of the simulations and
it indicates that for the RADARSAT-1 range parameters, the FrACF has good estimation
performance for SNRs above approximately 7 dB. The performance of the GLRT is better
for low SNR environments, as expected from [17]. To use the FRFT for autocorrelation
or other applications requires the signal to be compact in the fractional domain, mean-
ing all energy must be contained within a constant radius in the Wigner distribution [25].
To assure compactness, s(t) is over-sampled, increasing the number of samples N, from
1349 to 4096, which corresponds to a good size for FFT. The CRLB is provided using the
original number of samples which could explain the slightly improved MSE performance
for the FrACF method. The standard deviation of the FM rate estimate is provided in
Figure 4.3 and is in agreement with the MSE plot. This parameter is necessary for the
expected FM rate mismatch when using the estimated rate in SAR focusing applications.
36
Figure 4.2: Range FM rate estimation (N = 1349, K
r
= 721.4 GHz/s (see Table 4.1)).
Figure 4.3: Range standard deviation FM rate estimation.
37
Figure 4.4: Range bin FM rate estimation (N = 705, K
a
= 1808 Hz/s (see Table 4.2)).
Figure 4.5: Azimuth standard deviation azimuth FM rate estimation.
Figure 4.4 indicates that the performance for the azimuth estimation is similar to that
of the range: an abrupt change in estimation error for SNRs above 5 dB and, as expected,
a better estimation of performance for the GLRT in low SNR environments.
38
Figure 4.6 is a plot of the receiver operating characteristics (ROC) for the azimuth FM
rate or the rate of each range bin. The same detection statistic as provided in (4.17) was
used, with the exception that the maximum value of the FrACF estimate was used instead
of the sweep rate as depicted in Figure 4.1. To generate the ROC curves, 250 simulations
were performed for different SNRs and the curves were produced using [31],
P
D
=Q
_
Q
1
(P
FA
)
_
NA
2
/
2
_
(4.20)
where A = 1, N = 256 and
2
was estimated from the simulation runs for each SNR.
Figure 4.6: Azimuth FrACF ROC.
4.2.4 RADARSAT FRFT FM Rate Detection Resolution
For completeness, we wanted to include this section on the FM rate resolution ca-
pability of the FRFT, with resolution dened as the ability to distinguish between two
equally strong LFM signals in a noise-free environment [33]. We will not be address-
ing multi-component LFM signals any further but wanted to include this because of its
dependence on LFM accuracy and its use in SAR signal processing, specically in the
areas of moving target detection and inverse synthetic aperture radar (ISAR) [35, 36]. For
39
conventional SAR, the radar antenna moves in relation to the target, whereas for ISAR,
the antenna is stationary and the target moves to form the synthetic aperture.
It is noted in [35] that the ability to separate two chirp signals with very small dif-
ference in chip rate has not been well dened, but simulations reveal that resolution is
inversely proportional to the square of the signal length T
d
. A recent work [37] provides
an algorithm which we will use to estimate the minimum resolution but also cites the lack
of any comparative studies.
For multi-component signals of the form [37],
s(t) =
M

m=1
A
m
cos(2 f
0m
t +K
0m
t
2
) +w(t), T
d
/2 t T
d
/2 (4.21)
where M is the number of LFM components, A
m
is the amplitude of the mth LFM signal
and w(t) is zero-mean white Gaussian noise. The minimum resolution capability of the
FRFT is provided by,
K
min
=

2csc(cot
1
(
K
0
T
d
/f
s
))
T
2
d

(4.22)
where K
min
is the difference between the upper and lower lines of a parallelogram cre-
ated in the normalized TF plane by two LFM signals with the same center frequency f
0
.
The axis normalization is achieved by rst dening the scaling variable s =
_
t/f ,
then creating the normalized axis x = t/s and = f
s
s. The length of both axes is now
represented by the dimensionless value

tf depicted as x in Figure 4.7. Additional
details of TF axis normalization can be found in [23, 25].
40
Figure 4.7: Multi-component TF plane (normalized axis).
Figure 4.8 plots K
min
as provided by (4.22) for different up-sampling ratios.
Figure 4.8: Minimum detectable range resolution.
Figure 4.9 is a MATLAB simulation using the same range LFM values as in the previous
section to include the up-sampling of the signal to an FFT size of 4096. The multi-
component signal was created using (4.21) with M = 2, A
m
= 1 for all m, f
0m
= 0, for all
m and w(t) = 0 for all t, and for the FM rates, K
01
= K
r
and K
02
= K
r
+K
/
r
.
41
The FrACF plot in Figure 4.9 (left) shows good FMrate separability for K 3.2GHz/s
and the FrACF plot Figure 4.9 (right) shows the inability to separate FM rates for K
3.0GHz/s.
Figure 4.9: Multi-component detection, N = 4096.
42
CHAPTER 5
Fractional Focusing and the Chirp Scaling Algorithm
5.1 Fractional Chirp Scaling Algorithm (FrCSA)
The fractional Fourier transform (FRFT) has found widespread acceptance within the
synthetic aperture radar (SAR) signal processing community, mainly in operations of
linear frequency-modulated (LFM) chirp detection, high prole range imaging (inverse
synthetic aperture radar (ISAR)) [8, 38] as well as single channel SAR images for the de-
tection of moving targets [18]. Acommon thread to the mentioned algorithms is the detec-
tion of chirp type signals amongst random clutter. Most recently, fractional chirp scaling
algorithm (FrCSA) approaches have been investigated [2, 39] with improved signal-to-
noise ratio (SNR) and greater sidelobe reduction, when compared with previous imple-
mentations of the chirp scaling algorithm (CSA). Typical sidelobe mitigation strategies in
SAR image processing make use of weighting functions, however this method causes an
unwanted boarding of the impulse response width (IRW) and degradation of SNR. If the
sidelobes can be reduced without the use of weighting functions, then the IRW and SNR
degradations may be avoided, yielding improved performance for the FrCSA.
The new FRFT based CSA uses the FRFT to enhance the overall focusing capabilities
of the CSA [2]. A ow diagram of the algorithm is provided in Figure 5.1. The algorithm
makes use of a local optimization procedure (LOP) [2] that investigates all possible FRFT
rotation angles and selects the optimum value to be used throughout the algorithm.
43
Figure 5.1: Block diagram of FrCSA algorithm. Here R denotes range, and Az denotes
azimuth. Figure taken from [2].
The most important steps in the FrFT is in range transformation processing and the frac-
tional correlator-based range match lter [40]. A module combining these two steps gives
the FrCSA its unique strength relative to the Fourier transform (FT) based CSA in han-
dling signals, particularly at the far end of the scene.
5.2 Fractional Focusing with Contrast Maximization
The intent of this research is to understand how the FRFT can be used in SAR signal
processing, specically as part of the CSA algorithm, to improve image formation. Using
the relationship of the FRFT with time-frequency (TF) representations [12], we began by
analyzing the Wigner distributions (WDs) of range-bin samples and the corresponding
matched lter to understand how the rotational parameter of the FRFT could tune the
transform to produce an optimal pulse compression response. Figure 5.2 shows plots of
44
the match lter (upper plot) and the range-bin signal (lower plot) for range-bin 969. As
expected, the signals are conjugates of each other and matched ltering should concen-
trate the energy to discrete frequencies with narrow bandwidth. This method will only
apply to stationary targets. Moving targets will not be localized in frequency but walk
across frequencies [7], maintaining a slope or ramp, thereby maintaining bandwidth in
the TF plane.
Figure 5.2: Wigner distribution representations of matched lter and range-bin signals.
The upper plot in Figure 5.2 is the WD of the pulse after matched ltering, using the
classical FT, with the fractional order set to 1, (F
1
r
x)( f ), and the lower plot is the result of
tuning the transform with (F

r
x)( f

) where =0.001386 and is related to the rotational

parameter = /2. We see that with optimal parameter selection, the energy or target
after match ltering is concentrated in frequency with narrowbandwidth for the optimized
center frequency. As the WD indicates, the resulting signal is a frequency domain signal
45
and an inverse FT will rotate it by /2 concentrating the target in time.
Figure 5.3: (WD) or pulse after matched ltering: using classical FT (upper plot) and
optimized with FRFT optimization (lower plot).
Figure 5.4 is a comparison of the target response for range-bin 969, comparing the FT
and FRFT methods. We see the FRFT response has an improved peak-to-sidelobe ratio
(PSLR) with better concentration of the target energy. To obtain the optimized value =
0.001386, we computed the MLE in (4.11) with contrast maximization as the selection
criteria.
46
Figure 5.4: Comparison of target response for range-bin 969 using classical FT (F
1
r
)
(blue-line) and optimized FRFT (F

r
) (red-line).
Recent works have indicated that the FRFT can be used to correct mismatches between
the expected and the received signals FM rate [2, 8, 18]. This mismatch can be explained
as a nonlinear motion between the platform and target. To understand the optimization
capabilities of the FRFT with mismatches in the expected FM rate, we conducted a sweep
of the azimuth FM rate K
a
while maximizing the contrast using the FRFT. The same
contrast maximization method was provided in (4.11), with the exception that the FRFT
rotational parameter is also swept between
min
and
max
creating a 2-D search.
47
Figure 5.5: MATLAB mesh plot of 2-D contrast maximization search; sweeping K
a
and
.
Figure 5.6: MATLAB scaled image of 2-D search, highlighting maximum contrast at:
Ka = 1804.5 and = 1.0002.
Figures 5.5 and 5.6 indicate that the FRFT was able to maximize the contrast for each
FM rate, however the absolute maximum occurs with a K
a
mismatch that is less then 0.01
48
percent of the optimal FM rate. We also notice that the absolute maximum occurred for
the rotational value = 1.0002 indicating that for this signal, the FRFT enhanced the
PSLR beyond that of the MLE search methods.
5.3 Fractional Focusing and the CSA
To apply the fractional focusing method of the previous section, it is necessary to
modify the CSA in Figure 2.2 so FRFT optimizations can be implemented. There are
two common approaches for FRFT optimization [2, 34]. The rst approach chooses a
global FRFT order and uses it for all pulses or range-bins processed with the FRFT.
The second approach evaluates each pulse or range-bin and tunes the FRFT order for the
optimal response. This is referred to as the local optimization procedure (LOP) [2, 34].
While the rst global optimization approach is easier to implement and provides improved
results, the nely tuned LOP provides the best performance [34]. Figure 5.7 shows the
FRFT modied CSA which includes a pre-azimuth FRFT lter to correct for mismatches
between the received pulses and the range FM rate K
r
, and a modication of the azimuth
match lter to maximize contrast for each range-bin using the FRFT to optimize the
azimuth FM rate K
a
. Note that, for the LOP approach, the FRFT optimization values need
to be obtained using some approach; the approach we adopt in this work is discussed in
the next section. Note also that the forward and inverse FRFTs are not optimized. The
FRFT order is set to = 1 so the FRFT is the classical FT. This was done to provide easy
modication of the new CSA procedure and allow for direct comparison to the CSA in
Figure 2.2 without transform optimization.
49
Figure 5.7: Block diagram providing an overview of fractional focusing with CSA.
Figure 5.7 is a block diagram overview of the fractional CSA. The Rg-FrFT indicates
that the FRFT is applied in the range direction, and the Az-FrFT indicates that the FRFT is
applied in the azimuth direction. The IFrFTs indicate the inverse transform provided with
the inverse FRFT property (F

r
)
1
=F
()
r
. Secondary range compression and range cell
migration correction (RCMC) are applied in the range matched ltering step. Differential
RCMC is applied in the chirp scaling step.
The fractional CSA begins with a 2-D array of demodulated raw (unprocessed) phase-
history data s
r
(t, ) as given in Equation (2.2). The phase-history data is in the range-time
and azimuth-time domains, respectively indicated by t and .
The fractional focusing CSA steps [1, 2] are summarized as follows
1
,
1. The fractional lter is applied to each received pulse,
s
r
(t

, ) = (F

r
s
r
)(t

, ). Since this is only applied to the fast-time pulses, the

50
range-time variable t is updated to t

, indicating that the TF plane of the signal is

rotated by .
2. An azimuth FRFT is performed to transform the data into the range Doppler do-
main,
S
rd
(t

, f

) = (F
1
r
s
r
)(t

, ). With the FRFT order set to 1, the data is transformed

to the range-time, azimuth-frequency domain, also referred to as the range Doppler
domain; this is indicated by the change of to f

. Since this is equivalent to

applying the classical FT, the azimuth frequency index is not affected.
3. Chirp scaling is applied, using a phase multiply to equalize the range migration of
all targets, S
1
(t

, f

) = s
sc
(t, f

)S
rd
(t

, f

).
4. A range FRFT is computed to transform the data to the 2-D frequency domain,
S
2
( f
t

, f

) = (F
1
r
S
1
)(t

, f

). The range-time variable t

is updated to indicate
range-frequency f
t

with no additional fractional rotations.

5. A phase multiply is performed with a reference function, which applies range com-
pression, SRC, and bulk RCMC in the same operation,
S
3
( f
t

, f

) = S
MF
Range
( f
t
, f

)S
2
( f
t

, f

).
6. A range IFRFT is performed to transform the data back to the range Doppler do-
main, S
4
(t

, f

) = (F
1
r
S
3
)( f
t

, f

). The range Doppler is indicated with t

, f

.
7. A phase multiply is performed to apply azimuth compression with a time varying
match lter. A phase correction is also required as a result of the chirp scaling in
step 2, which is incorporated into the same phase multiply,
S
5
(t

, f

) = (F

r
S
MF
Azimuth
)(t

, f

)S
4
(t

, f

). Contrast maximization is achieved

by applying F

r
in the azimuth direction. The new azimuth-frequency variable is
updated to f

, indicating that the TF plane of the azimuth signal is rotated by .

8. The nal azimuth IFRFT is performed to transform the compressed data to the SAR
image domain, S
6
(t

) = (F
1
r
S
5
)(t

, f

). The data is now back in the range-

time, azimuth-time domains, with FRFT optimizations indicated by the updated
variables t

. Also, since this is the last step, the data is compressed in both
range and azimuth. This ensures that the data is suitable for additional processing
or detected and scaled for image display.
1
The subscripts , on the index variables, indicate that the TF plane for the signal data is in the
fractional domain or the data along that axis is rotated by a fractional operator ,= n, n N, or =
(/2)n, n N for = /2 or = /2.
51
CHAPTER 6
Application of Enhanced FRFT Focusing Technique Using RADARSAT Data
In Chapter 5, we presented the enhanced fractional Fourier transform (FRFT) tech-
nique based on enhancements of the chirp scaling algorithm (CSA). This fractional fo-
cusing method is a rst attempt at implementing the fractional CSA (FrCSA) provided in
[2]. It includes modications that were necessitated by the use of the fractional autocorre-
lation function (FrACF) to obtain locally optimized FRFT rotational parameters from real
SAR images. In this chapter, we apply the enhanced FRFT focusing (EFrF) technique to
SAR data obtained from RADARSAT scenes.
We will obtain performance comparisons between the standard CSA and the EFrF
implementation by setting the FRFT order = 1 for the original CSA steps requir-
ing fast Fourier transforms (FFTs). A challenge that has not been addressed with this
implementation is the generation of the optimal FRFT transform parameters. We will
present one approach; however, given the additivity of the rotations property of the FRFT
F
1+2
r
= F
1
r
F
2
r
, and the time-frequency plane, many optimized fractional CSA imple-
mentations are possible by combining standard SAR signal processing techniques with
the FRFT to obtain optimal transform parameters.
We will rst describe the test environment, followed by the generation of the opti-
mized transform parameters for a RADARSAT-1 [41, 42] scene, then provide results of
our fractional CSA implementation.
6.1 Processing Architecture
All processing in this thesis was accomplished on a Sun Ultra24 Workstation (Quad
x86) with the following conguration: (a) Operating System: SunOS 5.11; (b) Compiler:
52
Sun Studio 12; (c) Programming Language: FORTRAN90; (d) Parallel Processing: Open
Multi-Processing (OpenMP); (e) All FFTs: Sun Performance Library (ZFFTZ); and (f)
Polynomial regression: Netlib (DGETRF, DGETRI)
6.1.1 FRFT Algorithm Implementation
To enable FRFT processing on the above architecture, it was necessary to port the
MATLAB fracF routine in Section 3.3.1 to FORTRAN 90. It could have also been ported
to C or C++ but we wanted to work with FORTRAN since all the high performance algo-
rithm libraries are written in FORTRAN. Sufcient testing was conducted to verify that
the ported FORTRAN module provided the same results as the MATLAB fracF routine.
Testing consisted of comparing both the real and imagery parts of a rectangular function
processed at FRFT orders between 0.5 and 1.0.
The fracF is considered O(Nlog
2
N) where N is the number of data samples, however
its implementation compared to the FFT was noticeably slower. In addition to the two
chirp multiplications and the frequency domain multiplication to accomplish the convo-
lution, the fracF algorithm required two FFTs of length 2N and three FFTs of length 16N.
Since all of the lter and data arrays are N = 4096 or radix-2, an approximation of the
complexity is 4Nlog
2
2N +48Nlog
2
16N which is approximately 68 times higher than
Nlog
2
N for the standard FFT.
A comparison of performance on the test architecture was performed using a one di-
mensional (1-D) FFT on the 4096x4096 data array. The test was accomplish using a
single thread that the standard FFT completed in 0.97 seconds and the fracF procedure
completed in 88.38 seconds, which is still considerably faster than an implementation of
complexity O(N
2
).
53
6.2 RADARSAT-1 Scene Phase History Data
The RADARSAT-1 phase-history data was included with the book digital processing
of Synthetic Aperture Radar Data [1] and made available by Gordon Staples of Radarsat
International. The Canadian Space Agency holds the copyright, and it is provided on
condition that it only be used for educational purposes.
The data is provided in CEOS format, and MATLAB routines are made available for
extracting the CEOS format data from the CD. The MATLAB routines also make neces-
sary corrections to account for uctuations in the gain or automatic gain control (AGC)
during the collect.
Some of the RADARSAT-1 scene parameters are listed below. To account for the
minimum array size needed for convolution, good FFT sizes and minimum number of
samples to meet compact signal support requirements, all arrays with be over-sampled
(zero padded) to a length 4096. The raw phase-history samples used as input to the
CSA were 2048x2048, or 2048 pulses by 2048 range-bins. Due to interpolation methods
used in the fracF routine, all phase-history data and matched lters were symmetrically
zero padded keeping the original data in the center of the up-sampled arrays. Table 6.1
contains some RADARSAT-1 non-changing parameters that are used in the matched lter
and chirp scaling operations.
54
Table 6.1: RADARSAT-1 Fine-beam:2 parameters
Parameter Symbol Value Units
Sampling rate f
s
32.317 MHz
Pulse bandwidth 30.11 MHz
Range FM rate K
r
721.35 GHz/s
Pulse duration T
r
41.74 s
Number of range samples N
r
1349 samples
Center frequency f
0
5.30 GHz
Wavelength 0.05657 m
Pulse repetition frequency PRF 1256.98 Hz
The actual Vancouver scene is a small subset of an overall larger scene of phase-
history data provided in [1]. Table 6.2 contains a list of parameters that change throughout
the scene based on either range or time.
Table 6.2: RADARSAT-1 Vancouver scene parameters
Parameter Symbol Value Units
Slant range (scene center) R
0
974804 m
Effective radar velocity V
r
7052.2 m/s
Doppler centroid f

c
-8190 Hz
Azimuth FM rate K
a
1795 Hz/s
Number of azimuth samples N
a
705 samples
55
6.3 FM Rate Estimation Using Fractional Autocorrelation and Contrast Maximization
In Section 4, we characterized the performance of the fractional autocorrelation func-
tion (FrACF) detector against the Cramr-Rao lower bound (CRLB) in varying signal-
to-noise ratio (SNR) environments. To provide a measure of FrACF performance using
the received RADARSAT-1 data, we will need to provide an estimate of the received and
processed SNR. The SNR of a SAR signal is governed by the radar range equation and,
for an uncompressed pulse it is given by [1, 6, 43, 3],
SNR
image
=
P
av
G
2

(4)
3
R
3
kT
o
BLV
a
(6.1)
where P
av
is the average power, G is the antenna gain, is the target radar cross section
(RCS), R is the slant range, k is Boltzmanns constant, T
o
is 290 degrees Kelvin, B is the
receiver bandwidth, L is radar losses,
a
is the azimuth resolution and V is the platform
velocity. If clutter is being observed, then =
0

r
/cos, providing the clutter-to-
noise ratio (CNR) as
CNR =
P
av
G
2

r
2(4)
3
R
3
k
kT
o
BLV cos
(6.2)
where
0
is the backscatter coefcient, is the grazing angle and
r
is the range reso-
lution. Since we are working with previously collected data and the randomness of
0
in (6.2), it will be necessary to obtain an CNR estimate based on SAR imagery (speckle)
noise components.
The total noise in SAR imagery
N
, is comprised of two components, additive noise

n
which is the result of thermal noise and multiplicative which is the signal dependent
noise proportional to the average signal level
0
[6],

N
=
n
+(MNR)
0
(6.3)
56
where MNR is the multiplicative noise ratio, and the relationship of
n
to CNR is

n
=
0
/CNR. (6.4)
The principle contributors of MNR are the integrated sidelobe ratio (ISLR) of points in
the surrounding area, ambiguities in the image, quantization noise and image artifacts or
MNR = ISLR + QNR + AMBR, where QNR is the quantization ratio and AMBR is the
ambiguity ratio [6]. Figure 6.1 is a plot of image clutter for [MNR[ <[CNR[ showing the
effect of MNR on the ISLR of points used to produce the image clutter. As noted, MNR
is the ratio of the image no-return area (NRA), divided by the average image clutter [3].
Note that the total noise in the NRA is
N
NRA
=
n
+(QNR+AMBR)
0
.
Figure 6.1: Plot of image clutter for [MNR[ < [CNR[, showing the relationship of total
noise
N
with the total noise in the NRA,
N
NRA
. A similar graphic is provided in [3].
Since the LFM detection simulations assumed an additive noise model, we will use the
ratio of the total noise in the NRA with the total of the image clutter for CNR estimates
used to gauge detection performance. One reason for this approach is that,
N
NRA
can be
estimated directly form the received or processed data. Since this involves MNR, we must
57
address limitations of time-frequency (TF) LFM chirp rate detection methods for signals
with MNR > 1. Results of simulations in [44] indicate the instantaneous frequency (IF)
of signals corrupted with both multiplicative and complex additive white Gaussian noise
(AWGN), is that the IF estimation can be accomplished with time frequency methods
provided the standard deviation of the multiplicative noise is less than its mean. It is
noted in [45] that the MNR is the ratio of the standard deviation over the mean and that
the theoretical MNR of a single look amplitude SAR image is 0.5227. Since this is less
than 1.0, time frequency methods such as the FrACF can be used to estimate the IF or
LFM chirp rates.
The method that will be used to provide an estimate of the noise will be to identify re-
gions in the collected or processed imagery where there is little or no back-scatter energy
such as radar shadows or large areas of calm water [6, 3, 45]. Since the multiplicative
noise component is proportional to the average signal level
0
, the CNR will be estimated
as the ratio of the NRA to total image clutter. Figure 6.2 is the magnitude of received
phase-history data from RADARSAT-1 suitable for processing into a SAR map (image).
For this data, range is increasing from left to right and the satellite was moving from
top to bottom, transmitting and receiving pulses at the pulse repetition frequency (PRF).
Each received pulse is sampled at 32.317 MHz providing a 2-D array of phase-history
data. The received baseband signal data is given by (2.2). To provide a comparison of
detector performance in varying CNR regions, we will obtain the CNR in dB using,
CNR
dB
=10log
10
_
P
signal
P
noise
_
(6.5)
where P
noise
will be chosen from homogeneous regions in both the phase history data sets
depicted in Figure 6.2 and Figure 6.5 [6, 3, 45]. Since the homogeneous regions and
received pulses will not be the same length, (6.5) can be redened as,
CNR
dB
=10log
10
_
E[s
N
NRA
[
2
E[
N
NRA
[
2
_
(6.6)
58
where s is a 1-D array containing the pulse or range-bin samples and E() is statistical
expectation.
The detection was done with FrACF and the performance is provided in Figure 6.3
which is in agreement with the results of the simulations in that, the performance of
FrACF will depend on having good SNR. Both detection and CNR calculations can be
noisy so a smoothing lter was used when necessary to help identify trends in the plotted
data. When a smoothing lter was used, this was indicated on the plot as win(n) where n
is the length of the lter used in the MATLAB convolution (conv) routine.
Figure 6.2: RADARSAT-1 phase-history data (Vancouver Airport BC).
59
Figure 6.3: Range FM rate detection using FrACF.
Figure 6.4: Range SNR for RADARSAT-1 scene (Vancouver Airport BC).
Figure 6.5 is an image of the phase-history data in the range Doppler domain, after
range pulse compression (matched ltering) and range cell migration correction (RCMC)
and the data in this domain is provided by Equation ( 2.20).
60
Figure 6.5: RADARSAT-1 range Doppler (Vancouver Airport BC).
We see in Figure 6.6 that the azimuth LFM rate detector performed well mainly due to
good CNR in the range Doppler domain after range compression and range cell migration
correction.
61
Figure 6.6: Azimuth FM rate detection using the FrACF.
Figure 6.7: Azimuth CNR for RADARSAT-1 scene.
Comparing the range and azimuth FM rate detection performance in Figure 6.3 and Fig-
ure 6.6, we can see that the azimuth FM rate detection performed better in that there is
less noise and an overall trend noticeable in Figure 6.6, whereas in Figure 6.3 we no-
tice possible good LFM detection in the region where the CNR peaked between pulse
62
numbers 400 - 1000. The better estimation performance of the azimuth detector can be
attributed to performing the detection on the data in the range Doppler domain after range
compression. Since it is after range compression, the CNR is improved as indicated in
Figure 6.7. It should also be noted, that clear SAR imagery is dependent on having CNR
between 5 dB and 15 dB [3] and this threshold is evident in Figure 6.7.
Figure 6.8: Azimuth FRFT optimized parameters using contrast maximization.
Figure 6.8 is a plot of the optimized FRFT rotational parameters using contrast maximiza-
tion for the detection statistic in (4.11). Since this statistic is captured during the azimuth
matched ltering stage, the target energy is fully compressed. For this implementation,
the algorithm maximizes the contrast of the strongest target within each range bin. It is
also possible to force the maximization for specic spatial regions by bounding the re-
gions for which contrast is maximized. Both MATLAB and FORTRAN provided intrinsic
routines for selecting maximum values and their locations in multi-dimensional arrays.
To obtain a better estimate for the range FM rate, we performed the detection after
azimuth compression when the CNR is improved and, as expected, this did provide a
better noise estimate. Figure 6.9 is a plot of the CNR for each pulse using Equation (6.6)
63
on data that is compressed in both range an azimuth, in other words, a single look complex
(SLC) image. As shown in Figure 6.10, the CNR is well above the needed threshold of
5 dB as indicated in Figure 4.2. Figure 6.10 is a plot of the FRFT optimization values
obtained using FrACF.
Figure 6.9: Clutter-to-noise (CNR) ratio for each pulse from a single look complex (SLC)
image (compressed in both range and azimuth).
64
Figure 6.10: Pulse FRFT optimization values obtained using the FrACF.
Using the noisy estimates to correct mismatches in the range FM rate caused severe
streaking or high noise levels in the azimuth direction of the SLC image. Since we are
using the FRFT to correct for nonlinear motion between the satellite and target, it is not
possible for the uncompensated motion to change this rapidly for non-moving targets, so
the range FM estimates were used as input to a polynomial tting algorithm. The tted
polynomial will be used for the pulse FRFT optimization values and a plot of the values
is provided in Figure 6.8. Note that FRFT optimization values are the difference between
the matched lter and received data LFM chirp rates, and are corrections to the azimuth
dependent range FM rates needed to concentrate the frequency in the TF plane as demon-
strated in Figure 5.3. These corrections are provided as updates to the original FM rate
using the FRFT operator additivity property.
65
Figure 6.11: Pulse FRFT optimization values using FrACF; for =7e
5
, K
r
=2.8e
8
using Equation (4.19).
6.4 Fractional Auto Focus for Chirp Scaling Algorithm
The fractional focusing CSAwas implemented using the steps depicted in the overview
Figure 5.7.
6.4.1 Fast-time FRFT chirp optimization
The rst FRFT optimization was applied to each pulse using the values in Figure 6.11,
s
r
(t

, ) = (F

r
s
r
)(t

, ). This is done by applying a 1-D FRFT to each row of s

r
(t

, )
with
i
= FRFT
order
(i).
6.4.2 Azimuth FRFT
A azimuth FRFT is performed by applying a 1-D FRFT to each column or range-bin
with
i
= 1 for all range-bins, S
rd
(t

, f

) = (F
1
r
s
r
)(t

, ). Notice the frequency index f

66
indicating the data is now in the range-time azimuth-frequency domain.
6.4.3 Chirp scaling
Chirp scaling is applied using the f

dependent phase multiply,

S
1
(t

, f

) = s
sc
(t, f

)S
rd
(t

, f

). This is accomplished with a 1-D multiply applied to

each row of the 2-D data array. The phase multiply is provided with (2.16).
6.4.4 Range FRFT
A range FRFT is performed by applying a 1-D FRFT to each row or pulse with
i
= 1
for all pulses, S
2
( f
t

, f

) = (F
1
r
S
1
)(t

, f

). The phase-history data is now in the 2-D

frequency domain. The frequency indexes f
t

and f

indicate the data is now in the

range-frequency, azimuth-frequency domain.
6.4.5 Range matched lter
An f

dependent range matched lter is applied to remove the second and forth ex-
ponential terms of (2.18), S
3
( f
t

, f

) = S
MF
R
S
2
( f
t

, f

). The phase of matched lter is,

RangeMF
=
D( f

,V
r
)
K
m
D( f

re f
,V
r
)
f
2
t
+
4
c
_
1
D( f

,V
r
re f
)

1
D( f

re f
,V
r
re f
)
_
R
re f
f
t
(6.7)
A 1-D range matched lter if applied by multiplying each row of the S
2
( f
t

, f

) with
F
1
r
_
exp( j
RangeMF
)
_
.
Figure 6.12 is a plot of the phase and matched lter for pulse 2048.
67
Figure 6.12: Range matched lter phase and magnitude for pulse 2048.
6.4.6 Range IFRFT
A range IFRFT is performed by applying a 1-D FRFT to each row or pulse with

i
=1.0 for all pulses, S
4
(t

, f

) = (F
1
r
S
3
)( f
t

, f

). The phase-history data is now in

the range Doppler domain.
6.4.7 Azimuth matched lter
At

dependent azimuth matched lter is applied to remove the rst term of (2.18) and
(2.25), S
5
(t

, f

) = (F

r
S
MF
Az
S
4
)(t

, f

). The phase of matched lter is,

AzimuthMF
=
4R
0
f
0
D( f

,V
r
)
c
+
4
c
2
K
m
g
1
K
m
+g
1
_
R
0
D( f

,V
r
)

R
re f
D( f

,V
r
re f
)
+
cg
0
4g
1
_
2
+
g
2
0
4g
1
(6.8)
With the application of the matched lter, the FRFT optimization is applied to maximize
contrast using the optimization values displayed in Figure 6.8. The FRFT optimized 1-D
68
matched lter is applied to each column of S
4
(t

, f

) using, F

i
r
exp( j
AzimuthMF
). The
Netlib algorithms DGETRF and DGETRI were used as a polynomial tting routine to
produce the coefcients g
0
and g
1
.
6.4.8 Azimuth IFRFT
A azimuth IFRFT is performed by applying a 1-D FRFT to each column or range-bin
with
i
= 1.0 for all range-bins, S
6
(t

) = (F
1
r
S
5
)(t

, f

). The data is now com-

pressed in both range and azimuth with concentrated target energy. This product is often
referred to as a single look complex (SLC) image.
6.4.9 Results
The result of the CSA or fractional focusing CSA is an single look complex (SLC)
image. Each pixel of this image has an I and Q or real and imaginary component and
often, the dynamic range of the magnitude y(m, n) =
_
I(m, n)
2
+Q(m, n)
2
exceeds the
display capabilities of common display systems. There are many algorithms and theories
for remapping the pixels into an acceptable display range, many of which make use of log
functions and image statistics. The remap of the SLC image included was accomplished
using a nonlinear contrast stretch routine that remapped all pixel values between the min-
imum and maximum magnitude
_
_
I
2
+Q
2
_
values to an 8-bit display range of 0 to 255.
Also, it is customary to project the slant plane SLC image into the ground plane before
remapping. This slant-to-ground projection was not performed, hence the image is in the
slant plane.
An overview image (1/8 resolution) is provided in Figure 6.13, where azimuth-time
is increasing from the bottom to top and range increases from let to right.
69
Figure 6.13: RADARSAT-1 scene (Vancouver, BC).
Figure 6.14 is a full resolution chip from the center of the RADARSAT-1 scene used for
this evaluation. First to get an appreciation of what 8 meter resolution imagery looks
like, the runway 08L-26R marker was placed in the chip. According to the Vancouver
airport guide, that runway is 200 feet wide. Since we will be comparing performance
enhancements of the FRFT, no attempt was made to improve the overall response by
including weighting functions or multi-look methods in the processing (this is mentioned
because weighting functions such as the Taylor window are standard techniques used in
SAR signal processing for mitigating side-lobe levels.) There are two rectangles placed
70
around areas of bright return to indicate regions for calculating contrast. See Table 6.3 for
contrast values. EFrF is used to indicate fractional focusing in the CSA. The area Loc (1)
is the rectangle in the upper region of the chip (approximate row 2250) and Loc (2) is the
lower region (approximate row 2450). Also included in the table is a point target located
at row 2634, column 1967. This target is not located within the full resolution chip but is
noticeable in the 1/8 resolution scene.
Table 6.3: RADARSAT-1 Vancouver scene contrast comparisons
Location CSA FFrF
Loc (1) 3.797 4.013
Loc (2) 4.829 5.109
(2634,1967) 3.004 3.113
Figure 6.14: Full resolution chip of the RADARSAT-1 scene.
Figure 6.15 is a point target located at row 2424 column 2415. To produce the MAT-
LAB plots, the complex points are up-sampled by a factor of 16. The target on the left
is from CSA processing with original FT or FRFT order F
1
r
. The target on the right was
71
produced using the same algorithm and parameters except that the FRFT LOP parameters
or F

r
applied as indicated in Figure 5.7. We see a signicant improvement in the range
direction especially in the resolution; this is a result of correcting the FM rate mismatch.
In other words, the target is no longer walking across range bins but aligned in frequency
or time as depicted in Figure 5.3.
Figure 6.15: MATLAB point target analysis at target location (2424,2415). Target upsam-
pled by a factor on 16 and displayed with MATLAB routine imagesc.
Figure 6.16 provides a fast time comparison of point target at row 2424. For this target
it is easy to see an improved peak sidelobe ratio (PSLR) (approximately 1dB) along with
an overall improvement in the impulse response width (IRW) or resolution. It should be
noted, that it is customary to provide comparisons in dB, however we wanted to keep the
scaling between Figure 6.15 and Figure 6.16 comparable.
72
Figure 6.16: Point target fast-time comparison (azimuth pulse 2424).
73
CHAPTER 7
Conclusions and Future Work
7.1 LFM Detection Accuracy
From both the high resolution results presented in [16] and our simulation-error anal-
ysis results, we can conclude that fractional autocorrelation function (FrCSA) is an ef-
cient method for LFM detection with high performance in low SNR environments. For
RADARSAT-1 LFM chirp rate estimation, Figures 4.2 and 4.4 indicate that FrACF can be
used for estimation in SNR environments above 5 dB. One measure of SAR image quality
is quadratic phase error (QPE) which contributes to elevated sidelobe and broadening of
the impulse response width (IRW). A mismatch of FM rates between the matched lter
and returned signal will contribute to QPE thereby decreasing image quality. A rule of
thumb for high focusing [1] accuracy (less than 2 percent IRW boarding), is

K
K

1
TBP
(7.1)
where K is the FM rate, Kis the FM rate mismatch and TBP is the time bandwidth
product. This indicates that the range FM rate mismatch must be, K
r
0.57 GHz/s and
the standard deviation in Figure 4.3 indicates that this can be achieved in normal SNR
environments.
The detection performance for the RADARSAT-1 scene in Figures 6.3 and 6.6 are in
agreement with the MSE simulations indicating that the FrACF can be used for FM rate
detection in good SNR environments. One possible approach to enhance the detection
of the range FM rate could be to perform the detection after azimuth compression to en-
hance the SNR of each pulse. Depending on how this is implemented, it could result in
performing complete image formation processing twice.
74
7.2 Fractional Focusing and the CSA Implementation
While the concentration of this work has been the use of the FRFT to enhance the
focusing of the CSA, some of the obtained results require additional consideration.
7.2.1 Contrast Maximization
The results of testing contrast maximization with changing FM rate K
a
and changing
FRFT order indicate that the FRFT has the capability to enhance the contrast beyond
that of correcting the FM rate mismatch. This is depicted in Figure 5.6 and suggests that
matched ltering at optimal FRFT orders will increase the peak sidelobe ratio (PSLR)
compared to classical FT methods. The expected behavior was that the absolute maxi-
mum contrast would have occurred for an FRFT order = 1 when the chirp rate of the
matched lter matched that of the signal. Since the maximum value occurred for ,= 1,
additional testing is necessary to verify the possibility of an isolated result. We need to
test whether applying a matched lter and using the extra degree of freedom provided by
the FRFT [2] will allow the contrast to be maximized beyond that of FT matched ltering.
7.2.2 Enhanced Fractional Focusing (EFrF) using the CSA
The excellent results obtained and presented in Figures 6.15 and 6.16 suggest that the
initial focusing of the CSA was not optimal; however, the only difference between the
presented results was the value of the FRFT transform order. The CSA was produced
within the same program without the application or the range and azimuth FRFT updates,
indicating the optimal FRFT transform not only enhanced the PSLR but also improved
the resolution or IRW. Although the FRFT optimization was applied in both the range and
azimuth directions, the range or fast-time application provided an overall better enhance-
75
ment. Applying the FRFT before the rst azimuth FFT and getting signicant results,
suggests that motion compensation may not have been optimal for this scene.
7.3 Future Research
7.3.1 LFM detection using the Discrete Chirp-Fourier Transform
As noted in Chapter 6, the mapping of the continuous FRFT to a discrete implemen-
tation requires signicant up-sampling or interpolation, thus increasing the processing
time. Although its implementation is O(Nlog
2
N) where N is the number of samples, its
performance is considerably slower. A possible faster implementation for LFM detection
is the use of the discrete chirp-Fourier transform (DCFT) which uses the DFT to match
multiple chirps to multiple chirp components [46]. The denition of the DFT,
X(k) =
1

N
N1

n=0
x(n)W
nk
N
, 0 k N1 (7.2)
where W
N
= exp(
j2
/N), the DCFT is
X
c
(k, l) =
1

N
N1

n=0
x(n)W
ln
2
+kn
N
, 0 k, l N1 (7.3)
where k represents the constant frequencies and l the chirp rates. The author indicates
a relationship between the DCFT and FRFT exists in that the chirp rate parameter l is
related to the rotational parameter of the FRFT. Note, this method does not have a rela-
tionship with the ambiguity function similar to FrACF, meaning a 2-D search is necessary
to identify both k and l. It may be possible to minimize the 2-D search by constraining k
and l to their expected values and some initial testing needs to be done to understand the
performance.
7.3.2 Multi-component LFM Rate Resolution
While researching LFM detection using the FRFT, a question that was not answered
deals with the resolution of the sweep rate parameter used to create the detection statistic
76
in FrACF. The idea being, if we understood a bound on the LFM rate resolution, that
would provide a bound for the sweep rate resolution. The workaround for not having this
bound, was to understand the LFM rate accuracy needed to minimize QPE. We were able
to locate a recent work [37] where it is noted that there are no comparative works.
Based on the relationship of the FRFT with the WD, we tried locating a more gen-
eralized denition for chirp-rate resolution and time-frequency. As noted by Wang [35],
resolution dened as the ability to separate two chirp signals with very small difference
in chirp rate, has not been well dened.
Additional research is needed to identify additional work in the area of chirp rate res-
olution and the FRFT, or a more generalized denition for time-frequency.
7.3.3 LFM detection and instantaneous frequency
Since we used the signal SNR in our analysis, it should be noted that the multiplica-
tive noise ratio (MNR) is a common image quality metric used for speckle imagery. The
MNR is dened as the ratio of the image intensity in the no return area (NRA) divided by
the intensity in a bright surrounding area [3]. There is work [44] indicating that the in-
stantaneous frequency (IF) time-frequency estimation will be impacted by multiplicative
noise and not possible for MNR > 1.0.
Additional research is needed using the signal model in [44] in place of the model pro-
vided with Equation (4.8) to understand the LFM detection performance for SAR LFM
signals.
7.3.4 Matched ltering and time-bandwidth product
With the FRFT additivity of rotations property, it is possible to rotate the time-frequency
(TF) plane of a chirp, throughout a continuous range of fractional domains or angles using
77
the rotational parameter as depicted in Figure 3.3.
Figure 7.1: Chirp: time-bandwidth product for 1 1.
Figure 7.1 is a plot of a RADARSAT-1 chirp rotated throughout the range of the FRFT
using the additivity of rotations property and as expected, the time-bandwidth product
(TBP) is maximized for = /4, where = /2. Since the matched ltering pro-
cess will cancel the rotation by applying a conjugate chirp, a study should be done to
understand the results of applying the lter at the maximum TBP. As indicated in [42]
the matched lter compression gain is attributed to the TBP and by Equation (4.19), we
know that oversampling or zero padding the arrays used in pulse compression will cause a
clockwise rotation in the TF plane decreasing the TBP. By use of the FRFT rotational pa-
rameter, it may be possible to enhance the compression gain by rotating both the matched
lter and the signal to the maximum TBP.
78
7.3.5 Time-frequency methods for moving target and inverse synthetic aperture radar
(ISAR) processing
The direct relationship of the FRFT to linear canonical transforms makes the FRFT
an attractive transform for many TF applications. The focus of the work in this research
was to estimate uncompensated motion and enhance the CSA processing by correcting
the LFM chirp rates of stationary targets. If we were interested in processing the moving
target information, then there is recent work using the FRFT for this purpose with many
proposed methods [7, 8, 18, 36, 38]. As we demonstrated with the enhanced FRFT fo-
cusing (EFrF) technique using the CSA and RADARSAT-1 data, the implementation and
characteristics of the data determined how the FRFT was used. We suspect there will be
similar challenges in applying the FRFT to real data used for both moving target detection
and inverse synthetic aperture radar (ISAR) processing. If real data is available, this work
could be extended into these areas.
79
REFERENCES
[1] I. Cumming and F. Wong, digital processing of Synthetic Aperture Radar Data.
Norwood, MA: Artech House, 2005.
[2] A. S. Amein and J. J. Soraghan, A new chirp scaling algorithm based on the frac-
tional Fourier transform, IEEE Signal Processing Letters, vol. 12, pp. 705708,
October 2005.
[3] R. J. Sullivan, Microwave Radar: Imaging and Advanced Concepts. Norwood, MA:
Artech House, 2000.
[4] D. Munson and R. Visentin, A signal processing view of strip-mapping synthetic
aperture radar, IEEE Transactions on Acoustics, Speech, and Signal Processing,
vol. 37, pp. 21312147, December 1989.
[5] J. Curlander and R. McDonough, Synthetic Aperture Radar: Systems and Signal
Processing. New York, NY: Wiley, 1991.
[6] W. Carrara, R. Goodman, and R. M. Majewski, Spotlight Synthetic Aperture Radar:
Signal Processing Algorithms. Artech House, 1995.
[7] V. Chen and H. Ling, Time-Frequency Transforms for Radar Imaging and Signal
Analysis. Norwood, MA: Artech House, 2002.
[8] L. Du and G. Su, Adaptive inverse synthetic aperture radar imaging for nonuni-
formly moving targets, IEEE Geoscience and Remote Sensing Letters, vol. 2,
pp. 247249, July 2005.
[9] J. Kolman, Aperture weighting for maximumcontrast of SARimagery, 2008 IEEE
Radar Conference, pp. 16, May 2008.
[10] R. Raney, H. Runge, R. Balmer, I. Cumming, and F. H. Wong, Precision SAR pro-
cessing using chirp scaling, IEEE Transactions on Geoscience and Remote Sensing,
vol. 32, pp. 786799, July 1994.
[11] R. Namias, The fractional order Fourier transform and its application to quantum
mechanics, Journal of the Institute of Mathematics and Its Applications, vol. 25,
pp. 241265, April 1980.
80
[12] L. B. Almeida, The fractional Fourier transform and time-frequency representa-
tions, IEEE Transactions on Signal Processing, vol. 42, pp. 30843091, November
1994.
[13] O. Akay and G. F. Boudreaux-Bartels, Fractional convolution and correlation via
operator methods and application to detection of linear FM signals, IEEE Transac-
tions on Signal Processing, vol. 49, pp. 979993, May 2001.
[14] O. Akay and G. F. Boudreaux-Bartels, New fractional operators and their proper-
ties: A generalization of time and frequency operators, IEEE 2
nd
UK Symposium
on Applications of Time-Frequency and Time-Scale Methods, pp. 141144, August
1997.
[15] O. Akay and G. F. Boudreaux-Bartels, Unitary and Hermitian fractional operators
and their relation to the fractional Fourier transform, IEEE Signal Processing Let-
ters, vol. 5, pp. 312314, November 1998.
[16] O. Akay and G. Boudreaux-Bartels, Fractional autocorrelation and its application
to detection and estimation of linear FM signals, Proceedings of the IEEE-SP In-
ternational Symposium on Time-Frequency and Time-Scale Analysis, pp. 213216,
1998.
[17] O. Akay and E. Erzden, Employing fractional autocorrelation for fast detection
and sweep rate estimation of pulse compression radar waveforms, in Signal Pro-
cessing, vol. 89, pp. 24792489, Elsevier, December 2009.
[18] H.-B. Sun, Application of the fractional Fourier transform to moving target detec-
tion in airborne SAR, IEEE Transactions on Aerospace and Electronic Systems,
vol. 38, pp. 14161424, October 2002.
[19] A. Papoulis, Systems and Transforms with Applications in Optics. New York:
MaGraw-Hill, 1968.
[20] F. Hlawatsch and G. Boudreaux-Bartels, Linear and quadratic time-frequency sig-
nal representations, IEEE Signal Processing Magazine, vol. 9, pp. 2167, April
1992.
[21] R. Raney, A new and fundamental Fourier transform pair, International Geo-
science and Remote Sensing Symposium, pp. 106107, May 1992.
81
[22] A. Amein and J. Soraghan, Azimuth fractional transformation of the fractional
chirp scaling algorithm (FrCSA), IEEE Transactions on Geoscience and Remote
Sensing, vol. 44, pp. 28712879, October 2006.
[23] H. M. Ozaktas, Z. Zalevsky, and M. A. Kutay, The Fractional Fourier Transform:
With Applications in Optics and Signal Processing. Chicester, U.K.: Wiley, 2001.
[24] H. Ozaktas, N. Erkaya, and M. Kutay, Effect of fractional Fourier transformation
on time-frequency distributions belonging to the Cohen class, IEEE Signal Pro-
cessing Letters, vol. 3, pp. 4041, February 1996.
[25] H. M. Ozaktas, O. Arikan, M. A. Kutay, and G. Bozdagi, Digital computation of
the fractional Fourier transform, IEEE Transactions on Signal Processing, vol. 44,
pp. 21412150, September 1996.
[26] A. Bultheel and H. M. Sulbaran, Computation of the fractional Fourier transform,
Applied and Computational Harmonic Analysis, vol. 16, no. 3, pp. 182202, 2004.
[27] C. Candan, M. A. Kutay, and H. M. Ozaktas, The discrete fractional Fourier trans-
form, IEEE Transactions on Signal Processing, vol. 48, pp. 13291337, May 2000.
[28] A. Oppenheim and R. Schafer, Digital Signal Processing. Prentice Hall, 1975.
[29] A. Amein and J. Soraghan, Fractional chirp scaling algorithm: Mathematical
model, IEEE Transactions on Signal Processing, vol. 55, pp. 41624172, August
2007.
[30] P. Djuri c and S. Kay, Parameter estimation of chirp signals, IEEE Transactions on
Acoustics, Speech, and Signal Processing, vol. 38, pp. 21182126, December 1990.
[31] S. Kay, Fundamentals of Statistical Signal Processing:Detection Theory. Upper
Saddle River, NJ: Prentice-Hall, 1998.
[32] J. Mathews, Numerical Methods for Computer Science, Engineering, and Mathe-
matics. Englewood Cliffs, NJ: Prentice-Hall, 1987.
[33] M. Wang, A. Chan, and C. Chui, Linear frequency-modulated signal detection us-
ing Radon-ambiguity transform, IEEE Transactions on Signal Processing, vol. 46,
pp. 571586, March 1998.
82
[34] C. Capus and K. Brown, Short-time fractional Fourier methods for the time-
frequency representation of chirp signals, Journal of the Acoustical Society of
America, vol. 113, pp. 32533263, June 2003.
[35] M. Wang, A. Chan, and C. Chui, Wigner-Ville distribution decomposition via
wavelet packet transform, Proceedings of the IEEE-SP International Symposium
on Time-Frequency and Time-Scale Analysis, pp. 413416, June 1996.
[36] X. Chen and J. Guan, A fast FRFT based detection algorithm of multiple moving
targets in sea clutter, Proceedings IEEE International Radar Conference, pp. 402
406, May 2010.
[37] L. Feng, H. Yu, T. Rao, and W. Yue, Resolution ability of fractional Fourier trans-
form in multi-component LFM signal chirp-rate, 4th International Conference on
Wireless Communications, Networking and Mobile Computing (WiCOM), pp. 14,
October 2008.
[38] C. Min, F. Yaowen, J. Weidong, L. Xiang, and Z. Zhaowen, High resolution range
prole imaging of high speed moving targets based on fractional Fourier trans-
form, Proceedings of the SPIE - The International Society for Optical Engineering,
vol. 6786, pp. 67865419, November 2007.
[39] H. Li, B. Yu, W. Xi, and K. Wang, A fractional Fourier approach SAR imaging
method, Congress on Image and Signal Processing, vol. 5, pp. 215218, May 2008.
[40] D. Sazbon, Z. Zalevsky, E. Rivlin, and D. Mendlovic, Using Fourier/Mellin-based
correlators and their fractional versions in navigation task, in Journal of Pattern
Recognition Society, vol. 35, pp. 29932999, New York: Elsevier, 2002.
[41] S. Parashar, E. Langham, J. McNally, and S. Ahmed, RADARSAT mission require-
ments and concept, Canadian Journal of Remote Sensing, vol. 19, pp. 280288,
November 1993.
[42] A. Mahmood, RADARSAT-1 background mission monitoring of the arctic, IEEE
International Geoscience and Remote Sensing Symposium, p. 4, July 2005.
[43] A. Doerry, Performance limits for synthetic aperture radar - second addition, San-
dia Report SAND2006-0821, p. 70, 2006.
[44] B. Barkat and B. Boashash, IF estimation of linear FM signals corrupted by multi-
plicative and additive noise: A time-frequency approach, IEEE International Con-
83
ference on Acoustics, Speech, and Signal Processing, vol. 2, pp. 661664, June
2000.
[45] J. Lee, K. Hoppel, and S. Mango, Unsupervised estimation of speckel noise in
radar images, International Journal of Imaging Systems and Technology, vol. 4,
pp. 298305, August 1992.
[46] X. Xia, Discrete chirp-Fourier transform and its application to chirp rate estima-
tion, IEEE Transactions on Signal Processing, vol. 48, pp. 31223133, November
2000.
84