RELAXATIONBASED METHODS FOR SAR TARGET
FEATURE EXTRACTION AND IMAGE FORMATION
By
ZHAOQIANG BI
A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY.
UNIVERSITY OF FLORIDA
1999
This work is dedicated to my wife, Jing Wang, for her love and sharing my
happiness and depression in the past years; to my parents, who continue supporting
and encouraging me for my education.
ACKNOWLEDGEMENTS
I am sincerely grateful to my advisor and committee chairman, Dr. Jian
Li, not only for her academic guidance and insightful suggestions throughout the
development of this dissertation and providing an invaluable environment for the
research, but also for her financial support in the past years.
I would like to thank my graduate committee members, Dr. Jos6 C. Principe,
Dr. Fred J. Taylor, Dr. William W. Edmonson, and Dr. David C. Wilson, for their
time and interest in serving on my supervisory committee.
I would also like to thank Dr. ZhengShe Liu, Dr. Renbiao Wu, and the fellow
students, past and present, of the Spectral Analysis Laboratory (SAL) for their great
help.
Finally, I would like to gratefully acknowledge all the people who helped me
during my Ph.D program.
iii
TABLE OF CONTENTS
ACKNOWLEDGEMENTS ................... ........ iii
LIST OF TABLES ....... .................. ......... vi
LIST OF FIGURES .................... ............ vii
A BSTRACT . . . . . . . . . . . . . . . . . . x
CHAPTERS
1 INTRODUCTION ............................. 1
1.1 Background ................... ........... 1
1.2 Scope of the W ork .......................... 4
1.3 Contributions ................... ......... 8
1.4 Dissertation Outline ................... ...... 9
2 LITERATURE SURVEY ......................... 11
2.1 2D Super Resolution SAR Image Formation .......... 11
2.2 3D SAR Feature Extraction ...................... 15
2.3 SAR Motion Compensation ................. ... 16
3 PARAMETRIC METHODS FOR SAR IMAGING .. ......... 18
3.1 RELAX ................................ 18
3.1.1 Review of the RELAX Algorithm ............. 18
3.1.2 Image Formation ....................... 21
3.2 RELAXNLS .................... ......... 23
3.2.1 Review of the RELAXNLS Algorithm ........... 23
3.2.2 Image Formation ................. ..... . 27
3.3 Experimental Results ........................ 28
4 SEMIPARAMETRIC METHODS FOR SAR IMAGING ....... 35
4.1 Problem Formulation and Data Model ............... 35
4.2 Review of the APES Algorithm .......... ....... 37
4.3 Data Model Ambiguities and Their Effects on SAR Image Formation 38
4.3.1 Model Ambiguities ...................... 39
4.3.2 Image Formation ...... ................. 41
4.3.3 Model Ambiguity Effects on SAR Image Formation .... 43
4.4 The SPAR Algorithm ........................ 45
iv
4.4.1 Target Feature Extraction. . . . . . . .... 45
4.5 Modified RELAXNLS Algorithm . . . . . . ... 51
4.6 Cram6rRao Bound (CRB) of the SemiParametric Data Model 52
4.7 Numerical and Experimental Results . . . . . ..... 54
5 3D TARGET FEATURE EXTRACTION USING CLSAR ...... 63
5.1 CLSAR Data Model ......................... 63
5.1.1 1D Data Model for High Range Resolution Radar . .. 63
5.1.2 Full Synthetic Aperture Radar . . . . . . ... 64
5.1.3 Curvilinear SAR ....................... 67
5.2 The RELAX Algorithm ..................... . 68
5.2.1 Full Aperture .... .. ..... ... .. ... .. .. 68
5.2.2 Curvilinear Aperture . . . . . . . . ... 70
5.3 Performance Analysis of Parameter Estimation via CRB . . 71
5.4 Numerical and Experimental Results. . . . . . ... . 72
5.4.1 Performance Analysis of Different Curvilinear Apertures
via C R Bs . . . . . . . . . . . . . .. 72
5.4.2 Experimental Examples . . . . . . . . ... 74
6 AUTOFOCUS IN CLSAR ........................ 79
6.1 Aperture Error Effects ................... ... 80
6.2 The RelaxationBased Autofocus Algorithm (AUTORELAX) .83
6.2.1 Aperture Parameter Estimation .............. 83
6.2.2 Target Feature Extraction . . . . . . . ... 85
6.3 Experimental and Numerical Results . . . . . ..... 87
7 CONCLUSION ................... ......... 94
7.1 Summary ...... ... .............. ...... 94
7.2 Future W orks ................... ......... 96
APPENDIXES
A A SINGLE SCATTERER FEATURE EXTRACTION VIA SPAR .98
B APPROXIMATION OF THE RANGE R . . . . . . ... 100
C CALCULATING (x, y, z) FROM (x, y, ) ................ 102
REFERENCES ................................... .104
BIOGRAPHICAL SKETCH ...... . . ....... ..... 109
v
LIST OF TABLES
4.1 True parameters of the two dihedrals used in Figures 1 and 2 ... 44
5.1 Comparison of the CRBs (in dB) of the target features for the cases
of the full, circular, parabolic, Lshaped, Arc1, and Arc2 apertures
when M = 63, K = 1,w = = w= 0, a = 1, and 2 = 40. . . . 73
6.1 Estimation errors for the parameters of the highest scatterer in the
simulation example before and after autofocusing. . . . ... 89
vi
LIST OF FIGURES
1.1 Illustration of 2D radar imaging. . . . . . . . ... ... 2
1.2 Illustration of SAR imaging modes. (a) Spotlightmode SAR. (b)
Stripm ap SAR .... .. .... . .. . . .. .. ... ... 3
1.3 Possible apertures for a 3D SAR system, a) Full aperture. b) Parabolic
aperture. (c) Orthogonal subaperture. ( 0 and 0 denote the azimuth
and elevation angles, respectively.) . . . . . . . . . 6
1.4 Mesh plots of the modulus of the RCS obtained by using 2D FFT with
different apertures (range information suppressed for the illustration
purpose only). (a) Full aperture as shown in Figure protect 1.2(a). (b)
Curvilinear aperture as shown in Figure 1.2(c). . . . . . . 7
3.1 RELAX SAR Images by using the 40 x 40 ERIM data with K = 59. (a)
Unwindowed 2D FFT image. (b) Windowed 2D FFT image. (c) Un
windowed RELAX image without background clutter. (d) Windowed
RELAX image without background clutter. (e) Unwindowed RELAX
image with background clutter. (f) Windowed RELAX image with
background clutter. ........................... 31
3.2 RELAX Images by using the 20 x 20 ERIM data with K = 41. (a)
Unwindowed 2D FFT image. (b) Windowed 2D FFT image. (c) Un
windowed RELAX image without background clutter. (d) Windowed
RELAX image without background clutter. (e) Unwindowed RELAX
image with background clutter. (f) Windowed RELAX image with
background clutter. ... . . . . . . . . . 32
3.3 Target photo taken at 450 azimuth angle. . . . . . . ... 33
3.4 SAR images of the Slicy data (00 azimuth and 300 elevation angles). (a)
Original windowed 2D FFT image. (b) Windoed 2D FFT image with
spoiled resolution. (c) ~ (d): Windowed RELAX image without and
with background clutter with K = 36. (e) ~ (f): Windowed RELAX
NLS image without and with background clutter with K = 9. (c) ~
(f) are obatined using the data in (b). (The vertical and horizontal
axes are for range and crossrange, respectively.) . . . .... 34
vii
4.1 Ambiguity effect on the SAR image formation in the absence of range
estimation errors. (a) True windowed FFT SAR image. (b) Windowed
FFT image of the first "scatterer". (c) Windowed FFT image of the
second "scatterer". (d) Combined windowed FFT image of the two
"scatterers". (The vertical and horizontal axes are for range and cross
range, respectively.) ........ .................. 58
4.2 Ambiguity effect on the SAR image formation in the presence of range
estimation errors. (a) Windowed FFT image of the first "scatterer".
(b) Windowed FFT image of the second "scatterer". (c) Combined
windowed FFT image of the two "scatterers" with ( = 1 (without
extrapolation). (d) Combined windowed FFT image with ( = 2.
(The vertical and horizontal axes are for range and crossrange, re
spectively.) . . . . . . . . . . . . . . .. . 59
4.3 Comparison of SAR images formed using different algorithms for sim
ulated data at high SNR (u = 0.6). (a) True SAR image. (b) Win
dowed 2D FFT SAR image. (c) Hybrid SAR image. (d) SPAR SAR
image. (e) RELAXNLS SAR image. (f) Modified RELAXNLS SAR
image. (The vertical and horizontal axes are for range and crossrange,
respectively.) ... ... ........... ......... .. 60
4.4 Comparison of SAR images formed using different algorithms for sim
ulated data at low SNR (Uc = 6). (a) Windowed 2D FFT SAR image.
(b) Hybrid SAR image. (c) SPAR SAR image. (d) RELAXNLS SAR
image. (e) Modified RELAXNLS SAR image. (The vertical and hor
izontal axes are for range and crossrange, respectively.) ...... ..61
4.5 Comparison of SAR images obtained via different algorithms for the
Slicy data hb15533.015 (0 azimuth and 300 elevation angles). (a)
Windowed 2D FFT SAR image from the XPATCH data. (b) Win
dowed 2D FFT SAR image from the Slicy data. (c) Hybrid SAR
image. (d) SPAR SAR image. (e) RELAXNLS SAR image. (f) Mod
ified RELAXNLS SAR image. (c) ~ (f) are all obtained from the
data used in (b). (The vertical and horizontal axes are for range and
crossrange, respectively.) . . . . . . . ..... ....... 62
5.1 Broadside spotlightmode SAR data collection geometry. ...... . 65
5.2 Curvilinear apertures. a) The circular aperture. b) The Arc1 aper
ture. c) The Arc2 aperture ...................... . 73
5.3 The CRBs of the target parameters as a function of the circular aper
ture radius when K = 1, a = 1, w = C = c = 0, and a2 = 40. a) The
CRB of the complex amplitude a. b) The CRB of w. c) The CRB of
D. d) The CRB of C. ......................... .. 77
5.4 3D plots of K = 8 scatterers extracted by using RELAX with the
indoor experimental data. (a) Obtained with full aperture as shown in
Figure 1.2(a). (b) Obtained with curved aperture as shown in Figure
1.2(c). .... .. ..................... ......... .. 78
viii
6.1 (a) Curvilinear apertures for the experimental example. (b) True scat
terer distribution. (c) 3D plot of K = 20 scatterers extracted from
the initial aperture (dashed line) in (a) with RELAX. (d) 3D plot
of K = 20 scatterers extracted from the manually adjusted aperture
(solid line) in (a) with RELAX. ....... ............. . 90
6.2 Autofocused curvilinear aperture and scatterer distribution obtained
with AUTORELAX by autofocusing only in the elevation direction
and using K = 20 for the experimental example. (a) Autofocused
curvilinear aperture (dotted line). (b) Scatterer distribution. . . 91
6.3 (a) Manually adjusted (solid line) and autofocused (dotted line) curvi
linear apertures for the experimental example. (b) Fitting the manu
ally adjusted aperture to the autofocused aperture by adding a line to
the form er .. . . . . . . . . . . . . . . . 91
6.4 Autofocused curvilinear aperture and scatterer distribution obtained
with AUTORELAX by autofocusing in both the elevation and azimuth
directions and using K = 20 for the experimental example. (a) Auto
focused curvilinear aperture (dotted line). (b) Scatterer distribution. 92
6.5 (a) Curvilinear aperture for the simulation example. (b) True scatterer
distribution for the simulation example. (c) Scatterer distribution ob
tained from the initial aperture (dashed line) in (a) by using RELAX
with K = 20. (d) Scatterer distribution obtained via AUORELAX
with K = 20 and using the autofocused aperture (dotted line) shown
in (a). . . . . . . . . . . . . . . . . .. .. 93
ix
Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy.
RELAXATIONBASED METHODS FOR SAR TARGET
FEATURE EXTRACTION AND IMAGE FORMATION
By
Zhaoqiang Bi
May 1999
Chairman: Dr. Jian Li
Major Department: Electrical and Computer Engineering
Synthetic aperture radar (SAR) has been a mature but actively researched
technology due to its dayandnight and allweather capability of offering high res
olution imaging for both military and civilian applications. As the foundation of
automatic target detection and recognition, SAR imaging and autofocusing continue
to attract more research interest. Nonparametric spectral estimation methods are
robust methods for SAR image formation. However, nonparametric methods cannot
be used to significantly improve the resolution of the formed SAR images since they
generally do not fully exploit the characteristics of radar targets of interests even
when such information is available.
In this dissertation, efforts have been made to form super resolution two
dimensional (2D) SAR images via relaxationbased parametric methods. The
relaxationbased optimization methods have been proved to be quite useful in several
x
other applications, such as radio astronomy, microwave imaging, and spectral estima
tion. The relaxationbased methods are extended for super resolution SAR imaging
of radar targets consisting of only trihedrals or both trihedrals and dihedrals. We
have also devised a robust and computationally simple SPAR (SemiPARametric) al
gorithm for 2D SAR imaging based on a flexible semiparametric data model when
it is difficult to establish an accurate target data model in crossrange. Hence SPAR
takes advantages of both parametric and nonparametric spectral estimation meth
ods to form enhanced SAR images. Numerical and experimental examples have been
used to demonstrate the performances of the proposed algorithms. We have observed
that the relaxationbased parametric methods provide super resolution SAR images
when the assumed data model is valid; otherwise SPAR performs better.
Threedimensional (3D) target features, including the height information,
radar cross section (RCS), and 2D location (range and crossrange), provide quite
useful information for such applications as automatic target recognition. Thus, ef
forts have also been made in this dissertation to devise an effective relaxationbased
algorithm, referred to as AUTORELAX, for both 3D target feature extraction and
motion compensation via curvilinear SAR (CLSAR), a novel technology which is
still at its developing stage. The proposed AUTORELAX algorithm is shown to be
promising when evaluated by using both the experimental and simulated examples.
xi
CHAPTER 1
INTRODUCTION
1.1 Background
RADAR stands for RAdio Detection And Ranging. By now, radar has been
used not only to detect and locate targets, as was originally intended, but to gen
erate images of illuminated scenes without the restriction of weather and time as
well. Radar image is defined as a twodimensional (2D) mapping of the reflectivity
corresponding to the scene illuminated by a radar. More specifically, a radar image
is the reconstruction of the illuminated scene from the electromagnetic reflectivity,
which is usually treated as a 2D function in range and crossrange. The direction
in which a radar beam propagates is referred to as the range direction, while the
one perpendicular to the line of sight is called the crossrange direction, or the az
imuth direction (see Figure 1.1 for an illustration). Radar imaging [2] has played a
crucial role in both military applications, including reconnaissance and surveillance
[21], and civilian ones including geological and topographic map generation, weather
forecasting, scene classification, and target recognition [60, 65].
The quality of radar images depends greatly on the image resolution, which
is nominally referred to as the ability to distinguish two closelyspaced elements of a
target [48]. The higher the radar resolution, the more details of target features are
available for detection, characterization, and identification of targets of interest. As
such, much effort has been made to improve 2D imaging radar resolutions, which
include range and crossrange resolutions. High range resolution can be achieved
by using wide bandwidth signals, such as linear frequency modulated (FM) chirp
pulses [48], since range resolution is inversely proportional to the transmitted signal
1
2
Flight Path
^/
CrossRange
Figure 1.1: Illustration of 2D radar imaging.
bandwidth. The crossrange resolution 6, of a conventional radar or real aperture
radar (RAR) is rangedependent and is determined by
S= R, (1.1)
where R represents the range of the target; A and L, respectively, denote the wave
length and the size of the antenna aperture. To increase crossrange resolution (given
R), an obvious way is either to reduce the wavelength, or to increase the size of the
antenna. Using shorter wavelength requires more transmission power since electro
magnetic signals with smaller wavelengths are more likely absorbed by the atmo
sphere. In the meantime, it is not feasible to increase the antenna size too much,
especially when the radar is carried on board of an airplane or spacecraft. Hence
a conventional radar or RAR inherently has a limited crossrange resolution. It is
desired to achieve high crossrange resolution without reducing the wavelength and
exploiting large physical antennas. Synthetic aperture radar (SAR) is a solution and
can achieve high crossrange resolution independent of range.
The SAR technology was first proposed in the 1960s [59]. Instead of exploiting
large antennas and reducing wavelength, SAR makes use of a synthetic aperture to
achieve high crossrange resolution. A synthetic aperture is formed by coherently
3
Synthetic Aperture Synthetic Aperture
^ " / Flight Path Flight Path
 " ........ _.....
CrossRange
CrosRange ________   
(a) (b)
Figure 1.2: Illustration of SAR imaging modes. (a) Spotlightmode SAR. (b) Strip
map SAR.
processing radar returns from a sequence of locations when a nominallysized antenna,
which is carried on board of an aircraft, sequentially moves along the flight path
(crossrange direction). Consequently, a synthetic antenna with an effective antenna
length proportional to the synthetic aperture can be realized even with a small real
antenna. The longer the synthetic aperture, the finer the crossrange resolution
achieved. It has been demonstrated in [29] that the crossrange resolution of SAR
depends only on the wavelength A and the synthetic aperture interval. To achieve
such a desirable crossrange resolution, many techniques have been developed to
record data, compensate motion errors, and construct high quality SAR images ever
since the 1960s. (Detailed descriptions of these techniques are well documented in
many books, e.g. [29, 11]). SAR has grown so dramatically through advances in
digital signal processing, computing power, and microwave technology that it is one
of the most important techniques for high resolution radar image formation.
Spotlightmode and stripmap SAR are two commonly used SAR imaging
modes. In a spotlightmode SAR, antenna beams continuously steer to the same
terrain patch (scene) illuminated by a radar when the aircraft traverses the flight
path, while the antenna steering of a stripmap SAR is fixed relative to the flight
path when the aircraft moves along the flight path (see Figure 1.2 for the difference).
4
The crossrange resolution of stripmap SAR is limited by the azimuth beamwidth,
while that of spotlightmode SAR is independent of the azimuth beamwidth since the
antenna steers its beam to continuously illuminate the same scene. Hence Spotlight
mode SAR is more suitable for high resolution SAR image formation of a small scene
than stripmap SAR.
1.2 Scope of the Work
Spotlightmode SAR has been extensively applied to 2D high resolution im
age formation and target feature extraction of small scenes due to its finer resolutions
and readily controllable flight path, which is a straight line. The 2D target features
include the radar cross section (RCS) and the 2D location (range and crossrange) of
each scatterer. One topic of our work is super resolution 2D image formation from
phase history data collected by spotlightmode SAR. The super resolution herein is
defined as the resolution that is better than the Rayleigh limits determined by the
transmission wavelength and the length of the synthetic aperture of SAR. A conven
tional approach for 2D image formation from SAR phase history data is the fast
Fourier transform (FFT), which is computationally efficient and robust. However,
SAR images obtained via FFT suffer from poor resolution and high sidelobes. SAR
image formation can be considered as a spectral estimation problem since SAR images
are the reconstruction of scene reflectivity intensities versus space locations. Mod
ern spectral estimation methods including parametric and nonparametric methods
have found extensive applications in improving the resolution and accuracy of the
formed SAR images. In particular, parametric spectral estimation methods are more
suitable to form super resolution SAR images when parametric data models can be
used to model the target. The super resolution of the formed SAR images can be
achieved by extrapolating the phase history data beyond observations using an appro
priate data model and the parameter estimates. We will discuss how to form super
5
resolution SAR images by using the target features extracted via relaxationbased
methods including RELAX [39] and RELAXNLS [42] in Chapter 3. The relaxation
based optimization approach is one of the effective ways to minimize a nonlinear least
squares (NLS) cost function and has been proved to be quite useful in many applica
tions, such as radio astronomy [28], microwave imaging [66], spectral estimation [23],
and sinusoidal parameter estimation and target feature extraction [39, 42]. RELAX
was originally proposed for the parameter estimation of complex sinusoids in colored
noise. Compared with other NLS algorithms, RELAX is relatively robust against
data mismodeling errors [39]. We will show that RELAX can be used to enhance
SAR images of targets consisting of point scatterers. However, the point scatterer
data model may not be valid in practice since, for example, for many manmade
targets, such as vehicles and buildings, much of the returned energy is caused not
only by trihedral corner reflectors, but also by dihedral corner reflectors of a target
[36]. We also describe how to form super resolution SAR images of targets consisting
of both trihedrals and dihedrals using RELAXNLS [42]. RELAXNLS is based on
a mixed data model, which models a trihedral as a point scatterer and a dihedral as
a complex sinusoid with a constant phase and a variable amplitude described by a
sinc function in crossrange. We will demonstrate the super resolution property of
the RELAX and RELAXNLS algorithms with experimental data including MSTAR
and ERIM (Environmental Research Institute of Michigan) data.
Parametric methods may outperform nonparametric methods provided that
the assumed data models are valid. In the more likely cases where it is difficult to
establish a good parametric data model in crossrange due to a variety of physical
phenomena including glints, resonances, and motioninduced phase errors [15], a more
flexible data model is desired and nonparametric and parametric methods may be
combined together for 2D SAR target feature extraction and image formation. In
our work, we also establish a more flexible data model for manmade targets when
6
(a) e (b) 0 (c) 0
Figure 1.3: Possible apertures for a 3D SAR system. a) Full aperture. b) Parabolic
aperture. (c) Orthogonal subaperture. ( 0 and 4 denote the azimuth and elevation
angles, respectively.)
no accurate parametric data model is available in crossrange. Based on the flexible
data model, we present a robust and computationally simple semiparametric (SPAR)
algorithm for target feature extraction and SAR image formation by exploiting a
hybrid of nonparametric and parametric spectral estimation methodes.
It has been shown in [62, 16] that threedimensional (3D) target features
including the height information as well as the 2D target features provide very useful
information in certain applications, such as automatic target recognition (ATR). 3D
SAR image formation and feature extraction are becoming active research areas. It
is possible to obtain 3D information of an illuminated scene with Spotlightmode
SAR when a nonstraight line synthetic aperture is exploited, which is determined by
the azimuth angle 0 and elevation angle 0. In our work, we will also study 3D target
feature extraction and motion compensation (autofocus) using curvilinear SAR, in
which the trajectory of a spotlightmode SAR is a suitable curvilinear aperture.
Among existing 3D SAR imaging techniques, interferometric SAR (IFSAR)
is now a welldeveloped technique that plays an important role in 3D (as well as 2D)
imaging. However, IFSAR suffers from ambiguity problems [11, 40]. Theoretically,
3D target features and images can also be obtained when the trajectory of SAR is a
full aperture as shown in Figure 1.2(a) even though such an aperture is not feasible in
7
Height CrossRange Height CrossRange
(a) (b)
Figure 1.4: Mesh plots of the modulus of the RCS obtained by using 2D FFT with
different apertures (range information suppressed for the illustration purpose only).
(a) Full aperture as shown in Figure protect 1.2(a). (b) Curvilinear aperture as shown
in Figure 1.2(c).
practice. Nevertheless, a suitable subset of the full aperture can be used as a trajectory
of SAR to obtain 3D information. This technique is known as curvilinear SAR
(CLSAR). CLSAR traverses a curvilinear aperture path, which is a subset of the full
aperture (see, for example, Figures 1.2(b) and (c)), and collects a sparsely sampled
dataset, from which 3D information can be obtained. CLSAR avoids ambiguity
problems inherent in IFSAR systems, but suffers from severe high sidelobes when
used with FFT to form SAR images due to the limited size of measurements. This
observation is illustrated in Figure 1.2. (Figures 1.2(a) and (b) show the modulus
of the RCS of a single point scatterer obtained by applying 2D FFT to the data
collected via the full aperture as shown in Figure 1.2(a) and the curvilinear aperture
as shown in Figure 1.2(c).) Due to its high sidelobes, CLSAR is of little interest in
SAR imaging. Yet CLSAR can be used with parametric spectral estimation methods
to extract 3D target features of small targets consisting of a finite number of isolated
point scatterers. In Chapter 5, we will establish a CLSAR data model and describe
how the RELAX algorithm [39] can be used for 3D target feature extraction using
CLSAR with different curvilinear apertures.
8
Motion error compensation also plays a crucial role in accurately extracting
target features using CLSAR. Motion errors in 2D SAR systems can be approximated
as phase errors in crossrange since the azimuth angle 0 is very small and phase errors
are mainly introduced by the uncertainty in the distance from the radar to the center
of the illuminated scene [29]. However, it is not the case for CLSAR since curvilinear
aperture determined by azimuth angle 0 and elevation angle is critical in obtaining
3D information of the target. Hence motion errors in 0 and may not be negligible.
Motion compensation (autofocus) in CLSAR turns out to be more complicated than
that in 2D SAR systems. In Chapter 6, we investigate the effects of aperture errors
on 3D target feature extraction and present a relaxationbased autofocus algorithm,
referred to as AUTORELAX, to extract target features in the presence of aperture
errors using CLSAR.
1.3 Contributions
The significance of this research lies in that it examines the possibility of super
resolution image formation and autofocus of small targets via spotlightmode SAR
system by using relaxationbased parametric and semiparametric spectral estimation
methods. Super resolution SAR images obtained via the proposed methods can
be further used for SAR image segmentation and automatic target detection and
recognition. This research makes contributions in the following two areas:
(I) 2D super resolution SAR image formation
(1) We have extended RELAX to form enhanced SAR images of targets consisting
of point scatterers;
(2) We have extended RELAXNLS for super resolution SAR image formation of
targets consisting of both trihedrals and dihedrals;
9
(3) We have devised a novel, robust, and computationally simple SemiPARametric
(SPAR) algorithm for SAR target feature extraction and image formation based
on a flexible data model;
(4) Experimental results using ERIM and MSTAR Slicy data show that the relaxation
based parametric approaches including RELAX and RELAXNLS and the semi
parametric algorithm SPAR, can yield SAR images with higher resolution than
FFT based methods. Also SPAR is more robust and computationally simple
than RELAXNLS.
(II) 3D target feature extraction and motion compensation via CLSAR
(1) We have established a 3D data model;
(2) We have performed Cram6rRao Bound (CRB) analyses of parameter estima
tion using various flight trajectories, which provide more insight into the per
formances associated with different curvilinear apertures;
(3) The effects of aperture errors on target feature estimation have been investi
gated and a relaxationbased approach for autofocus and target feature extrac
tion using CLSAR has been introduced.
1.4 Dissertation Outline
The remainder of the dissertation is organized as follows. Chapter 2 gives a
literature survey of emphasis on the topics considered, including 2D SAR imaging, 3
D SAR target feature extraction, and motion compensation (autofocus). In Chapter
3, we describe how to form super resolution 2D SAR images via the relaxation
based parametric spectral estimation algorithms including RELAX and RELAX
NLS. A novel semiparametric (SPAR) method is introduced in Chapter 4 for SAR
image formation when an accurate parametric data model in crossrange is difficult
10
to establish. In Chapter 4, we also present a modified RELAXNLS algorithm by
using the initial conditions provided by SPAR to save the amount of computations
needed by the original RELAXNLS algorithm. In Chapter 5, a CLSAR data model
is established and RELAX is introduced for 3D target feature extraction via CLSAR.
Chapter 6 investigates the ambiguity problems inherent in the proposed CLSAR data
model when aperture errors are present. A relaxationbased autofocus algorithm for
3D target feature extraction in the presence of aperture errors is also presented in
Chapter 6. Finally, Chapter 7 give the summary and conclusions of this research and
outlines the future work.
CHAPTER 2
LITERATURE SURVEY
In this chapter, we give a brief review of the topics related to the work, namely
super resolution 2D SAR image formation, 3D SAR target feature extraction, and
SAR motion compensation.
2.1 2D Super Resolution SAR Image Formation
Since SAR image formation can be considered as a spectral estimation prob
lem, modern spectral estimation methods find extensive applications in improving
the resolution and accuracy of the formed SAR images. Excellent texts on spectral
estimation are available [63, 30, 46] and historical and modern perspectives on the
general topics of spectral analysis can be found in [54, 31]. A brief review of super res
olution SAR image formation via nonparametric and parametric spectral estimation
methods is given below.
Nonparametric Spectral Estimators
An important class of nonparametric spectral estimators are the adaptive
filterbank approaches which can be used to form SAR images with better resolu
tion and lower sidelobes than the FFT based methods [15]. The adaptive sidelobe
reduction (ASR) approach [14] and the widely used Capon method [10] as well as
its reduced rank variations [5, 6] are members of this class. A recently introduced
matchedfilter bank based complex spectral estimator that is of interest is the APES
(Amplitude and Phase Estimation of a Sinusoid) method [38]. Despite their different
derivations, it was found that both Capon and APES are members of the general
class of the matchedfilterbank (MAFI) approaches to complex spectral estimation
11
12
[37]. Empirically, it was observed in [38] that Capon appears to give biased spec
tral estimates whereas APES is unbiased. This observation was later proved in [37]
by using the secondorder Taylor expansion technique. Also the statistical perfor
mance of the Capon estimator degrades rapidly as the filter length increases, while
that of APES remains unaffected. The study in [37] gives a compelling reason for
preferring APES to Capon in most spectral estimation applications including SAR
imaging. In cases where an appropriate parametric data model of targets is not
available, APES is a good candidate for SAR imaging which provides much lower
sidelobes and narrower spectral peaks than Capon and the conventional FFT based
methods. However, significant resolution improvement cannot be obtained by using
nonparametric methods since they do not fully exploit the characteristics of targets,
especially when such information is available.
Parametric Spectral Estimators
When parametric data models can be used to describe radar targets, paramet
ric spectral estimators provide attractive alternatives for super resolution SAR image
formation. The Burg's maximum entropy (ME) method [9] was applied to radar im
age enhancement by Mancill in 1977 [44]. Unlike FFT methods which assume that
the signal beyond the observations is zero, ME methods make linear prediction of the
signal beyond the observation window so that they can achieve better resolution than
FFT methods. Identical to ME methods in principles, the autoregressive (AR) model
based methods [15, 13, 20, 19, 26] have attracted a lot of research attention. However,
the high resolution property of the AR based methods are highly dependent on the
presence of observation noise [31]. As the signaltonoise ratio (SNR) decreases, so
does the resolution. On the other hand, the AR based methods are more appropriate
for SAR image formation of distributed scatterers rather than point scatterers since
the AR spectral estimator by nature is a continuous spectral estimator.
13
In recent years, many sophisticated and high resolution methods based on
point scatterer data model, such as eigendecomposition based methods [4], including
MUSIC (MUltiple SIgnal Classification) [58] and ESPRIT (Estimation of Signal Pa
rameters via Rotational Invariance Technology) [56], and relaxationbased nonlinear
least squares (NLS) method such as RELAX [39], have also been applied to super
resolution SAR image formation. The eigendecomposition based methods are signal
or noise subspace based approaches and generally are not suitable for SAR image
formation since they require that the number of data samples be much larger than
that of the dominant point scatterers and the process of whitening the noise sub
space (used in MUSIC) will destroy the spatial inhomogeneities of the terrain clutter
[15]. The RELAX algorithm [39] was originally proposed for sinusoidal parameter
estimation in colored noise. RELAX obtains the parameter estimates by minimizing
the NLS cost function via a sequence of FFTs, which is conceptually and computa
tionally simple and hence robust. RELAX coincides with the maximum likelihood
(ML) method when the noise is white and is also asymptotically (for large numbers
of data samples) statistical efficient when the noise is an AR or autoregressive mov
ing average (ARMA) process that can be described by a finite number of unknowns
[63]. Due to the iterations performed in each step of RELAX, a set of more accurate
initial estimates are available, which make RELAX converge to the global minimum
with a high probability. RELAX can also be referred to as SUPER CLEAN when
compared with two other approximate relaxation methods including CLEAN [28] and
MCLEAN (more clean) [23]. CLEAN was first proposed in radio astronomy and later
used in microwave imaging [66]. If no iteration is performed in each step of RELAX,
then it is equivalent to CLEAN. RELAX becomes MCLEAN if only one iteration
is performed in each step of RELAX. Hence RELAX is a computationally efficient
and robust algorithm with good statistical performance and is suitable for super
resolution SAR image formation of targets consisting of point scatterers (trihedrals).
14
Recently, RELAXNLS [42] was introduced for the feature extraction of targets
consisting of both trihedrals and dihedrals based on a mixed data model rather than
the point scatterer data model. An ideal trihedral acts as a triplebounce reflector
and can be modeled as a strong point scatterer [53]. The point scatterer data model,
however, is no longer valid in crossrange for an ideal dihedral corner reflector, which
acts as a doublebounce reflector, since its radar cross section (RCS) is described by
a sinc function, i.e., sinc(0) = sin(0)/0, where 0 denotes the angle between the radar
beam and the line perpendicular to the dihedral.
Super resolution SAR images obtained via modern spectral estimation meth
ods provide more details required for the detection and automatic target recognition
(ATR) than the conventional FFT based methods. Quantitative results showing the
advantages of using these sophisticated spectral estimation methods in SAR appli
cations begin to appear in the literature. For example, it was shown in [51, 50] that
a significant improvement in ATR performance can be achieved by using SAR im
ages with enhanced resolution obtained via the reducedrank variations of the Capon
method.
Compared with nonparametric spectral estimators, high resolution paramet
ric spectral estimators provide great potential for SAR imaging with better resolution
and lower sidelobes when the assumed data model is valid. On the other hand, para
metric methods are more sensitive to data model mismatch than nonparametric
methods. Hence, practically only robust super resolution parametric methods can be
used for SAR image formation. We will use robust relaxationbased algorithms in
cluding RELAX and RELAXNLS to form SAR images of different manmade targets
in our work. In cases where an appropriate data model of manmade targets is not
available in crossrange, robust nonparametric spectral estimators with good perfor
mances such as APES and high resolution parametric methods may work together to
obtain satisfactory results. To the best of our knowledge, a hybrid of nonparametric
15
and parametric methods for feature extraction and image formation of SAR targets
has not been addressed in the literature before.
2.2 3D SAR Feature Extraction
One relatively mature technology for 3D SAR imaging and feature extrac
tion is IFSAR [24, 55, 69, 27, 70] which exploits two different apertures. A simple
implementation of a IFSAR is to use two antenna displaced in the crosstrack plane
to obtain two coherent and parallel measurements. The relative height information
can be estimated from the phase difference between the two measurements. The use
of interferometric techniques with SAR imagery to provide height information makes
IFSAR suitable for 3D imaging since it inherits the finer resolutions (in range and
crossrange) associated with the spotlightmode SAR and also has reasonable height
resolution. Many 2D modern spectral estimation algorithms including Capon [10],
APES [38], and MUSIC [58] have been generalized for 3D imaging and target feature
extraction using IFSAR [40, 47]. An NLS method that outperforms MUSIC in speed
and estimation accuracy was also reported in [40]. However, such a system cannot
resolve more than one target scatterer at the same projected range and crossrange
but at different heights [40].
Another relatively new technology for 3D imaging is curvilinear SAR. CLSAR
traverses a curvilinear aperture path and avoids the height ambiguity problem inher
ent in IFSAR systems. CLSAR images are useless in practice due to high sidelobes.
However, CLSAR can be used in conjunction with modern spectral estimation meth
ods to accurately extract 3D features of small targets, which consist of a finite number
of isolated point scatterers. CLSAR is still at its exploratory stage of development
and only a few studies were reported in the literature. In [33], a coherent CLEAN
algorithm was presented by Knaell to effectively eliminate sidelobes in CLSAR im
ages. The CLEAN algorithm was originally proposed in radio astronomy [28] and
16
later was used in microwave imaging [66]. Encouraged by the initial results in [33],
Knaell later presented in [34] a correlation technique which works in conjunction with
the coherent CLEAN algorithm to reduce sidelobe interference. Furthermore, Knaell
in [35] compared the performances of the coherent CLEAN method and two other
methods including a maximum likelihood (ML) method and a sidelobe leakage reduc
tion algorithm [12] in extracting 3D target features by using CLSAR datasets. We
will discuss how RELAX (SUPER CLEAN) can be extended for 3D target feature
extraction using CLSAR in Chapter 5.
2.3 SAR Motion Compensation
Aperture errors exist due to atmospheric turbulence and platform uncertainty.
These undesirable errors may significantly degrade the quality of SAR images in sev
eral ways. For example, aperture errors may cause image geometric distortion, spu
rious targets, loss of resolution, and decrease in image contrast [32, 25] depending on
their natures. Therefore, effective motion compensation methods are indispensable
for SAR imaging and target feature extraction in practice. In 2D SAR systems, mo
tion errors mainly cause phase errors in crossrange [29]. Autofocus techniques that
compensate for phase errors using radar phase history data have been extensively
exploited by 2D SAR systems. Of the existing autofocus algorithms, mapdrift [45]
and phase difference [68] algorithms are used to compensate for phase errors asso
ciated with certain order polynomials, whereas the phasegradient autofocus (PGA)
algorithm [17, 18, 67] is independent of the phase error model, thereby is more ro
bust. Autofocus algorithms based on super resolution spectral estimation methods
also begin to appear in the literature [43, 52]. All the aforementioned 2D autofocus
algorithms assume that phase errors change in crossrange only and are constant for
all range bins. This assumption is generally valid for the 2D SAR systems where
17
significant phase errors are caused by motion measurement errors, such as alongtrack
velocity errors or lineofsight acceleration errors [11].
As in 2D SAR systems, motion error compensation is also indispensable for
3D SAR systems. Autofocus in CLSAR, however, becomes more complicated since
the aperture errors can no longer be approximated as phase errors only in crossrange
since the errors introduced in 0 (azimuth angle) and 0 (elevation angle) may not be
negligible and the extracted 3D target features are sensitive to the shape of the
curvilinear aperture (see the discussion in Chapter 6). To the best of our knowledge,
autofocus for CLSAR has not been reported in the literature, partly because CLSAR
is a relatively new technology.
CHAPTER 3
PARAMETRIC METHODS FOR SAR IMAGING
When an appropriate data model is available for radar targets, parametric
spectral estimation algorithms, which are less sensitive to the model mismatch, offer
the potential to significantly improve the resolution of the formed SAR images. In
this chapter, we consider forming enhanced SAR images with robust relaxationbased
methods including RELAX and RELAXNLS.
3.1 RELAX
When the dominant features of targets can be approximated as point scatter
ers, enhanced SAR images can be formed by using the target features extracted via
RELAX and the point scatterer data model. We first briefly introduce the RELAX
algorithm [39] and then discuss how to form SAR images using RELAX.
3.1.1 Review of the RELAX Algorithm
Assume that there are K dominant point scatterers or trihedral corner re
flectors in a target. An ideal trihedral acts as a triplebounce reflector and can be
modeled as a strong point scatterer [53] since its radar cross section (RCS) can be
considered as a constant due to the very small angle variation of the radar beam in
SAR. Then the target data model in the presence of unknown noise has the form:
K
y(n,h)= akej(ckn+Wkf) +e(n,'), n= 0,1,. *,N1, = 0,1, 1,
k=1
(3.1)
where N and N denote the numbers of data samples in range and crossrange, respec
tively; ak and {(k, Wk}, respectively, denote the unknown complex amplitude and 2D
18
19
frequencies of the kth sinusoid or point scatterer, k = 1, 2, K; {e(n, n)} denotes
the unknown noise sequence. Note that {wk, wk} corresponds to the 2D location of
the kth scatterer and ak is determined by the RCS and range of the kth scatterer.
The NLS estimates {&k, k, Wk }kk= of {ok, Wk, Wk}k= can be obtained by minimizing
the following NLS criterion [39]:
K 2
C, ({ak,wk, kI}k ) = WN(Wk)& ) (3.2)
k=1 F
where (.)T denotes the transpose; I IF denotes the Frobenius norm; Y is an N x N
matrix with y(n, n) being its (nn)th element;
WN(k) = 1 ewk ... e(1) ; (3.3)
r IT
g(wk) = 1 ejk ... eik(1) (3.4)
RELAX minimizes the NLS cost function C1 in (3.2) by using a complete relaxation
based search, i.e., for each K, the intermediate number of scatterers, only the param
eters of one scatterer vary while all others are fixed at their most recently determined
values. Let
K
Y =Y i N (i) i), (3.5)
i=l,ik
where {&i, i, 'i}i=,iW k are assumed to have been estimated. Then {&k, Ck, &k} are
obtained by
Wk (wk)Yk w k=kI (3.6)
ak ~ NN Wk=WkWk=Wk
where ()H and (.)*, respectively, denote the conjugate transpose and complex con
jugate, and
{(Wk, k} = argmax wH(wk)YkW (0k)12, (3.7)
WkWk
which indicates that {Ck, wk} can be obtained as the location of the dominant peak
of the periodogram wH(wk)YkW*f (k) 2 /NN, which can be efficiently calculated
via FFT. Then, &k can be calculated from the complex amplitude of the peak of
20
wH(wk)YkW* (Jk) 2 /NN. The RELAX algorithm is summarized as follows:
Step (1): Assume K = 1. Obtain {fk, k, k k=1 from Y by using (3.7) and (3.6).
Step (2): Assume K = 2. Compute Y2 with (3.5) by using {fk,2 k,&k}k=1 ob
tained in Step (1). Obtain {fk,Wk, &kk}k=2 from Y2. Next, compute Y1 by using
{kjk, lk, &k}k=2 and then redetermine {k, Wk, &k}k=1 from Y1.
Iterate the previous two substeps until "practical convergence" is achieved.
Step (3): Assume = 3. Compute Y3 by using {Ok, k, &k =1 obtained in Step
(2). Obtain {&k, k, k&k}k=3 from Y3. Next, compute Y1 by using {Ck, Wk, &k k=2 and
redetermine {jk, k, &k} k=1 from Y1. Then compute Y2 by using {Ck, Wk, kjck=1,3
and redetermine {Wkc, k, k, }k=2 from Y2.
Iterate the previous three substeps until "practical convergence".
Remaining Steps: Continue similarly until K is equal to K, which can be deter
mined by using generalized Akaike information criterion (GAIC) to be discussed later
on.
The "practical convergence" in the iterations may be determined by checking
the relative change e of the cost function C1 ({fk, k, &k}=l) in (3.2) between two
consecutive iterations.
It has been demonstrated in [39] that RELAX is related to CLEAN [28], which
is only an approximate relaxationbased method that minimizes the NLS criterion.
RELAX turns out to be CLEAN when no iteration is performed in each step of
RELAX. Due to the iterations performed in each step of RELAX, a more accurate
initial can be obtained, which makes RELAX converge to the global minimum with a
high probability. Meanwhile, RELAX is asymptotically (for large number of samples)
efficient, while CLEAN gives biased estimates, especially when two scatterers are
closely spaced. The RELAX algorithm, thereby, can also be referred to as SUPER
CLEAN.
21
We consider using GAIC (see [39] and the references therein for details) to
determine K, the number of sinusoids, by assuming the unknown noise and clutter
being white. The estimate K of K can be determined as an integer that minimizes
the following GAIC cost function:
(N1N1
GAICk = NNln 6 (n, 2)2 + y7n[ln(NN)](4k + 1), (3.8)
n=o n=0o
where
k
((n,n)=y(n,)) &kej3(kn+ k'n), n=0,1,,N 1, n=0,1,.,1,
k=1
(3.9)
4k + 1 denotes the total number of unknown realvalued parameters (of which 4K
are for the sinusoids and 1 is for the white noise), and y is a parameter of user choice.
3.1.2 Image Formation
We can form SAR images by using the point scatterer data model and the
target features extracted via RELAX. SAR images obtained via RELAX are referred
to as RELAX SAR images or simply RELAX images. The steps of using RELAX
for SAR image formation are outlined as follows.
Step 1: Obtain the parameter estimates of {ok, Wk, kk K= via RELAX by using the
measured phase history data (see Section 3.1.1 for details).
Step 2: Generate ys(ns, ns), the simulated phase history data of large dimensions,
from the estimated parameters and based on the data model in (3.1):
K
ys(n7, n,) = eakej((kn.+ks8), n, = 0,1, .. N 1, ns = 0,1, .. CN 1,(3.10)
k=1
where ( (( > 1) denotes the extrapolation factor, which is a parameter of user
choice, and &k, jk, and &k, respectively, denote the estimates of 0k, Wk, and ak,
k = 1,2,,K.
Step 3: Form RELAX SAR images containing only the dominant target features by
22
applying the normalized FFT to the simulated phase history data {y,(ns, s,)}, i.e.,
by computing
S(N1 (N1
C2NNE E ys(nsns)eEi(w9+^8). (3.11)
n,=O fi,=0
To suppress the sidelobes, the normalized FFT can be applied to the windowed
sequence {w( (ns, ft,)y,(ns, is)}, where the window sequence w,(n,, ns) satisfies
(N1 (N1
E W,(n.,Iu) = (2NN. (3.12)
n,=O is=0
The RELAX SAR images containing both the dominant target features and back
ground clutter may be appreciated since, for example, the shadow information may
be desired for automatic target recognition. If so, the normalized FFT can be applied
to the sum of the simulated phase history data {ys(n,, i,)} and {(26(n, u)} with zero
padding to have dimensions (N and (N, where 6(n, n) denotes the estimated back
ground clutter and is determined by (3.9) with K replaced by K. Note that scaling
the e(n, n) by a factor of (2 is needed when the background clutter is included in
the RELAX SAR images since both of its dimensions are 1/( times of those of the
simulated phase history data. Similarly, the normalized FFT can also be applied
to the sum of {ws(ns, s)ys(nris)} and {(C2W(n,, )6(n, )} with zero padding to
have dimensions (N and (N to suppress the sidelobes, where the window sequence
we(n, n) satisfies
N1 N1
S we (n, n) = NN. (3.13)
n=0 n=0
Note that since we cannot model the clutter effectively, its resolution cannot be
improved. We remark that the resolution of the RELAX SAR images is determine
by RELAX and ( > 1 is just used to demonstrate the super resolution property of
the RELAX algorithm for target feature extraction.
23
3.2 RELAXNLS
When targets consist of both trihedral (point scatterers) and dihedral corner
reflectors, we can form enhanced SAR images by using the RELAXNLS algorithm
[42]. We first give a brief review of the RELAXNLS algorithm and then discuss SAR
image formation via RELAXNLS.
3.2.1 Review of the RELAXNLS Algorithm
RELAXNLS [42] is devised to identify and obtain target features of multiple
corner reflectors (trihedral and dihedral) based on a mixed parametric data model.
An ideal dihedral corner reflector acts as a doublebounce reflector and the RCS
achieves the maximum when the radar beam is perpendicular to the dihedral and
falls off as a function of sinc(0) = sin(0)/9 when the angle between the radar beam
and the line perpendicular to the dihedral increases [49, 57]. Then dihedrals can be
modeled as
Kd
Sd(n, n) = adsinc[rbk(n Tk)]ej( dk n+dk), n = 0, 1, . N1, n = 0, 1, N,
k=1
(3.14)
where N and N denote the numbers of the available data samples in range and
crossrange, respectively; adk, {Wdki, dk}, and bk, k = 1, 2,  Kd, are, respectively,
proportional to the maximal RCS, the central location, and the length of the kth
dihedral corner reflector; Tk, k = 1, 2,. Kd, denotes the peak location of the data
sequence in crossrange and is determined by the orientation of the kth dihedral
corner reflector; finally, Kd is the number of the dihedral corners. Assume that a
target consists of Kd dihedrals and Kt trihedrals and K = Kd + Kt. Then the target
data model in the presence of noise and clutter has the form:
y(n, n) = sd(n, ) + st(n, n) + e(n, n), n = 0,1, N 1, = 0, 1, V 1,
(3.15)
24
where
Kt
St(n, i) = Zatej(tke +t, n =O,1, ,N 1, = 0,1, *, N 1, (3.16)
k=1
with {atk}k= and {Wtktk}=,K respectively, denoting the unknown complex am
plitudes and 2D unknown frequencies of the Kt trihedrals. The RELAXNLS al
gorithm proposed in [42] can be used to identify and effectively estimate the tar
get features{adk, bk, Wdk, Wd, k 7k=l} and {a, tk, tk 1k by minimizing the following
NLS criterion:
C2 ({atkWtk, tk }k=l, dk, bk, Wdk) WdkTk}i) = YSt Sd (3.17)
where [Y]j = y(i,j), [St]ij = st(i,j), and [Sd]ij = Sd(i,j). RELAXNLS converts
the problem of multiple corner reflector identification and feature extraction to that
of a sequence of single reflector identification and feature extraction by using the
relaxation based optimization approach. The identification and feature extraction of
a single corner reflector in RELAXNLS is described as follows:
Feature Extraction of a Single Trihedral Corner Reflector
Let Yt denote the data of a single trihedral in the presence of unknown noise
Et. Then
Yt = atWN(Wt)W (Wt) + Et, (3.18)
where at and {wt, et}, respectively, are proportional to the RCS and 2D location
of the trihedral. The NLS estimates {&t, t, tt} of {at, wt, @t} can be obtained by
minimizing the following cost function
Ct (at, WtCt) = IYt atwNy(Wt)WN^(wt)I F, (3.19)
and {&t, LCt, wt}, respectively, are given by
= W *(wt)Ytw ')
N l7 (3.20)
NN
25
and
{LJt, t} = argmax w(w) (3.21)
Thus {fLt, w't} corresponds to the peak location of IwH(wt)Ytw* (act)2, which can be
calculated efficiently via 2D FFT, and &t corresponds to the complex height of the
normalized 2D FFT of Yt at {Pt, Wt}.
Feature Extraction of a Single Dihedral Corner Reflector
Let Yd denote the data of a single dihedral in the presence of unknown noise
Ed. Then
Yd = adWN(d)gT + Ed, (3.22)
where {Wd, Wd ad, b, 7} are the target features of a dihedral corner reflector, and
== g(o) g(1) .. (N 1) (3.23)
with g(n) = sinc[b7r(r n)]ej"a1, n = 0,1, ,N 1. The NLS estimates
{ad, dd, bW, ,} of {ad, Wd, C,b, T} are determined by minimizing the following cost
function:
Cd (a, Wd, Wd, db, r) = YYd adWN(Wd)gT (3.24)
It can be shown that &d is determined by
w( = wYdg* (3.25)
and {Wd, d, b, } can be obtained by maximizing the cost function
C3 (Wd, Wd, b, T) = Iw(wN)Ydg* 2 (3.26)
RELAXNLS maximizes C3 in (3.26) by utilizing an alternating maximization ap
proach [8], i.e., alternatively updating {Od, wd, 6, r} while fixing other parameter esti
mates at their recently determined values. Note that fourdimensional search over the
parameter space is performed to estimate the dihedral corner reflector parameters.
26
Identification of Corner Reflector
It has been shown in [42] that Cd is much less than Ct when the true corner
reflector is a dihedral and Ct is almost the same as Cd when the true corner reflector
is a trihedral. Also b 26f when a trihedral is mistakenly considered as a dihedral,
where 5f is the smallest positive solution to
sin(Nflr) N
 (3.27)
sin(6f7r) 2
Then a corner reflector can be identified by checking the estimate b and the cost
functions Ct in (3.19) (with {at, t,&t} replaced by {a&tt, ,t}) and Cd in (3.24)
(with {Wd, Ow, ad, b, r} replaced by {jd, Wd, ad, b, r}. A corner reflector is identified as
a dihedral if C > 0.1 and b > 26.
Cd >
Steps of RELAXNLS are outlined as follows:
Step (1): Assume K = 1. Identify and obtain the parameters of a single
corner reflector.
Step (2): Assume K = 2. Subtract the signal obtained in Step (1) from
{y(n, )} and then identify and obtain parameters of the second corner reflector.
Subtract the second signal from {y(n, n)} and reidentify and redetermine the pa
rameters of the first signal.
Iterate the previous two substeps until "practical convergence".
Step (3): Assume K = 3. Subtract the two signals obtained in Step (2) from
{y(n, n)} and then identify and obtain the parameters of the third signal. Subtract
the second and third signals from {y(n, n)} and reidentify and redetermine the first
signal. Subtract the first signal (most recently determined) and the third signal from
{y(n, A)} and then reidentify and redetermine the second signal.
Iterate the previous three substeps until "practical convergence".
Remaining Steps: Continue similarly until K = K (K can be determined
by using GAIC to be discussed later on).
27
The "practical convergence" in the iterations can also be determined by check
ing the relative change of the cost function C2 in (3.17) between two consecutive
iterations.
We can also determine K, the total number of trihedrals and dihedrals, by
extending the GAIC discussed in Section 3.1 and assuming white noise and clutter.
The estimate K of K can be determined as an integer that minimizes the following
extended GAIC cost function:
(N1N1
GAICk = NNlnn (n, )2 + 7ln[ln(NN)](4t + 6K + 1), (3.28)
n=0 n=0
where
kd kt
e(n, n) = y(n, f) S & dsinc[7b(k rtk)]ejk+dk) ckej), (3.29)
k=1 k=1
with n = 0, 1, l N 1 and n = 0, 1, N 1; 7, as before, is a parameter of user
choice; K = Kt + d with kt and Kd denoting the numbers of trihedral and dihedral
corners, respectively, determined by RELAXNLS given K; finally, 4Kt + 6Kd + 1
is the total number of unknown realvalued parameters (of which 4Kt and 6Kd,
respectively, are for the trihedral and dihedral corners and 1 is for the white noise).
3.2.2 Image Formation
We can form enhanced SAR images of targets consisting of both trihedrals
and dihedrals by using the extracted target features via RELAXNLS and the mixed
data model in (3.15). The SAR images obtained via RELAXNLS are referred to
as RELAXNLS SAR images or simply RELAXNLS images. To form RELAXNLS
images, the dominant target features are first extracted by using the RELAXNLS
algorithm (see Section 3.2.1 for details) and then Steps 2 and 3 discussed in Section
3.1.2 are used except that the simulated phase history data of large dimensions used
in Step 2 is now determined by
Kd Kt
y,(ns, ,) = C & cdkSinc[7rbk (s k)]ej(2dk dkf)t + E &tk e(atk + n), (3.30)
k=1 k=l
28
where n, = 0, 1,  , (N 1 and i, = 0, 1, .. CN 1 with ( being an extrapolation
factor (C > 1); &dk, dk), Wdk, bk, and 7k, respectively, denote the estimates of adk,
Wdk, 7dk, bk, and Tk, k = 1, 2, *, Kd; finally, &t,, Wtk, and 2 t, respectively, denote
the estimates of atk, wik, and wtk, k = 1,2, *. Kt. The clutter estimate 6(n, n)
used in Step 3 is now determined by (3.29) with kt and kd replaced by Kt and Kd,
respectively. Note that scaling the simulated data for the dihedrals by a factor of C
in (3.30) is needed since the sinc function goes to zero as n, increases or decreases
away from Tk. Note also that C is a parameter of user choice and ( > 1 is just
used to demonstrate the super resolution property of RELAXNLS. Finally, since the
Fourier transforms of the sinc functions of sufficient lengths do not result in sidelobes,
when ys(ns, n,) in (3.30) is windowed, the second term of y,(ns, n,) is multiplied by
wz(ns, in) and the first term is multiplied by wi(n,), where wl(n,) is a 1D window
sequence satisfying
(N1
E w(n,) = (N. (3.31)
n,=o
3.3 Experimental Results
In this section, we demonstrate the image formation performances of RELAX
and RELAXNLS with the experimental data. In the following examples, the extrap
olation factor C = 2 is used and GAIC with y = 4 and y = 18 are used to determine
K for RELAX and RELAXNLS, respectively. Kaiser windows with shape parameter
/ = 6 are used whenever needed.
First consider SAR image formation via 2D FFT and RELAX by using a
portion of the 2D data corresponding to some roof rims collected by one of the two
apertures of the ERIM's (Environment Research Institute of Michigan's) DCS inter
ferometric synthetic aperture radar (IFSAR). Figures 3.1(a) and (b), respectively,
show the unwindowed and windowed 2D FFT images obtained by zeropadding the
40 x 40 phase history data. The unwindowed and windowed RELAX images with and
29
without the background clutter are illustrated in Figures 3.1 (c) ~ (f) with K = 59.
Note that the RELAX images have a higher resolution than the FFT images for the
dominant scatterers. Utilizing only 25% of the 40 x 40 phase history data, i.e., using
a 20 x 20 phase history data, we form the 2D FFT SAR images shown in Figures
3.2(a) and (b). Figures 3.2(c) ~ (f) show the RELAX images with K = 41. Compar
ing Figures 3.2(f) and 3.2(b), we note that the two images are quite similar although
the former uses only 25% of the data used by the latter.
Consider next an example of the MSTAR Slicy data consisting of both trihe
dral and dihedral corner reflectors collected by the SNL using the STARLOS sensor.
The field data was collected by a spotlightmode SAR with a carrier frequency 9.559
GHz and bandwidth 0.591 GHz. The radar was about 5 kilometers away from the
ground target shown in Figure 3.3. The SAR images are obtained when the tar
get is illuminated by the radar from the azimuth angle 0 and elevation angle 30.
The range and crossrange resolutions of the original data (54 x 54) are 0.3 meters
and 0.32 meters, respectively, and the windowed 2D FFT SAR image (obtained
via Taylor window) is shown in Figure 3.4(a). To demonstrate the super resolution
property of RELAX and RELAXNLS, we use part of the original data, whose di
mension is 32 x 32. Figure 3.4(b) shows the windowed 2D FFT images with the
spoiled resolution 0.51 x 0.54 meters obtained from this reduced size data. RELAX
and RELAXNLS are also applied to this 32 x 32 phase history data matrix. Figures
3.4(c) and (d), respectively, show the windowed RELAX SAR images with and with
out estimated background clutter with K = 36. We note that RELAX images have a
better resolution than the FFT images for the dominant trihedrals and resemble the
FFT images for dihedrals even though the data model in the crossrange dimension
used by RELAX is not correct for this example. The windowed RELAXNLS SAR
images with and without estimated background clutter with K = 9 are illustrated
in Figures 3.4(e) and (f), respectively. We note again that the RELAXNLS images
30
not only have a better resolution than the 2D FFT images, but also more closely
reflects the characteristics of the photo shown in Figure 3.3.
31
i . .
(a) (b)
410
(c) (d)
(e) (f)
Figure 3.1: RELAX SAR Images by using the 40 x 40 ERIM data with = 59.
(a) Unwindowed 2D FFT image. (b) Windowed 2D FFT image. (c) Unwindowed
RELAX image without background clutter. (d) Windowed RELAX image without
background clutter. (e) Unwindowed RELAX image with background clutter. (f)
Windowed RELAX image with background clutter.
32
(a) (b)
(c) (d)
(e) (f)
Figure 3.2: RELAX Images by using the 20 x 20 ERIM data with K = 41. (a)
Unwindowed 2D FFT image. (b) Windowed 2D FFT image. (c) Unwindowed
RELAX image without background clutter. (d) Windowed RELAX image without
background clutter. (e) Unwindowed RELAX image with background clutter. (f)
Windowed RELAX image with background clutter.
33
Figure 3.3: Target photo taken at 45 azimuth angle.
34
(a) (b)
(c) (d)
* 40 A*
(e) (f)
Figure 3.4: SAR images of the Slicy data (0 azimuth and 300 elevation angles).
(a) Original windowed 2D FFT image. (b) Windoed 2D FFT image with spoiled
resolution. (c) ~ (d): Windowed RELAX image without and with background clutter
with K = 36. (e) ~ (f): Windowed RELAXNLS image without and with background
clutter with K = 9. (c) ~ (f) are obatined using the data in (b). (The vertical and
horizontal axes are for range and crossrange, respectively.)
CHAPTER 4
SEMIPARAMETRIC METHODS FOR SAR IMAGING
4.1 Problem Formulation and Data Model
Compared with parametric spectral estimation methods, nonparametric spec
tral estimation methods are robust, but cannot be used to significantly improve the
resolution of the formed SAR images. As can be seen from the previous chapter that
relaxationbased parametric methods can be used to generate super resolution SAR
images when the data model for targets of interest is appropriate. However, the super
resolution property of parametric methods will be compromised when the appropriate
data model is not available. In this chapter, we consider the case where it is difficult
to establish a good parametric data model in crossrange for target scatterers due
to a variety of physical phenomena including glints, resonances, and motioninduced
phase errors, etc. Instead of modeling the target of interest as 2D sinusoid (point
scatterer) in both range and crossrange, we apply herein a semiparametric model
and simply assume that the data model in crossrange due to one or more corner re
flectors is a complex sinusoid with an arbitrary amplitude and a constant phase. Due
to the flexibility of this data model, feature extraction and image formation meth
ods based on this data model should be more robust compared with those based on
the approximated dihedral and trihedral data models in Chaper 3. The robustness,
however, comes at the cost of small resolution improvement in crossrange over the
conventional fast Fourier transform methods due to assuming little in crossrange. To
preserve the robustness, we use a refined nonparametric spectral estimation method
such as APES (Amplitude and Phase Estimation of a Sinusoid) algorithm [38] to
improve the crossrange resolution.
35
36
The received signal reflected from a target scatterer herein is modeled as:
s(n,h)= x()ejOej2(fnf), n= 0,1,,N 1, n = 0,1, ,N 1, (4.1)
where N and N denote the dimensions of the available data samples in range and
crossrange, respectively; x(n) is an arbitrary unknown realvalued function of h
determined by the radar cross section (RCS) of the scatterer; 0 is a constant phase;
finally, {f, f} is the frequency pair proportional to the 2D location (range and cross
range) of the scatterer. This data model is essentially semiparametric since little
parameterization is done in crossrange.
Assume that a target consists of K scatterers. Then the target data model in
the presence of noise has the form:
K
y(n, ) = Z xk()e k e3j2n(fkn+kf) +e(n, n), n = 0,1,  N1, n = 0,1, N1,
k=1
(4.2)
where {xk ( 0)}1 denotes the realvalued amplitude function of n for the kth scat
terer; Ok and {fk, fk}, respectively, are the constant phase and the frequency pair
of the kth scatterer; finally, {e(n, )} denotes the unknown 2D noise and clutter
sequence.
Since SAR images are often used in SAR applications, our problem of interest
herein is to estimate the target parameters {(k, Xk( ( k, fk= from the 2
D data sequence {y(n, n)} and then to form high resolution SAR images with the
estimated target parameters. Before presenting the effects of the semiparametric
data model in (4.2) on SAR image formation and the SPAR (SemiPARametric)
algorithm for feature extraction and image formation, we first briefly introduce the
APES algorithm.
37
4.2 Review of the APES Algorithm
The APES [38] algorithm belongs to the adaptive filterbank class, in which
the adaptive finite impulse response (FIR) filter is designed based on not only the data
to be processed, but also the interference at the frequency of interest. APES mimics
the maximum likelihood estimator of a complex sinusoid in circularly symmetric
zeromean Gaussian noise. It has been shown in [37] that APES has better statistical
performance than Capon [10], another popularly used adaptive filterbank approach
for SAR imagery.
Let {zn}[=0 denote the available 1D data sequence and have the form:
Z = a(w)exp{jnw} +eN(w), n= 0,1,. N 1, (4.3)
where a(w) denotes the complex amplitude of a sinusoid with frequence w and en(w)
denotes the unmodeled noise and interference at w. Then APES can be used to obtain
the complex amplitude estimate &(w) of a(w) from {zn}o1. It has been shown in
[38] that the leastsquares estimate of a(w) using the forward and backward FIR
filter output has the form:
&() = [hH(w)Z(w) + /(w)ZH(w)h(w)], (4.4)
where h(w) denotes the impulse response of an Mtap FIR filter and has the following
form:
h(w)= hi(w) h2(w) .. hM(w) ; (4.5)
Z(w) and Z(w), respectively, denote the normalized Fourier transforms of {Zi} M
and {zi} i.e., the Fourier transforms of {zi}N and {i,}m divided by (N
M + 1), where zi and i, respectively, denote the overlapping subvectors of the data
vector z = [ zO z ZN1 ] and the complex conjugated and backwardordered
data sequence z = z*,_ Z_2 ... z* and have the form
Zi= zi z i+ Zi+M1 i=0,1,,NM, (4.6)
38
and
ii= z*,_ z,_i2 iM i = 0,1, N M, (4.7)
[ ZNi1 ZNi2i
respectively; finally, /(w) = exp{j(N 1)w}. It has been shown in Appendix C
of [38] that by using a certain conjugate symmetric condition, &(w) in (4.4) can be
reduced to
&(w) = hH(w)Z(w). (4.8)
The adaptive FIR filter hAPES(W) corresponding to APES is derived in Ap
pendix A of [38] and has the form:
Q (w)a(w)
hAPES() = a( ( (4.9)
aH wQ19a^)
where the steering vector a(w) is given by
rT
a(w) ei ... ei(M1) (4.10)
and Q(w) denotes the estimate of the covariance matrix of the noise and is given by:
Q(w) = R Z(w)ZH(w), (4.11)
with
Z(w) = [ Z(w) Z(w)] (4.12)
and R = (R + R)/2 denoting the average of the forward and backward sample
covariance matrices R and R. The forward sample covariance matrix R is obtained
from zi as follow:
NM
R = NM 1 (4.13)
i=O
and the backward sample covariance matrix R can be obtained similarly from zi.
4.3 Data Model Ambiguities and Their Effects on SAR Image Formation
Target feature extraction methods devised based on the data model in (4.2)
are robust against data mismodeling errors due to the model flexibility. However,
39
there are ambiguity problems associated with the semiparametric data model. In
this section, we first analyze the possible data model ambiguities and then illustrate
their effects on SAR target feature extraction and image formation. The discussions
below will motivate the introduction of the SPAR algorithm, which will be presented
in detail in Chapter 4.4.
4.3.1 Model Ambiguities
Due to the flexibility of the data model in (4.2), there are various types of
ambiguities that may impact the feature extraction of each scatterer. Below we list
several types of the ambiguities inherent in the data model.
Type 1: Single scatterer
From (4.1), we note that ambiguity exists between and x(n) since
X(n)eo = x(f)ej(O+") ()e (4.14)
where r(n) = x(n) and 4 = + 7r. Ambiguity also exists between f and x(i) since
(n)e = (1)x(i)ej2"(f05)" )e2, (4.15)
where 1(n) = (1)"z(n) and f = f0.5. The above two types of ambiguities cannot
be resolved.
Type 2: Two identical scatterers located in the same range
Let fa and fb, respectively, denote the crossrange locations of two identical scatterers
and let 0a and Ob, respectively, denote their phases. Then
x(n) [e(ji+i2 + ei+i n)J 2x(fi) cos [r(h fb)f + 2 ] ej 2
e 7(ao+)^, (4.16)
which indicates that two identical corner reflectors (trihedrals or dihedrals) located
in the same range but different crossrange positions fa and fb, respectively, can be
modeled by (4.1) as one "scatterer" located at (fa + fb)/2 in crossrange with x(n)
40
modulated by 2 cos[7r(fa fb)n + (0a b)/2]. Thus the data model in (4.1) cannot
be used to describe each of the two corner reflectors in this case.
Type 3: Two different scatterers located in the same range
Assume that two different corner reflectors with parameters {4i, {xi()} f, fi }i=
are located in the same range. Then the target model in the absence of noise has the
form:
2
r(n,n)= Exi()ej'ej2 (fn++hn), n= 0,1,,N 1, = 0,1, ,N 1.
i=1
(4.17)
With straightforward calculations, we can rewrite (4.17) as
2
r(n, n) = Zi (A)eje2 ej2i"fifn ), (4.18)
i=1
where 2 = 1 + 7i/2 with 1 denoting an arbitrary phase, fi = f2 f f, = f2 = f
with f denoting an arbitrary crossrange location,
2
x,(n) = E i() cos[27r(f f)n + ( 1)], (4.19)
i=1
and
2
x2(n) = Xi(n) sin[2zr(f f)n + (Oi 01)]. (4.20)
Note from (4.18) that { i, {i()}N 'fi, fi =1 are the ambiguous features of
{<, {xi()}=o J, Ji=l
Type 4: Multiple scatterers located in the same range
When more than two scatterers are located in the same range, the data model in the
absence of noise can still be written as (4.18) except that
L
Zi(") = xi(A) cos[27r(fi f) + (q 1)], (4.21)
i=1
and
L
C2(l) = i(n) sin[2r(f f)[ + ( <1)], (4.22)
i=1
41
where L > 2 denotes the number of scatterers located in the same range. Thus the
L scatterers located in the same range are considered as two "scatterers" when using
the data model in (4.2).
Before we discuss the impact of the model ambiguities on SAR image for
mation, we first describe how the image formation is done if we have the estimated
model parameters.
4.3.2 Image Formation
Assume for now that we have extracted the target features based on any of
the ambiguous data models. For notational convenience, we will use the notation
used in (4.2). Since the target data model in range is a sum of several complex
sinusoids with constant amplitudes and phases, we can use the estimated sinusoidal
parameters to simulate a data matrix with a larger dimension in range and then use
FFT to demonstrate the super resolution property of the feature extraction algorithm
we shall present. Yet we cannot extrapolate the estimate {xk(n)} of {xk(n)} since it is
assumed to be an arbitrary unknown realvalued function of n and hence FFT cannot
be used to obtain SAR images with enhanced resolution in crossrange. Instead, we
use 1D APES [37, 38, 41] in crossrange when forming SAR images since APES
belongs to the class of matched filterbank spectral estimators and provides lower
sidelobes, narrower spectral peaks, and more accurate spectral estimates than FFT.
Let {s, (n, )} denote the simulated data sequence with a larger dimension in
range based on the estimated target features {k, {k(n)} =0, fk, fk}= Of
{(k, {xk(n)} 0 fk, k= where K denotes the estimate of the scatterer number
K. Then
k
s9(n, F) = 'k()ejke j2"(fk +kn), = 0,1,. CN 1, = 0,1, .., N 1,
k=1
(4.23)
42
where ( denotes an extrapolation factor (( > 1) and is a parameter of user choice.
Note that the super resolution property of the soformed SAR images is determined
by the feature extraction algorithm and C 2 1 is only used to demonstrate the super
resolution property of the target feature extraction algorithm. The estimated noise
and clutter data matrix is
(n, ) =y(n, ) ,(n, ), n = 0, 1, N 1, n = 0,1, N 1, (4.24)
which is also important in many SAR applications since, for example, important
target information such as the target shadow information is contained in 6(n, h). We
cannot extrapolate 6(n, n) in either range or crossrange since no parametric data
model is available for ((n, i).
To obtain SAR images with low sidelobes in range via 1D FFT, we apply 1D
windows to ,s(n, A) and &(n, n) in range. We obtain a new (CN) x N data matrix Y
as follows:
y(n, n) = s,(n, n)ws(n) + (e(n, n)we(n), n = 0, 1, ., N 1, n = 0,1, 1 ., V 1,
y(n,f) = s,(n,f)w,(n), n = N,N + 1, ,CN 1, n = 0,1, ,N 1,
(4.25)
where y(n, n) denotes the (n, n)th element of Y and w,(n) and we(n) are 1D windows
of lengths (N and N, respectively, satisfying
(N1
E w,(n) = CN, (4.26)
n=O
and
N1
we,(n) = N. (4.27)
n=0
Note that scaling 6(n, A) in y(n, n) by a factor of ( is necessary since the range
dimension of ,s(n, n) is ( times of that of 6(n, n). The steps needed for SAR image
formation are as follows:
Step (1): Form Y from y(n,n) by using (4.23), (4.24), and (4.25).
43
Step (2): Apply the normalized 1D FFT to each column of Y to obtain
an intermediate matrix and then apply 1D APES to each row of the intermediate
matrix. (See [41] for the efficient implementation of APES.) Note that the normalized
1D FFT has the form
1N1
S y(n, )ej2"fn, = 0,1, ,N 1. (4.28)
n=O
4.3.3 Model Ambiguity Effects on SAR Image Formation
All of the aforementioned types of ambiguities will have no effect on the SAR
image formation if no parameter estimation errors exist since the scatterers will be
perfectly reconstructed by using any of the possible ambiguous data models. For
example, when there are two identical scatterers located in the same range, the data
model in (4.2) can still be used for SAR image formation since the two scatterers are
now described as one "scatterer" with (4.16), which still fits the data model of (4.2)
with K = 1. Hence the original SAR image can still be reconstructed by using the
parameters of the one "scatterer" described by the right side of (4.16).
In the presence of parameter estimation errors due to the presence of noise
and clutter, however, Types 1 and 2 ambiguities discussed in Section 3.1 will have
little effect on SAR image formation, whereas Types 3 and 4 ambiguities can result in
artifact problems for the high resolution SAR image formation. Generally speaking,
the higher the SNR, the more accurate the parameter estimates and hence the less
significant the artifact problem. The effect of Type 3 ambiguity on SAR image
formation in the presence of estimation errors is illustrated in Figures 4.1 and 4.2.
(The effects are similar for Type 4 ambiguity.) Figure 4.1 is obtained by assuming
no parameter estimation errors. Figure 4.1(a) shows the FFT image of a target
consisting of two dihedrals of different lengths located in the same range. We use
Xk(n) = aksinc[bk7r(in Tk)], k = 1,2, n = 0,1, ... 31, (4.29)
44
k ak k fk fk bk Tk
k=l 9.6 0 0.1 0.3 0.3 18.6
k=2 6.4 0 0.1 0.1 0.2 18.6
Table 4.1: True parameters of the two dihedrals used in Figures 1 and 2.
to simulate the dihedrals, where ak and bk, respectively, are proportional to the
maximal RCS and the length of the kth dihedral corner reflector and Tk denotes the
peak location of the data sequence and is determined by the orientation of the kth
dihedral. The size of the simulated data matrix is 32 x 32 (i.e., N = N = 32).
The parameters for the two dihedrals are given in Table 1. An ambiguous set of
target features can be obtained by choosing fi = f2 = 0.1, fl = f2 = f = 0.2,
01 = 0, and 02 = 7r/2 in (4.19) and (4.20). The windowed FFT SAR images of
the two "scatterers" are shown in Figures 4.1(b) and (c), respectively, which differ
considerably from the two dihedral scatterers in Figure 4.1(a). The combined SAR
image of the two "scatterers" is given in Figure 4.1(d), which is exactly the same as
the true image shown in Figure 4.1(a). However, due to the presence of noise and
clutter, parameter estimation errors are inevitable. The errors in range are the main
cause of the artifact problem in the high resolution SAR image formation. In Figure
4.2, we assume that all parameters are accurate except that f2 = f2 + 0.01 = 0.11.
Figures 4.2(a) and (b), respectively, show the windowed FFT images of the two
aforementioned "scatterers" in the presence of estimation errors and Figure 4.2(c)
shows the combined SAR image. By comparing Figures 4.1(a) and 4.2(c), we note
that an extra line (artifact) shows up next to the short dihedral. The reason is that
due to the estimation errors, f, 5 f2. Hence the two "scatterers" in Figures 4.2(a)
and (b) are not exactly in the same range and cannot be "combined" perfectly to
obtain the two dihedral lines in Figure 4.1(a). This problem becomes even worse
when super resolution SAR images are formed via data extrapolation in range. The
45
larger the extrapolation factor (, the more significant the artifact problem since the
difference between fl and f2 is exaggerated C times. Figures 4.2(d) shows the SAR
image obtained with ( = 2. By comparing Figures 4.2(d) and (c) (here ( = 1,
without extrapolation), we note that the artifact next to the short dihedral becomes
more significant. Severe artifacts may exist at low SNR since the accuracy of the
parameter estimates is poor. The SPAR algorithm we present below attempts to
avoid this problem by using windows to isolate the multiple scatterers located in the
same range.
4.4 The SPAR Algorithm
SPAR can be summarized by the following two steps:
Step 1: Scatterer Isolation based Target Feature Extraction: See Chapter
4.4.1 below for details.
Step 2: SAR Image Formation: See Chapter 4.3.2 for details, where the
estimated target features are obtained by using Step 1.
4.4.1 Target Feature Extraction
The basic idea behind SPAR is to extract the features of each scatterer sepa
rately. Before we present the target feature extraction algorithm, we first summarize
the steps needed for the feature extraction of a single scatterer as a preparation. The
generalized Akaike information criterion (GAIC) is also introduced to estimate K,
the number of scatterers, at the end of this subsection.
Feature Extraction of a Single Scatterer
The data model of a single scatterer in the presence of noise has the form:
ys(nr ) = s(n,) + e,(n,n), n = 0,1,. *, N 1, n = 0,1, * , N 1, (4.30)
46
where s(n, ) is given in (4.1) and {es(n, )} denotes the unknown 2D noise and
clutter sequence. Let
WN(f)= 1 ej2rf ... e2if(N1) (4.31)
and
(f) = 1 ej2l ... ej2 (R1) (4.32)
where (.)T denotes the transpose. Let D(f) denote the following diagonal matrix:
D(f) = diag 1, ej2f, ..., ej2r(N) (4.33)
Define
x = [(0) x(1) .. x(SN 1) (4.34)
Let Y, be an N x N matrix with its (n, n)th element being y,(n, n). Then we can
rewrite (4.30) as:
Y, = ejOG(x, f, f) + E,, (4.35)
where
G(x, f, f) = WN(f)xD(f), (4.36)
and E, denotes an N x N matrix with e,(n, f) being its (n, n)th element. Let y,,
S= 0, 1, N 1, denote the nth column of Y, and define
,(f)= Y w(f), (4.37)
where (.)* denotes the complex conjugate. Then the NLS estimates f{, If, ff of
{x, 4, f, /} are (see Appendix A for the detailed derivations):
S= tRe [ejy,(f) 0 wL(f)], (4.38)
where Re(x) denotes the real part of x and 0 denotes the Hadamard matrix product
or the elementwise product of two matrices;
= arg [y *(f)]2 j2(21) (4.39)
I=0
47
where arg(x) denotes the argument of a complex variable x; finally,
{f, f} = argmax C4(f, f), (4.40)
where
N1 N1
C4(f,f) = y yw(f)2 N [] ej2(2f)" (4.41)
ft=0 1=0
The steps needed to obtain the NLS estimates of a single scatterer are summarized
as follows:
Step (I): Use (4.37) to obtain y(f) and obtain the cost function C4(f, f)
according to (4.41). Determine {f, f} by maximizing C4(f, f) using the method given
in Appendix A.
Step (II): Calculate according to (4.39) with {f, } replaced by {f, }
obtained in Step (I).
Step (III): Calculate x via (4.38) with {(, f,f} replaced by ({, f,f} deter
mined in Steps (I) and (II), respectively.
Feature Extraction of Multiple Scatterers
When a target consists of multiple scatterers, we can obtain the NLS esti
mates of the target features based on (4.2) by using a relaxationbased optimization
approach. Let
xk= xk(0) xk(1) .,k (N1) (4.42)
and let Y and E be N x N matrices with their (n, i)th elements being y(n, n) and
e(n, a), respectively. Then we can rewrite (4.2) as
K
Y = E ei Gk(xk, fk, fk) + E, (4.43)
k=1
where Gk(Xk, fk, fk) has the same form as the G(x, f, f) in (4.36) except that x, f,
and f are replaced by xk, fk, and fk, respectively. Let yn, f = 0, 1, N 1, be the
48
Ath column of Y. Then the estimates {Ik, Ak, f, k}j=1 of {k, Xkk, fk k=1 can be
obtained by minimizing the following NLS cost function:
K 2
C5 ({k, xk, Ak, fk}k=) = Y e3 Gk(xk, fk,k) (4.44)
k=1 F
where 1 pIF denotes the Frobenius norm [22]. The minimization of
C5 ({, Xkf, fk, f}=) in (4.44) is a very complicated optimization problem. The
proposed SPAR algorithm performs a complete relaxationbased search by letting
only the parameters of one scatterer vary and freezing all others at their most re
cently determined values for each assumed number of scatterers K. Let
K
Yk = Y e Gi(i, fi, f), (4.45)
i=l,i k
and assume that {i, ^i, fi, i}l,1,i k are given. Then the NLS estimates
{k, Xfk, fkk=1 f {k, Xk, k, fk}k= can be obtained by minimizing
C6 (k, Xk,fk, fk), where
C6(k,Xk, ffk) = Yk ejkk (Xk, fk, k) F, (4.46)
and using the method presented in the previous subsection for the feature extraction
of a single scatterer. However, when multiple scatterers are located in the same
range, the minimization of C6 k, Xk, fk, fk) has numerous ambiguous solutions that
may lead to the artifact problem in the high resolution SAR image formation.
SPAR attempts to avoid the ambiguity problem by isolating out the most
dominant scatterer in Yk by using a 2D rectangular window, which is determined
from and applied to the 2D FFT of Yk. The isolation process has the following
steps:
Step (i): Obtain Vk, the 2D FFT of Yk, without zero padding.
Step (ii): Determine the 2D window w(n, n) from Vk. We first locate the
peak location (n+, l+) of the magnitude of Vk. We then fix n to n+ and search for
49
the interval nl <_ n+ < n2 so that the magnitude of Vk is above a certain threshold,
say Ti, within the interval. Similarly, we can fix n to n+ and search for the interval
nil < n+ < in2. Then the N x N rectangular window w(n, n) has unit value for
nl < n < n2 and nl < n < n2 and zero value elsewhere. The threshold Ti we use in
our numerical and experimental examples is 10% of the peak value of the magnitude
of Vk.
Step (iii): Determine Yk by applying 2D inverse FFT (IFFT) to Vk 0 W,
where the (n, n)th element of W is w(n, n).
Instead of minimizing C6(k, Xk, fk, fk), we now minimize
C7(k,Xk, fkk)== Yk e 3kGk(xk,fk, fk) 2, (4.47)
where Yk is used to replace Yk in C6G(k, k, fk, fk), by using the method presented
in Section 4.1.1.
With the above preparations, now we provide the steps of the scatterer isola
tion and relaxation based NLS algorithm, which are the substeps of Step 1 of SPAR.
Step I: Assume K = 1. Calculate Y from Y by using the isolation process.
Obtain {&, xk, fk, fk}k=1 from Y.
Step II: Assume K = 2. Compute Y2 with (4.45) by using {(k, k, fk, fk}k=1
obtained in Step I. Calculate Y2 from Y2. Obtain {(,k, k, k, fk k k=2 from Y2. Next
compute Y1 with (4.45) by using {k, Xk, k, fk}kk=2, calculate Y1 from Y1, and then
redetermine { k, fk, fk}kk=1 from Y'.
Iterate the previous two substeps until "practical convergence" is achieved (to
be discussed later on).
Step III: Assume K = 3. Compute Y3 with (4.45) by using {Jk, Xk, k, fk} =1
obtained in Step II. Calculate Y3 from Y3. Obtain {(, k, fk, 1fk}k=3 from Y3. Next,
compute Y1 with (4.45) by using {(k, Xk, fk, fk} =2, calculate Y1 from Y1, and then
redetermine {(,Ck,^ k, fk) }k= from Y1. Then compute Y2 with (4.45) by using
50
{ kc,x k, fk}ck=1,3, calculate Y2 from Y2, and then redetermine {(k,k, xk, f fk}k=2
from Y2.
Iterate the previous three substeps until "practical convergence".
Remaining Steps: Continue similarly until K is equal to the desired or
estimated number of scatterers.
The "practical convergence" in the iterations of the above relaxationbased
NLS algorithm may be determined by checking the relative change e of the cost
function C5 ({I kf, fk, fk}k=i) in (4.44) between two consecutive iterations. Our
numerical and experimental examples show that the algorithm usually converges in
a few iterations.
We can determine K, the number of scatterers in (4.2), by extending the
generalized Akaike information criterion (see [61] for details). By assuming that the
noise is white, the estimate K of K is determined as an integer that minimizes the
following extended GAIC cost function:
GAICk = NNln E 1: (n, n)2) + yn[lny(NR)][((N + 3) + 1], (4.48)
Sn=0 ni=0
where y is a constant of user choice,
K
ek(n, n) = y(n, n) xSk(n)ekej2 (fk+f), n = 0,1, .., Nl, f = 0,1, .., ,1,
k=1
(4.49)
and K(N + 3) + 1 is the total number of realvalued unknown parameters (of which
K(N + 3) is for the scatterers and 1 is for the white noise variance).
Note that the NLS estimates of {(k, Xk, fk, kK=1 can also be estimated from
{Yk}k=1 determined in (4.45), rather than {Y}1,k=, via the above relaxationbased
optimization algorithm. We refer to this approach as the Hybrid method. When
multiple scatterers are located in the same range, Hybrid may be computationally
more efficient than SPAR since Hybrid does not isolate the scatterers so that multiple
scatterers located in the same range can be more efficiently described as at most two
51
"scatterers". However, the Hybrid SAR images may suffer from more severe artifact
problem than SPAR, especially at low SNR. When no multiple scatterers are located
in the same range, SPAR and Hybrid perform similarly.
4.5 Modified RELAXNLS Algorithm
RELAXNLS [42] is a parametric approach for the feature extraction of targets
consisting of both trihedrals and dihedrals. It is based on a mixed data model in
which x(n) is modeled as a realvalued constant for a trihedral or a sinc function of f
for a dihedral. Like SPAR, RELAXNLS extracts the target features by minimizing
an NLS cost function via a relaxationbased approach. However, RELAXNLS is
computationally expensive since a 4D search over the parameter space is required
for dihedral corner reflectors. Since SPAR is more robust and computationally more
efficient than RELAXNLS, the former can be used to provide the initial conditions
needed by the latter.
Let {(k, Xk, k kfk}k=1 denote the parameter estimates obtained via SPAR
according to the data model in (4.2), where K is the estimated number of scatterers
obtained via GAIC. The SPAR estimates {0k, xk, fk, fk}k= cannot be used directly
as initial conditions for RELAXNLS. The initial conditions are obtained by applying
the first step of RELAXNLS [42] to each Uk, k = 1, 2, ... K, where
K
Uk = ejiGi,(xii,,fi), k = 1,2,,K. (4.50)
i=l,i k
Once we have the initial conditions, we can use the last step of RELAXNLS [42] to
obtain the dihedral and trihedral parameter estimates, which are then used for SAR
image formation [7]. We refer to this approach as the modified RELAXNLS.
52
4.6 CrambrRao Bound (CRB) of the SemiParametric Data Model
Let Y be an N x N matrix with its (n, n)th element being y(n, n) with
K
y(n, 5) = E k (n)ejkej2(fkn+n) +e(n, ), n = 0,1, **,N1, = 0,1, N1,
k=1
(4.51)
where K denotes the number of scatterers; {xk(n)}0=o denotes the realvalued am
plitude function of f for the kth scatterer; 'k and {fk, fk}, respectively, are the
constant phase and the frequency pair of the kth scatterer; finally, {e(n, f)} denotes
the unknown 2D noise and clutter sequence. Then we can rewrite (4.51) as:
K
Y e= eJk (fk)xTD(fk) + E, (4.52)
k=l
where E denotes an N x N matrix with e(n, f) being its (n, f)th element,
WN(fk)= 1 e j2fk e2 jfk(N) (4.53)
(fk) = 1 ej2fk .. ej2*(7 ) (4.54)
with (.)T denoting the transpose, D(fk) denoting the following diagonal matrix:
D(fk) =diag{ 1, ej2f *.., ej2k(N1) (4.55)
and
Xk = [ xk(0) xk(1) *.. xk(i1) (4.56)
Let y = vec(Y) = [y yT ... y ]T with yn(A = 1,2,..,N) being the Ath
column of Y. Then the data model can be rewritten as follows:
y = A/ + e, (4.57)
where
A= [ [(fl) 0 X1] WN(fl) *.. [N(JK) YXKI WN(fK) (4.58)
53
with 0 and , respectively, denoting the elementwise product and Kronecker prod
uct of two vectors,
p = e 1 ... eOK (4.59)
and e = vec(E).
Next, we derive the CRB matrix for the data model in (4.57) where the noise
covariance matrix is assumed to be arbitrary and unknown. Let Q = E{eeH} be
the noise covariance matrix with (.)H denoting the conjugate transpose. Then the
extended SlepianBangs formula for the ijth element of the Fisher information matrix
(FIM) has the form [3, 64]:
{FIM}ij = tr (Q Q'Q, + 2Re [(HAH) Q1 (At)j] (4.60)
where X' denotes the derivative of X with respect to the ith unknown parameter,
tr(X) denotes the trace of X, and Re(X) denotes the real part of X. Note that FIM
is block diagonal since Q does not depend on the parameters in (Api), and (Api)
does not depend on the elements of Q. Hence the CRB matrix for the target features
of interest can be calculated from the second term on the right side of (5.31). Let
S= O T fT T xe (4.61)
where
0= 01 02 O/K ] '
f= [f f2 fK ]
]T
L= f 2 ... / K 7
and
xe= [xT xi .. xK (4.62)
Let FI, F2, and F3 denote NN x K matrices whose kth columns are determined,
respectively, by
jei"" {[CW (fk) 0 Xk] WN(fk)} (4.63)
54
j27re k { [k~(f) 0 Xkj (WN(f) O dN)} (4.64)
with
dN = 0 1 N1 (4.65)
and
j27re k {[R (fk) 0 Xk 0 dN] 0 WN(fk)} (4.66)
with
d = 0 1 .. 1 (4.67)
Let F4 denote an NN x KN matrix whose [(k 1)N + n]th column is determined
by
e jk {[wf,,0 (f)]r wN(fk)}, k=1,2,...,K,=l, ,2,...,N, (4.68)
where wn is an N x 1 vector with the nth element being unit and all other elements
being zeros. Let
F= F1 F2 F3 F4], (4.69)
Then the CRB matrix for the parameter vector qr is given by:
CRB(l) = [2Re(FHQF)]1. (4.70)
4.7 Numerical and Experimental Results
We demonstrate and compare the SAR image formation performances of
SPAR and the modified RELAXNLS algorithm with both simulated and experi
mental examples. The algorithms are also compared with Hybrid and RELAXNLS.
In the following examples, the dimensions of the original SAR phase history data
matrix are N = N = 32 and GAIC with y = 5.5 is used to determine K for the
relaxationbased feature extraction algorithms of SPAR and Hybrid. The threshold
Ti used in the isolation process of SPAR is 10% of the peak value. The maximization
of C4(f, f) in (4.41) is done in two steps. First, initial frequency estimates f and f
55
are obtained via 1D FFT with zeropadding to a total length of 128 in range and to a
total length of 64 in crossrange. Next, these initial estimates are refined by using the
FMIN function in MATLAB alternately, i.e., by updating f while fixing f at its most
recently determined value and vice versa, until "practical convergence", which is de
termined by checking the relative change of the cost function C4(f, f). We have used
103 to determine the convergence of this fine search as well as the relaxationbased
algorithm. The extrapolation factor ( = 8 is used in range for SPAR and Hybrid
and in both range and crossrange for RELAXNLS and the modified RELAXNLS.
Both 1D and 2D Kaiser windows with shape parameter 3 = 6 are used whenever
needed.
First consider a simulated example with high SNR. The SAR phase history
data matrix is simulated by assuming that there are four trihedrals and three di
hedrals in the presence of zeromean white complex Gaussian noise with variance
a, = 0.6. The amplitude functions for the four trihedrals are generated as follows:
Xk(n) = k =1,2,3, = 0, 1, ,N1, (4.71)
and
x4(t})=2, =0, 1, ,N (4.72)
The amplitude functions for the three dihedrals are
x5(n) = 9.6sinc[0.3r(f 18.6)], n = 0, 1, N 1, (4.73)
and
Xk(r) = 6.4sinc[0.27r( 18.6)], k =6,7, f = 0,1, N 1, (4.74)
where sinc(x) = sin(x)/x. Figure 4.3(a) shows the modulus of the true SAR image.
Note that two of the dihedrals are located in the same crossrange and are closely
spaced in range and two of them are located in the same range. Of the four trihe
drals, two of them are closely spaced in range and the other two are located in the
56
same range. In this example, Hybrid and SPAR have the same estimated number of
scatterers, K = 7. Figure 4.3(b) shows the windowed 2D FFT SAR image obtained
by applying the normalized 2D FFT to the windowed data matrix. SAR images
formed via Hybrid, SPAR, RELAXNLS and the modified RELAXNLS algorithms
are shown in Figures 4.3(c) through 4.3(f), respectively. We note that at high SNR,
the Hybrid image is similar to the SPAR image. Both of the parametric RELAXNLS
and the modified RELAXNLS algorithms outperform their semiparametric coun
terparts SPAR and Hybrid since the data model used by the parametric methods is
correct rather than approximate. For this example, our simulations show that the
ratios between the MATLAB flops needed by Hybrid, SPAR, the modified RELAX
NLS, and RELAXNLS over the flops needed by the windowed FFT method are 27.4,
28.4, 50.1, and 70.8, respectively. Note that both SPAR and Hybrid are computa
tionally more efficient than RELAXNLS and the modified RELAXNLS, with the
modified RELAXNLS being more efficient than RELAXNLS.
Consider next a simulated example with low SNR. The SAR phase history
data is the same as in the above example except that the noise variance is increased
to a = 6. SAR images obtained by using the windowed 2D FFT, Hybrid, SPAR,
RELAXNLS, and the modified RELAXNLS are shown in Figures 4.4(a) through
4.4(e), respectively. Note that the artifact problem starts to show up in the Hybrid
image in Figure 4.4(b) due to large parameter estimation errors. By comparing
Figures 4.4(c) with (b), it can be seen that SPAR can effectively mitigate the artifact
problem.
Finally, consider an experimental example of SAR image formation by using
the MSTAR Slicy data collected by imaging an object consisting of both trihedral
and dihedral corner reflectors, which is shown in Figure 3.3. The data was collected
by the Sandia National Laboratory (SNL) using the STARLOS sensor. The field
data was collected by a spotlightmode SAR with a carrier frequency of 9.559 GHz
57
and bandwidth of 0.591 GHz. The radar was about 5 kilometers away from the
ground object. The data was collected when the object was illuminated by the radar
from approximately the azimuth angle 0 and elevation angle 30. To crossvalidate
the experimental results given below, XPATCH [1], a high frequency electromagnetic
scattering prediction code for complex 3D objects, was used to generate very high
resolution phase history data for the object shown in Figure 3.3. The data generated
by XPATCH has a resolution of 0.038 meters in both range and crossrange, and the
corresponding windowed FFT SAR image is shown in Figure 4.5(a). (We have used
the log scale for all of the images shown in Figure 4.5.) The original experimental
Slicy data has a resolution of 0.3 meters in range and 0.32 meters in crossrange.
The 32 x 32 data matrix we used to demonstrate the performance of our algorithms
has a spoiled resolution of 0.51 meters in range and 0.54 meters in crossrange. The
windowed 2D FFT SAR image of this data matrix is shown in Figure 4.5(b). Figure
4.5(c) shows the Hybrid image with K = 7 (obtained via GAIC). Figure 4.5(d)
shows the SPAR image with K = 7 (obtained via GAIC). Figures 4.5(e) and (f)
show the SAR images obtained via RELAXNLS and the modified RELAXNLS
with K = 7, respectively. Note that Hybrid has a more severe artifact problem
than SPAR. Note also that the SPAR image shown in Figure 4.5(d) appears to fit
Figure 4.5(a) and the characteristics of the object in Figure 3.3 well. However, the
parametric algorithms are not as robust as SPAR since the parametric images shown
in Figures 4.5(e) and (f) do not fit Figure 4.5(a) as well with one of the scatterers
misidentified and mislocated. For this experimental example, the ratios between the
MATLAB flops needed by Hybrid, SPAR, the modified RELAXNLS, and RELAX
NLS over the flops needed by the windowed FFT method are 29.2, 16.7, 32.8, and
43.1, respectively. Note that SPAR can sometimes be faster than Hybrid!
58
(a) (b)
(c) (d)
Figure 4.1: Ambiguity effect on the SAR image formation in the absence of range
estimation errors. (a) True windowed FFT SAR image. (b) Windowed FFT image
of the first "scatterer". (c) Windowed FFT image of the second "scatterer". (d)
Combined windowed FFT image of the two "scatterers". (The vertical and horizontal
axes are for range and crossrange, respectively.)
59
(a) (b)
(c) (d)
Figure 4.2: Ambiguity effect on the SAR image formation in the presence of range
estimation errors. (a) Windowed FFT image of the first "scatterer". (b) Windowed
FFT image of the second "scatterer". (c) Combined windowed FFT image of the
two "scatterers" with C = 1 (without extrapolation). (d) Combined windowed FFT
image with ( = 2. (The vertical and horizontal axes are for range and crossrange,
respectively.)
60
(a) (b)
(c) (d)
(e) (f)
Figure 4.3: Comparison of SAR images formed using different algorithms for simu
lated data at high SNR (oa = 0.6). (a) True SAR image. (b) Windowed 2D FFT
SAR image. (c) Hybrid SAR image. (d) SPAR SAR image. (e) RELAXNLS SAR
image. (f) Modified RELAXNLS SAR image. (The vertical and horizontal axes are
for range and crossrange, respectively.)
61
(a) (b)
(c) (d)
(e)
Figure 4.4: Comparison of SAR images formed using different algorithms for simu
lated data at low SNR (ou = 6). (a) Windowed 2D FFT SAR image. (b) Hybrid
SAR image. (c) SPAR SAR image. (d) RELAXNLS SAR image. (e) Modified
RELAXNLS SAR image. (The vertical and horizontal axes are for range and cross
range, respectively.)
62
(a) (b)
(c) (d)
(e) (f)
Figure 4.5: Comparison of SAR images obtained via different algorithms for the
Slicy data hb15533.015 (0 azimuth and 300 elevation angles). (a) Windowed 2D
FFT SAR image from the XPATCH data. (b) Windowed 2D FFT SAR image from
the Slicy data. (c) Hybrid SAR image. (d) SPAR SAR image. (e) RELAXNLS
SAR image. (f) Modified RELAXNLS SAR image. (c) ~ (f) are all obtained from
the data used in (b). (The vertical and horizontal axes are for range and crossrange,
respectively.)
CHAPTER 5
3D TARGET FEATURE EXTRACTION USING CLSAR
5.1 CLSAR Data Model
We describe how to obtain 3D target features via the full synthetic aperture
shown in Figure 1.2(a), which makes a preparation for the discussion on CLSAR. We
start by establishing onedimensional (1D) point scatterer data model of high range
resolution radar as a preparation for the establishment of 3D data model herein.
5.1.1 1D Data Model for High Range Resolution Radar
The range resolution of a radar is determined by the radar bandwidth. To
achieve high resolution in range, the radar must transmit wide band pulses, which
are often linear FM chirp pulses [29, 49]. A normalized chirp pulse can be written as
s(t) = ej(2fot+7t), ItI To/2, (5.1)
where fo denotes the carrier frequency, 27 denotes the FM rate, and To denotes the
width of the pulse. We assume that fo, 7, and To are known. The signal returned by
a scatterer of a target has the form
r(t) = 6,ej[2fo(t7)+(t7)2, (5.2)
where 6, is determined by the RCS of the scatterer and r denotes the roundtrip
time delay. The demodulated signal d(t) is obtained by mixing r(t) with s*(t o)
for some given T0 = 2Ro/c, where (.)* denotes the complex conjugate, c denotes the
propagation speed of the transmitted electromagnetic wave, and Ro denotes the range
distance from radar to the center of the illuminated patch.
d(t) = 6,ej2(rfo7~o)(T7To) e 7(r0)2 ej27(y70)t. (5.3)
63
64
The term ej3('T0)2 in (5.3) is usually close to a constant for all Tmin T, T0 Tmax,
where Tmax and Tmin correspond to the maximum and minimum values, respectively,
of the roundtrip time delays between the scatterers of a target and the radar. This
term can also be partially removed [49]. Let D(w) denote the Fourier transform of
d(t). Then the inverse Fourier transform of D(w)eT will have the term ejr('ro)2
removed. Yet this removal can only be approximate since d(t) is not known for all
t and hence D(w) is not known exactly. The closer e7j(T7o)2 is to a constant for
Tmin < T < Tmax, the better its removal. With this removal, we have
d(t) = 6,ej2(fo77Y)(T7)ej2(7t )t, (5.4)
which is a complex sinusoid with frequency 27(77o) and amplitude 6,ej2(foy770)( ).
We know Tmax and Tmin approximately since we assume that the altitude, antenna
beamwidth, and grazing angle of the radar are known. We also assume that (Tmax 
Tmin) < To. Then for To/2 + Tma, < t < To/2 + Tmin, the scatterers of the target at
different ranges correspond to different frequencies of the signal d(t), while the RCS's
of the scatterers are proportional to the amplitudes of the corresponding sinusoids.
The ranges and RCS's of the target scatterers are the 1D target features.
5.1.2 Full Synthetic Aperture Radar
A broadside data collection geometry in a spotlightmode SAR is shown in
Figure 5.1 [29]. The XYZ coordinate system is centered on a small patch of ground,
where a target is located. The ground is illuminated by a narrow radio frequency
(RF) beam from the moving radar that rotates (with radius Ro) around the coordinate
origin. In Figure 5.1, R denotes the distance between the radar and a scatterer at
the position (x, y, z), and 0 and are the azimuth and elevation angles of the radar
relative to the XYZ coordinate system. We assume that 0, , and Ro are known.
The range R of the scatterer located at (x, y, z) can be written as
R = [(Ro cos cos 0 x)2 + (Ro sin cos y)2 + (R sin )2 1/2 (5.5)
65
Radar
R Scatterer Z
(x,y,z)
RO
IX 0
Y
Figure 5.1: Broadside spotlightmode SAR data collection geometry.
It has been shown in Appendix 3B that R can be approximated as
R Ro x cos 0 cos y sin 0 cos q sin (5.6)
where
o = Ro + [(x2 z2) sin 0o cos o 2xz cos2 Q0] ( o) (5.7)
Ro = Ro + (5.7)
2Ro
(y2 + 2) cos + 2xz sin (5.8
x= x +2R' (5.8)
2Ro
xy cos o0 + yz sin (o
y = Y + (5.9)
Ro
and
(x2 + y2) sin (o
z = z 2 (5.10)
2Ro
Note that the second terms of the right sides of (5.8), (5.9), and (5.10) are due to
the range and elevation curvature effects and can be neglected for large Ro. Let TO
= 2Ro. Since 7 = 2, then from (5.4), we have
C C
d(t, 9, ) = x,y,z exp Ij4(7rf f++t) (Ro R o cos cos y sin 0 cos
i sin )] (5.11)
66
where 5,y,z is proportional to the RCS of the scatterer located at (x, y, z). For
4(7rfo 770 + yt)(Ro Ro)/c < 27, where It T0o < To, we can write (5.11) as
d(t, 0, ) 6Xy,zej(tx+ty+itz), (5.12)
where
4(7rfo 770 + 7t) cos 0 cos (5
tx = (5.13)
S4(7rfo 770 + yt) sin 0 cos (5
ty = (5.14)
C
and
S 4(r fo 7To + yt) sin (5.15)
c
Note that d(t, 0, ) is a 3D complex sinusoid. The frequencies of the 3D sinusoid
correspond to the 3D location (J, y, 2) of the scatterer, while the amplitude is propor
tional to its RCS. Note that (2, y, 5) is not the true location (x, y, z) of the scatterer,
but is close to (x, y, z) for large Ro. (See the Appendix 3C on how to calculate (x, y, z)
from (2, y, 2).) When a target has multiple scatterers, d(t, 0, ) in (5.12) will be a sum
of sinusoids. The 3D locations and RCS's of the target scatterers are the 3D target
features. Since usually the samples on the t, 0, and 4 axes are uniformly spaced, the
samples of tx, ty, and tz occur at the points of a polar grid. Hence PolartoCartesian
interpolation may be needed for the data samples to occur at rectangular grid points.
(See Section 5.4.2 for an alternative approach.)
After PolartoCartesian interpolation and sampling, the signal obtained by
the 3D full aperture SAR in the presence of noise can be written as:
K
y(n, 5) = akej(ikn+k+k ") + e(n, i, i), (5.16)
k=1
where n = 0,1,*, N 1, n = 0,1, N 1, and n = 0,1, .., N 1, with
N, N, and N denoting the numbers of available data samples in the three dimen
sions; K denotes the number of sinusoids; ak and {Wk, ck, CO}) denote the unknown
67
complex amplitude and 3D frequencies of the kth sinusoid, k = 1,2, K; fi
nally, {e(n, i, n)} denotes the unknown noise sequence. The sinusoidal frequencies
{Wk = 27rfk, k = 27rfk,, k = 27rfk} correspond to the 3D location of the kth
scatterer of a radar target and ak is determined by its RCS.
Let
Y = [y(0,0,0) y (1,0,0) y(N 1,0,0) y(0,1,0) y(l,l,0) ...
y(Nl,1,0) ...... y(0,N1,Nl) y(1,N1,N1)
T
... y(N1,N1,7N1)
(5.17)
Let e be defined similarly from e(n, n, n) as y from y(n, f, n). Then the data model
in (5.16) can be rewritten as:
y = Aa + e, (5.18)
where
S= a, C2 * K (5.19)
and
A = a[, i 0 a ... aK K aK (5.20)
where denotes the Kronecker product [22] and
1 1 1
ejwk ejCk eiwk
ak = a, k = k = (5.21)
ej(Y1)wk ej( 1)k ej(rl)1)k
5.1.3 Curvilinear SAR
Assume that the curvilinear aperture used in CLSAR consists of M different
radar viewing angles. Then the received data vector y, in CLSAR, is an (MN) x 1
subvector of y in (5.18). Let Ic denote an M x (NN) matrix with each column and
68
row containing only one nonzero unit element corresponding to the locations of the
available data samples. Then
Yc = Aca + e, (5.22)
where
Ac= {Ic(ai 9 )} a *.. {Ic(aa K a )} K aK (5.23)
The unknown sinusoidal parameters {wk, wk, k, Ok, ak K=1 are the features of our
interest and are to be estimated from the y, collected by CLSAR. In the following
sections, we will consider parametric 3D target feature extraction using CLSAR.
5.2 The RELAX Algorithm
The RELAX algorithm [39] can be extended to extract 3D target features. We
first consider using RELAX with the full aperture for 3D target feature extraction
and then extend this approach to the case of curvilinear apertures.
5.2.1 Full Aperture
The RELAX algorithm [39] minimizes the following NLS cost function:
C4 = ly Aal2, (5.24)
where  denotes the Euclidean norm. When the noise e(n, i, n) is the zeromean
white Gaussian random process, the NLS estimates obtained with RELAX coincide
with the ML estimates of the target features. When the noise is colored, the NLS
estimates are no longer the ML estimates, but they still possess excellent statistical
performance [39].
We present below the relaxationbased minimization approach that leads to a
conceptually and computationally simple method. For each fixed K, the intermediate
number of scatterers, we perform a complete relaxationbased search by letting only
the parameters of one scatterer vary and freezing all others at their most recently
determined values. In this way, we will also take advantage of the fact that the
69
parameter estimates for the first K 1 scatterers can be used to initialize the search
for the parameters of the Kth one. We first briefly make a preparation for the RELAX
approach. Let
K
yk =y Y di (i, ( ), (5.25)
i=1,i k
where fi, aj, and ai are formed, similarly to those in Equation (5.21), from ji, wi,
and ij, respectively, and {ci, S, 2i, di} l,ifk are assumed to have been estimated.
Then minimizing C4 in (5.24) with respect to ak yields the estimate &k of ak:
[ak 0 k 0 aka Yk
k = [ k ak]Hyk (5.26)
NNN
Wk =Wk ,k =Wk ,Wk =Wk
and
2
Wk, Wk, k = arg max [a&k ( ak a k (5.27)
Wk ,Wk ,Wk
Hence {Wk, wk, wk} can be obtained as the location of the dominant peak of the 3D
periodogram,
2
[ak O ak ak]H Y
NNN
which can be computed efficiently with 3D FFT. Note that padding with zeros
for the 3D FFT is necessary to achieve high accuracy. An alternative approach is
to find an approximate location corresponding to the global maximum with the 3D
FFT without much zeropadding and then use the approximate location as the initial
condition to find a more accurate position via a multidimensional search method, such
as the FMINV function in PVWAVE. We used the latter approach in our examples
presented in Section 5.4.2. Note that &k is easily computed from the complex height
of the peak of [ak afk 0 ak]H yk/(NNN).
With the above preparations, we now proceed to present the steps of the
RELAX algorithm for 3D target feature extraction with the full aperture SAR.
Step (1): Assume K = 1. Obtain {Lk,WkWk,kk}k=1 from y by using (5.27) and
(5.26).
70
Step (2): Assume K = 2. Compute y2 with (5.25) by using {)k, k,Wk, 7 k}k=1
obtained in Step (1). Obtain { k, k, k, ,k}k=2 from Y2. Next, compute yl by using
{Wk,i k, k, & k}k=2 and then redetermine {W(k, Wk, k, &k}k=1 from yl.
Iterate the previous two substeps until "practical convergence" is achieved (to
be discussed later on).
Step (3): Assume K = 3. Compute Y3 by using {Ok, W k, Wk, ak} =1 obtained in Step
(2). Obtain { k, k, &k}k k=3 from y3. Next, compute yj by using {,k, w, wk, &k, l =2
and redetermine {Ik, Wk, k,&k}k=1 from yi. Then compute Y2 by using
{wk, Wk, Wk, &k}k=1,3 and redetermine {Wkk, Wk, k, k}k=2 from Y2.
Iterate the previous three substeps until "practical convergence".
Remaining Steps: Continue similarly until K is equal to the desired or estimated
number of sinusoids. (Whenever K is unknown, it can be estimated from the available
data, for instance, by using the generalized Akaike information criterion (GAIC) rules
which are particularly tailored to the RELAX method of parameter estimation. See,
e.g., [39].)
The "practical convergence" in the iterations of the above RELAX method
may be determined by checking the relative change e of the cost function
C4 ({CJk, W k, &k }=1) in (5.24) between two consecutive iterations. Our numerical
examples show that the iterations usually converge in a few steps.
5.2.2 Curvilinear Aperture
The RELAX algorithm for the curvilinear aperture is similar to that of the
full aperture except that Equations (5.26) and (5.27) above are replaced by
k [{Ic(ak ak)}M ak]H Yck
k =, (5.28)
MN
wk =cWk =W k ,Wktk =k
and
I 7 2
{Ck, kk} = arg max [{Ic(ak 0 ak)} ak] yCk (5.29)
Wk ,Wk ,Wk
71
respectively, where
Yck =Yc i [{Ic( o )} (5.30)
i=l,i k
Let 'k, be similar to Yk except that the elements in Yk that are missing in Yck are
replaced with zeros in YCk. Then the right hand side of (5.29) can also be computed
by applying 3D FFT to Sc.
5.3 Performance Analysis of Parameter Estimation via CRB
Consider first the case of the full aperture. Let Q = E{eeH} be the noise
covariance matrix, which is arbitrary and unknown. The extended SlepianBangs
formula for the ijth element of the Fisher information matrix (FIM) has the form
[3, 64]:
{FIMj}i = tr (Q'QQ Q;) + 2Re [(aHAH Q (Aa)] (5.31)
where X' denotes the derivative of X with respect to the ith unknown parameter,
tr(X) denotes the trace of X, and Re(X) denotes the real part of X. Note that FIM
is block diagonal since Q does not depend on the parameters in (Aa), and (Aa)
does not depend on the elements of Q. Hence the CRB matrix for the target features
of interest can be calculated from the second term on the right side of (5.31). Let
[ = Re(ca) ImT(a) WT ~T T], (5.32)
where
W = W1 U02 '" WK (5.33)
r YT
S= 1 2 * K (5.34)
and
C = 01 o2 ... I K (5.35)
72
Let
F= A jA D,) Do, DI ,] (5.36)
where A is defined in (5.20), the kth columns of D,, Dc, and D@ are 9O[k 0 ak 0
ak]/owk, 9[ak 0 ak 0 akl]/&k, and O[ak 0 ak 0 ak]/aw3k, respectively, and
S=diag{ al, o2, *.** K }. (5.37)
Then the CRB matrix for the parameter vector l is given by:
CRB() = [2Re(FHQ1F)]1. (5.38)
For the case of curvilinear aperture, the CRB matrix for the target parameters
is similar to the one in (5.38) except that the A in (5.38) is now replaced by Ac.
5.4 Numerical and Experimental Results
We first use CRBs to investigate the performances of different curved apertures
for target feature extraction and then demonstrate with an experimental example the
performance of the RELAX algorithm for 3D target feature extraction using CLSAR.
5.4.1 Performance Analysis of Different Curvilinear Apertures via CRBs
Since CRBs are the best unbiased performance any asymptotic estimator can
achieve, we not only compare CRBs of target features when different curved apertures
are used for 3D target feature extraction, but compare them with the CRB of the
full aperture. Without loss of generality, let us consider the case of a single scatter
with a = 1 and w = D = c = 0. The additive noise is assumed to be zeromean
white Gaussian with variance a2 = 40. The full aperture data is generated according
to (5.16) with N = N = N = 32.
Consider first the example of M = 63. The curved apertures we consider
include the parabolic one in Figure 1.2(b), the Lshaped one in Figure 1.2(c), and
the ones in Figure 5.2. For the arc apertures, the Arc1 aperture is a quarter of a
73
4 .** e. A **.
Radius % *
.
'*
S
(a) (b) (c)
Figure 5.2: Curvilinear apertures, a) The circular aperture. b) The Arc1 aperture.
c) The Arc2 aperture.
Aperture CRB(a) CRB(w) CRB(&) CRB(D)
Full 5.9875 51.4511 51.4511 51.4511
Circular 5.3663 39.3415 40.8469 40.8788
Parabolic 5.8693 39.3415 40.4915 39.6521
Lshaped 4.8584 39.3415 38.2414 38.2414
Arc1 11.7482 39.3415 31.7272 31.7272
Acr2 17.2473 39.3415 26.8015 27.4636
Table 5.1: Comparison of the CRBs (in dB) of the target features for the cases of the
full, circular, parabolic, Lshaped, Arc1, and Arc2 apertures when M = 63, K = 1,
w = = = 0, a = 1, and a2 = 40.
circle whose center is at the upper right corner of the full aperture shown in Figure
1.2(a) and whose radius is 31, as shown in Figure 5.2(b), and the Arc2 aperture
is one half of a parabolic aperture whose vertex is at the lower right corner of the
full aperture and who starts at the upper left corner of the full aperture shown in
Figure 1.2(a), as shown in Figure 5.2(c). These curvilinear apertures are subsets
of the full aperture in Figure 1.2(a) and are made to be as large as possible. For
example, the radius of the circular aperture is 15.5 in this example. Table 1 shows
the CRBs of the target parameters. As expected, the CRB for w is the same for
74
all of the curvilinear apertures since w is in the range direction, which is normal
to the plane of the curvilinear apertures. Comparing the CRBs for the Lshaped,
Arc1, and Arc2 apertures, we note that as expected, the more curved the aperture
is, the lower the CRBs for the target features. Since the Arc2 aperture is the least
curved, it has the largest CRBs. When the aperture becomes one straight line and
hence no longer curved, the CRBs go to infinity since we can no longer extract 3D
target features. Note also that from the top to bottom of Table 5.4.1, the CRBs
for & and & increase since the aperture length decreases. The CRB for a, however,
does not always increase. Consider next the circular aperture in Figure 5.2(a) when
w = & = C = 0. Figure 5.4.2 shows the CRBs of the target parameters as a function
of the radius of the circular aperture for different M. Note that as expected, the
larger the radius and/or M, the lower CRBs of the target parameters. Note also
that, as expected, the CRB for w does not change with the radius for the fixed
M since again w is in the range direction, which is orthogonal to the plane of the
curvilinear aperture.
5.4.2 Experimental Examples
Before we present an experimental example to demonstrate the performance
of the RELAX algorithm for 3D target feature extraction with a CLSAR, we first
introduce an alternative to the PolartoCartesian interpolation. Instead of Polar
toCartesian interpolation, we created an N x N x N rectangular grid and mapped
the data sample at each (tx, ty, tz) to the nearest grid point, where t,, ty, and tz are
functions of t, 0, and (see Equations (5.13) to (5.15)). Let N denote the number of
available data samples. Then for both full and curvilinear apertures, the data model
can be written as
y, = I,Aa + e, (5.39)
75
where Ir denotes an N x (NNN) matrix with each column and row containing
only one nonzero unit element corresponding to the locations of the available data
samples in the rectangular grid. Note that the larger the dimensions of the rectangular
grid, the more accurate the mapping approximation. For large enough dimensions,
however, the noise will become the dominant source of error. For our examples, we
chose N = N = N = 128. As compared to PolartoCartesian interpolation, the grid
mapping approach avoids the interpolation step needed by the former, which could
be complicated for an arbitrary curvilinear aperture. The former approach, however,
is computationally more efficient due to its smaller data dimensions.
The steps of applying RELAX to y, are similar to those for the case of full
aperture except that Equations (5.26) and (5.27) above are replaced by
&k = [,( ak (9 ak)] Y (5.40)
Wk =Wk ,k =Wk ,Wk =Wk
and
2
{jk, k, k}) = arg max Ir[I(a k ( ak)]Hrk 2, (5.41)
Wk ,Uk @Wk
respectively, where
K
Yrk = yr C iIra, (ai a) (5.42)
i=1,i k
Let ,frV be similar to Yk except that the elements in Yk that are missing in yk are
replaced with zeros in k,,. Then the right hand side of (5.41) can also be computed
by applying 3D FFT to kYk. Since the dimensions of the rectangular grid are larger
than the dimensions of the original data due to zerofilling, the 3D FFT of rk is
approximately periodic and we should limit our attentions to only one period. For our
examples, the former is four times as large as the latter and hence the peak searching
in (5.41) is limited to the frequency intervals [ 27r x (0.125) 27r x 0.125 ]. Note
that the computational advantages of FFT over DFT (discrete Fourier transform)
and backprojection diminish as the dimensions of the rectangular grid increases,
especially for curvilinear apertures.
76
We now present an experimental example to demonstrate the performances
of the RELAX algorithm for 3D target feature extraction with a CLSAR. We use
e = 0.01 to test the convergence of the RELAX algorithm in this example. The
indoor data used herein was obtained by the Radar Signature Branch, Naval Air
Warfare Center, Mugu, California. The radar carrier frequency is 9.968 GHz and the
bandwidth 1.524 GHz. The 32 x 32 x 32 data set was obtained with the full aperture
shown in Figure 1.2(a) with 32 samples in each dimension and the angular increments
were 0.28'. The target consists of K = 8 corner reflectors with a cubic configuration
and was about 15 meters away from the radar. The scatterers were about 0.5 meters
apart.
Figure 5.4.2 shows the extracted scatterers when K = 8 is used with RELAX.
Figure 5.4.2(a) is obtained when the full aperture shown in Figure 1.2(a) is used to
extract the 3D target features and Figure 5.4.2(b) is obtained when the Lshaped
aperture shown in Figure 1.2(c) is used. The centers of the circles denote the locations
of the extracted scatterers in 3D space and the radius of each circle is proportional
to the modulus of the RCS of the corresponding scatterer. The triangles show the
projections of the scatterer locations onto the horizontal plane and their sizes are also
scaled according to the RCS's of the scatterers. Note that Figure 5.4.2(b) is similar
to Figure 5.4.2(a) even though the former is obtained by using only 6.15% amount
of data used by the latter.
77
20I I I I 36
Full Aperture
15" .. M=32 3
1 "0,... M=64 .......................................................4
""' "'^ ............... ' M=80 ... ____.,
1042
S .... .44 Full Aperture
S46M=
S0 .. M4
U S  M=a0
0 4 8
.50
55
52
1 54 I I I
4 6 8 10 12 14 16 4 6 8 10 12 14 16
Radius Radius
(a) (b)
25 25
Full Aperture Full Aperture
.. M=32 ]...... 1=M32
M30 *' "". M 30 ., M 32
...  M=80 .* **.....  M.80
0 35 351 ..
S................
a 4o ,' 40
4 0
45 45
50 50
5sI I I 55. I
4 6 8 10 12 14 16 6 8 10 12 14 16
Radius Radius
(c) (d)
Figure 5.3: The CRBs of the target parameters as a function of the circular aperture
radius when K = 1, a = 1, w = C = C = 0, and a2 = 40. a) The CRB of the
complex amplitude a. b) The CRB of w. c) The CRB of a. d) The CRB of C.
78
0.4
0.2 
E 0.2 4 
06
A
(a)
0.4
0.2
E 0.2 
(b)
0.6
Figure 5.4: 3D plots of K 8 scatterers extracted by using RELAX with the indoor
experimental data. (a) Obtained with full aperture as shown in Figure 1.2(a). (b)
Obtained with curved aperture as shown in Figure 1.2(c).
0.4 ^ ^ '= ^
^ ^ ^ c>^
Obtained with curved aperture as shown in Figure 1.2(c).
CHAPTER 6
AUTOFOCUS IN CLSAR
To analyze the effect of CLSAR aperture errors, we first rewrite the CLSAR
data model in (5.12) with respect to radar viewing angles. Assume that a curvilin
ear aperture consists of M different viewing angles and let {Mm, 0m}=I denote the
elevation and azimuth angle pairs of the M look angles of the radar. Let y(n, m),
n = 0, 1, ... N 1, denote the onedimensional (1D) data samples obtained after
dechirping from the mth viewing angle of the radar. Let
4(7rfo Yro + Ytn)
4t = (rfo,, n 0, 1, .,N 1, (6.1)
c
where t, denotes the time samples. Let tm (n), tym (n), and t, (n),respectively, denote
the time samples of the mth look angle, where
txm (n) = tn cos(0m) cos(m), (6.2)
tym (n) = in sin(0m) cos(0m), (6.3)
and
tzm (n) = t, sin(0m), (6.4)
with m = 1, 2, *, M and n = 0, 1, *, N 1. Then y(n, m) has the form
K
y(n, m) = O ake2[fkm (n)+fkum ()+fk*"()l z(n,m)+e(n,m), m= 1,2, , M,
k=l
(6.5)
where e(n, m) denotes the noise and clutter and K is the number of scatterers.
We have assumed that {Om, em}m=1 and Ro(m) = Ro, m = 1,2, , M, are
known exactly in the CLSAR data (6.5). For a practical curvilinear SAR system, how
ever, the radar positions relative to the XYZ coordinate system may not be known
79
80
exactly due to atmospheric turbulence and platform position uncertainty. In 2D SAR
imaging, it is generally assumed that the errors in {0m, 0m}m=1 are negligible and the
errors in {Ro(m)}m=1 cause phase errors along the synthetic aperture. In CLSAR,
however, the errors in {Om, 0m}m=1 may no longer be negligible since the aperture
shape is critical for 3D target feature extraction. Our problem of interest herein is to
compensate for the curvilinear aperture errors in {Ro(m)}= and {Om, 0m}4m=i and
extract the 3D target parameters {ok, fk, fk, fkK= from {y(n, m)}, n = 0, 1,  , N
1, m = 1, 2, *. M. We first analyze aperture errors in CLSAR and then present an
autofocus algorithm, referred to as AUTORELAX, for 3D target feature extraction
in the presence of aperture errors.
6.1 Aperture Error Effects
We will consider the approximations and the ambiguity problems in our data
model in the presence of aperture errors. We omit the dependence on n in the
following analysis For the sake of notation convenience. For the broadside data
collection geometry shown in Figure 5.1, 0m is very small. For very small 0m, we have
sin(0m) m 0m and cos(0m) 1. These approximations also hold for the true look
angles of the radar. Then according to (6.2), (6.3), and (6.4), respectively, we have
o 0
txm t(cos 0m), (6.6)
0 0 o
ty,~ t(cos 0m) Om, (6.7)
and
o 0
tzm, t(sin a), (6.8)
where {Om, )a}M=1 denote the true look angles of the radar and
= 4(7rfo 7ro + 7t)
c
Consider first the errors in {m}mf=1 when {0m =0m,Ro(m) = Ro}m=r. Let
(m =eL +Adm. (6.9)
81
Let
to = 4rf (6.10)
c
It can be shown that for very small A0m and 7rfo > 7(t To), where t To < To/2,
we have
0 0
t, m t cos Om to(sin 0o)Aom, (6.11)
tym i t(cos 0o) Om, (6.12)
and
o O
tm tisin Cm +io(cos 0o)AOm, (6.13)
where 0o is the average of all 0m, m = 1, 2,  M, and hence is a constant. Then
{ [ r o0
ak exp j2rf(fktxm ky + fktz) + ak exp j27r [fktcos +fkto cos 0 0Om
+/ftsin ,mI } exp j27r fko sin +fi0o cos 0o Am (6.14)
Equation (6.14) shows that if A4m is a constant, then the phase error due to Aom
and the phase of ak are ambiguous. Hence the phase of ak can never be determined
exactly in the presence of Aom. If Aom is a linear function of Om, then fk and the
linear phase error due to Aom are ambiguous and cannot be determined exactly.
0
Similarly, we can analyze the errors in {0m}fI=i when {qm =m, Ro(m)
Ro0}m=1. Let
0~ =0m +AZM. (6.15)
0
For very small Om and AOm and for 7rfo > (t TO), where It Tro < To/2, we have
0 0
tX' __t COS 0m, (6.16)
tyu too cos 0o (Om +A0t), (6.17)
and
Stsin (6.18)
tZ t sin m (6.18)
82
Then
Ck exp j2r(fktxm + fktym + fktzm) aik exp j27r fktcos (m +fkt0 COS 00o m
+fktsin m} exp j27rfkto cos 0o A (6.19)
which shows that a constant AOm also results in the ambiguity between the phase
error due to AOm and the phase of ak. Hence the phase of ak can never be determined
0
exactly in the presence of AOm. Also, if A0m is a linear function of Om, then fk and
the linear phase error due to AOm are ambiguous and cannot be determined exactly.
0 0
Finally, consider the errors in Ro(m) when Om =0m and Om =Om, m =
1,2,,M. Let
Ro(m) = Ro + ARo(m). (6.20)
Replacing Ro in (5.7) with Ro(m) and for small ARo(m), large Ro(m), and 7rfo >
y(t To), where t Tol < To/2, we have from (5.11)
d(t, 0, () 6xy,zeJ(tXt +Yt+itz)ejtjoAR(m). (6.21)
Hence the errors in {Ro(m)}mM=1 result in phase errors along the synthetic aperture,
which is consistent with the analysis in [29, 43]. Then the z(n, m) in (6.5) should,
for this case, be replaced by
x(n, m)= "cgkej2[fkt )+fktym ()+ fktzm()] }e'J, (6.22)
where rim = toARo(m) is the phase error caused by ARo(m). Note that this phase
error differs from the phase errors in (6.14) and (6.19) since it does not depend on the
parameters of the kth scatterer and is easier to deal with. If ARo(m) is a constant,
then the phase error due to ARo(m) and the phase of Ok are ambiguous. Hence
the phase of ak can never be determined exactly in the presence of ARo(m). If
0
ARo(m) is a linear function of Om, then fk and the linear phase error due to ARo(m)
is ambiguous and cannot be determined exactly.
83
6.2 The RelaxationBased Autofocus Algorithm (AUTORELAX)
The AUTORELAX algorithm obtains the estimates {Om, mn, )m}=1 and
K 0
<^ o o
{J&k, fc, fk, k}k=l, respectively, of the true values {Om, 7, m}m=l and
{ak, fk, fk, fk}k=1 by minimizing the following NLS criterion:
K M N1
C5 O{k, fk k k ,{ m, m 7m}, m =1l y(n,m) x(n, m)2, (6.23)
m=1 n=o
where
x(n, m) = ackeJ2 [fktm(n)+fktym(n)+fktzm(n)] } m, (6.24)
Sk=l 1
with tm (n), tym(n), and tz((n) defined in (6.2), (6.3), and (6.4), respectively. Note
that as shown below, we will determine Om and (m via a search method. Hence there
is no need to use the approximations such as those in (6.11), (6.12), and (6.13) for
txm (n), tm (n), and tzm (n), respectively. We use an alternating optimization approach
to minimize C5 in (6.23) and the AUTORELAX algorithm is outlined as follows:
Step 1: Extract the target features {&k, ,k fk fk}k= with the RELAX algorithm
from an initial curvilinear aperture by assuming that {im = 0}m= .
Step 2: Update the curvilinear aperture {0m, im}m=i and the phase error {7m}m=
with Steps (1) (4) (to be discussed in Section 6.2.1). (If {7rm}=, is known to be
negligible, then only update {Om, m}m=1 with Step (1).)
Step 3: Redetermine the target parameters with the RELAX algorithm by using
the curvilinear aperture {Om, im}m=1 and {r7m}m=l obtained in Step 2.
Step 4: Repeat Steps 2 and 3 until "practical convergence".
The "practical convergence" of AUTORELAX is determined similarly to that
of the RELAX algorithm by checking the relative change of the cost function C5 in
(6.23) between two consecutive iterations.
6.2.1 Aperture Parameter Estimation
Assume that the target feature estimates {&k, fk, fk, fk}k=i are given. Then
Om, (m, and fm can be determined by minimizing the following Cm with respect to
84
{Om, m, rim}, where
N1
Cm,(Om,rm, rm) = E y(n,mm) (nm)12, n=0,,N1, m=l,2,,M,
n=0
(6.25)
where xl(n, m) has the same form as x(n, m) in (6.24) except that {aLk, fk, fk, kf=1
are replaced by {&k, k, fk, ffk}k=l1 To simplify the optimization of Cm, we determine
{Om, em}m=1 and {rm}m=, iteratively as follows:
Step (1): Obtain {Om, ,m}, by minimizing the following Cmi with respect to {Om, mr},
where
N1
Cmi (0m, Om) = I Uy(n, m) l(n, m)2 n = 0, 1, m = 1,  M,
n=0
(6.26)
where 1(n, m) has the same form as z(n, m) in (6.5) except that {ak, fk, k,fk}k=1
are replaced by {&k, fk, fk, fk}K=1. When there are errors in both the elevation and
0 0
azimuth directions, we can estimate Cm and Om by the alternating minimization
approach, i.e., by iteratively fixing the estimate km of 0m and minimizing Cm, with
0
respect to Om, and then fixing the estimate 0m of 0m and minimizing Cm, with respect
to Om until "practical convergence". The "practical convergence" in the alternating
minimization approach is determined by the relative change of the cost function Cmi.
In the numerical examples, we terminate the iteration when the relative change of the
cost function Cmi between two consecutive iterations is less than 103. When there
are errors in only one angular direction, i.e., either 0m or qm is to be determined,
then the minimization of Cmi is a simple onedimensional search problem.
Step (2): Determine phase errors {r m}m=" by minimizing the following cost func
tions Cm2:
N1
Cm2 (m) = y(n, m) 2(n, m)iem 2 (6.27)
n=0
85
where
K
z2(n, m) = &kej"r2[fklr t(n)+k t'm (n)+fktm(n), n = 0, 1, ., N 1, m = 1, 2, , M,
k=l
(6.28)
and im (n), ym (n), and tzm(n) are the same as tx,(n), ty (n), and tzm(n), respec
tively, except that {Om, qm}m=1 are replaced by {Om, qm^} 1, obtained in Step (1).
tivelyera, obtained in Step (1).
This step is similar to [43] and we have
im = angle(Hmym), m 1,2,. ,M, (6.29)
where
Ym= y(0,m) y(l,m) ", y(Nl,m) (6.30)
and
2m 2 (0, m) 2(1, ) = 2(Nl ,m) (6.31)
Note from (6.29) that we do not need the search over parameter space to determine
{ m}m=1 and hence the errors in {Ro(m)}m=1 are easier to deal with than those in
{Om, m}m=1
Step (3): Repeat Step (1) by replacing y(n,m) with y(n,m) = y(n,m)ej"m,
m = 1, 2, M, where {rim}m=1 are determined in Step (2).
Step (4): Repeat Steps (2) and (3) until "practical convergence", which is deter
mined by the relative change of the cost function Cm in (6.25) between two consec
utive iterations. In the numerical examples, we terminate the repetition of Steps (2)
and (3) when the relative change of Cm is less than 103 between two consecutive
iterations.
Note that if the errors in {Ro(m)}m=1 are known to be negligible, then Step
(1) alone is sufficient for the autofocusing.
6.2.2 Target Feature Extraction
Assume aperture parameter estimates {Om, im, 'lm}m= are given. Then the
problem becomes the target feature extraction problem considered in Chapter 5. As
86
shown in the previous chapter, the estimates {&k, fk, kfk}=i of {Ok, fk, k k=
can be obtained with the RELAX algorithm by equivalently minimizing the following
NLS criterion:
SK ) M N1
F ({k, ,kk, k=1 = (Yn, m) 3(n,m)2, (6.32)
m=1 n=o0
where y(n,m) = y(n,m)ej1m and 23(n,m) has the same form as z(n,m) in (6.5)
except that tm(n), tyu(n), and tzm(n) are replaced by i, (n), ym(n), and i (n),
respectively, which are determined by {Om, )m}=i instead of {Om, Om}m=1. Let
M N1
Fk (&k, A, 1k, .k) = 9k (n, km) i2k[fk mHR)++fm)+fkzm")1 (6.33)
m=1 n=O
where t (n), ym (n), and tm (n) are the same as t, (n), tu (n), and tz (n), respec
tively, except that {0m, 0m}m=1 are replaced by {0m, m}1=, and
K
k(n, m) = (n, m) 4 j2[if(6(")+ .i()+./^(n)], (6.34)
i= ,ifk
Then minimizing Fk in (6.33) with respect to ck yields the estimate &k of ak:
mM IN01k (m)ej2rfk am {n)+ k v^mn) +hk
Zk =k(n, m)e=j2r[o fk (n)+fk(,(n)+fkZm(n)] ,. =, (6.35)
MN fk=fk,fk=fk,fk=fk
and
M NI 2
{fk,, fk k = arg max k (n, m)ei2 [f (n)+Af(m)+ft()] (6.36)
fk,fkfk m=1 n=O
Herein, we omit the detailed steps used for the 3D target feature extraction
since they are the same as those in the previous chapter except that yk in (5.25)
is replaced by {(k(n,m)} in (6.34). The "practical convergence" in the iterations
of the RELAX algorithm now is determined by checking the relative change of the
cost function F ({&k, k, k fk}k=,) in (6.32) between two consecutive iterations. To
speed up the RELAX algorithm via utilizing FFT, a Cartesian grid mapping approach
is introduced in Chapter 5 to approximately map {tf (n), iym (n), m (n)}M= on a
87
Cartesian grid. When the Cartesian grid is fine enough, the errors introduced by the
approximation are negligible. In this chapter, we use a coarse Cartesian grid and the
FFT method to obtain an initial estimate of the parameters and then minimize (6.36)
via a multidimensional search method, such as the FMINV function in PVWAVE.
The latter approach uses less computer memory.
6.3 Experimental and Numerical Results
We first present an experimental example to show the performance of the
AUTORELAX algorithm. The field data was obtained by the Deployable Signature
Measurement System (DSMS), Carderock Division, Naval Surface Warfare Center,
Bethesda, Maryland. The radar was carried on board of a helicopter. The radar
carrier frequency is 9.449 GHz and the bandwidth 0.498 GHz. The data set was
obtained with a curved aperture not exactly known but is roughly the same as the
dash line shown in Figure 6.1(a), where there are 64 look angles and 64 samples per
look angle. The radar was about 300 meters away from the ground target. The ground
target consists of 13 corner reflectors on the ground plane and 7 corner reflectors
mounted on a wooden tripod that is about 2.65 meters tall. The true distribution of
the scatterers is shown in Figure 6.1(b), where the centers of the squares denote the
locations of the scatterers in 3D space and the length of each square is proportional
to the modulus of the RCS of the corresponding scatterer. The triangles show the
projections of the scatterer locations onto the ground plane and their sizes are also
scaled to be proportional to the RCS's of the scatterers.
Figures 6.1(c) and (d), respectively, show the scatterer distribution obtained
with the RELAX algorithm from the initial and an manually adjusted apertures
shown in Figure 6.1(a). We note that as compared to Figure 6.1 (b), the results in
Figure 6.1(d) are obviously better than those in Figure 6.1(c).
We now first consider autofocusing only in the elevation direction by as
suming that no errors exist in both the azimuth angles {0m}B=1 and the distances
88
{Ro(m))}m,. The AUTORELAX algorithm converges after 6 iterations. Figure 6.2
shows the target scatterers extracted by AUTORELAX with the search interval for
{Aqm}m=1 being 0.34'. Comparing Figures 6.2(b) and 6.1(d) with Figure 6.1(b),
we find that AUTORELAX works well and the AUTORELAX results are slightly
better than those obtained by using the manually adjusted aperture. Figure 6.3 shows
that the manually adjusted and autofocused apertures fit quite well after adding a
line to the former. Since the results in Figure 6.2(b) are better than those in Figure
6.1(d), it appears that a linear phase error (in Om as a function of Om) was introduced
in the manually adjusted aperture. The linear phase error can cause shifts in fk and
the amount of shift is different for different scatterer k, as can be seen from (6.14).
Consider next autofocusing in both the elevation and azimuth directions by as
suming that no errors exist in {Ro(m)}^m=. The AUTORELAX algorithm converges
after 7 iterations. Figure 6.4 shows the results obtained with the search intervals for
{Aom}m= and {Alm}m= being 0.340 and 0.0060, respectively. Compared with
the results obtained by autofocusing only in the elevation direction, we find that aut
ofocusing in both directions provides little further improvement to the accuracy of
the estimated target parameters in this example. Hence for this example, it appears
that the aperture errors mainly occur in the elevation direction.
For this experimental example, we have also used AUTORELAX to extract
target features when we assumed that errors exist in both {Om, Om}m= and
{Ro(m)}M=1. Again, as compared with Figure 6.2(b), we have noticed little change
in the extracted target parameters. It could be that DSMS has already done a good
job compensating for the errors in {Ro(m)I}^M= with some traditional method so that
they are now negligible as compared with the errors in {(m} i=1
Finally, we use a simulation example to show that the accuracy of target
feature extraction via CLSAR is very sensitive to the accuracy of the curvilinear
aperture. We assume that there are 20 scatterers with a similar distribution to that
89
Estimation Error Magnitude Range (f) Crossrange (f) Height (f)
AUTORELAX 0.015 0.00288 0.00179 0.00348
RELAX 0.296 0.31544 0.01858 0.24444
Table 6.1: Estimation errors for the parameters of the highest scatterer in the simu
lation example before and after autofocusing.
in the experimental example above. The true aperture is the dot line shown in Figure
6.5(a) and the scatterer distribution is shown in Figure 6.5(b). Here we simulate the
case where the aperture errors exist only in the elevation direction. Figure 6.5(c)
shows the scatterer distribution obtained with the RELAX algorithm from the initial
aperture (dash line) shown in Figure 6.5(a). We note that the scatterer distribution
in Figure 6.5(c) is quite different from the true one in Figure 6.5(b). When used
with the simulation data, AUTORELAX converges after four iterations. The auto
focused aperture is the dot line shown in Figure 6.5(a) and the scatterer distribution
obtained with AUTORELAX is illustrated in Figure 6.5(d). We see that the scat
terer distribution obtained by AUTORELAX is almost the same as the true one. We
also notice that the shape of the autofocused aperture is closer to the true one than
the initial aperture. The constant difference between the autofocused and the true
apertures will cause phase errors in {Oak K=1, which cannot be eliminated due to the
ambiguity problems discussed in Section 6.1.To quantitatively illustrate the accuracy
improvement of the parameter estimates via AUTORELAX, we compare the esti
mation errors of the highest scatterer in Figure 6.5(b) before and after autofocusing.
Table 6.3 shows the differences between the true values and their estimates obtained
via RELAX and AUTORELAX, respectively.

TABLE OF CONTENTS
ACKNOWLEDGEMENTS iii
LIST OF TABLES vi
LIST OF FIGURES vii
ABSTRACT x
CHAPTERS
1 INTRODUCTION 1
1.1 Background 1
1.2 Scope of the Work 4
1.3 Contributions 8
1.4 Dissertation Outline 9
2 LITERATURE SURVEY 11
2.1 2D Super Resolution SAR Image Formation 11
2.2 3D SAR Feature Extraction 15
2.3 SAR Motion Compensation 16
3 PARAMETRIC METHODS FOR SAR IMAGING 18
3.1 RELAX 18
3.1.1 Review of the RELAX Algorithm 18
3.1.2 Image Formation 21
3.2 RELAXNLS 23
3.2.1 Review of the RELAXNLS Algorithm 23
3.2.2 Image Formation 27
3.3 Experimental Results 28
4 SEMIPARAMETRIC METHODS FOR SAR IMAGING 35
4.1 Problem Formulation and Data Model 35
4.2 Review of the APES Algorithm 37
4.3 Data Model Ambiguities and Their Effects on SAR Image Formation 38
4.3.1 Model Ambiguities 39
4.3.2 Image Formation 41
4.3.3 Model Ambiguity Effects on SAR Image Formation .... 43
4.4 The SPAR Algorithm 45
IV
22
applying the normalized FFT to the simulated phase history data {ys(ns, ns)}, ie.,
by computing
CN1C1
75jvjv E E (3.11)
ns= 0 na=0
To suppress the sidelobes, the normalized FFT can be applied to the windowed
sequence {ws(ns,s)ys(ns,s)}, where the window sequence ws(ns,s) satisfies
CAT1 C 1
y; y ws(ns,ns) = Â£2_/VIV. (3.12)
ns=0 s=0
The RELAX SAR. images containing both the dominant target features and back
ground clutter may be appreciated since, for example, the shadow information may
be desired for automatic target recognition. If so, the normalized FFT can be applied
to the sum of the simulated phase history data {ya(ns, ns)} and {Â£2(n, )} with zero
padding to have dimensions (N and Â£1V, where (n, ) denotes the estimated back
ground clutter and is determined by (3.9) with K replaced by K. Note that scaling
the (n, ) by a factor of Â£2 is needed when the background clutter is included in
the RELAX SAR images since both of its dimensions are 1 /Â£ times of those of the
simulated phase history data. Similarly, the normalized FFT can also be applied
to the sum of {ws(ns, s)ys(ns, s)} and {(2we(n, )(n, )} with zero padding to
have dimensions (N and Â£ to suppress the sidelobes, where the window sequence
we(n,) satisfies
Nl 1
y y We{n,) = N. (3.13)
n=0 =0
Note that since we cannot model the clutter effectively, its resolution cannot be
improved. We remark that the resolution of the RELAX SAR images is determine
by RELAX and Â£ > 1 is just used to demonstrate the super resolution property of
the RELAX algorithm for target feature extraction.
48
nth column of Y. Then the estimates {fa, xfc, fk, fk}k=1 of {fa, xfc, fk, fk}k=1 can be
obtained by minimizing the following NLS cost function:
C5 ({fa,^k, fk, fk}k=i) =
K
YÂ£e*G k{xk,fkJk)
k=1
(4.44)
where  \\p denotes the Frobenius norm [22], The minimization of
C5 {{fa, xfc, fk, fk}k=1) in (4.44) is a very complicated optimization problem. The
proposed SPAR algorithm performs a complete relaxationbased search by letting
only the parameters of one scatterer vary and freezing all others at their most re
cently determined values for each assumed number of scatterers K. Let
K
Y*=Y Â£ ,(SÂ¡,/Â¡,/i),
i=l,i^k
(4.45)
and assume that {fa, x, /, fi}?=iare given. Then the NLS estimates
{fa,^k,fk,fa}k=i of {h,^k,fk,fk}k=i can be obtained by minimizing
C6{fa,Xk,fk,fk), where
C6{fa,xkJk,fk) = Yfce^Gfc(xfc,/fc,/fc)^
(4.46)
and using the method presented in the previous subsection for the feature extraction
of a single scatterer. However, when multiple scatterers are located in the same
range, the minimization of C${fa, x.k, fk, fk) has numerous ambiguous solutions that
may lead to the artifact problem in the high resolution SAR image formation.
SPAR attempts to avoid the ambiguity problem by isolating out the most
dominant scatterer in Yk by using a 2D rectangular window, which is determined
from and applied to the 2D FFT of Y*,. The isolation process has the following
steps:
Step (i): Obtain Yk, the 2D FFT of Yk, without zero padding.
Step (ii): Determine the 2D window w(n,n) from Vfc. We first locate the
peak location (n+,+) of the magnitude of Vfc. We then fix to + and search for
65
Radar
Figure 5.1: Broadside spotlightmode SAR data collection geometry.
It has been shown in Appendix 3B that R can be approximated as
Rq x cos 9 cos 0 y sin 9 cos 0 z sin 0,
where
o o +
[(x2 z2) sin0o cos 0O 2xz cos2 0O](0 0o)
X X +
y = y +
2o
(y2 + z2) cos 0o + 2a;zsin 0O
2o :
xy cos 0o + yz sin 0O
o
and
z = z
(x1 + y2) sin
2 ft,
(5.6)
(5.7)
(5.8)
(5.9)
(5.10)
Note that the second terms of the right sides of (5.8), (5.9), and (5.10) are due to
the range and elevation curvature effects and can be neglected for large 0. Let r0
= Since r = , then from (5.4), we have
d(t, 9, 0) = SXty,z exp
j4(7t/q7TQ+7)
(o o x cos 9 cos 0 y sin 9 cos 0
zsin 0)],
(5.11)
39
there are ambiguity problems associated with the semiparametric data model. In
this section, we first analyze the possible data model ambiguities and then illustrate
their effects on SAR target feature extraction and image formation. The discussions
below will motivate the introduction of the SPAR algorithm, which will be presented
in detail in Chapter 4.4.
4.3.1 Model Ambiguities
Due to the flexibility of the data model in (4.2), there are various types of
ambiguities that may impact the feature extraction of each scatterer. Below we list
several types of the ambiguities inherent in the data model.
Type 1: Single scatterer
From (4.1), we note that ambiguity exists between
x{)e:i = a;()e7^+7r^ = Â£()e^, (414)
where x(n) = x(n) and = ^ + 7r. Ambiguity also exists between / and x(n) since
x{)ej2*f = (l)x{)ej2nV5) = SfriW2*?"
x(ri)e3
(4.15)
where x{n) = (l)nx() and / = / 0.5. The above two types of ambiguities cannot
be resolved.
Type 2: Two identical scatterers located in the same range
Let fa and fb, respectively, denote the crossrange locations of two identical scatterers
and let 4>a and 4>b, respectively, denote their phases. Then
x(n)
e0a+t27r/an) __ e0,+j27T/i,n)
2x(n) cos [7T(fa fb)n +
eMfa+h)n, (4T6)
which indicates that two identical corner reflectors (trihedrals or dihedrals) located
in the same range but different crossrange positions fa and fb, respectively, can be
modeled by (4.1) as one scatterer located at (fa + fb)/2 in crossrange with x{n)
CHAPTER 1
INTRODUCTION
1.1 Background
RADAR stands for RAdio Detection And Ranging. By now, radar has been
used not only to detect and locate targets, as was originally intended, but to gen
erate images of illuminated scenes without the restriction of weather and time as
well. Radar image is defined as a twodimensional (2D) mapping of the reflectivity
corresponding to the scene illuminated by a radar. More specifically, a radar image
is the reconstruction of the illuminated scene from the electromagnetic reflectivity,
which is usually treated as a 2D function in range and crossrange. The direction
in which a radar beam propagates is referred to as the range direction, while the
one perpendicular to the line of sight is called the crossrange direction, or the az
imuth direction (see Figure 1.1 for an illustration). Radar imaging [2] has played a
crucial role in both military applications, including reconnaissance and surveillance
[21], and civilian ones including geological and topographic map generation, weather
forecasting, scene classification, and target recognition [60, 65].
The quality of radar images depends greatly on the image resolution, which
is nominally referred to as the ability to distinguish two closelyspaced elements of a
target [48]. The higher the radar resolution, the more details of target features are
available for detection, characterization, and identification of targets of interest. As
such, much effort has been made to improve 2D imaging radar resolutions, which
include range and crossrange resolutions. High range resolution can be achieved
by using wide bandwidth signals, such as linear frequency modulated (FM) chirp
pulses [48], since range resolution is inversely proportional to the transmitted signal
1
APPENDIX A
A SINGLE SCATTERER FEATURE EXTRACTION VIA SPAR
The estimates of {, x, /, /} can be obtained by minimizing the following NLS
cost function:
Â£
Ci(fxJ,/) = Y.e*G(x,/,/)
JV1
= Iys" ~ x()e27Tfn+
=0
After simple calculations, we can rewrite (A.l) as
1 f r
C2 (, x, /, /) = ] lly^ll2 + 7V x() ~ Re (
r,(2 nfn+4>)yT^N{f)
(A.l)
Tl0
1
e 3'(2'/+*)ylu'M))
*1 2
(A.2)
Minimizing (A.2) with respect to x(n) yields
x(n) = He
v N
j(2irfn+) T
y nv*N(f)
, n = 0,1, , N 1.
(A.3)
Hence
x = Re{e"[Yja.*w(/)]0^(/)}
= ^[e2*y,(/)0u.s(/)].
(A.4)
Inserting (A.3) into (A.2), we obtain the NLS estimates {(f>,/,/} of {0,/,/} by
equivalently maximizing the following cost function:
1
n=0
= \ x {y>(/)l2 + R* [(yf,^(/))2 eJ[2'(2,>+2*1]} (A.5)
98
9
(3) We have devised a novel, robust, and computationally simple SemiPARametric
(SPAR) algorithm for SAR target feature extraction and image formation based
on a flexible data model;
(4) Experimental results using ERIM and MSTAR Slicy data show that the relaxation
based parametric approaches including RELAX and RELAXNLS and the semi
parametric algorithm SPAR, can yield SAR images with higher resolution than
FFT based methods. Also SPAR is more robust and computationally simple
than RELAXNLS.
(II) 3D target feature extraction and motion compensation via CLSAR
(1) We have established a 3D data model;
(2) We have performed CramrRao Bound (CRB) analyses of parameter estima
tion using various flight trajectories, which provide more insight into the per
formances associated with different curvilinear apertures;
(3) The effects of aperture errors on target feature estimation have been investi
gated and a relaxationbased approach for autofocus and target feature extrac
tion using CLSAR has been introduced.
1.4 Dissertation Outline
The remainder of the dissertation is organized as follows. Chapter 2 gives a
literature survey of emphasis on the topics considered, including 2D SAR imaging, 3
D SAR target feature extraction, and motion compensation (autofocus). In Chapter
3, we describe how to form super resolution 2D SAR images via the relaxation
based parametric spectral estimation algorithms including RELAX and RELAX
NLS. A novel semiparametric (SPAR) method is introduced in Chapter 4 for SAR
image formation when an accurate parametric data model in crossrange is difficult
87
Cartesian grid. When the Cartesian grid is fine enough, the errors introduced by the
approximation are negligible. In this chapter, we use a coarse Cartesian grid and the
FFT method to obtain an initial estimate of the parameters and then minimize (6.36)
via a multidimensional search method, such as the FMINV function in PVWAVE.
The latter approach uses less computer memory.
6.3 Experimental and Numerical Results
We first present an experimental example to show the performance of the
AUTORELAX algorithm. The field data was obtained by the Deployable Signature
Measurement System (DSMS), Carderock Division, Naval Surface Warfare Center,
Bethesda, Maryland. The radar was carried on board of a helicopter. The radar
carrier frequency is 9.449 GHz and the bandwidth 0.498 GHz. The data set was
obtained with a curved aperture not exactly known but is roughly the same as the
dash line shown in Figure 6.1(a), where there are 64 look angles and 64 samples per
look angle. The radar was about 300 meters away from the ground target. The ground
target consists of 13 corner reflectors on the ground plane and 7 corner reflectors
mounted on a wooden tripod that is about 2.65 meters tall. The true distribution of
the scatterers is shown in Figure 6.1(b), where the centers of the squares denote the
locations of the scatterers in 3D space and the length of each square is proportional
to the modulus of the RCS of the corresponding scatterer. The triangles show the
projections of the scatterer locations onto the ground plane and their sizes are also
scaled to be proportional to the RCSs of the scatterers.
Figures 6.1(c) and (d), respectively, show the scatterer distribution obtained
with the RELAX algorithm from the initial and an manually adjusted apertures
shown in Figure 6.1(a). We note that as compared to Figure 6.1 (b), the results in
Figure 6.1(d) are obviously better than those in Figure 6.1(c).
We now first consider autofocusing only in the elevation direction by as
suming that no errors exist in both the azimuth angles {#m}m=i and the distances
(a)
(b)
(c) (d)
Figure 4.1: Ambiguity effect on the SAR image formation in the absence of range
estimation errors, (a) True windowed FFT SAR image, (b) Windowed FFT image
of the first scatterer. (c) Windowed FFT image of the second scatterer. (d)
Combined windowed FFT image of the two scatterers. (The vertical and horizontal
axes are for range and crossrange, respectively.)
REFERENCES
[1] D. J. Andersh, M. Hazlett, S. W. Lee, D. D. Reeves, D. P. Sullivan, and Y. Chu.
XPATCH: a highfrequency electromagnetic scattering prediction code and en
vironment for complex threedimensional objects. IEEE Antennas and Propa
gation Magazine, 36(1):6569, February 1994.
[2] D. A. Ausherman, A. Kozma, J. L. Walker, H. M. Jones, and E. C. Poggio.
Developments in radar imaging. IEEE Transactions on Aerospace and Electronic
Systems, 20(4):363400, July 1984.
[3] W. J. Bangs. Array processing with generalized beamformers. Ph.D. dissertation,
Yale University, New Haven, CT, 1971.
[4] Sergio Barbarossa, Luigi Marsili, and Gaetano Mungari. SAR superresolution
imaging by signal subspace projection techniques. AEU International Journal
of Electronics and Communications, 50(2): 133138, March 1996.
[5] G. R. Benitz. Adaptive highdefinition imaging. SPIE Proceedings on Optical
Engineering in Aerospace Sensing, Orlando, FL, pages 106119, April 1994.
[6] G. R. Benitz. High definition vector imaging for synthetic aperture radar. Pro
ceedings of the 31st Asilomar Conference on Signals, Systems and Computers,
Pacific Grove, CA, November 1997.
[7] Zhaoqiang Bi, Jian Li, and ZhengShe Liu. Super resolution SAR image for
mation via parametric spectral estimation methods. IEEE Transactions on
Aerospace and Electronic Systems, 35(1) :267281, January 1999.
[8] B. D. Bunday. Basic Optimization Methods. Edward Arnold Ltd, London, 1984.
[9] J. P. Burg. Maximum entropy spectral analysis. Proceedings of the 31th Meeting
Society of Exploration Geophysicists, Oklahoma City, OK, Oct. 1967.
[10] J. Capon. High resolution frequencywavenumber spectrum analysis. Proceedings
of the IEEE, 57:14081418, August 1969.
[11] Walter G. Carrara, Ron S. Goodman, and Ronald M. Majewski. Spotlight Syn
thetic Aperture Radar Signal Processing Algorithms. Artech House Inc., Nor
wood, MA, 1995.
[12] I. J. Clarke. Supervised Interpretation of Sampled Data Using Efficient Imple
mentations of HigherRank Spectrum Estimation, Vol. II. Prentice Hall, Engle
wood Cliff, NJ, 1991.
[13] S. R. DeGraaf. SAR imaging via modern 2D spectral estimation methods. SPIE
Proceedings on Optical Engineering in Aerospace Sensing, 2230:3647, Orlando,
FL, April 1994.
104
8
Motion error compensation also plays a crucial role in accurately extracting
target features using CLSAR. Motion errors in 2D SAR systems can be approximated
as phase errors in crossrange since the azimuth angle 6 is very small and phase errors
are mainly introduced by the uncertainty in the distance from the radar to the center
of the illuminated scene [29]. However, it is not the case for CLSAR since curvilinear
aperture determined by azimuth angle 9 and elevation angle is critical in obtaining
3D information of the target. Hence motion errors in 6 and (j) may not be negligible.
Motion compensation (autofocus) in CLSAR turns out to be more complicated than
that in 2D SAR systems. In Chapter 6, we investigate the effects of aperture errors
on 3D target feature extraction and present a relaxationbased autofocus algorithm,
referred to as AUTORELAX, to extract target features in the presence of aperture
errors using CLSAR.
1.3 Contributions
The significance of this research lies in that it examines the possibility of super
resolution image formation and autofocus of small targets via spotlightmode SAR
system by using relaxationbased parametric and semiparametric spectral estimation
methods. Super resolution SAR images obtained via the proposed methods can
be further used for SAR image segmentation and automatic target detection and
recognition. This research makes contributions in the following two areas:
(I) 2D super resolution SAR image formation
(1) We have extended RELAX to form enhanced SAR images of targets consisting
of point scatterers;
(2) We have extended RELAXNLS for super resolution SAR image formation of
targets consisting of both trihedrals and dihedrals;
103
(x,y,z) = (4.9991,3.9346,2.0477). Note that in both examples, (x,y,z) are very
close to (x,y,z) and hence can be used to approximate (x,y,z).
PAGE 1
5(/$;$7,21%$6(' 0(7+2'6 )25 6$5 7$5*(7 )($785( (;75$&7,21 $1' ,0$*( )250$7,21 %\ =+$24,$1* %, $ ',66(57$7,21 35(6(17(' 72 7+( *5$'8$7( 6&+22/ 2) 7+( 81,9(56,7< 2) )/25,'$ ,1 3$57,$/ )8/),//0(17 2) 7+( 5(48,5(0(176 )25 7+( '(*5(( 2) '2&725 2) 3+,/2623+< 81,9(56,7< 2) )/25,'$
PAGE 2
7KLV ZRUN LV GHGLFDWHG WR P\ ZLIH LQJ :DQJ IRU KHU ORYH DQG VKDULQJ P\ KDSSLQHVV DQG GHSUHVVLRQ LQ WKH SDVW \HDUV WR P\ SDUHQWV ZKR FRQWLQXH VXSSRUWLQJ DQG HQFRXUDJLQJ PH IRU P\ HGXFDWLRQ
PAGE 3
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f IRU WKHLU JUHDW KHOS )LQDOO\ ZRXOG OLNH WR JUDWHIXOO\ DFNQRZOHGJH DOO WKH SHRSOH ZKR KHOSHG PH GXULQJ P\ 3K' SURJUDP
PAGE 4
7$%/( 2) &217(176 $&.12:/('*(0(176 LLL /,67 2) 7$%/(6 YL /,67 2) ),*85(6 YLL $%675$&7 [ &+$37(56 ,1752'8&7,21 %DFNJURXQG 6FRSH RI WKH :RUN &RQWULEXWLRQV 'LVVHUWDWLRQ 2XWOLQH /,7(5$785( 6859(< 6XSHU 5HVROXWLRQ 6$5 ,PDJH )RUPDWLRQ 6$5 )HDWXUH ([WUDFWLRQ 6$5 0RWLRQ &RPSHQVDWLRQ 3$5$0(75,& 0(7+2'6 )25 6$5 ,0$*,1* 5(/$; 5HYLHZ RI WKH 5(/$; $OJRULWKP ,PDJH )RUPDWLRQ 5(/$;1/6 5HYLHZ RI WKH 5(/$;1/6 $OJRULWKP ,PDJH )RUPDWLRQ ([SHULPHQWDO 5HVXOWV 6(0,3$5$0(75,& 0(7+2'6 )25 6$5 ,0$*,1* 3UREOHP )RUPXODWLRQ DQG 'DWD 0RGHO 5HYLHZ RI WKH $3(6 $OJRULWKP 'DWD 0RGHO $PELJXLWLHV DQG 7KHLU (IIHFWV RQ 6$5 ,PDJH )RUPDWLRQ 0RGHO $PELJXLWLHV ,PDJH )RUPDWLRQ 0RGHO $PELJXLW\ (IIHFWV RQ 6$5 ,PDJH )RUPDWLRQ 7KH 63$5 $OJRULWKP ,9
PAGE 5
7DUJHW )HDWXUH ([WUDFWLRQ 0RGLILHG 5(/$;1/6 $OJRULWKP &UDPU5DR %RXQG &5%f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f $SHUWXUH 3DUDPHWHU (VWLPDWLRQ 7DUJHW )HDWXUH ([WUDFWLRQ ([SHULPHQWDO DQG 1XPHULFDO 5HVXOWV &21&/86,21 6XPPDU\ )XWXUH :RUNV $33(1',;(6 $ $ 6,1*/( 6&$77(5(5 )($785( (;75$&7,21 9,$ 63$5 % $3352;,0$7,21 2) 7+( 5$1*( 5 & &$/&8/$7,1* [\]f )520 [\]f 5()(5(1&(6 %,2*5$3+,&$/ 6.(7&+ OJ Y
PAGE 6
/,67 2) 7$%/(6 7UXH SDUDPHWHUV RI WKH WZR GLKHGUDOV XVHG LQ )LJXUHV DQG &RPSDULVRQ RI WKH &5%V LQ G%f RI WKH WDUJHW IHDWXUHV IRU WKH FDVHV RI WKH IXOO FLUFXODU SDUDEROLF /VKDSHG $UF DQG $UF DSHUWXUHV ZKHQ 0 XM 4 4 D DQG D (VWLPDWLRQ HUURUV IRU WKH SDUDPHWHUV RI WKH KLJKHVW VFDWWHUHU LQ WKH VLPXODWLRQ H[DPSOH EHIRUH DQG DIWHU DXWRIRFXVLQJ YL
PAGE 7
/,67 2) ),*85(6 ,OOXVWUDWLRQ RI UDGDU LPDJLQJ ,OOXVWUDWLRQ RI 6$5 LPDJLQJ PRGHV Df 6SRWOLJKWPRGH 6$5 Ef 6WULSPDS 6$5 3RVVLEOH DSHUWXUHV IRU D 6$5 V\VWHP Df )XOO DSHUWXUH Ef 3DUDEROLF DSHUWXUH Ff 2UWKRJRQDO VXEDSHUWXUH DQG I! GHQRWH WKH D]LPXWK DQG HOHYDWLRQ DQJOHV UHVSHFWLYHO\f 0HVK SORWV RI WKH PRGXOXV RI WKH 5&6 REWDLQHG E\ XVLQJ ))7 ZLWK GLIIHUHQW DSHUWXUHV UDQJH LQIRUPDWLRQ VXSSUHVVHG IRU WKH LOOXVWUDWLRQ SXUSRVH RQO\f Df )XOO DSHUWXUH DV VKRZQ LQ )LJXUH SURWHFW Df Ef &XUYLOLQHDU DSHUWXUH DV VKRZQ LQ )LJXUH Ff 5(/$; 6$5 ,PDJHV E\ XVLQJ WKH [ (5,0 GDWD ZLWK fÂ§ Df 8QZLQGRZHG ))7 LPDJH Ef :LQGRZHG ))7 LPDJH Ff 8Qn ZLQGRZHG 5(/$; LPDJH ZLWKRXW EDFNJURXQG FOXWWHU Gf :LQGRZHG 5(/$; LPDJH ZLWKRXW EDFNJURXQG FOXWWHU Hf 8QZLQGRZHG 5(/$; LPDJH ZLWK EDFNJURXQG FOXWWHU If :LQGRZHG 5(/$; LPDJH ZLWK EDFNJURXQG FOXWWHU 5(/$; ,PDJHV E\ XVLQJ WKH [ (5,0 GDWD ZLWK fÂ§ Df 8QZLQGRZHG ))7 LPDJH Ef :LQGRZHG ))7 LPDJH Ff 8Qn ZLQGRZHG 5(/$; LPDJH ZLWKRXW EDFNJURXQG FOXWWHU Gf :LQGRZHG 5(/$; LPDJH ZLWKRXW EDFNJURXQG FOXWWHU Hf 8QZLQGRZHG 5(/$; LPDJH ZLWK EDFNJURXQG FOXWWHU If :LQGRZHG 5(/$; LPDJH ZLWK EDFNJURXQG FOXWWHU 7DUJHW SKRWR WDNHQ DW r D]LPXWK DQJOH 6$5 LPDJHV RI WKH 6OLF\ GDWD r D]LPXWK DQG r HOHYDWLRQ DQJOHVf Df 2ULJLQDO ZLQGRZHG ))7 LPDJH Ef :LQGRHG ))7 LPDJH ZLWK VSRLOHG UHVROXWLRQ Ff a Gf :LQGRZHG 5(/$; LPDJH ZLWKRXW DQG ZLWK EDFNJURXQG FOXWWHU ZLWK Hf a If :LQGRZHG 5(/$; 1/6 LPDJH ZLWKRXW DQG ZLWK EDFNJURXQG FOXWWHU ZLWK Ff a If DUH REDWLQHG XVLQJ WKH GDWD LQ Ef 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f YLL
PAGE 8
$PELJXLW\ HIIHFW RQ WKH 6$5 LPDJH IRUPDWLRQ LQ WKH DEVHQFH RI UDQJH HVWLPDWLRQ HUURUV Df 7UXH ZLQGRZHG ))7 6$5 LPDJH Ef :LQGRZHG ))7 LPDJH RI WKH ILUVW fVFDWWHUHUf Ff :LQGRZHG ))7 LPDJH RI WKH VHFRQG fVFDWWHUHUf Gf &RPELQHG ZLQGRZHG ))7 LPDJH RI WKH WZR fVFDWWHUHUVf 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVn UDQJH UHVSHFWLYHO\f $PELJXLW\ HIIHFW RQ WKH 6$5 LPDJH IRUPDWLRQ LQ WKH SUHVHQFH RI UDQJH HVWLPDWLRQ HUURUV Df :LQGRZHG ))7 LPDJH RI WKH ILUVW fVFDWWHUHUf Ef :LQGRZHG ))7 LPDJH RI WKH VHFRQG fVFDWWHUHUf Ff &RPELQHG ZLQGRZHG ))7 LPDJH RI WKH WZR fVFDWWHUHUVf ZLWK e ZLWKRXW H[WUDSRODWLRQf Gf &RPELQHG ZLQGRZHG ))7 LPDJH ZLWK e 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHn VSHFWLYHO\f &RPSDULVRQ RI 6$5 LPDJHV IRUPHG XVLQJ GLIIHUHQW DOJRULWKPV IRU VLPn XODWHG GDWD DW KLJK 615 D f Df 7UXH 6$5 LPDJH Ef :LQn GRZHG ))7 6$5 LPDJH Ff +\EULG 6$5 LPDJH Gf 63$5 6$5 LPDJH Hf 5(/$;1/6 6$5 LPDJH If 0RGLILHG 5(/$;1/6 6$5 LPDJH 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f &RPSDULVRQ RI 6$5 LPDJHV IRUPHG XVLQJ GLIIHUHQW DOJRULWKPV IRU VLPn XODWHG GDWD DW ORZ 615 D f Df :LQGRZHG ))7 6$5 LPDJH Ef +\EULG 6$5 LPDJH Ff 63$5 6$5 LPDJH Gf 5(/$;1/6 6$5 LPDJH Hf 0RGLILHG 5(/$;1/6 6$5 LPDJH 7KH YHUWLFDO DQG KRUn L]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f &RPSDULVRQ RI 6$5 LPDJHV REWDLQHG YLD GLIIHUHQW DOJRULWKPV IRU WKH 6OLF\ GDWD KEO r D]LPXWK DQG r HOHYDWLRQ DQJOHVf Df :LQGRZHG ))7 6$5 LPDJH IURP WKH ;3$7&+ GDWD Ef :LQn GRZHG ))7 6$5 LPDJH IURP WKH 6OLF\ GDWD Ff +\EULG 6$5 LPDJH Gf 63$5 6$5 LPDJH Hf 5(/$;1/6 6$5 LPDJH If 0RGn LILHG 5(/$;1/6 6$5 LPDJH Ff a If DUH DOO REWDLQHG IURP WKH GDWD XVHG LQ Ef 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f %URDGVLGH VSRWOLJKWPRGH 6$5 GDWD FROOHFWLRQ JHRPHWU\ &XUYLOLQHDU DSHUWXUHV Df 7KH FLUFXODU DSHUWXUH Ef 7KH $UF DSHUn WXUH Ff 7KH $UF DSHUWXUH 7KH &5%V RI WKH WDUJHW SDUDPHWHUV DV D IXQFWLRQ RI WKH FLUFXODU DSHUn WXUH UDGLXV ZKHQ D Xf DQG R Df 7KH &5% RI WKH FRPSOH[ DPSOLWXGH D Ef 7KH &5% RI X Ff 7KH &5% RI &R Gf 7KH &5% RI *R SORWV RI VFDWWHUHUV H[WUDFWHG E\ XVLQJ 5(/$; ZLWK WKH LQGRRU H[SHULPHQWDO GDWD Df 2EWDLQHG ZLWK IXOO DSHUWXUH DV VKRZQ LQ )LJXUH Df Ef 2EWDLQHG ZLWK FXUYHG DSHUWXUH DV VKRZQ LQ )LJXUH Ff YLLL
PAGE 9
Df &XUYLOLQHDU DSHUWXUHV IRU WKH H[SHULPHQWDO H[DPSOH Ef 7UXH VFDW WHUHU GLVWULEXWLRQ Ff SORW RI VFDWWHUHUV H[WUDFWHG IURP WKH LQLWLDO DSHUWXUH GDVKHG OLQHf LQ Df ZLWK 5(/$; Gf SORW RI VFDWWHUHUV H[WUDFWHG IURP WKH PDQXDOO\ DGMXVWHG DSHUWXUH VROLG OLQHf LQ Df ZLWK 5(/$; $XWRIRFXVHG FXUYLOLQHDU DSHUWXUH DQG VFDWWHUHU GLVWULEXWLRQ REWDLQHG ZLWK $8725(/$; E\ DXWRIRFXVLQJ RQO\ LQ WKH HOHYDWLRQ GLUHFWLRQ DQG XVLQJ IRU WKH H[SHULPHQWDO H[DPSOH Df $XWRIRFXVHG FXUYLOLQHDU DSHUWXUH GRWWHG OLQHf Ef 6FDWWHUHU GLVWULEXWLRQ Df 0DQXDOO\ DGMXVWHG VROLG OLQHf DQG DXWRIRFXVHG GRWWHG OLQHf FXUYLn OLQHDU DSHUWXUHV IRU WKH H[SHULPHQWDO H[DPSOH Ef )LWWLQJ WKH PDQXn DOO\ DGMXVWHG DSHUWXUH WR WKH DXWRIRFXVHG DSHUWXUH E\ DGGLQJ D OLQH WR WKH IRUPHU $XWRIRFXVHG FXUYLOLQHDU DSHUWXUH DQG VFDWWHUHU GLVWULEXWLRQ REWDLQHG ZLWK $8725(/$; E\ DXWRIRFXVLQJ LQ ERWK WKH HOHYDWLRQ DQG D]LPXWK GLUHFWLRQV DQG XVLQJ fÂ§ IRU WKH H[SHULPHQWDO H[DPSOH Df $XWRn IRFXVHG FXUYLOLQHDU DSHUWXUH GRWWHG OLQHf Ef 6FDWWHUHU GLVWULEXWLRQ Df &XUYLOLQHDU DSHUWXUH IRU WKH VLPXODWLRQ H[DPSOH Ef 7UXH VFDWWHUHU GLVWULEXWLRQ IRU WKH VLPXODWLRQ H[DPSOH Ff 6FDWWHUHU GLVWULEXWLRQ REn WDLQHG IURP WKH LQLWLDO DSHUWXUH GDVKHG OLQHf LQ Df E\ XVLQJ 5(/$; ZLWK Gf 6FDWWHUHU GLVWULEXWLRQ REWDLQHG YLD $825(/$; ZLWK DQG XVLQJ WKH DXWRIRFXVHG DSHUWXUH GRWWHG OLQHf VKRZQ LQ Df ,;
PAGE 10
$EVWUDFW RI 'LVVHUWDWLRQ 3UHVHQWHG WR WKH *UDGXDWH 6FKRRO RI WKH 8QLYHUVLW\ RI )ORULGD LQ 3DUWLDO )XOILOOPHQW RI WKH 5HTXLUHPHQWV IRU WKH 'HJUHH RI 'RFWRU RI 3KLORVRSK\ 5(/$;$7,21%$6(' 0(7+2'6 )25 6$5 7$5*(7 )($785( (;75$&7,21 $1' ,0$*( )250$7,21 %\ =KDRTLDQJ %L 0D\ &KDLUPDQ 'U LDQ /L 0DMRU 'HSDUWPHQW (OHFWULFDO DQG &RPSXWHU (QJLQHHULQJ 6\QWKHWLF DSHUWXUH UDGDU 6$5f KDV EHHQ D PDWXUH EXW DFWLYHO\ UHVHDUFKHG WHFKQRORJ\ GXH WR LWV GD\DQGQLJKW DQG DOOZHDWKHU FDSDELOLW\ RI RIIHULQJ KLJK UHVn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f 6$5 LPDJHV YLD UHOD[DWLRQEDVHG SDUDPHWULF PHWKRGV 7KH UHOD[DWLRQEDVHG RSWLPL]DWLRQ PHWKRGV KDYH EHHQ SURYHG WR EH TXLWH XVHIXO LQ VHYHUDO [
PAGE 11
RWKHU DSSOLFDWLRQV VXFK DV UDGLR DVWURQRP\ PLFURZDYH LPDJLQJ DQG VSHFWUDO HVWLPDn WLRQ 7KH UHOD[DWLRQEDVHG PHWKRGV DUH H[WHQGHG IRU VXSHU UHVROXWLRQ 6$5 LPDJLQJ RI UDGDU WDUJHWV FRQVLVWLQJ RI RQO\ WULKHGUDOV RU ERWK WULKHGUDOV DQG GLKHGUDOV :H KDYH DOVR GHYLVHG D UREXVW DQG FRPSXWDWLRQDOO\ VLPSOH 63$5 6HPL3$5DPHWULFf DOn JRULWKP IRU 6$5 LPDJLQJ EDVHG RQ D IOH[LEOH VHPLSDUDPHWULF GDWD PRGHO ZKHQ LW LV GLIILFXOW WR HVWDEOLVK DQ DFFXUDWH WDUJHW GDWD PRGHO LQ FURVVUDQJH +HQFH 63$5 WDNHV DGYDQWDJHV RI ERWK SDUDPHWULF DQG QRQSDUDPHWULF VSHFWUDO HVWLPDWLRQ PHWKn RGV WR IRUP HQKDQFHG 6$5 LPDJHV 1XPHULFDO DQG H[SHULPHQWDO H[DPSOHV KDYH EHHQ XVHG WR GHPRQVWUDWH WKH SHUIRUPDQFHV RI WKH SURSRVHG DOJRULWKPV :H KDYH REVHUYHG WKDW WKH UHOD[DWLRQEDVHG SDUDPHWULF PHWKRGV SURYLGH VXSHU UHVROXWLRQ 6$5 LPDJHV ZKHQ WKH DVVXPHG GDWD PRGHO LV YDOLG RWKHUZLVH 63$5 SHUIRUPV EHWWHU 7KUHHGLPHQVLRQDO 'f WDUJHW IHDWXUHV LQFOXGLQJ WKH KHLJKW LQIRUPDWLRQ UDGDU FURVV VHFWLRQ 5&6f DQG ORFDWLRQ UDQJH DQG FURVVUDQJHf SURYLGH TXLWH XVHIXO LQIRUPDWLRQ IRU VXFK DSSOLFDWLRQV DV DXWRPDWLF WDUJHW UHFRJQLWLRQ 7KXV HIn IRUWV KDYH DOVR EHHQ PDGH LQ WKLV GLVVHUWDWLRQ WR GHYLVH DQ HIIHFWLYH UHOD[DWLRQEDVHG DOJRULWKP UHIHUUHG WR DV $8725(/$; IRU ERWK WDUJHW IHDWXUH H[WUDFWLRQ DQG PRWLRQ FRPSHQVDWLRQ YLD FXUYLOLQHDU 6$5 &/6$5f D QRYHO WHFKQRORJ\ ZKLFK LV VWLOO DW LWV GHYHORSLQJ VWDJH 7KH SURSRVHG $8725(/$; DOJRULWKP LV VKRZQ WR EH SURPLVLQJ ZKHQ HYDOXDWHG E\ XVLQJ ERWK WKH H[SHULPHQWDO DQG VLPXODWHG H[DPSOHV [L
PAGE 12
&+$37(5 ,1752'8&7,21 %DFNJURXQG 5$'$5 VWDQGV IRU 5$GLR 'HWHFWLRQ $QG 5DQJLQJ %\ QRZ UDGDU KDV EHHQ XVHG QRW RQO\ WR GHWHFW DQG ORFDWH WDUJHWV DV ZDV RULJLQDOO\ LQWHQGHG EXW WR JHQn HUDWH LPDJHV RI LOOXPLQDWHG VFHQHV ZLWKRXW WKH UHVWULFWLRQ RI ZHDWKHU DQG WLPH DV ZHOO 5DGDU LPDJH LV GHILQHG DV D WZRGLPHQVLRQDO 'f PDSSLQJ RI WKH UHIOHFWLYLW\ FRUUHVSRQGLQJ WR WKH VFHQH LOOXPLQDWHG E\ D UDGDU 0RUH VSHFLILFDOO\ D UDGDU LPDJH LV WKH UHFRQVWUXFWLRQ RI WKH LOOXPLQDWHG VFHQH IURP WKH HOHFWURPDJQHWLF UHIOHFWLYLW\ ZKLFK LV XVXDOO\ WUHDWHG DV D IXQFWLRQ LQ UDQJH DQG FURVVUDQJH 7KH GLUHFWLRQ LQ ZKLFK D UDGDU EHDP SURSDJDWHV LV UHIHUUHG WR DV WKH UDQJH GLUHFWLRQ ZKLOH WKH RQH SHUSHQGLFXODU WR WKH OLQH RI VLJKW LV FDOOHG WKH FURVVUDQJH GLUHFWLRQ RU WKH D]n LPXWK GLUHFWLRQ VHH )LJXUH IRU DQ LOOXVWUDWLRQf 5DGDU LPDJLQJ >@ KDV SOD\HG D FUXFLDO UROH LQ ERWK PLOLWDU\ DSSOLFDWLRQV LQFOXGLQJ UHFRQQDLVVDQFH DQG VXUYHLOODQFH >@ DQG FLYLOLDQ RQHV LQFOXGLQJ JHRORJLFDO DQG WRSRJUDSKLF PDS JHQHUDWLRQ ZHDWKHU IRUHFDVWLQJ VFHQH FODVVLILFDWLRQ DQG WDUJHW UHFRJQLWLRQ > @ 7KH TXDOLW\ RI UDGDU LPDJHV GHSHQGV JUHDWO\ RQ WKH LPDJH UHVROXWLRQ ZKLFK LV QRPLQDOO\ UHIHUUHG WR DV WKH DELOLW\ WR GLVWLQJXLVK WZR FORVHO\VSDFHG HOHPHQWV RI D WDUJHW >@ 7KH KLJKHU WKH UDGDU UHVROXWLRQ WKH PRUH GHWDLOV RI WDUJHW IHDWXUHV DUH DYDLODEOH IRU GHWHFWLRQ FKDUDFWHUL]DWLRQ DQG LGHQWLILFDWLRQ RI WDUJHWV RI LQWHUHVW $V VXFK PXFK HIIRUW KDV EHHQ PDGH WR LPSURYH LPDJLQJ UDGDU UHVROXWLRQV ZKLFK LQFOXGH UDQJH DQG FURVVUDQJH UHVROXWLRQV +LJK UDQJH UHVROXWLRQ FDQ EH DFKLHYHG E\ XVLQJ ZLGH EDQGZLGWK VLJQDOV VXFK DV OLQHDU IUHTXHQF\ PRGXODWHG )0f FKLUS SXOVHV >@ VLQFH UDQJH UHVROXWLRQ LV LQYHUVHO\ SURSRUWLRQDO WR WKH WUDQVPLWWHG VLJQDO
PAGE 13
)OLJKW 3DWK Z &URVV5DQJH )LJXUH ,OOXVWUDWLRQ RI UDGDU LPDJLQJ EDQGZLGWK 7KH FURVVUDQJH UHVROXWLRQ & RI D FRQYHQWLRQDO UDGDU RU UHDO DSHUWXUH UDGDU 5$5f LV UDQJHGHSHQGHQW DQG LV GHWHUPLQHG E\ f ZKHUH 5 UHSUHVHQWV WKH UDQJH RI WKH WDUJHW $ DQG / UHVSHFWLYHO\ GHQRWH WKH ZDYHn OHQJWK DQG WKH VL]H RI WKH DQWHQQD DSHUWXUH 7R LQFUHDVH FURVVUDQJH UHVROXWLRQ JLYHQ 5f DQ REYLRXV ZD\ LV HLWKHU WR UHGXFH WKH ZDYHOHQJWK RU WR LQFUHDVH WKH VL]H RI WKH DQWHQQD 8VLQJ VKRUWHU ZDYHOHQJWK UHTXLUHV PRUH WUDQVPLVVLRQ SRZHU VLQFH HOHFWURn PDJQHWLF VLJQDOV ZLWK VPDOOHU ZDYHOHQJWKV DUH PRUH OLNHO\ DEVRUEHG E\ WKH DWPRn VSKHUH ,Q WKH PHDQWLPH LW LV QRW IHDVLEOH WR LQFUHDVH WKH DQWHQQD VL]H WRR PXFK HVSHFLDOO\ ZKHQ WKH UDGDU LV FDUULHG RQ ERDUG RI DQ DLUSODQH RU VSDFHFUDIW +HQFH D FRQYHQWLRQDO UDGDU RU 5$5 LQKHUHQWO\ KDV D OLPLWHG FURVVUDQJH UHVROXWLRQ ,W LV GHVLUHG WR DFKLHYH KLJK FURVVUDQJH UHVROXWLRQ ZLWKRXW UHGXFLQJ WKH ZDYHOHQJWK DQG H[SORLWLQJ ODUJH SK\VLFDO DQWHQQDV 6\QWKHWLF DSHUWXUH UDGDU 6$5f LV D VROXWLRQ DQG FDQ DFKLHYH KLJK FURVVUDQJH UHVROXWLRQ LQGHSHQGHQW RI UDQJH 7KH 6$5 WHFKQRORJ\ ZDV ILUVW SURSRVHG LQ WKH V >@ ,QVWHDG RI H[SORLWLQJ ODUJH DQWHQQDV DQG UHGXFLQJ ZDYHOHQJWK 6$5 PDNHV XVH RI D V\QWKHWLF DSHUWXUH WR DFKLHYH KLJK FURVVUDQJH UHVROXWLRQ $ V\QWKHWLF DSHUWXUH LV IRUPHG E\ FRKHUHQWO\
PAGE 14
)LJXUH ,OOXVWUDWLRQ RI 6$5 LPDJLQJ PRGHV Df 6SRWOLJKWPRGH 6$5 Ef 6WULS PDS 6$5 SURFHVVLQJ UDGDU UHWXUQV IURP D VHTXHQFH RI ORFDWLRQV ZKHQ D QRPLQDOO\VL]HG DQWHQQD ZKLFK LV FDUULHG RQ ERDUG RI DQ DLUFUDIW VHTXHQWLDOO\ PRYHV DORQJ WKH IOLJKW SDWK FURVVUDQJH GLUHFWLRQf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f 6$5 KDV JURZQ VR GUDPDWLFDOO\ WKURXJK DGYDQFHV LQ GLJLWDO VLJQDO SURFHVVLQJ FRPSXWLQJ SRZHU DQG PLFURZDYH WHFKQRORJ\ WKDW LW LV RQH RI WKH PRVW LPSRUWDQW WHFKQLTXHV IRU KLJK UHVROXWLRQ UDGDU LPDJH IRUPDWLRQ 6SRWOLJKWPRGH DQG VWULSPDS 6$5 DUH WZR FRPPRQO\ XVHG 6$5 LPDJLQJ PRGHV ,Q D VSRWOLJKWPRGH 6$5 DQWHQQD EHDPV FRQWLQXRXVO\ VWHHU WR WKH VDPH WHUUDLQ SDWFK VFHQHf LOOXPLQDWHG E\ D UDGDU ZKHQ WKH DLUFUDIW WUDYHUVHV WKH IOLJKW SDWK ZKLOH WKH DQWHQQD VWHHULQJ RI D VWULSPDS 6$5 LV IL[HG UHODWLYH WR WKH IOLJKW SDWK ZKHQ WKH DLUFUDIW PRYHV DORQJ WKH IOLJKW SDWK VHH )LJXUH IRU WKH GLIIHUHQFHf
PAGE 15
7KH FURVVUDQJH UHVROXWLRQ RI VWULSPDS 6$5 LV OLPLWHG E\ WKH D]LPXWK EHDPZLGWK ZKLOH WKDW RI VSRWOLJKWPRGH 6$5 LV LQGHSHQGHQW RI WKH D]LPXWK EHDPZLGWK VLQFH WKH DQWHQQD VWHHUV LWV EHDP WR FRQWLQXRXVO\ LOOXPLQDWH WKH VDPH VFHQH +HQFH 6SRWOLJKWn PRGH 6$5 LV PRUH VXLWDEOH IRU KLJK UHVROXWLRQ 6$5 LPDJH IRUPDWLRQ RI D VPDOO VFHQH WKDQ VWULSPDS 6$5 6FRSH RI WKH :RUN 6SRWOLJKWPRGH 6$5 KDV EHHQ H[WHQVLYHO\ DSSOLHG WR KLJK UHVROXWLRQ LPn DJH IRUPDWLRQ DQG WDUJHW IHDWXUH H[WUDFWLRQ RI VPDOO VFHQHV GXH WR LWV ILQHU UHVROXWLRQV DQG UHDGLO\ FRQWUROODEOH IOLJKW SDWK ZKLFK LV D VWUDLJKW OLQH 7KH WDUJHW IHDWXUHV LQFOXGH WKH UDGDU FURVV VHFWLRQ 5&6f DQG WKH ORFDWLRQ UDQJH DQG FURVVUDQJHf RI HDFK VFDWWHUHU 2QH WRSLF RI RXU ZRUN LV VXSHU UHVROXWLRQ LPDJH IRUPDWLRQ IURP SKDVH KLVWRU\ GDWD FROOHFWHG E\ VSRWOLJKWPRGH 6$5 7KH VXSHU UHVROXWLRQ KHUHLQ LV GHILQHG DV WKH UHVROXWLRQ WKDW LV EHWWHU WKDQ WKH 5D\OHLJK OLPLWV GHWHUPLQHG E\ WKH WUDQVPLVVLRQ ZDYHOHQJWK DQG WKH OHQJWK RI WKH V\QWKHWLF DSHUWXUH RI 6$5 $ FRQYHQn WLRQDO DSSURDFK IRU LPDJH IRUPDWLRQ IURP 6$5 SKDVH KLVWRU\ GDWD LV WKH IDVW )RXULHU WUDQVIRUP ))7f ZKLFK LV FRPSXWDWLRQDOO\ HIILFLHQW DQG UREXVW +RZHYHU 6$5 LPDJHV REWDLQHG YLD ))7 VXIIHU IURP SRRU UHVROXWLRQ DQG KLJK VLGHOREHV 6$5 LPDJH IRUPDWLRQ FDQ EH FRQVLGHUHG DV D VSHFWUDO HVWLPDWLRQ SUREOHP VLQFH 6$5 LPDJHV DUH WKH UHFRQVWUXFWLRQ RI VFHQH UHIOHFWLYLW\ LQWHQVLWLHV YHUVXV VSDFH ORFDWLRQV 0RGn HUQ VSHFWUDO HVWLPDWLRQ PHWKRGV LQFOXGLQJ SDUDPHWULF DQG QRQSDUDPHWULF PHWKRGV KDYH IRXQG H[WHQVLYH DSSOLFDWLRQV LQ LPSURYLQJ WKH UHVROXWLRQ DQG DFFXUDF\ RI WKH IRUPHG 6$5 LPDJHV ,Q SDUWLFXODU SDUDPHWULF VSHFWUDO HVWLPDWLRQ PHWKRGV DUH PRUH VXLWDEOH WR IRUP VXSHU UHVROXWLRQ 6$5 LPDJHV ZKHQ SDUDPHWULF GDWD PRGHOV FDQ EH XVHG WR PRGHO WKH WDUJHW 7KH VXSHU UHVROXWLRQ RI WKH IRUPHG 6$5 LPDJHV FDQ EH DFKLHYHG E\ H[WUDSRODWLQJ WKH SKDVH KLVWRU\ GDWD EH\RQG REVHUYDWLRQV XVLQJ DQ DSSURn SULDWH GDWD PRGHO DQG WKH SDUDPHWHU HVWLPDWHV :H ZLOO GLVFXVV KRZ WR IRUP VXSHU
PAGE 16
UHVROXWLRQ 6$5 LPDJHV E\ XVLQJ WKH WDUJHW IHDWXUHV H[WUDFWHG YLD UHOD[DWLRQEDVHG PHWKRGV LQFOXGLQJ 5(/$; >@ DQG 5(/$;1/6 >@ LQ &KDSWHU 7KH UHOD[DWLRQ EDVHG RSWLPL]DWLRQ DSSURDFK LV RQH RI WKH HIIHFWLYH ZD\V WR PLQLPL]H D QRQOLQHDU OHDVW VTXDUHV 1/6f FRVW IXQFWLRQ DQG KDV EHHQ SURYHG WR EH TXLWH XVHIXO LQ PDQ\ DSSOLFDn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f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
PAGE 17
Df Ff H )LJXUH 3RVVLEOH DSHUWXUHV IRU D 6$5 V\VWHP Df )XOO DSHUWXUH Ef 3DUDEROLF DSHUWXUH Ff 2UWKRJRQDO VXEDSHUWXUH DQG c! GHQRWH WKH D]LPXWK DQG HOHYDWLRQ DQJOHV UHVSHFWLYHO\f QR DFFXUDWH SDUDPHWULF GDWD PRGHO LV DYDLODEOH LQ FURVVUDQJH %DVHG RQ WKH IOH[LEOH GDWD PRGHO ZH SUHVHQW D UREXVW DQG FRPSXWDWLRQDOO\ VLPSOH VHPLSDUDPHWULF 63$5f DOJRULWKP IRU WDUJHW IHDWXUH H[WUDFWLRQ DQG 6$5 LPDJH IRUPDWLRQ E\ H[SORLWLQJ D K\EULG RI QRQSDUDPHWULF DQG SDUDPHWULF VSHFWUDO HVWLPDWLRQ PHWKRGHV ,W KDV EHHQ VKRZQ LQ > @ WKDW WKUHHGLPHQVLRQDO 'f WDUJHW IHDWXUHV LQFOXGLQJ WKH KHLJKW LQIRUPDWLRQ DV ZHOO DV WKH WDUJHW IHDWXUHV SURYLGH YHU\ XVHIXO LQIRUPDWLRQ LQ FHUWDLQ DSSOLFDWLRQV VXFK DV DXWRPDWLF WDUJHW UHFRJQLWLRQ $75f 6$5 LPDJH IRUPDWLRQ DQG IHDWXUH H[WUDFWLRQ DUH EHFRPLQJ DFWLYH UHVHDUFK DUHDV ,W LV SRVVLEOH WR REWDLQ LQIRUPDWLRQ RI DQ LOOXPLQDWHG VFHQH ZLWK 6SRWOLJKWPRGH 6$5 ZKHQ D QRQVWUDLJKW OLQH V\QWKHWLF DSHUWXUH LV H[SORLWHG ZKLFK LV GHWHUPLQHG E\ WKH D]LPXWK DQJOH DQG HOHYDWLRQ DQJOH M! ,Q RXU ZRUN ZH ZLOO DOVR VWXG\ WDUJHW IHDWXUH H[WUDFWLRQ DQG PRWLRQ FRPSHQVDWLRQ DXWRIRFXVf XVLQJ FXUYLOLQHDU 6$5 LQ ZKLFK WKH WUDMHFWRU\ RI D VSRWOLJKWPRGH 6$5 LV D VXLWDEOH FXUYLOLQHDU DSHUWXUH $PRQJ H[LVWLQJ 6$5 LPDJLQJ WHFKQLTXHV LQWHUIHURPHWULF 6$5 ,)6$5f LV QRZ D ZHOOGHYHORSHG WHFKQLTXH WKDW SOD\V DQ LPSRUWDQW UROH LQ DV ZHOO DV 'f LPDJLQJ +RZHYHU ,)6$5 VXIIHUV IURP DPELJXLW\ SUREOHPV > @ 7KHRUHWLFDOO\ WDUJHW IHDWXUHV DQG LPDJHV FDQ DOVR EH REWDLQHG ZKHQ WKH WUDMHFWRU\ RI 6$5 LV D IXOO DSHUWXUH DV VKRZQ LQ )LJXUH Df HYHQ WKRXJK VXFK DQ DSHUWXUH LV QRW IHDVLEOH LQ
PAGE 18
Df Ef )LJXUH 0HVK SORWV RI WKH PRGXOXV RI WKH 5&6 REWDLQHG E\ XVLQJ ))7 ZLWK GLIIHUHQW DSHUWXUHV UDQJH LQIRUPDWLRQ VXSSUHVVHG IRU WKH LOOXVWUDWLRQ SXUSRVH RQO\f Df )XOO DSHUWXUH DV VKRZQ LQ )LJXUH SURWHFW Df Ef &XUYLOLQHDU DSHUWXUH DV VKRZQ LQ )LJXUH Ff SUDFWLFH 1HYHUWKHOHVV D VXLWDEOH VXEVHW RI WKH IXOO DSHUWXUH FDQ EH XVHG DV D WUDMHFWRU\ RI 6$5 WR REWDLQ LQIRUPDWLRQ 7KLV WHFKQLTXH LV NQRZQ DV FXUYLOLQHDU 6$5 &/6$5f &/6$5 WUDYHUVHV D FXUYLOLQHDU DSHUWXUH SDWK ZKLFK LV D VXEVHW RI WKH IXOO DSHUWXUH VHH IRU H[DPSOH )LJXUHV Ef DQG Fff DQG FROOHFWV D VSDUVHO\ VDPSOHG GDWDVHW IURP ZKLFK LQIRUPDWLRQ FDQ EH REWDLQHG &/6$5 DYRLGV DPELJXLW\ SUREOHPV LQKHUHQW LQ ,)6$5 V\VWHPV EXW VXIIHUV IURP VHYHUH KLJK VLGHOREHV ZKHQ XVHG ZLWK ))7 WR IRUP 6$5 LPDJHV GXH WR WKH OLPLWHG VL]H RI PHDVXUHPHQWV 7KLV REVHUYDWLRQ LV LOOXVWUDWHG LQ )LJXUH )LJXUHV Df DQG Ef VKRZ WKH PRGXOXV RI WKH 5&6 RI D VLQJOH SRLQW VFDWWHUHU REWDLQHG E\ DSSO\LQJ ))7 WR WKH GDWD FROOHFWHG YLD WKH IXOO DSHUWXUH DV VKRZQ LQ )LJXUH Df DQG WKH FXUYLOLQHDU DSHUWXUH DV VKRZQ LQ )LJXUH Fff 'XH WR LWV KLJK VLGHOREHV &/6$5 LV RI OLWWOH LQWHUHVW LQ 6$5 LPDJLQJ @ FDQ EH XVHG IRU WDUJHW IHDWXUH H[WUDFWLRQ XVLQJ &/6$5 ZLWK GLIIHUHQW FXUYLOLQHDU DSHUWXUHV
PAGE 19
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f PD\ QRW EH QHJOLJLEOH 0RWLRQ FRPSHQVDWLRQ DXWRIRFXVf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f VXSHU UHVROXWLRQ 6$5 LPDJH IRUPDWLRQ f :H KDYH H[WHQGHG 5(/$; WR IRUP HQKDQFHG 6$5 LPDJHV RI WDUJHWV FRQVLVWLQJ RI SRLQW VFDWWHUHUV f :H KDYH H[WHQGHG 5(/$;1/6 IRU VXSHU UHVROXWLRQ 6$5 LPDJH IRUPDWLRQ RI WDUJHWV FRQVLVWLQJ RI ERWK WULKHGUDOV DQG GLKHGUDOV
PAGE 20
f :H KDYH GHYLVHG D QRYHO UREXVW DQG FRPSXWDWLRQDOO\ VLPSOH 6HPL3$5DPHWULF 63$5f DOJRULWKP IRU 6$5 WDUJHW IHDWXUH H[WUDFWLRQ DQG LPDJH IRUPDWLRQ EDVHG RQ D IOH[LEOH GDWD PRGHO f ([SHULPHQWDO UHVXOWV XVLQJ (5,0 DQG 067$5 6OLF\ GDWD VKRZ WKDW WKH UHOD[DWLRQ EDVHG SDUDPHWULF DSSURDFKHV LQFOXGLQJ 5(/$; DQG 5(/$;1/6 DQG WKH VHPL SDUDPHWULF DOJRULWKP 63$5 FDQ \LHOG 6$5 LPDJHV ZLWK KLJKHU UHVROXWLRQ WKDQ ))7 EDVHG PHWKRGV $OVR 63$5 LV PRUH UREXVW DQG FRPSXWDWLRQDOO\ VLPSOH WKDQ 5(/$;1/6 ,,f WDUJHW IHDWXUH H[WUDFWLRQ DQG PRWLRQ FRPSHQVDWLRQ YLD &/6$5 f :H KDYH HVWDEOLVKHG D GDWD PRGHO f :H KDYH SHUIRUPHG &UDPU5DR %RXQG &5%f DQDO\VHV RI SDUDPHWHU HVWLPDn WLRQ XVLQJ YDULRXV IOLJKW WUDMHFWRULHV ZKLFK SURYLGH PRUH LQVLJKW LQWR WKH SHUn IRUPDQFHV DVVRFLDWHG ZLWK GLIIHUHQW FXUYLOLQHDU DSHUWXUHV f 7KH HIIHFWV RI DSHUWXUH HUURUV RQ WDUJHW IHDWXUH HVWLPDWLRQ KDYH EHHQ LQYHVWLn JDWHG DQG D UHOD[DWLRQEDVHG DSSURDFK IRU DXWRIRFXV DQG WDUJHW IHDWXUH H[WUDFn WLRQ XVLQJ &/6$5 KDV EHHQ LQWURGXFHG 'LVVHUWDWLRQ 2XWOLQH 7KH UHPDLQGHU RI WKH GLVVHUWDWLRQ LV RUJDQL]HG DV IROORZV &KDSWHU JLYHV D OLWHUDWXUH VXUYH\ RI HPSKDVLV RQ WKH WRSLFV FRQVLGHUHG LQFOXGLQJ 6$5 LPDJLQJ 6$5 WDUJHW IHDWXUH H[WUDFWLRQ DQG PRWLRQ FRPSHQVDWLRQ DXWRIRFXVf ,Q &KDSWHU ZH GHVFULEH KRZ WR IRUP VXSHU UHVROXWLRQ 6$5 LPDJHV YLD WKH UHOD[DWLRQ EDVHG SDUDPHWULF VSHFWUDO HVWLPDWLRQ DOJRULWKPV LQFOXGLQJ 5(/$; DQG 5(/$; 1/6 $ QRYHO VHPLSDUDPHWULF 63$5f PHWKRG LV LQWURGXFHG LQ &KDSWHU IRU 6$5 LPDJH IRUPDWLRQ ZKHQ DQ DFFXUDWH SDUDPHWULF GDWD PRGHO LQ FURVVUDQJH LV GLIILFXOW
PAGE 21
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
PAGE 22
&+$37(5 /,7(5$785( 6859(< ,Q WKLV FKDSWHU ZH JLYH D EULHI UHYLHZ RI WKH WRSLFV UHODWHG WR WKH ZRUN QDPHO\ VXSHU UHVROXWLRQ 6$5 LPDJH IRUPDWLRQ 6$5 WDUJHW IHDWXUH H[WUDFWLRQ DQG 6$5 PRWLRQ FRPSHQVDWLRQ 6XSHU 5HVROXWLRQ 6$5 ,PDJH )RUPDWLRQ 6LQFH 6$5 LPDJH IRUPDWLRQ FDQ EH FRQVLGHUHG DV D VSHFWUDO HVWLPDWLRQ SUREn OHP PRGHUQ VSHFWUDO HVWLPDWLRQ PHWKRGV ILQG H[WHQVLYH DSSOLFDWLRQV LQ LPSURYLQJ WKH UHVROXWLRQ DQG DFFXUDF\ RI WKH IRUPHG 6$5 LPDJHV ([FHOOHQW WH[WV RQ VSHFWUDO HVWLPDWLRQ DUH DYDLODEOH > @ DQG KLVWRULFDO DQG PRGHUQ SHUVSHFWLYHV RQ WKH JHQHUDO WRSLFV RI VSHFWUDO DQDO\VLV FDQ EH IRXQG LQ > @ $ EULHI UHYLHZ RI VXSHU UHVn ROXWLRQ 6$5 LPDJH IRUPDWLRQ YLD QRQSDUDPHWULF DQG SDUDPHWULF VSHFWUDO HVWLPDWLRQ PHWKRGV LV JLYHQ EHORZ 1RQSDUDPHWULF 6SHFWUDO (VWLPDWRUV $Q LPSRUWDQW FODVV RI QRQSDUDPHWULF VSHFWUDO HVWLPDWRUV DUH WKH DGDSWLYH ILOWHUEDQN DSSURDFKHV ZKLFK FDQ EH XVHG WR IRUP 6$5 LPDJHV ZLWK EHWWHU UHVROXn WLRQ DQG ORZHU VLGHOREHV WKDQ WKH ))7 EDVHG PHWKRGV >@ 7KH DGDSWLYH VLGHOREH UHGXFWLRQ $65f DSSURDFK >@ DQG WKH ZLGHO\ XVHG &DSRQ PHWKRG >@ DV ZHOO DV LWV UHGXFHG UDQN YDULDWLRQV > @ DUH PHPEHUV RI WKLV FODVV $ UHFHQWO\ LQWURGXFHG PDWFKHGILOWHU EDQN EDVHG FRPSOH[ VSHFWUDO HVWLPDWRU WKDW LV RI LQWHUHVW LV WKH $3(6 $PSOLWXGH DQG 3KDVH (VWLPDWLRQ RI D 6LQXVRLGf PHWKRG >@ 'HVSLWH WKHLU GLIIHUHQW GHULYDWLRQV LW ZDV IRXQG WKDW ERWK &DSRQ DQG $3(6 DUH PHPEHUV RI WKH JHQHUDO FODVV RI WKH PDWFKHGILOWHUEDQN 0$),f DSSURDFKHV WR FRPSOH[ VSHFWUDO HVWLPDWLRQ
PAGE 23
>@ (PSLULFDOO\ LW ZDV REVHUYHG LQ >@ WKDW &DSRQ DSSHDUV WR JLYH ELDVHG VSHFn WUDO HVWLPDWHV ZKHUHDV $3(6 LV XQELDVHG 7KLV REVHUYDWLRQ ZDV ODWHU SURYHG LQ >@ E\ XVLQJ WKH VHFRQGRUGHU 7D\ORU H[SDQVLRQ WHFKQLTXH $OVR WKH VWDWLVWLFDO SHUIRUn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n ULF VSHFWUDO HVWLPDWRUV SURYLGH DWWUDFWLYH DOWHUQDWLYHV IRU VXSHU UHVROXWLRQ 6$5 LPDJH IRUPDWLRQ 7KH %XUJfV PD[LPXP HQWURS\ 0(f PHWKRG >@ ZDV DSSOLHG WR UDGDU LPn DJH HQKDQFHPHQW E\ 0DQFLOO LQ >@ 8QOLNH ))7 PHWKRGV ZKLFK DVVXPH WKDW WKH VLJQDO EH\RQG WKH REVHUYDWLRQV LV ]HUR 0( PHWKRGV PDNH OLQHDU SUHGLFWLRQ RI WKH VLJQDO EH\RQG WKH REVHUYDWLRQ ZLQGRZ VR WKDW WKH\ FDQ DFKLHYH EHWWHU UHVROXWLRQ WKDQ ))7 PHWKRGV ,GHQWLFDO WR 0( PHWKRGV LQ SULQFLSOHV WKH DXWRUHJUHVVLYH $5f PRGHO EDVHG PHWKRGV > @ KDYH DWWUDFWHG D ORW RI UHVHDUFK DWWHQWLRQ +RZHYHU WKH KLJK UHVROXWLRQ SURSHUW\ RI WKH $5 EDVHG PHWKRGV DUH KLJKO\ GHSHQGHQW RQ WKH SUHVHQFH RI REVHUYDWLRQ QRLVH >@ $V WKH VLJQDOWRQRLVH UDWLR 615f GHFUHDVHV VR GRHV WKH UHVROXWLRQ 2Q WKH RWKHU KDQG WKH $5 EDVHG PHWKRGV DUH PRUH DSSURSULDWH IRU 6$5 LPDJH IRUPDWLRQ RI GLVWULEXWHG VFDWWHUHUV UDWKHU WKDQ SRLQW VFDWWHUHUV VLQFH WKH $5 VSHFWUDO HVWLPDWRU E\ QDWXUH LV D FRQWLQXRXV VSHFWUDO HVWLPDWRU
PAGE 24
,Q UHFHQW \HDUV PDQ\ VRSKLVWLFDWHG DQG KLJK UHVROXWLRQ PHWKRGV EDVHG RQ SRLQW VFDWWHUHU GDWD PRGHO VXFK DV HLJHQGHFRPSRVLWLRQ EDVHG PHWKRGV >@ LQFOXGLQJ 086,& 0XOWLSOH 6LJQDO &ODVVLILFDWLRQf >@ DQG (635,7 (VWLPDWLRQ RI 6LJQDO 3Dn UDPHWHUV YLD 5RWDWLRQDO ,QYDULDQFH 7HFKQRORJ\f >@ DQG UHOD[DWLRQEDVHG QRQOLQHDU OHDVW VTXDUHV 1/6f PHWKRG VXFK DV 5(/$; >@ KDYH DOVR EHHQ DSSOLHG WR VXSHU UHVROXWLRQ 6$5 LPDJH IRUPDWLRQ 7KH HLJHQGHFRPSRVLWLRQ EDVHG PHWKRGV DUH VLJQDO RU QRLVH VXEVSDFH EDVHG DSSURDFKHV DQG JHQHUDOO\ DUH QRW VXLWDEOH IRU 6$5 LPDJH IRUPDWLRQ VLQFH WKH\ UHTXLUH WKDW WKH QXPEHU RI GDWD VDPSOHV EH PXFK ODUJHU WKDQ WKDW RI WKH GRPLQDQW SRLQW VFDWWHUHUV DQG WKH SURFHVV RI ZKLWHQLQJ WKH QRLVH VXEn VSDFH XVHG LQ 086,&f ZLOO GHVWUR\ WKH VSDWLDO LQKRPRJHQHLWLHV RI WKH WHUUDLQ FOXWWHU >@ 7KH 5(/$; DOJRULWKP >@ ZDV RULJLQDOO\ SURSRVHG IRU VLQXVRLGDO SDUDPHWHU HVWLPDWLRQ LQ FRORUHG QRLVH 5(/$; REWDLQV WKH SDUDPHWHU HVWLPDWHV E\ PLQLPL]LQJ WKH 1/6 FRVW IXQFWLRQ YLD D VHTXHQFH RI ))7V ZKLFK LV FRQFHSWXDOO\ DQG FRPSXWDn WLRQDOO\ VLPSOH DQG KHQFH UREXVW 5(/$; FRLQFLGHV ZLWK WKH PD[LPXP OLNHOLKRRG 0/f PHWKRG ZKHQ WKH QRLVH LV ZKLWH DQG LV DOVR DV\PSWRWLFDOO\ IRU ODUJH QXPEHUV RI GDWD VDPSOHVf VWDWLVWLFDO HIILFLHQW ZKHQ WKH QRLVH LV DQ $5 RU DXWRUHJUHVVLYH PRYn LQJ DYHUDJH $50$f SURFHVV WKDW FDQ EH GHVFULEHG E\ D ILQLWH QXPEHU RI XQNQRZQV >@ 'XH WR WKH LWHUDWLRQV SHUIRUPHG LQ HDFK VWHS RI 5(/$; D VHW RI PRUH DFFXUDWH LQLWLDO HVWLPDWHV DUH DYDLODEOH ZKLFK PDNH 5(/$; FRQYHUJH WR WKH JOREDO PLQLPXP ZLWK D KLJK SUREDELOLW\ 5(/$; FDQ DOVR EH UHIHUUHG WR DV 683(5 &/($1 ZKHQ FRPSDUHG ZLWK WZR RWKHU DSSUR[LPDWH UHOD[DWLRQ PHWKRGV LQFOXGLQJ &/($1 >@ DQG 0&/($1 PRUH FOHDQf >@ &/($1 ZDV ILUVW SURSRVHG LQ UDGLR DVWURQRP\ DQG ODWHU XVHG LQ PLFURZDYH LPDJLQJ >@ ,I QR LWHUDWLRQ LV SHUIRUPHG LQ HDFK VWHS RI 5(/$; WKHQ LW LV HTXLYDOHQW WR &/($1 5(/$; EHFRPHV 0&/($1 LI RQO\ RQH LWHUDWLRQ LV SHUIRUPHG LQ HDFK VWHS RI 5(/$; +HQFH 5(/$; LV D FRPSXWDWLRQDOO\ HIILFLHQW DQG UREXVW DOJRULWKP ZLWK JRRG VWDWLVWLFDO SHUIRUPDQFH DQG LV VXLWDEOH IRU VXSHU UHVROXWLRQ 6$5 LPDJH IRUPDWLRQ RI WDUJHWV FRQVLVWLQJ RI SRLQW VFDWWHUHUV WULKHGUDOVf
PAGE 25
5HFHQWO\ 5(/$;1/6 >@ ZDV LQWURGXFHG IRU WKH IHDWXUH H[WUDFWLRQ RI WDUJHWV FRQVLVWLQJ RI ERWK WULKHGUDOV DQG GLKHGUDOV EDVHG RQ D PL[HG GDWD PRGHO UDWKHU WKDQ WKH SRLQW VFDWWHUHU GDWD PRGHO $Q LGHDO WULKHGUDO DFWV DV D WULSOHERXQFH UHIOHFWRU DQG FDQ EH PRGHOHG DV D VWURQJ SRLQW VFDWWHUHU >@ 7KH SRLQW VFDWWHUHU GDWD PRGHO KRZHYHU LV QR ORQJHU YDOLG LQ FURVVUDQJH IRU DQ LGHDO GLKHGUDO FRUQHU UHIOHFWRU ZKLFK DFWV DV D GRXEOHERXQFH UHIOHFWRU VLQFH LWV UDGDU FURVV VHFWLRQ 5&6f LV GHVFULEHG E\ D VLQH IXQFWLRQ LH VLQFf VLQf ZKHUH GHQRWHV WKH DQJOH EHWZHHQ WKH UDGDU EHDP DQG WKH OLQH SHUSHQGLFXODU WR WKH GLKHGUDO 6XSHU UHVROXWLRQ 6$5 LPDJHV REWDLQHG YLD PRGHUQ VSHFWUDO HVWLPDWLRQ PHWKn RGV SURYLGH PRUH GHWDLOV UHTXLUHG IRU WKH GHWHFWLRQ DQG DXWRPDWLF WDUJHW UHFRJQLWLRQ $75f WKDQ WKH FRQYHQWLRQDO ))7 EDVHG PHWKRGV 4XDQWLWDWLYH UHVXOWV VKRZLQJ WKH DGYDQWDJHV RI XVLQJ WKHVH VRSKLVWLFDWHG VSHFWUDO HVWLPDWLRQ PHWKRGV LQ 6$5 DSSOLn FDWLRQV EHJLQ WR DSSHDU LQ WKH OLWHUDWXUH )RU H[DPSOH LW ZDV VKRZQ LQ > @ WKDW D VLJQLILFDQW LPSURYHPHQW LQ $75 SHUIRUPDQFH FDQ EH DFKLHYHG E\ XVLQJ 6$5 LPn DJHV ZLWK HQKDQFHG UHVROXWLRQ REWDLQHG YLD WKH UHGXFHGUDQN YDULDWLRQV RI WKH &DSRQ PHWKRG &RPSDUHG ZLWK QRQSDUDPHWULF VSHFWUDO HVWLPDWRUV KLJK UHVROXWLRQ SDUDPHWn ULF VSHFWUDO HVWLPDWRUV SURYLGH JUHDW SRWHQWLDO IRU 6$5 LPDJLQJ ZLWK EHWWHU UHVROXWLRQ DQG ORZHU VLGHOREHV ZKHQ WKH DVVXPHG GDWD PRGHO LV YDOLG 2Q WKH RWKHU KDQG SDUDn PHWULF PHWKRGV DUH PRUH VHQVLWLYH WR GDWD PRGHO PLVPDWFK WKDQ QRQSDUDPHWULF PHWKRGV +HQFH SUDFWLFDOO\ RQO\ UREXVW VXSHU UHVROXWLRQ SDUDPHWULF PHWKRGV FDQ EH XVHG IRU 6$5 LPDJH IRUPDWLRQ :H ZLOO XVH UREXVW UHOD[DWLRQEDVHG DOJRULWKPV LQn FOXGLQJ 5(/$; DQG 5(/$;1/6 WR IRUP 6$5 LPDJHV RI GLIIHUHQW PDQPDGH WDUJHWV LQ RXU ZRUN ,Q FDVHV ZKHUH DQ DSSURSULDWH GDWD PRGHO RI PDQPDGH WDUJHWV LV QRW DYDLODEOH LQ FURVVUDQJH UREXVW QRQSDUDPHWULF VSHFWUDO HVWLPDWRUV ZLWK JRRG SHUIRUn PDQFHV VXFK DV $3(6 DQG KLJK UHVROXWLRQ SDUDPHWULF PHWKRGV PD\ ZRUN WRJHWKHU WR REWDLQ VDWLVIDFWRU\ UHVXOWV 7R WKH EHVW RI RXU NQRZOHGJH D K\EULG RI QRQSDUDPHWULF
PAGE 26
DQG SDUDPHWULF PHWKRGV IRU IHDWXUH H[WUDFWLRQ DQG LPDJH IRUPDWLRQ RI 6$5 WDUJHWV KDV QRW EHHQ DGGUHVVHG LQ WKH OLWHUDWXUH EHIRUH 6$5 )HDWXUH ([WUDFWLRQ 2QH UHODWLYHO\ PDWXUH WHFKQRORJ\ IRU 6$5 LPDJLQJ DQG IHDWXUH H[WUDFn WLRQ LV ,)6$5 > @ ZKLFK H[SORLWV WZR GLIIHUHQW DSHUWXUHV $ VLPSOH LPSOHPHQWDWLRQ RI D ,)6$5 LV WR XVH WZR DQWHQQD GLVSODFHG LQ WKH FURVVWUDFN SODQH WR REWDLQ WZR FRKHUHQW DQG SDUDOOHO PHDVXUHPHQWV 7KH UHODWLYH KHLJKW LQIRUPDWLRQ FDQ EH HVWLPDWHG IURP WKH SKDVH GLIIHUHQFH EHWZHHQ WKH WZR PHDVXUHPHQWV 7KH XVH RI LQWHUIHURPHWULF WHFKQLTXHV ZLWK 6$5 LPDJHU\ WR SURYLGH KHLJKW LQIRUPDWLRQ PDNHV ,)6$5 VXLWDEOH IRU LPDJLQJ VLQFH LW LQKHULWV WKH ILQHU UHVROXWLRQV LQ UDQJH DQG FURVVUDQJHf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n HQW LQ ,)6$5 V\VWHPV &/6$5 LPDJHV DUH XVHOHVV LQ SUDFWLFH GXH WR KLJK VLGHOREHV +RZHYHU &/6$5 FDQ EH XVHG LQ FRQMXQFWLRQ ZLWK PRGHUQ VSHFWUDO HVWLPDWLRQ PHWKn RGV WR DFFXUDWHO\ H[WUDFW IHDWXUHV RI VPDOO WDUJHWV ZKLFK FRQVLVW RI D ILQLWH QXPEHU RI LVRODWHG SRLQW VFDWWHUHUV &/6$5 LV VWLOO DW LWV H[SORUDWRU\ VWDJH RI GHYHORSPHQW DQG RQO\ D IHZ VWXGLHV ZHUH UHSRUWHG LQ WKH OLWHUDWXUH ,Q >@ D FRKHUHQW &/($1 DOJRULWKP ZDV SUHVHQWHG E\ .QDHOO WR HIIHFWLYHO\ HOLPLQDWH VLGHOREHV LQ &/6$5 LPn DJHV 7KH &/($1 DOJRULWKP ZDV RULJLQDOO\ SURSRVHG LQ UDGLR DVWURQRP\ >@ DQG
PAGE 27
ODWHU ZDV XVHG LQ PLFURZDYH LPDJLQJ >@ (QFRXUDJHG E\ WKH LQLWLDO UHVXOWV LQ >@ .QDHOO ODWHU SUHVHQWHG LQ >@ D FRUUHODWLRQ WHFKQLTXH ZKLFK ZRUNV LQ FRQMXQFWLRQ ZLWK WKH FRKHUHQW &/($1 DOJRULWKP WR UHGXFH VLGHOREH LQWHUIHUHQFH )XUWKHUPRUH .QDHOO LQ >@ FRPSDUHG WKH SHUIRUPDQFHV RI WKH FRKHUHQW &/($1 PHWKRG DQG WZR RWKHU PHWKRGV LQFOXGLQJ D PD[LPXP OLNHOLKRRG 0/f PHWKRG DQG D VLGHOREH OHDNDJH UHGXFn WLRQ DOJRULWKP >@ LQ H[WUDFWLQJ WDUJHW IHDWXUHV E\ XVLQJ &/6$5 GDWDVHWV :H ZLOO GLVFXVV KRZ 5(/$; 683(5 &/($1f FDQ EH H[WHQGHG IRU WDUJHW IHDWXUH H[WUDFWLRQ XVLQJ &/6$5 LQ &KDSWHU 6$5 0RWLRQ &RPSHQVDWLRQ $SHUWXUH HUURUV H[LVW GXH WR DWPRVSKHULF WXUEXOHQFH DQG SODWIRUP XQFHUWDLQW\ 7KHVH XQGHVLUDEOH HUURUV PD\ VLJQLILFDQWO\ GHJUDGH WKH TXDOLW\ RI 6$5 LPDJHV LQ VHYn HUDO ZD\V )RU H[DPSOH DSHUWXUH HUURUV PD\ FDXVH LPDJH JHRPHWULF GLVWRUWLRQ VSXn ULRXV WDUJHWV ORVV RI UHVROXWLRQ DQG GHFUHDVH LQ LPDJH FRQWUDVW > @ GHSHQGLQJ RQ WKHLU QDWXUHV 7KHUHIRUH HIIHFWLYH PRWLRQ FRPSHQVDWLRQ PHWKRGV DUH LQGLVSHQVDEOH IRU 6$5 LPDJLQJ DQG WDUJHW IHDWXUH H[WUDFWLRQ LQ SUDFWLFH ,Q 6$5 V\VWHPV PRn WLRQ HUURUV PDLQO\ FDXVH SKDVH HUURUV LQ FURVVUDQJH >@ $XWRIRFXV WHFKQLTXHV WKDW FRPSHQVDWH IRU SKDVH HUURUV XVLQJ UDGDU SKDVH KLVWRU\ GDWD KDYH EHHQ H[WHQVLYHO\ H[SORLWHG E\ 6$5 V\VWHPV 2I WKH H[LVWLQJ DXWRIRFXV DOJRULWKPV PDSGULIW >@ DQG SKDVH GLIIHUHQFH >@ DOJRULWKPV DUH XVHG WR FRPSHQVDWH IRU SKDVH HUURUV DVVRn FLDWHG ZLWK FHUWDLQ RUGHU SRO\QRPLDOV ZKHUHDV WKH SKDVHJUDGLHQW DXWRIRFXV 3*$f DOJRULWKP > @ LV LQGHSHQGHQW RI WKH SKDVH HUURU PRGHO WKHUHE\ LV PRUH URn EXVW $XWRIRFXV DOJRULWKPV EDVHG RQ VXSHU UHVROXWLRQ VSHFWUDO HVWLPDWLRQ PHWKRGV DOVR EHJLQ WR DSSHDU LQ WKH OLWHUDWXUH > @ $OO WKH DIRUHPHQWLRQHG DXWRIRFXV DOJRULWKPV DVVXPH WKDW SKDVH HUURUV FKDQJH LQ FURVVUDQJH RQO\ DQG DUH FRQVWDQW IRU DOO UDQJH ELQV 7KLV DVVXPSWLRQ LV JHQHUDOO\ YDOLG IRU WKH 6$5 V\VWHPV ZKHUH
PAGE 28
VLJQLILFDQW SKDVH HUURUV DUH FDXVHG E\ PRWLRQ PHDVXUHPHQW HUURUV VXFK DV DORQJWUDFN YHORFLW\ HUURUV RU OLQHRIVLJKW DFFHOHUDWLRQ HUURUV >@ $V LQ 6$5 V\VWHPV PRWLRQ HUURU FRPSHQVDWLRQ LV DOVR LQGLVSHQVDEOH IRU 6$5 V\VWHPV $XWRIRFXV LQ &/6$5 KRZHYHU EHFRPHV PRUH FRPSOLFDWHG VLQFH WKH DSHUWXUH HUURUV FDQ QR ORQJHU EH DSSUR[LPDWHG DV SKDVH HUURUV RQO\ LQ FURVVUDQJH VLQFH WKH HUURUV LQWURGXFHG LQ D]LPXWK DQJOHf DQG c! HOHYDWLRQ DQJOHf PD\ QRW EH QHJOLJLEOH DQG WKH H[WUDFWHG WDUJHW IHDWXUHV DUH VHQVLWLYH WR WKH VKDSH RI WKH FXUYLOLQHDU DSHUWXUH VHH WKH GLVFXVVLRQ LQ &KDSWHU f 7R WKH EHVW RI RXU NQRZOHGJH DXWRIRFXV IRU &/6$5 KDV QRW EHHQ UHSRUWHG LQ WKH OLWHUDWXUH SDUWO\ EHFDXVH &/6$5 LV D UHODWLYHO\ QHZ WHFKQRORJ\
PAGE 29
&+$37(5 3$5$0(75,& 0(7+2'6 )25 6$5 ,0$*,1* :KHQ DQ DSSURSULDWH GDWD PRGHO LV DYDLODEOH IRU UDGDU WDUJHWV SDUDPHWULF VSHFWUDO HVWLPDWLRQ DOJRULWKPV ZKLFK DUH OHVV VHQVLWLYH WR WKH PRGHO PLVPDWFK RIIHU WKH SRWHQWLDO WR VLJQLILFDQWO\ LPSURYH WKH UHVROXWLRQ RI WKH IRUPHG 6$5 LPDJHV ,Q WKLV FKDSWHU ZH FRQVLGHU IRUPLQJ HQKDQFHG 6$5 LPDJHV ZLWK UREXVW UHOD[DWLRQEDVHG PHWKRGV LQFOXGLQJ 5(/$; DQG 5(/$;1/6 5(/$; :KHQ WKH GRPLQDQW IHDWXUHV RI WDUJHWV FDQ EH DSSUR[LPDWHG DV SRLQW VFDWWHUn HG HQKDQFHG 6$5 LPDJHV FDQ EH IRUPHG E\ XVLQJ WKH WDUJHW IHDWXUHV H[WUDFWHG YLD 5(/$; DQG WKH SRLQW VFDWWHUHU GDWD PRGHO :H ILUVW EULHIO\ LQWURGXFH WKH 5(/$; DOJRULWKP >@ DQG WKHQ GLVFXVV KRZ WR IRUP 6$5 LPDJHV XVLQJ 5(/$; 5HYLHZ RI WKH 5(/$; $OJRULWKP $VVXPH WKDW WKHUH DUH GRPLQDQW SRLQW VFDWWHUHUV RU WULKHGUDO FRUQHU UHn IOHFWRUV LQ D WDUJHW $Q LGHDO WULKHGUDO DFWV DV D WULSOHERXQFH UHIOHFWRU DQG FDQ EH PRGHOHG DV D VWURQJ SRLQW VFDWWHUHU >@ VLQFH LWV UDGDU FURVV VHFWLRQ 5&6f FDQ EH FRQVLGHUHG DV D FRQVWDQW GXH WR WKH YHU\ VPDOO DQJOH YDULDWLRQ RI WKH UDGDU EHDP LQ 6$5 7KHQ WKH WDUJHW GDWD PRGHO LQ WKH SUHVHQFH RI XQNQRZQ QRLVH KDV WKH IRUP \Q f fÂ§ DNHAXONQNQn! HQ f Q f f f 1 fÂ§ IL f f f Â“ fÂ§ N f ZKHUH 1 DQG Â“ GHQRWH WKH QXPEHUV RI GDWD VDPSOHV LQ UDQJH DQG FURVVUDQJH UHVSHFn WLYHO\ RÂN DQG ^XMN Â` UHVSHFWLYHO\ GHQRWH WKH XQNQRZQ FRPSOH[ DPSOLWXGH DQG
PAGE 30
IUHTXHQFLHV RI WKH IFWK VLQXVRLG RU SRLQW VFDWWHUHU N f Â‘ .@ ^HQ f` GHQRWHV WKH XQNQRZQ QRLVH VHTXHQFH 1RWH WKDW ^FRNFRN` FRUUHVSRQGV WR WKH ORFDWLRQ RI WKH NWK VFDWWHUHU DQG DN LV GHWHUPLQHG E\ WKH 5&6 DQG UDQJH RI WKH IFWK VFDWWHUHU 7KH 1/6 HVWLPDWHV ^DNXMN/2N`N rI ^DN ANfN L FDQ EH REWDLQHG E\ PLQLPL]LQJ WKH IROORZLQJ 1/6 FULWHULRQ >@ &L ^DNXMNX!N`A f < fÂ§ < FY1XNfXMIUXMOfDN N f ZKHUH Â‘f7 GHQRWHV WKH WUDQVSRVH __ f ??S GHQRWHV WKH )UREHQLXV QRUP < LV DQ 1 [ 1 PDWUL[ ZLWK \Q f EHLQJ LWV QfWK HOHPHQW A1ANf fÂ§ HMXN 7 HMXN1Of f Q 7 A1ANf fÂ§ HAN HANA1 E f f 5(/$; PLQLPL]HV WKH 1/6 FRVW IXQFWLRQ &? LQ f E\ XVLQJ D FRPSOHWH UHOD[DWLRQ EDVHG VHDUFK LH IRU HDFK WKH LQWHUPHGLDWH QXPEHU RI VFDWWHUHUV RQO\ WKH SDUDPn HWHUV RI RQH VFDWWHUHU YDU\ ZKLOH DOO RWKHUV DUH IL[HG DW WKHLU PRVW UHFHQWO\ GHWHUPLQHG YDOXHV /HW a f f
PAGE 31
?XMbX!Nf@ WKDW 5(/$; LV UHODWHG WR &/($1 >@ ZKLFK LV RQO\ DQ DSSUR[LPDWH UHOD[DWLRQEDVHG PHWKRG WKDW PLQLPL]HV WKH 1/6 FULWHULRQ 5(/$; WXUQV RXW WR EH &/($1 ZKHQ QR LWHUDWLRQ LV SHUIRUPHG LQ HDFK VWHS RI 5(/$; 'XH WR WKH LWHUDWLRQV SHUIRUPHG LQ HDFK VWHS RI 5(/$; D PRUH DFFXUDWH LQLWLDO FDQ EH REWDLQHG ZKLFK PDNHV 5(/$; FRQYHUJH WR WKH JOREDO PLQLPXP ZLWK D KLJK SUREDELOLW\ 0HDQZKLOH 5(/$; LV DV\PSWRWLFDOO\ IRU ODUJH QXPEHU RI VDPSOHVf HIILFLHQW ZKLOH &/($1 JLYHV ELDVHG HVWLPDWHV HVSHFLDOO\ ZKHQ WZR VFDWWHUHUV DUH FORVHO\ VSDFHG 7KH 5(/$; DOJRULWKP WKHUHE\ FDQ DOVR EH UHIHUUHG WR DV 683(5 &/($1
PAGE 32
:H FRQVLGHU XVLQJ *$,& VHH >@ DQG WKH UHIHUHQFHV WKHUHLQ IRU GHWDLOVf WR GHWHUPLQH WKH QXPEHU RI VLQXVRLGV E\ DVVXPLQJ WKH XQNQRZQ QRLVH DQG FOXWWHU EHLQJ ZKLWH 7KH HVWLPDWH RI FDQ EH GHWHUPLQHG DV DQ LQWHJHU WKDW PLQLPL]HV WKH IROORZLQJ *$,& FRVW IXQFWLRQ 1fÂ§O Â“fÂ§ ? *$,&A 1Â“?Q (( _HQ Qf_ M M?Q>OQ11f@. Of f ZKHUH A fÂ§ Q Qf \Q Qf fÂ§ A $MWHAnFQ:IFQ? Q f f f 1 fÂ§ Q f f f 1 fÂ§ N f GHQRWHV WKH WRWDO QXPEHU RI XQNQRZQ UHDOYDOXHG SDUDPHWHUV RI ZKLFK DUH IRU WKH VLQXVRLGV DQG LV IRU WKH ZKLWH QRLVHf DQG LV D SDUDPHWHU RI XVHU FKRLFH ,PDJH )RUPDWLRQ :H FDQ IRUP 6$5 LPDJHV E\ XVLQJ WKH SRLQW VFDWWHUHU GDWD PRGHO DQG WKH WDUJHW IHDWXUHV H[WUDFWHG YLD 5(/$; 6$5 LPDJHV REWDLQHG YLD 5(/$; DUH UHIHUUHG WR DV 5(/$; 6$5 LPDJHV RU VLPSO\ 5(/$; LPDJHV 7KH VWHSV RI XVLQJ 5(/$; IRU 6$5 LPDJH IRUPDWLRQ DUH RXWOLQHG DV IROORZV 6WHS 2EWDLQ WKH SDUDPHWHU HVWLPDWHV RI ^RA XN Â`A \LD 5(/$; E\ XVLQJ WKH PHDVXUHG SKDVH KLVWRU\ GDWD VHH 6HFWLRQ IRU GHWDLOVf 6WHS *HQHUDWH \V^QVVf WKH VLPXODWHG SKDVH KLVWRU\ GDWD RI ODUJH GLPHQVLRQV IURP WKH HVWLPDWHG SDUDPHWHUV DQG EDVHG RQ WKH GDWD PRGHO LQ f QV f f f &$7 V Â‘ Â‘ f Â“ f \ Â‘^QVQVf A DNH MXLNQVXMNQVf IF ZKHUH e e f GHQRWHV WKH H[WUDSRODWLRQ IDFWRU ZKLFK LV D SDUDPHWHU RI XVHU FKRLFH DQG !IF DQG X!N UHVSHFWLYHO\ GHQRWH WKH HVWLPDWHV RI DN XfN DQG ZIF N f f f Le 6WHS )RUP 5(/$; 6$5 LPDJHV FRQWDLQLQJ RQO\ WKH GRPLQDQW WDUJHW IHDWXUHV E\
PAGE 33
DSSO\LQJ WKH QRUPDOL]HG ))7 WR WKH VLPXODWHG SKDVH KLVWRU\ GDWD ^\VQV QVf` LH E\ FRPSXWLQJ &1&Â“ MYMY ( ( f f QV QD 7R VXSSUHVV WKH VLGHOREHV WKH QRUPDOL]HG ))7 FDQ EH DSSOLHG WR WKH ZLQGRZHG VHTXHQFH ^ZVQVVf\VQVVf` ZKHUH WKH ZLQGRZ VHTXHQFH ZVQVVf VDWLVILHV &$7 &Â“ fÂ§ \ \ ZVQVQVf eB9,9 f QV V 7KH 5(/$; 6$5 LPDJHV FRQWDLQLQJ ERWK WKH GRPLQDQW WDUJHW IHDWXUHV DQG EDFNn JURXQG FOXWWHU PD\ EH DSSUHFLDWHG VLQFH IRU H[DPSOH WKH VKDGRZ LQIRUPDWLRQ PD\ EH GHVLUHG IRU DXWRPDWLF WDUJHW UHFRJQLWLRQ ,I VR WKH QRUPDOL]HG ))7 FDQ EH DSSOLHG WR WKH VXP RI WKH VLPXODWHG SKDVH KLVWRU\ GDWD ^\DQV QVf` DQG ^eQ f` ZLWK ]HUR SDGGLQJ WR KDYH GLPHQVLRQV 1 DQG e9 ZKHUH Q f GHQRWHV WKH HVWLPDWHG EDFNn JURXQG FOXWWHU DQG LV GHWHUPLQHG E\ f ZLWK UHSODFHG E\ 1RWH WKDW VFDOLQJ WKH Q f E\ D IDFWRU RI e LV QHHGHG ZKHQ WKH EDFNJURXQG FOXWWHU LV LQFOXGHG LQ WKH 5(/$; 6$5 LPDJHV VLQFH ERWK RI LWV GLPHQVLRQV DUH e WLPHV RI WKRVH RI WKH VLPXODWHG SKDVH KLVWRU\ GDWD 6LPLODUO\ WKH QRUPDOL]HG ))7 FDQ DOVR EH DSSOLHG WR WKH VXP RI ^ZVQV Vf\VQV Vf` DQG ^ZHQ fQ f` ZLWK ]HUR SDGGLQJ WR KDYH GLPHQVLRQV 1 DQG eÂ“ WR VXSSUHVV WKH VLGHOREHV ZKHUH WKH ZLQGRZ VHTXHQFH ZHQf VDWLVILHV 1O Â“ \ \ :H^Qf 1Â“ f Q 1RWH WKDW VLQFH ZH FDQQRW PRGHO WKH FOXWWHU HIIHFWLYHO\ LWV UHVROXWLRQ FDQQRW EH LPSURYHG :H UHPDUN WKDW WKH UHVROXWLRQ RI WKH 5(/$; 6$5 LPDJHV LV GHWHUPLQH E\ 5(/$; DQG e LV MXVW XVHG WR GHPRQVWUDWH WKH VXSHU UHVROXWLRQ SURSHUW\ RI WKH 5(/$; DOJRULWKP IRU WDUJHW IHDWXUH H[WUDFWLRQ
PAGE 34
5(/$;1/6 :KHQ WDUJHWV FRQVLVW RI ERWK WULKHGUDO SRLQW VFDWWHUHUVf DQG GLKHGUDO FRUQHU UHIOHFWRUV ZH FDQ IRUP HQKDQFHG 6$5 LPDJHV E\ XVLQJ WKH 5(/$;1/6 DOJRULWKP >@ :H ILUVW JLYH D EULHI UHYLHZ RI WKH 5(/$;1/6 DOJRULWKP DQG WKHQ GLVFXVV 6$5 LPDJH IRUPDWLRQ YLD 5(/$;1/6 5HYLHZ RI WKH 5(/$;1/6 $OJRULWKP 5(/$;1/6 >@ LV GHYLVHG WR LGHQWLI\ DQG REWDLQ WDUJHW IHDWXUHV RI PXOWLSOH FRUQHU UHIOHFWRUV WULKHGUDO DQG GLKHGUDOf EDVHG RQ D PL[HG SDUDPHWULF GDWD PRGHO $Q LGHDO GLKHGUDO FRUQHU UHIOHFWRU DFWV DV D GRXEOHERXQFH UHIOHFWRU DQG WKH 5&6 DFKLHYHV WKH PD[LPXP ZKHQ WKH UDGDU EHDP LV SHUSHQGLFXODU WR WKH GLKHGUDO DQG IDOOV RII DV D IXQFWLRQ RI VLQFf VLQf ZKHQ WKH DQJOH EHWZHHQ WKH UDGDU EHDP DQG WKH OLQH SHUSHQGLFXODU WR WKH GLKHGUDO LQFUHDVHV > @ 7KHQ GLKHGUDOV FDQ EH PRGHOHG DV .G VGQQf A DGNVPF^QENQ7Nf@H^8GNQ8LNQ?Q f f f 1O f f f 1 N f ZKHUH 1 DQG Â“ GHQRWH WKH QXPEHUV RI WKH DYDLODEOH GDWD VDPSOHV LQ UDQJH DQG FURVVUDQJH UHVSHFWLYHO\ DGN ^XMGNGN` DQG EN N f f f .G DUH UHVSHFWLYHO\ SURSRUWLRQDO WR WKH PD[LPDO 5&6 WKH FHQWUDO ORFDWLRQ DQG WKH OHQJWK RI WKH $WK GLKHGUDO FRUQHU UHIOHFWRU UN N fÂ§ f f f .G GHQRWHV WKH SHDN ORFDWLRQ RI WKH GDWD VHTXHQFH LQ FURVVUDQJH DQG LV GHWHUPLQHG E\ WKH RULHQWDWLRQ RI WKH IFWK GLKHGUDO FRUQHU UHIOHFWRU ILQDOO\ .G LV WKH QXPEHU RI WKH GLKHGUDO FRUQHUV $VVXPH WKDW D WDUJHW FRQVLVWV RI .G GLKHGUDOV DQG .W WULKHGUDOV DQG .G .W 7KHQ WKH WDUJHW GDWD PRGHO LQ WKH SUHVHQFH RI QRLVH DQG FOXWWHU KDV WKH IRUP \Q IWf VGQ f VÂQ f HQ f Q f f f 1 fÂ§ f f f Â“ fÂ§ f
PAGE 35
ZKHUH .W VWQ Qf \AMDWNHAXWNQPNQ? Q f f f 1 f f f Â“ f N ZLWK DQG ^D+NWN`N Y UHVSHFWLYHO\ GHQRWLQJ WKH XQNQRZQ FRPSOH[ DPn SOLWXGHV DQG XQNQRZQ IUHTXHQFLHV RI WKH .W WULKHGUDOV 7KH 5(/$;1/6 DOn JRULWKP SURSRVHG LQ >@ FDQ EH XVHG WR LGHQWLI\ DQG HIIHFWLYHO\ HVWLPDWH WKH WDUn JHW IHDWXUHV^DGN EN /RGN 2GN 7N`NLL DQG ^DWN XWN WN E\ PLQLPL]LQJ WKH IROORZLQJ 1/6 FULWHULRQ & A^DWN IF O! LDGN EN8GN GN U$`IF LA fÂ§ __ < fÂ§ 6W fÂ§ 6G??) f ZKHUH ><@\ \LMf >6W@rM VWLMf DQG >6G@LM VGLMf 5(/$;1/6 FRQYHUWV WKH SUREOHP RI PXOWLSOH FRUQHU UHIOHFWRU LGHQWLILFDWLRQ DQG IHDWXUH H[WUDFWLRQ WR WKDW RI D VHTXHQFH RI VLQJOH UHIOHFWRU LGHQWLILFDWLRQ DQG IHDWXUH H[WUDFWLRQ E\ XVLQJ WKH UHOD[DWLRQ EDVHG RSWLPL]DWLRQ DSSURDFK 7KH LGHQWLILFDWLRQ DQG IHDWXUH H[WUDFWLRQ RI D VLQJOH FRUQHU UHIOHFWRU LQ 5(/$;1/6 LV GHVFULEHG DV IROORZV )HDWXUH ([WUDFWLRQ RI D 6LQJOH 7ULKHGUDO &RUQHU 5HIOHFWRU /HW
PAGE 36
DQG ^WWW` DUJPD[ XbXMWfIfUU fÂ§ f@HnZLQ f f f Â“ fÂ§ 7KH 1/6 HVWLPDWHV ^Â£UIFcfUIÂUI E I` RI ^DGWRG &RE U` DUH GHWHUPLQHG E\ PLQLPL]LQJ WKH IROORZLQJ FRVW IXQFWLRQ &UI DG XG LG Uf __ @ LH DOWHUQDWLYHO\ XSGDWLQJ ^XMG X!G E I` ZKLOH IL[LQJ RWKHU SDUDPHWHU HVWLn PDWHV DW WKHLU UHFHQWO\ GHWHUPLQHG YDOXHV 1RWH WKDW IRXUGLPHQVLRQDO VHDUFK RYHU WKH SDUDPHWHU VSDFH LV SHUIRUPHG WR HVWLPDWH WKH GLKHGUDO FRUQHU UHIOHFWRU SDUDPHWHUV
PAGE 37
,GHQWLILFDWLRQ RI &UQHU 5HIOHFWRU ,W KDV EHHQ VKRZQ LQ >@ WKDW &G LV PXFK OHVV WKDQ &W ZKHQ WKH WUXH FRUQHU UHIOHFWRU LV D GLKHGUDO DQG &W LV DOPRVW WKH VDPH DV &G ZKHQ WKH WUXH FRUQHU UHIOHFWRU LV D WULKHGUDO $OVR E m I ZKHQ D WULKHGUDO LV PLVWDNHQO\ FRQVLGHUHG DV D GLKHGUDO ZKHUH 6I LV WKH VPDOOHVW SRVLWLYH VROXWLRQ WR VP1ILUf 1 aan f VLQUf 7KHQ D FRUQHU UHIOHFWRU FDQ EH LGHQWLILHG E\ FKHFNLQJ WKH HVWLPDWH E DQG WKH FRVW IXQFWLRQV &W LQ f ZLWK ^DWXfW RMW` UHSODFHG E\ ^DWXfWXfW`f DQG &G LQ f ZLWK ^F2GL &EG 2G E U` UHSODFHG E\ ^MÂ XLG EI` $ FRUQHU UHIOHFWRU LV LGHQWLILHG DV D GLKHGUDO LI FG DQG E I 6WHSV RI 5(/$;1/6 DUH RXWOLQHG DV IROORZV 6WHS f $VVXPH fÂ§ ,GHQWLI\ DQG REWDLQ WKH SDUDPHWHUV RI D VLQJOH FRUQHU UHIOHFWRU 6WHS f $VVXPH 6XEWUDFW WKH VLJQDO REWDLQHG LQ 6WHS f IURP ^\Qf` DQG WKHQ LGHQWLI\ DQG REWDLQ SDUDPHWHUV RI WKH VHFRQG FRUQHU UHIOHFWRU 6XEWUDFW WKH VHFRQG VLJQDO IURP ^\QQf` DQG UHLGHQWLI\ DQG UHGHWHUPLQH WKH SDn UDPHWHUV RI WKH ILUVW VLJQDO ,WHUDWH WKH SUHYLRXV WZR VXEVWHSV XQWLO fSUDFWLFDO FRQYHUJHQFHf 6WHS f $VVXPH 6XEWUDFW WKH WZR VLJQDOV REWDLQHG LQ 6WHS f IURP ^\Qf` DQG WKHQ LGHQWLI\ DQG REWDLQ WKH SDUDPHWHUV RI WKH WKLUG VLJQDO 6XEWUDFW WKH VHFRQG DQG WKLUG VLJQDOV IURP ^\Q f` DQG UHLGHQWLI\ DQG UHGHWHUPLQH WKH ILUVW VLJQDO 6XEWUDFW WKH ILUVW VLJQDO PRVW UHFHQWO\ GHWHUPLQHGf DQG WKH WKLUG VLJQDO IURP ^\QQf` DQG WKHQ UHLGHQWLI\ DQG UHGHWHUPLQH WKH VHFRQG VLJQDO ,WHUDWH WKH SUHYLRXV WKUHH VXEVWHSV XQWLO fSUDFWLFDO FRQYHUJHQFHf 5HPDLQLQJ 6WHSV &RQWLQXH VLPLODUO\ XQWLO fÂ§ . FDQ EH GHWHUPLQHG E\ XVLQJ *$,& WR EH GLVFXVVHG ODWHU RQf
PAGE 38
7KH fSUDFWLFDO FRQYHUJHQFHf LQ WKH LWHUDWLRQV FDQ DOVR EH GHWHUPLQHG E\ FKHFNn LQJ WKH UHODWLYH FKDQJH RI WKH FRVW IXQFWLRQ & LQ f EHWZHHQ WZR FRQVHFXWLYH LWHUDWLRQV :H FDQ DOVR GHWHUPLQH WKH WRWDO QXPEHU RI WULKHGUDOV DQG GLKHGUDOV E\ H[WHQGLQJ WKH *$,& GLVFXVVHG LQ 6HFWLRQ DQG DVVXPLQJ ZKLWH QRLVH DQG FOXWWHU 7KH HVWLPDWH RI FDQ EH GHWHUPLQHG DV DQ LQWHJHU WKDW PLQLPL]HV WKH IROORZLQJ H[WHQGHG *$,& FRVW IXQFWLRQ 1Â“ ? (( ,HQ Qf_ OQ>OQ$*9f@L* .G f f 7LfÂ§2 ZKHUH .G HQ Qf \Q Qf GGVLQF>UIHQ INf?H M^AGNQLKGNQf .W Â£WNHM^&MWrQeOWrf f N O N ZLWK Q f Â‘ f 1 fÂ§ DQG f f f Â“ fÂ§ DV EHIRUH LV D SDUDPHWHU RI XVHU FKRLFH .W .cL ZLWK .W DQG .G GHQRWLQJ WKH QXPEHUV RI WULKHGUDO DQG GLKHGUDO FRUQHUV UHVSHFWLYHO\ GHWHUPLQHG E\ 5(/$;1/6 JLYHQ .? ILQDOO\ .W .G LV WKH WRWDO QXPEHU RI XQNQRZQ UHDOYDOXHG SDUDPHWHUV RI ZKLFK .W DQG .GL UHVSHFWLYHO\ DUH IRU WKH WULKHGUDO DQG GLKHGUDO FRUQHUV DQG LV IRU WKH ZKLWH QRLVHf ,PDJH )RUPDWLRQ :H FDQ IRUP HQKDQFHG 6$5 LPDJHV RI WDUJHWV FRQVLVWLQJ RI ERWK WULKHGUDOV DQG GLKHGUDOV E\ XVLQJ WKH H[WUDFWHG WDUJHW IHDWXUHV YLD 5(/$;1/6 DQG WKH PL[HG GDWD PRGHO LQ f 7KH 6$5 LPDJHV REWDLQHG YLD 5(/$;1/6 DUH UHIHUUHG WR DV 5(/$;1/6 6$5 LPDJHV RU VLPSO\ 5(/$;1/6 LPDJHV 7R IRUP 5(/$;1/6 LPDJHV WKH GRPLQDQW WDUJHW IHDWXUHV DUH ILUVW H[WUDFWHG E\ XVLQJ WKH 5(/$;1/6 DOJRULWKP VHH 6HFWLRQ IRU GHWDLOVf DQG WKHQ 6WHSV DQG GLVFXVVHG LQ 6HFWLRQ DUH XVHG H[FHSW WKDW WKH VLPXODWHG SKDVH KLVWRU\ GDWD RI ODUJH GLPHQVLRQV XVHG LQ 6WHS LV QRZ GHWHUPLQHG E\ .G .W \VQV Vf &;@ mGIFVLQF0IFmV INf`HÂƒ8GNQVfGNQ$ A f N IF O
PAGE 39
ZKHUH QV f f f e1 fÂ§ DQG ILV f f f Â“ fÂ§ ZLWK e EHLQJ DQ H[WUDSRODWLRQ IDFWRU e f Â£GN &RGN X!GN K DQG IN UHVSHFWLYHO\ GHQRWH WKH HVWLPDWHV RI DGN 8MGN 4GN EN DQG 7N N f f f .G? ILQDOO\ Â£ÂIH ZLIF DQG XWN UHVSHFWLYHO\ GHQRWH WKH HVWLPDWHV RI DWN FRWN DQG 4WN N f f f .W 7KH FOXWWHU HVWLPDWH Q f XVHG LQ 6WHS LV QRZ GHWHUPLQHG E\ f ZLWK .W DQG .G UHSODFHG E\ .W DQG .G UHVSHFWLYHO\ 1RWH WKDW VFDOLQJ WKH VLPXODWHG GDWD IRU WKH GLKHGUDOV E\ D IDFWRU RI e LQ f LV QHHGHG VLQFH WKH VLQH IXQFWLRQ JRHV WR ]HUR DV QV LQFUHDVHV RU GHFUHDVHV DZD\ IURP UN 1RWH DOVR WKDW e LV D SDUDPHWHU RI XVHU FKRLFH DQG e LV MXVW XVHG WR GHPRQVWUDWH WKH VXSHU UHVROXWLRQ SURSHUW\ RI 5(/$;1/6 )LQDOO\ VLQFH WKH )RXULHU WUDQVIRUPV RI WKH VLQH IXQFWLRQV RI VXIILFLHQW OHQJWKV GR QRW UHVXOW LQ VLGHOREHV ZKHQ \VQVVf LQ f LV ZLQGRZHG WKH VHFRQG WHUP RI \VMLVILVf LV PXOWLSOLHG E\ ZVQVVf DQG WKH ILUVW WHUP LV PXOWLSOLHG E\ ZLQVf ZKHUH Z?LQVf LV D ZLQGRZ VHTXHQFH VDWLVI\LQJ &$7 Z^QDf 1 f OV ([SHULPHQWDO 5HVXOWV ,Q WKLV VHFWLRQ ZH GHPRQVWUDWH WKH LPDJH IRUPDWLRQ SHUIRUPDQFHV RI 5(/$; DQG 5(/$;1/6 ZLWK WKH H[SHULPHQWDO GDWD ,Q WKH IROORZLQJ H[DPSOHV WKH H[WUDSn RODWLRQ IDFWRU e LV XVHG DQG *$,& ZLWK DQG DUH XVHG WR GHWHUPLQH IRU 5(/$; DQG 5(/$;1/6 UHVSHFWLYHO\ .DLVHU ZLQGRZV ZLWK VKDSH SDUDPHWHU fÂ§ DUH XVHG ZKHQHYHU QHHGHG )LUVW FRQVLGHU 6$5 LPDJH IRUPDWLRQ YLD ))7 DQG 5(/$; E\ XVLQJ D SRUWLRQ RI WKH GDWD FRUUHVSRQGLQJ WR VRPH URRI ULPV FROOHFWHG E\ RQH RI WKH WZR DSHUWXUHV RI WKH (5,0fV (QYLURQPHQW 5HVHDUFK ,QVWLWXWH RI 0LFKLJDQfVf '&6 LQWHUn IHURPHWULF V\QWKHWLF DSHUWXUH UDGDU ,)6$5f )LJXUHV Df DQG Ef UHVSHFWLYHO\ VKRZ WKH XQZLQGRZHG DQG ZLQGRZHG ))7 LPDJHV REWDLQHG E\ ]HURSDGGLQJ WKH [ SKDVH KLVWRU\ GDWD 7KH XQZLQGRZHG DQG ZLQGRZHG 5(/$; LPDJHV ZLWK DQG
PAGE 40
ZLWKRXW WKH EDFNJURXQG FOXWWHU DUH LOOXVWUDWHG LQ )LJXUHV Ff a If ZLWK 1RWH WKDW WKH 5(/$; LPDJHV KDYH D KLJKHU UHVROXWLRQ WKDQ WKH ))7 LPDJHV IRU WKH GRPLQDQW VFDWWHUHUV 8WLOL]LQJ RQO\ b RI WKH [ SKDVH KLVWRU\ GDWD LH XVLQJ D [ SKDVH KLVWRU\ GDWD ZH IRUP WKH ))7 6$5 LPDJHV VKRZQ LQ )LJXUHV Df DQG Ef )LJXUHV Ff a If VKRZ WKH 5(/$; LPDJHV ZLWK fÂ§ &RPSDUn LQJ )LJXUHV If DQG Ef ZH QRWH WKDW WKH WZR LPDJHV DUH TXLWH VLPLODU DOWKRXJK WKH IRUPHU XVHV RQO\ b RI WKH GDWD XVHG E\ WKH ODWWHU &RQVLGHU QH[W DQ H[DPSOH RI WKH 067$5 6OLF\ GDWD FRQVLVWLQJ RI ERWK WULKHn GUDO DQG GLKHGUDO FRUQHU UHIOHFWRUV FROOHFWHG E\ WKH 61/ XVLQJ WKH 67$5/26 VHQVRU 7KH ILHOG GDWD ZDV FROOHFWHG E\ D VSRWOLJKWPRGH 6$5 ZLWK D FDUULHU IUHTXHQF\ *+] DQG EDQGZLGWK *+] 7KH UDGDU ZDV DERXW NLORPHWHUV DZD\ IURP WKH JURXQG WDUJHW VKRZQ LQ )LJXUH 7KH 6$5 LPDJHV DUH REWDLQHG ZKHQ WKH WDUn JHW LV LOOXPLQDWHG E\ WKH UDGDU IURP WKH D]LPXWK DQJOH r DQG HOHYDWLRQ DQJOH r 7KH UDQJH DQG FURVVUDQJH UHVROXWLRQV RI WKH RULJLQDO GDWD [ f DUH PHWHUV DQG PHWHUV UHVSHFWLYHO\ DQG WKH ZLQGRZHG ))7 6$5 LPDJH REWDLQHG YLD 7D\ORU ZLQGRZf LV VKRZQ LQ )LJXUH Df 7R GHPRQVWUDWH WKH VXSHU UHVROXWLRQ SURSHUW\ RI 5(/$; DQG 5(/$;1/6 ZH XVH SDUW RI WKH RULJLQDO GDWD ZKRVH GLn PHQVLRQ LV [ )LJXUH Ef VKRZV WKH ZLQGRZHG ))7 LPDJHV ZLWK WKH VSRLOHG UHVROXWLRQ [ PHWHUV REWDLQHG IURP WKLV UHGXFHG VL]H GDWD 5(/$; DQG 5(/$;1/6 DUH DOVR DSSOLHG WR WKLV [ SKDVH KLVWRU\ GDWD PDWUL[ )LJXUHV Ff DQG Gf UHVSHFWLYHO\ VKRZ WKH ZLQGRZHG 5(/$; 6$5 LPDJHV ZLWK DQG ZLWKn RXW HVWLPDWHG EDFNJURXQG FOXWWHU ZLWK :H QRWH WKDW 5(/$; LPDJHV KDYH D EHWWHU UHVROXWLRQ WKDQ WKH ))7 LPDJHV IRU WKH GRPLQDQW WULKHGUDOV DQG UHVHPEOH WKH ))7 LPDJHV IRU GLKHGUDOV HYHQ WKRXJK WKH GDWD PRGHO LQ WKH FURVVUDQJH GLPHQVLRQ XVHG E\ 5(/$; LV QRW FRUUHFW IRU WKLV H[DPSOH 7KH ZLQGRZHG 5(/$;1/6 6$5 LPDJHV ZLWK DQG ZLWKRXW HVWLPDWHG EDFNJURXQG FOXWWHU ZLWK fÂ§ DUH LOOXVWUDWHG LQ )LJXUHV Hf DQG If UHVSHFWLYHO\ :H QRWH DJDLQ WKDW WKH 5(/$;1/6 LPDJHV
PAGE 41
QRW RQO\ KDYH D EHWWHU UHVROXWLRQ WKDQ WKH ))7 LPDJHV EXW DOVR PRUH FORVHO\ UHIOHFWV WKH FKDUDFWHULVWLFV RI WKH SKRWR VKRZQ LQ )LJXUH
PAGE 42
Hf If )LJXUH 5(/$; 6$5 ,PDJHV E\ XVLQJ WKH [ (5,0 GDWD ZLWK Df 8QZLQGRZHG ))7 LPDJH Ef :LQGRZHG ))7 LPDJH Ff 8QZLQGRZHG 5(/$; LPDJH ZLWKRXW EDFNJURXQG FOXWWHU Gf :LQGRZHG 5(/$; LPDJH ZLWKRXW EDFNJURXQG FOXWWHU Hf 8QZLQGRZHG 5(/$; LPDJH ZLWK EDFNJURXQG FOXWWHU If :LQGRZHG 5(/$; LPDJH ZLWK EDFNJURXQG FOXWWHU
PAGE 43
Df Ef r m Â‘ Hf If )LJXUH 5(/$; ,PDJHV E\ XVLQJ WKH [ (5,0 GDWD ZLWK Df 8QZLQGRZHG ))7 LPDJH Ef :LQGRZHG ))7 LPDJH Ff 8QZLQGRZHG 5(/$; LPDJH ZLWKRXW EDFNJURXQG FOXWWHU Gf :LQGRZHG 5(/$; LPDJH ZLWKRXW EDFNJURXQG FOXWWHU Hf 8QZLQGRZHG 5(/$; LPDJH ZLWK EDFNJURXQG FOXWWHU If :LQGRZHG 5(/$; LPDJH ZLWK EDFNJURXQG FOXWWHU
PAGE 44
)LJXUH 7DUJHW SKRWR WDNHQ DW r D]LPXWK DQJOH
PAGE 45
Df Ef Y : &f Gf Hf If )LJXUH 6$5 LPDJHV RI WKH 6OLF\ GDWD r D]LPXWK DQG r HOHYDWLRQ DQJOHVf Df 2ULJLQDO ZLQGRZHG ))7 LPDJH Ef :LQGRHG ))7 LPDJH ZLWK VSRLOHG UHVROXWLRQ Ff a Gf :LQGRZHG 5(/$; LPDJH ZLWKRXW DQG ZLWK EDFNJURXQG FOXWWHU ZLWK Hf a If :LQGRZHG 5(/$;1/6 LPDJH ZLWKRXW DQG ZLWK EDFNJURXQG FOXWWHU ZLWK Ff a If DUH REDWLQHG XVLQJ WKH GDWD LQ Ef 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f
PAGE 46
&+$37(5 6(0,3$5$0(75,& 0(7+2'6 )25 6$5 ,0$*,1* 3UREOHP )RUPXODWLRQ DQG 'DWD 0RGHO &RPSDUHG ZLWK SDUDPHWULF VSHFWUDO HVWLPDWLRQ PHWKRGV QRQSDUDPHWULF VSHFn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f LQ ERWK UDQJH DQG FURVVUDQJH ZH DSSO\ KHUHLQ D VHPLSDUDPHWULF PRGHO DQG VLPSO\ DVVXPH WKDW WKH GDWD PRGHO LQ FURVVUDQJH GXH WR RQH RU PRUH FRUQHU UHn IOHFWRUV LV D FRPSOH[ VLQXVRLG ZLWK DQ DUELWUDU\ DPSOLWXGH DQG D FRQVWDQW SKDVH 'XH WR WKH IOH[LELOLW\ RI WKLV GDWD PRGHO IHDWXUH H[WUDFWLRQ DQG LPDJH IRUPDWLRQ PHWKn RGV EDVHG RQ WKLV GDWD PRGHO VKRXOG EH PRUH UREXVW FRPSDUHG ZLWK WKRVH EDVHG RQ WKH DSSUR[LPDWHG GLKHGUDO DQG WULKHGUDO GDWD PRGHOV LQ &KDSHU 7KH UREXVWQHVV KRZHYHU FRPHV DW WKH FRVW RI VPDOO UHVROXWLRQ LPSURYHPHQW LQ FURVVUDQJH RYHU WKH FRQYHQWLRQDO IDVW )RXULHU WUDQVIRUP PHWKRGV GXH WR DVVXPLQJ OLWWOH LQ FURVVUDQJH 7R SUHVHUYH WKH UREXVWQHVV ZH XVH D UHILQHG QRQSDUDPHWULF VSHFWUDO HVWLPDWLRQ PHWKRG VXFK DV $3(6 $PSOLWXGH DQG 3KDVH (VWLPDWLRQ RI D 6LQXVRLGf DOJRULWKP >@ WR LPSURYH WKH FURVVUDQJH UHVROXWLRQ
PAGE 47
7KH UHFHLYHG VLJQDO UHIOHFWHG IURP D WDUJHW VFDWWHUHU KHUHLQ LV PRGHOHG DV VQ f [fH"O!fr8QIQ? Q f f f 1 fÂ§ f f f Â“ fÂ§ f ZKHUH 1 DQG Â“ GHQRWH WKH GLPHQVLRQV RI WKH DYDLODEOH GDWD VDPSOHV LQ UDQJH DQG FURVVUDQJH UHVSHFWLYHO\ [f LV DQ DUELWUDU\ XQNQRZQ UHDOYDOXHG IXQFWLRQ RI GHWHUPLQHG E\ WKH UDGDU FURVV VHFWLRQ 5&6f RI WKH VFDWWHUHU LV D FRQVWDQW SKDVH ILQDOO\ ^ ` LV WKH IUHTXHQF\ SDLU SURSRUWLRQDO WR WKH ORFDWLRQ UDQJH DQG FURVVn UDQJHf RI WKH VFDWWHUHU 7KLV GDWD PRGHO LV HVVHQWLDOO\ VHPLSDUDPHWULF VLQFH OLWWOH SDUDPHWHUL]DWLRQ LV GRQH LQ FURVVUDQJH $VVXPH WKDW D WDUJHW FRQVLVWV RI VFDWWHUHUV 7KHQ WKH WDUJHW GDWD PRGHO LQ WKH SUHVHQFH RI QRLVH KDV WKH IRUP \Q f A ;N^fHANf5ANQUNA HQ f Q f f f 1 fÂ§ Â‘ f f Â“ fÂ§ N f ZKHUH ^Af`A GHQRWHV WKH UHDOYDOXHG DPSOLWXGH IXQFWLRQ RI IRU WKH NWK VFDWn WHUHU FIfN DQG ^ININ` UHVSHFWLYHO\ DUH WKH FRQVWDQW SKDVH DQG WKH IUHTXHQF\ SDLU RI WKH $WK VFDWWHUHU ILQDOO\ ^HQf` GHQRWHV WKH XQNQRZQ QRLVH DQG FOXWWHU VHTXHQFH 6LQFH 6$5 LPDJHV DUH RIWHQ XVHG LQ 6$5 DSSOLFDWLRQV RXU SUREOHP RI LQWHUHVW KHUHLQ LV WR HVWLPDWH WKH WDUJHW SDUDPHWHUV ^ILN ^[Nf`ba4 IN IN`N IURP WKH GDWD VHTXHQFH ^\Q f` DQG WKHQ WR IRUP KLJK UHVROXWLRQ 6$5 LPDJHV ZLWK WKH HVWLPDWHG WDUJHW SDUDPHWHUV %HIRUH SUHVHQWLQJ WKH HIIHFWV RI WKH VHPLSDUDPHWULF GDWD PRGHO LQ f RQ 6$5 LPDJH IRUPDWLRQ DQG WKH 63$5 6HPL3$5DPHWULFf DOJRULWKP IRU IHDWXUH H[WUDFWLRQ DQG LPDJH IRUPDWLRQ ZH ILUVW EULHIO\ LQWURGXFH WKH $3(6 DOJRULWKP
PAGE 48
5HYLHZ RI WKH $3(6 $OJRULWKP 7KH $3(6 >@ DOJRULWKP EHORQJV WR WKH DGDSWLYH ILOWHUEDQN FODVV LQ ZKLFK WKH DGDSWLYH ILQLWH LPSXOVH UHVSRQVH ),5f ILOWHU LV GHVLJQHG EDVHG RQ QRW RQO\ WKH GDWD WR EH SURFHVVHG EXW DOVR WKH LQWHUIHUHQFH DW WKH IUHTXHQF\ RI LQWHUHVW $3(6 PLPLFV WKH PD[LPXP OLNHOLKRRG HVWLPDWRU RI D FRPSOH[ VLQXVRLG LQ FLUFXODUO\ V\PPHWULF ]HURPHDQ *DXVVLDQ QRLVH ,W KDV EHHQ VKRZQ LQ >@ WKDW $3(6 KDV EHWWHU VWDWLVWLFDO SHUIRUPDQFH WKDQ &DSRQ >@ DQRWKHU SRSXODUO\ XVHG DGDSWLYH ILOWHUEDQN DSSURDFK IRU 6$5 LPDJHU\ /HW ^]Q`A,^J GHQRWH WKH DYDLODEOH GDWD VHTXHQFH DQG KDYH WKH IRUP ]Q DFRfH[S^MQX!` Q f f f 1 fÂ§ f ZKHUH DXf GHQRWHV WKH FRPSOH[ DPSOLWXGH RI D VLQXVRLG ZLWK IUHTXHQFH OR DQG HQXf GHQRWHV WKH XQPRGHOHG QRLVH DQG LQWHUIHUHQFH DW X 7KHQ $3(6 FDQ EH XVHG WR REWDLQ WKH FRPSOH[ DPSOLWXGH HVWLPDWH DFRf RI DXf IURP ^]X`A=T ,W KDV EHHQ VKRZQ LQ >@ WKDW WKH OHDVWVTXDUHV HVWLPDWH RI D^Xf XVLQJ WKH IRUZDUG DQG EDFNZDUG ),5 ILOWHU RXWSXW KDV WKH IRUP tOXf KDf=Zf FRf=+ XMfKXMf f ZKHUH KX!f GHQRWHV WKH LPSXOVH UHVSRQVH RI DQ 0WDS ),5 ILOWHU DQG KDV WKH IROORZLQJ IRUP KZf O 7 KLXf KFRf K0XMf f =X!f DQG =FXf UHVSHFWLYHO\ GHQRWH WKH QRUPDOL]HG )RXULHU WUDQVIRUPV RI ^]L`A0 DQG ^]LH WKH )RXULHU WUDQVIRUPV RI ^]M`A40 DQG ^]Â`AT0 GLYLGHG E\ 1 fÂ§ 0 f ZKHUH ]Â DQG ]UHVSHFWLYHO\ GHQRWH WKH RYHUODSSLQJ VXEYHFWRUV RI WKH GDWD YHFWRU ] ] ] L GDWD VHTXHQFH ] ]QL =1 =1 DQG WKH FRPSOH[ FRQMXJDWHG DQG EDFNZDUGRUGHUHG DQG KDYH WKH IRUP = 7 =L =L0fÂ§ L f f f 1 fÂ§ 0 f
PAGE 49
DQG ]1fÂ§LfÂ§O ]1L L f f f 1 fÂ§ 0 f 1L0 UHVSHFWLYHO\ ILQDOO\ Xf fÂ§ H[S^fÂ§M1 fÂ§ OfX!` ,W KDV EHHQ VKRZQ LQ $SSHQGL[ & RI >@ WKDW E\ XVLQJ D FHUWDLQ FRQMXJDWH V\PPHWULF FRQGLWLRQ DXf LQ f FDQ EH UHGXFHG WR DXMf K+ ORf=XMf f 7KH DGDSWLYH ),5 ILOWHU K$3(V^Xf FRUUHVSRQGLQJ WR $3(6 LV GHULYHG LQ $Sn SHQGL[ $ RI >@ DQG KDV WKH IRUP 4 FfDDf cÂ‘$3(6 0 ZKHUH WKH VWHHULQJ YHFWRU DR!f LV JLYHQ E\ f DXf 7 O JMZ HM0fÂ§OfW f DQG 4RLf GHQRWHV WKH HVWLPDWH RI WKH FRYDULDQFH PDWUL[ RI WKH QRLVH DQG LV JLYHQ E\ 4Zf 5 =Zf=IIX!f f ZLWK =0 A > =Zf =+@ f DQG 5 5 5f GHQRWLQJ WKH DYHUDJH RI WKH IRUZDUG DQG EDFNZDUG VDPSOH FRYDULDQFH PDWULFHV 5 DQG 5 7KH IRUZDUG VDPSOH FRYDULDQFH PDWUL[ 5 LV REWDLQHG IURP =L DV IROORZ 1fÂ§0 5 1 fÂ§ \ 0 O ]fÂ§f L =M]I f DQG WKH EDFNZDUG VDPSOH FRYDULDQFH PDWUL[ 5 FDQ EH REWDLQHG VLPLODUO\ IURP ] 'DWD 0RGHO $PELJXLWLHV DQG 7KHLU (IIHFWV RQ 6$5 ,PDJH )RUPDWLRQ 7DUJHW IHDWXUH H[WUDFWLRQ PHWKRGV GHYLVHG EDVHG RQ WKH GDWD PRGHO LQ f DUH UREXVW DJDLQVW GDWD PLVPRGHOLQJ HUURUV GXH WR WKH PRGHO IOH[LELOLW\ +RZHYHU
PAGE 50
WKHUH DUH DPELJXLW\ SUREOHPV DVVRFLDWHG ZLWK WKH VHPLSDUDPHWULF GDWD PRGHO ,Q WKLV VHFWLRQ ZH ILUVW DQDO\]H WKH SRVVLEOH GDWD PRGHO DPELJXLWLHV DQG WKHQ LOOXVWUDWH WKHLU HIIHFWV RQ 6$5 WDUJHW IHDWXUH H[WUDFWLRQ DQG LPDJH IRUPDWLRQ 7KH GLVFXVVLRQV EHORZ ZLOO PRWLYDWH WKH LQWURGXFWLRQ RI WKH 63$5 DOJRULWKP ZKLFK ZLOO EH SUHVHQWHG LQ GHWDLO LQ &KDSWHU 0RGHO $PELJXLWLHV 'XH WR WKH IOH[LELOLW\ RI WKH GDWD PRGHO LQ f WKHUH DUH YDULRXV W\SHV RI DPELJXLWLHV WKDW PD\ LPSDFW WKH IHDWXUH H[WUDFWLRQ RI HDFK VFDWWHUHU %HORZ ZH OLVW VHYHUDO W\SHV RI WKH DPELJXLWLHV LQKHUHQW LQ WKH GDWD PRGHO 7\SH 6LQJOH VFDWWHUHU )URP f ZH QRWH WKDW DPELJXLW\ H[LVWV EHWZHHQ I DQG [Qf VLQFH [^fHLW! fÂ§ DfHAUA efHA f ZKHUH [Qf fÂ§ [Qf DQG A U $PELJXLW\ DOVR H[LVWV EHWZHHQ DQG [Qf VLQFH [^fHMrI Of[^fHMQ9rf 6IUL:r" [ULfH f ZKHUH [^Qf fÂ§OfQ[f DQG fÂ§ 7KH DERYH WZR W\SHV RI DPELJXLWLHV FDQQRW EH UHVROYHG 7\SH 7ZR LGHQWLFDO VFDWWHUHUV ORFDWHG LQ WKH VDPH UDQJH /HW ID DQG IE UHVSHFWLYHO\ GHQRWH WKH FURVVUDQJH ORFDWLRQV RI WZR LGHQWLFDO VFDWWHUHUV DQG OHW !D DQG !E UHVSHFWLYHO\ GHQRWH WKHLU SKDVHV 7KHQ [Qf HDWUDQf B_B HM7LQf [Qf FRV >7ID IEfQ H0IDKfQ 7f ZKLFK LQGLFDWHV WKDW WZR LGHQWLFDO FRUQHU UHIOHFWRUV WULKHGUDOV RU GLKHGUDOVf ORFDWHG LQ WKH VDPH UDQJH EXW GLIIHUHQW FURVVUDQJH SRVLWLRQV ID DQG IE UHVSHFWLYHO\ FDQ EH PRGHOHG E\ f DV RQH fVFDWWHUHUf ORFDWHG DW ID IEf LQ FURVVUDQJH ZLWK [^Qf
PAGE 51
PRGXODWHG E\ FRV>UD fÂ§ IEfQ M!D fÂ§ !Ef@ 7KXV WKH GDWD PRGHO LQ f FDQQRW EH XVHG WR GHVFULEH HDFK RI WKH WZR FRUQHU UHIOHFWRUV LQ WKLV FDVH 7\SH 7ZR GLIIHUHQW VFDWWHUHUV ORFDWHG LQ WKH VDPH UDQJH $VVXPH WKDW WZR GLIIHUHQW FRUQHU UHIOHFWRUV ZLWK SDUDPHWHUV ^AÂf`A I Â`" DUH ORFDWHG LQ WKH VDPH UDQJH 7KHQ WKH WDUJHW PRGHO LQ WKH DEVHQFH RI QRLVH KDV WKH IRUP UQ Qf ;L^QfHALH3LUAQALQ? Q f Â‘ Â‘ 1 fÂ§ Q f f f 1 fÂ§ Â f :LWK VWUDLJKWIRUZDUG FDOFXODWLRQV ZH FDQ UHZULWH f DV UQf ;M f HAn HnU$ Q}f f ZKHUH M!L W ZLWK GHQRWLQJ DQ DUELWUDU\ SKDVH I I II I I ZLWK GHQRWLQJ DQ DUELWUDU\ FURVVUDQJH ORFDWLRQ DQG ÂLQf A[LQf FRV>YUL IfQ AÂ !f@ Â L ÂQf VLQ>UÂ B fmt a Lf@ f f 1RWH IURP f WKDW ^Â ^ÂÂQf`AU Â`Â DUH WKH DPELJXRXV IHDWXUHV RI ^FMf8 ^[Lf`baÂIIL`L Y 7\SH 0XOWLSOH VFDWWHUHUV ORFDWHG LQ WKH VDPH UDQJH :KHQ PRUH WKDQ WZR VFDWWHUHUV DUH ORFDWHG LQ WKH VDPH UDQJH WKH GDWD PRGHO LQ WKH DEVHQFH RI QRLVH FDQ VWLOO EH ZULWWHQ DV f H[FHSW WKDW / Lf A]ÂfFRV>UM If !Â tf@ f DQG ;Qf UÂ IfQ t Lf@ f
PAGE 52
ZKHUH / GHQRWHV WKH QXPEHU RI VFDWWHUHUV ORFDWHG LQ WKH VDPH UDQJH 7KXV WKH / VFDWWHUHUV ORFDWHG LQ WKH VDPH UDQJH DUH FRQVLGHUHG DV WZR fVFDWWHUHUVf ZKHQ XVLQJ WKH GDWD PRGHO LQ f %HIRUH ZH GLVFXVV WKH LPSDFW RI WKH PRGHO DPELJXLWLHV RQ 6$5 LPDJH IRUn PDWLRQ ZH ILUVW GHVFULEH KRZ WKH LPDJH IRUPDWLRQ LV GRQH LI ZH KDYH WKH HVWLPDWHG PRGHO SDUDPHWHUV ,PDJH )RUPDWLRQ $VVXPH IRU QRZ WKDW ZH KDYH H[WUDFWHG WKH WDUJHW IHDWXUHV EDVHG RQ DQ\ RI WKH DPELJXRXV GDWD PRGHOV )RU QRWDWLRQDO FRQYHQLHQFH ZH ZLOO XVH WKH QRWDWLRQ XVHG LQ f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f` GHQRWH WKH VLPXODWHG GDWD VHTXHQFH ZLWK D ODUJHU GLPHQVLRQ LQ UDQJH EDVHG RQ WKH HVWLPDWHG WDUJHW IHDWXUHV ^MfN ^[Nf`Aa4 IN IN`A O RI ^fN ^[Nf`A=4 IF IN`N L ZKHUH GHQRWHV WKH HVWLPDWH RI WKH VFDWWHUHU QXPEHU 7KHQ VVQ Qf fÂ§ ;NfHANN8? Q f f Â‘ e1 fÂ§ f f f Â“ fÂ§ IF L f
PAGE 53
ZKHUH e GHQRWHV DQ H[WUDSRODWLRQ IDFWRU e f DQG LV D SDUDPHWHU RI XVHU FKRLFH 1RWH WKDW WKH VXSHU UHVROXWLRQ SURSHUW\ RI WKH VRIRUPHG 6$5 LPDJHV LV GHWHUPLQHG E\ WKH IHDWXUH H[WUDFWLRQ DOJRULWKP DQG e LV RQO\ XVHG WR GHPRQVWUDWH WKH VXSHU UHVROXWLRQ SURSHUW\ RI WKH WDUJHW IHDWXUH H[WUDFWLRQ DOJRULWKP 7KH HVWLPDWHG QRLVH DQG FOXWWHU GDWD PDWUL[ LV Q Qf \Q f fÂ§ VVQ f Q f f f 1 fÂ§ f f f Â“ fÂ§ f ZKLFK LV DOVR LPSRUWDQW LQ PDQ\ 6$5 DSSOLFDWLRQV VLQFH IRU H[DPSOH LPSRUWDQW WDUJHW LQIRUPDWLRQ VXFK DV WKH WDUJHW VKDGRZ LQIRUPDWLRQ LV FRQWDLQHG LQ Q f :H FDQQRW H[WUDSRODWH Q f LQ HLWKHU UDQJH RU FURVVUDQJH VLQFH QR SDUDPHWULF GDWD PRGHO LV DYDLODEOH IRU Q f 7R REWDLQ 6$5 LPDJHV ZLWK ORZ VLGHOREHV LQ UDQJH YLD ))7 ZH DSSO\ ZLQGRZV WR VVQ f DQG Q f LQ UDQJH :H REWDLQ D QHZ e
PAGE 54
6WHS f $SSO\ WKH QRUPDOL]HG ))7 WR HDFK FROXPQ RI < WR REWDLQ DQ LQWHUPHGLDWH PDWUL[ DQG WKHQ DSSO\ $3(6 WR HDFK URZ RI WKH LQWHUPHGLDWH PDWUL[ 6HH >@ IRU WKH HIILFLHQW LPSOHPHQWDWLRQ RI $3(6f 1RWH WKDW WKH QRUPDOL]HG ))7 KDV WKH IRUP U1 f 0RGHO $PELJXLW\ (IIHFWV RQ 6$5 ,PDJH )RUPDWLRQ $OO RI WKH DIRUHPHQWLRQHG W\SHV RI DPELJXLWLHV ZLOO KDYH QR HIIHFW RQ WKH 6$5 LPDJH IRUPDWLRQ LI QR SDUDPHWHU HVWLPDWLRQ HUURUV H[LVW VLQFH WKH VFDWWHUHUV ZLOO EH SHUIHFWO\ UHFRQVWUXFWHG E\ XVLQJ DQ\ RI WKH SRVVLEOH DPELJXRXV GDWD PRGHOV )RU H[DPSOH ZKHQ WKHUH DUH WZR LGHQWLFDO VFDWWHUHUV ORFDWHG LQ WKH VDPH UDQJH WKH GDWD PRGHO LQ f FDQ VWLOO EH XVHG IRU 6$5 LPDJH IRUPDWLRQ VLQFH WKH WZR VFDWWHUHUV DUH QRZ GHVFULEHG DV RQH fVFDWWHUHUf ZLWK f ZKLFK VWLOO ILWV WKH GDWD PRGHO RI f ZLWK +HQFH WKH RULJLQDO 6$5 LPDJH FDQ VWLOO EH UHFRQVWUXFWHG E\ XVLQJ WKH SDUDPHWHUV RI WKH RQH fVFDWWHUHUf GHVFULEHG E\ WKH ULJKW VLGH RI f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f )LJXUH LV REWDLQHG E\ DVVXPLQJ QR SDUDPHWHU HVWLPDWLRQ HUURUV )LJXUH Df VKRZV WKH ))7 LPDJH RI D WDUJHW FRQVLVWLQJ RI WZR GLKHGUDOV RI GLIIHUHQW OHQJWKV ORFDWHG LQ WKH VDPH UDQJH :H XVH [NQf DNVPF>ENQQ UNf@ N Q f f r f
PAGE 55
N 2LN !N IN ,N EN 7N N O N 7DEOH 7UXH SDUDPHWHUV RI WKH WZR GLKHGUDOV XVHG LQ )LJXUHV DQG WR VLPXODWH WKH GLKHGUDOV ZKHUH DQG EUHVSHFWLYHO\ DUH SURSRUWLRQDO WR WKH PD[LPDO 5&6 DQG WKH OHQJWK RI WKH NWK GLKHGUDO FRUQHU UHIOHFWRU DQG UGHQRWHV WKH SHDN ORFDWLRQ RI WKH GDWD VHTXHQFH DQG LV GHWHUPLQHG E\ WKH RULHQWDWLRQ RI WKH NWK GLKHGUDO 7KH VL]H RI WKH VLPXODWHG GDWD PDWUL[ LV [ LH 1 Â“ f 7KH SDUDPHWHUV IRU WKH WZR GLKHGUDOV DUH JLYHQ LQ 7DEOH $Q DPELJXRXV VHW RI WDUJHW IHDWXUHV FDQ EH REWDLQHG E\ FKRRVLQJ IL fÂ§ I I I !L DQG c! fÂ§ W LQ f DQG f 7KH ZLQGRZHG ))7 6$5 LPDJHV RI WKH WZR fVFDWWHUHUVf DUH VKRZQ LQ )LJXUHV Ef DQG Ff UHVSHFWLYHO\ ZKLFK GLIIHU FRQVLGHUDEO\ IURP WKH WZR GLKHGUDO VFDWWHUHUV LQ )LJXUH Df 7KH FRPELQHG 6$5 LPDJH RI WKH WZR fVFDWWHUHUVf LV JLYHQ LQ )LJXUH Gf ZKLFK LV H[DFWO\ WKH VDPH DV WKH WUXH LPDJH VKRZQ LQ )LJXUH Df +RZHYHU GXH WR WKH SUHVHQFH RI QRLVH DQG FOXWWHU SDUDPHWHU HVWLPDWLRQ HUURUV DUH LQHYLWDEOH 7KH HUURUV LQ UDQJH DUH WKH PDLQ FDXVH RI WKH DUWLIDFW SUREOHP LQ WKH KLJK UHVROXWLRQ 6$5 LPDJH IRUPDWLRQ ,Q )LJXUH ZH DVVXPH WKDW DOO SDUDPHWHUV DUH DFFXUDWH H[FHSW WKDW )LJXUHV Df DQG Ef UHVSHFWLYHO\ VKRZ WKH ZLQGRZHG ))7 LPDJHV RI WKH WZR DIRUHPHQWLRQHG fVFDWWHUHUVf LQ WKH SUHVHQFH RI HVWLPDWLRQ HUURUV DQG )LJXUH Ff VKRZV WKH FRPELQHG 6$5 LPDJH %\ FRPSDULQJ )LJXUHV Df DQG Ff ZH QRWH WKDW DQ H[WUD OLQH DUWLIDFWf VKRZV XS QH[W WR WKH VKRUW GLKHGUDO 7KH UHDVRQ LV WKDW GXH WR WKH HVWLPDWLRQ HUURUV A I +HQFH WKH WZR fVFDWWHUHUVf LQ )LJXUHV Df DQG Ef DUH QRW H[DFWO\ LQ WKH VDPH UDQJH DQG FDQQRW EH fFRPELQHGf SHUIHFWO\ WR REWDLQ WKH WZR GLKHGUDO OLQHV LQ )LJXUH Df 7KLV SUREOHP EHFRPHV HYHQ ZRUVH ZKHQ VXSHU UHVROXWLRQ 6$5 LPDJHV DUH IRUPHG YLD GDWD H[WUDSRODWLRQ LQ UDQJH 7KH
PAGE 56
ODUJHU WKH H[WUDSRODWLRQ IDFWRU e WKH PRUH VLJQLILFDQW WKH DUWLIDFW SUREOHP VLQFH WKH GLIIHUHQFH EHWZHHQ I[ DQG I LV H[DJJHUDWHG e WLPHV )LJXUHV Gf VKRZV WKH 6$5 LPDJH REWDLQHG ZLWK e %\ FRPSDULQJ )LJXUHV Gf DQG Ff KHUH e ZLWKRXW H[WUDSRODWLRQf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n UDWHO\ %HIRUH ZH SUHVHQW WKH WDUJHW IHDWXUH H[WUDFWLRQ DOJRULWKP ZH ILUVW VXPPDUL]H WKH VWHSV QHHGHG IRU WKH IHDWXUH H[WUDFWLRQ RI D VLQJOH VFDWWHUHU DV D SUHSDUDWLRQ 7KH JHQHUDOL]HG $NDLNH LQIRUPDWLRQ FULWHULRQ *$,&f LV DOVR LQWURGXFHG WR HVWLPDWH WKH QXPEHU RI VFDWWHUHUV DW WKH HQG RI WKLV VXEVHFWLRQ )HDWXUH ([WUDFWLRQ RI D 6LQJOH 6FDWWHUHU 7KH GDWD PRGHO RI D VLQJOH VFDWWHUHU LQ WKH SUHVHQFH RI QRLVH KDV WKH IRUP \VQ f VQ f HVQ f Q f f f ,9 fÂ§ f f f Â“ fÂ§ f
PAGE 57
ZKHUH VQQf LV JLYHQ LQ f DQG ^HVQQf` GHQRWHV WKH XQNQRZQ QRLVH DQG FOXWWHU VHTXHQFH /HW 8Q8f fÂ§ HMLUI 7 HMQI1Of f DQG :IIf ?7 HMQI HM7UI1Of f ZKHUH f7 GHQRWHV WKH WUDQVSRVH /HW 'f GHQRWH WKH IROORZLQJ GLDJRQDO PDWUL[ 'f GLDJ_O H.r:Lff 'HILQH 7 7 [ fÂ§ [f [Of Â‘ Â‘ Â‘ [Â“ fÂ§ f Â‘ f /HW Hf\Vf ZAf@ f ZKHUH 5H Df GHQRWHV WKH UHDO SDUW RI [ DQG k GHQRWHV WKH +DGDPDUG PDWUL[ SURGXFW RU WKH HOHPHQWZLVH SURGXFW RI WZR PDWULFHV L ÂDUJ Â >\@!0f@9f f
PAGE 58
ZKHUH DUJDf GHQRWHV WKH DUJXPHQW RI D FRPSOH[ YDULDEOH [ ILQDOO\ ^` DUJPD[ & f ILI f ZKHUH 7KH VWHSV QHHGHG WR REWDLQ WKH 1/6 HVWLPDWHV RI D VLQJOH VFDWWHUHU DUH VXPPDUL]HG DV IROORZV 6WHS ,f 8VH f WR REWDLQ \Vf DQG REWDLQ WKH FRVW IXQFWLRQ &AI If DFFRUGLQJ WR f 'HWHUPLQH ^ ` E\ PD[LPL]LQJ & f XVLQJ WKH PHWKRG JLYHQ LQ $SSHQGL[ $ 6WHS ,,f &DOFXODWH DFFRUGLQJ WR f ZLWK ^` UHSODFHG E\ ^` REWDLQHG LQ 6WHS ,f 6WHS ,,,f &DOFXODWH [ YLD f ZLWK ^` UHSODFHG E\ ^` GHWHUn PLQHG LQ 6WHSV ,f DQG ,,f UHVSHFWLYHO\ )HDWXUH ([WUDFWLRQ RI 0XOWLSOH 6FDWWHUHUV :KHQ D WDUJHW FRQVLVWV RI PXOWLSOH VFDWWHUHUV ZH FDQ REWDLQ WKH 1/6 HVWLn PDWHV RI WKH WDUJHW IHDWXUHV EDVHG RQ f E\ XVLQJ D UHOD[DWLRQEDVHG RSWLPL]DWLRQ DSSURDFK /HW 7 ;IF ;IFf [NOf fff [N1 f f DQG OHW < DQG ( EH 1 [ 1 PDWULFHV ZLWK WKHLU QQfWK HOHPHQWV EHLQJ \QQf DQG HQ f UHVSHFWLYHO\ 7KHQ ZH FDQ UHZULWH f DV f N ZKHUH *N[N IN INf KDV WKH VDPH IRUP DV WKH *[f LQ f H[FHSW WKDW [ DQG DUH UHSODFHG E\ [IF IN DQG IN UHVSHFWLYHO\ /HW \ f f f Â“ fÂ§ EH WKH
PAGE 59
QWK FROXPQ RI < 7KHQ WKH HVWLPDWHV ^ID [IF IN IN`N RI ^ID [IF IN IN`N FDQ EH REWDLQHG E\ PLQLPL]LQJ WKH IROORZLQJ 1/6 FRVW IXQFWLRQ & ^IDAN IN IN`N Lf @ 7KH PLQLPL]DWLRQ RI & ^^ID [IF IN IN`N f LQ f LV D YHU\ FRPSOLFDWHG RSWLPL]DWLRQ SUREOHP 7KH SURSRVHG 63$5 DOJRULWKP SHUIRUPV D FRPSOHWH UHOD[DWLRQEDVHG VHDUFK E\ OHWWLQJ RQO\ WKH SDUDPHWHUV RI RQH VFDWWHUHU YDU\ DQG IUHH]LQJ DOO RWKHUV DW WKHLU PRVW UHn FHQWO\ GHWHUPLQHG YDOXHV IRU HDFK DVVXPHG QXPEHU RI VFDWWHUHUV /HW
PAGE 60
WKH LQWHUYDO Q? Q Q VR WKDW WKH PDJQLWXGH RI 9IH LV DERYH D FHUWDLQ WKUHVKROG VD\ ZLWKLQ WKH LQWHUYDO 6LPLODUO\ ZH FDQ IL[ Q WR Q DQG VHDUFK IRU WKH LQWHUYDO L 7KHQ WKH 1 [ Â“ UHFWDQJXODU ZLQGRZ ZQILf KDV XQLW YDOXH IRU QL Q Q DQG ? DQG ]HUR YDOXH HOVHZKHUH 7KH WKUHVKROG 7Â ZH XVH LQ RXU QXPHULFDO DQG H[SHULPHQWDO H[DPSOHV LV b RI WKH SHDN YDOXH RI WKH PDJQLWXGH RI 9IF 6WHS LLLf 'HWHUPLQH
PAGE 61
^AIF Frc IF IF`r L FDOFXODWH < IURP < DQG WKHQ UHGHWHUPLQH ^IF [IH IN IF`IF IURP < ,WHUDWH WKH SUHYLRXV WKUHH VXEVWHSV XQWLO fSUDFWLFDO FRQYHUJHQFHf 5HPDLQLQJ 6WHSV &RQWLQXH VLPLODUO\ XQWLO LV HTXDO WR WKH GHVLUHG RU HVWLPDWHG QXPEHU RI VFDWWHUHUV 7KH fSUDFWLFDO FRQYHUJHQFHf LQ WKH LWHUDWLRQV RI WKH DERYH UHOD[DWLRQEDVHG 1/6 DOJRULWKP PD\ EH GHWHUPLQHG E\ FKHFNLQJ WKH UHODWLYH FKDQJH H RI WKH FRVW IXQFWLRQ & I^IF;IFIFIF`LQ f EHWZHHQ WZR FRQVHFXWLYH LWHUDWLRQV 2XU QXPHULFDO DQG H[SHULPHQWDO H[DPSOHV VKRZ WKDW WKH DOJRULWKP XVXDOO\ FRQYHUJHV LQ D IHZ LWHUDWLRQV :H FDQ GHWHUPLQH WKH QXPEHU RI VFDWWHUHUV LQ f E\ H[WHQGLQJ WKH JHQHUDOL]HG $NDLNH LQIRUPDWLRQ FULWHULRQ VHH >@ IRU GHWDLOVf %\ DVVXPLQJ WKDW WKH QRLVH LV ZKLWH WKH HVWLPDWH L< RI L< LV GHWHUPLQHG DV DQ LQWHJHU WKDW PLQLPL]HV WKH IROORZLQJ H[WHQGHG *$,& FRVW IXQFWLRQ 1Â“ ? (( ?SFQQf?? \?Q>?Q\11f@>.1 f @ f Q ZKHUH LV D FRQVWDQW RI XVHU FKRLFH N IFLQ f \Q ffÂ§;N^QfHANHffn.ANQANQ? Q f f f 79fÂ§ f Â‘ f 79fÂ§ IF L f DQG 7I79 f LV WKH WRWDO QXPEHU RI UHDOYDOXHG XQNQRZQ SDUDPHWHUV RI ZKLFK .79 f LV IRU WKH VFDWWHUHUV DQG LV IRU WKH ZKLWH QRLVH YDULDQFHf 1RWH WKDW WKH 1/6 HVWLPDWHV RI ^MfN [IF IN IN`N L FDQ DOVR EH HVWLPDWHG IURP GHWHUPLQHG LQ f UDWKHU WKDQ ^
PAGE 62
fVFDWWHUHUVf +RZHYHU WKH +\EULG 6$5 LPDJHV PD\ VXIIHU IURP PRUH VHYHUH DUWLIDFW SUREOHP WKDQ 63$5 HVSHFLDOO\ DW ORZ 615 :KHQ QR PXOWLSOH VFDWWHUHUV DUH ORFDWHG LQ WKH VDPH UDQJH 63$5 DQG +\EULG SHUIRUP VLPLODUO\ 0RGLILHG 5(/$;1/6 $OJRULWKP 5(/$;1/6 >@ LV D SDUDPHWULF DSSURDFK IRU WKH IHDWXUH H[WUDFWLRQ RI WDUJHWV FRQVLVWLQJ RI ERWK WULKHGUDOV DQG GLKHGUDOV ,W LV EDVHG RQ D PL[HG GDWD PRGHO LQ ZKLFK [Qf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`N L GHQRWH WKH SDUDPHWHU HVWLPDWHV REWDLQHG YLD 63$5 DFFRUGLQJ WR WKH GDWD PRGHO LQ f ZKHUH LV WKH HVWLPDWHG QXPEHU RI VFDWWHUHUV REWDLQHG YLD *$,& 7KH 63$5 HVWLPDWHV ^ID ID IN IN`N FDQQRW EH XVHG GLUHFWO\ DV LQLWLDO FRQGLWLRQV IRU 5(/$;1/6 7KH LQLWLDO FRQGLWLRQV DUH REWDLQHG E\ DSSO\LQJ WKH ILUVW VWHS RI 5(/$;1/6 >@ WR HDFK 8IF N fÂ§ f f f ZKHUH N 8IF < fÂ§ e Hrr* LIUMLMLf N Â‘ Â‘ Â‘ f Â OÂA]N 2QFH ZH KDYH WKH LQLWLDO FRQGLWLRQV ZH FDQ XVH WKH ODVW VWHS RI 5(/$;1/6 >@ WR REWDLQ WKH GLKHGUDO DQG WULKHGUDO SDUDPHWHU HVWLPDWHV ZKLFK DUH WKHQ XVHG IRU 6$5 LPDJH IRUPDWLRQ >@ :H UHIHU WR WKLV DSSURDFK DV WKH PRGLILHG 5(/$;1/6
PAGE 63
&UDPU5DR %RXQG &5%f RI WKH 6HPL3DUDPHWULF 'DWD 0RGHO /HW < EH DQ 1 [ Â“ PDWUL[ ZLWK LWV Q fWK HOHPHQW EHLQJ \QILf ZLWK \Q Qf A [N^QfHLW!N HQ Qf Q fÂ§ f f f 1 fÂ§ Q f f f 1 fÂ§ N f ZKHUH GHQRWHV WKH QXPEHU RI VFDWWHUHUV ^[Nf`b 4 GHQRWHV WKH UHDOYDOXHG DPn SOLWXGH IXQFWLRQ RI IRU WKH FWK VFDWWHUHU TrfN DQG ^ININ` UHVSHFWLYHO\ DUH WKH FRQVWDQW SKDVH DQG WKH IUHTXHQF\ SDLU RI WKH eWK VFDWWHUHU ILQDOO\ ^HQ f` GHQRWHV WKH XQNQRZQ QRLVH DQG FOXWWHU VHTXHQFH 7KHQ ZH FDQ UHZULWH f DV < eHrf:MYrf[e'IFf ( f N ZKHUH ( GHQRWHV DQ 1 [ 1 PDWUL[ ZLWK HQ Qf EHLQJ LWV Q QfWK HOHPHQW Q W !1^INf fÂ§ ZYIFf O HAN HAcN^1?f JMUIF HMLUIK1Of f 7 f ZLWK fU GHQRWLQJ WKH WUDQVSRVH 'INf GHQRWLQJ WKH IROORZLQJ GLDJRQDO PDWUL[ 'rf GLDJ_ O HA8Â“Lf ? DQG ;F r 7 [Nf [NOf [N1 f f f /HW \ YHF \I \" \/ @7 ZLWK \Q^Q fÂ§ f f Â‘ 1f EHLQJ WKH QWK FROXPQ RI < 7KHQ WKH GDWD PRGHO FDQ EH UHZULWWHQ DV IROORZV \ $L H f ZKHUH $ >ZmOf 2;O@ p ZZLf ZLMWfk;.@p::Lf f
PAGE 64
ZLWK k DQG J! UHVSHFWLYHO\ GHQRWLQJ WKH HOHPHQWZLVH SURGXFW DQG .URQHFNHU SURGn XFW RI WZR YHFWRUV + f DQG H YHF(f 1H[W ZH GHULYH WKH &5% PDWUL[ IRU WKH GDWD PRGHO LQ f ZKHUH WKH QRLVH FRYDULDQFH PDWUL[ LV DVVXPHG WR EH DUELWUDU\ DQG XQNQRZQ /HW 4 fÂ§ (^HH+` EH WKH QRLVH FRYDULDQFH PDWUL[ ZLWK f+ GHQRWLQJ WKH FRQMXJDWH WUDQVSRVH 7KHQ WKH H[WHQGHG 6OHSLDQ%DQJV IRUPXOD IRU WKH LMWK HOHPHQW RI WKH )LVKHU LQIRUPDWLRQ PDWUL[ ),0f KDV WKH IRUP > @ ^),0` WU Tn4n&7n4f 5H >$f 4B $LfM@ f ZKHUH ;f GHQRWHV WKH GHULYDWLYH RI ; ZLWK UHVSHFW WR WKH ÂWK XQNQRZQ SDUDPHWHU WU;f GHQRWHV WKH WUDFH RI ; DQG 5H;f GHQRWHV WKH UHDO SDUW RI ; 1RWH WKDW ),0 LV EORFN GLDJRQDO VLQFH 4 GRHV QRW GHSHQG RQ WKH SDUDPHWHUV LQ $[f DQG $[f GRHV QRW GHSHQG RQ WKH HOHPHQWV RI 4 +HQFH WKH &5% PDWUL[ IRU WKH WDUJHW IHDWXUHV RI LQWHUHVW FDQ EH FDOFXODWHG IURP WKH VHFRQG WHUP RQ WKH ULJKW VLGH RI f /HW f ZKHUH DQG 9 I7 I7 [H If K f I IL Â f f f Â. I K K f Â‘ f ,N [H 7 7 ; 7 fff ;. W L 7 f /HW )L ) DQG ) GHQRWH 11 [ PDWULFHV ZKRVH $WK FROXPQV DUH GHWHUPLQHG UHVSHFWLYHO\ E\ MrN ^>X}Lf[P}Lff f
PAGE 65
MQHW!N ^>XIWINf k;IF@ J! X1INf k G1f` f ZLWK DQG 1 L 7 79 MLUHAN ^>ZAMIFf k;IF 2G\A@ pZMYIFf` f f ZLWK O9 Q 7 fff 79 fÂ§ f /HW ) GHQRWH DQ 7979 [ .Â“ PDWUL[ ZKRVH >N fÂ§ f79 @WK FROXPQ LV GHWHUPLQHG E\ HAN ^>ZMM k G!\YIFf@ 6! X"\YIHf` f N . Â‘ f f 79 f ZKHUH ZQ LV DQ 79 [ YHFWRU ZLWK WKH WK HOHPHQW EHLQJ XQLW DQG DOO RWKHU HOHPHQWV EHLQJ ]HURV /HW ) )L ) ) ) f 7KHQ WKH &5% PDWUL[ IRU WKH SDUDPHWHU YHFWRU LV JLYHQ E\ &5%LMf >5H)II4)f@B f 1XPHULFDO DQG ([SHULPHQWDO 5HVXOWV :H GHPRQVWUDWH DQG FRPSDUH WKH 6$5 LPDJH IRUPDWLRQ SHUIRUPDQFHV RI 63$5 DQG WKH PRGLILHG 5(/$;1/6 DOJRULWKP ZLWK ERWK VLPXODWHG DQG H[SHULn PHQWDO H[DPSOHV 7KH DOJRULWKPV DUH DOVR FRPSDUHG ZLWK +\EULG DQG 5(/$;1/6 ,Q WKH IROORZLQJ H[DPSOHV WKH GLPHQVLRQV RI WKH RULJLQDO 6$5 SKDVH KLVWRU\ GDWD PDWUL[ DUH 79 79 DQG *$,& ZLWK LV XVHG WR GHWHUPLQH IRU WKH UHOD[DWLRQEDVHG IHDWXUH H[WUDFWLRQ DOJRULWKPV RI 63$5 DQG +\EULG 7KH WKUHVKROG 7L XVHG LQ WKH LVRODWLRQ SURFHVV RI 63$5 LV b RI WKH SHDN YDOXH 7KH PD[LPL]DWLRQ RI &f LQ f LV GRQH LQ WZR VWHSV )LUVW LQLWLDO IUHTXHQF\ HVWLPDWHV DQG
PAGE 66
DUH REWDLQHG YLD ))7 ZLWK ]HURSDGGLQJ WR D WRWDO OHQJWK RI LQ UDQJH DQG WR D WRWDO OHQJWK RI LQ FURVVUDQJH 1H[W WKHVH LQLWLDO HVWLPDWHV DUH UHILQHG E\ XVLQJ WKH )0,1 IXQFWLRQ LQ 0$7/$% DOWHUQDWHO\ LH E\ XSGDWLQJ ZKLOH IL[LQJ DW LWV PRVW UHFHQWO\ GHWHUPLQHG YDOXH DQG YLFH YHUVD XQWLO fSUDFWLFDO FRQYHUJHQFHf ZKLFK LV GHn WHUPLQHG E\ FKHFNLQJ WKH UHODWLYH FKDQJH RI WKH FRVW IXQFWLRQ &AIIf :H KDYH XVHG WR GHWHUPLQH WKH FRQYHUJHQFH RI WKLV ILQH VHDUFK DV ZHOO DV WKH UHOD[DWLRQEDVHG DOJRULWKP 7KH H[WUDSRODWLRQ IDFWRU e LV XVHG LQ UDQJH IRU 63$5 DQG +\EULG DQG LQ ERWK UDQJH DQG FURVVUDQJH IRU 5(/$;1/6 DQG WKH PRGLILHG 5(/$;1/6 %RWK DQG .DLVHU ZLQGRZV ZLWK VKDSH SDUDPHWHU c DUH XVHG ZKHQHYHU QHHGHG )LUVW FRQVLGHU D VLPXODWHG H[DPSOH ZLWK KLJK 615 7KH 6$5 SKDVH KLVWRU\ GDWD PDWUL[ LV VLPXODWHG E\ DVVXPLQJ WKDW WKHUH DUH IRXU WULKHGUDOV DQG WKUHH GLn KHGUDOV LQ WKH SUHVHQFH RI ]HURPHDQ ZKLWH FRPSOH[ *DXVVLDQ QRLVH ZLWK YDULDQFH RQ 7KH DPSOLWXGH IXQFWLRQV IRU WKH IRXU WULKHGUDOV DUH JHQHUDWHG DV IROORZV [N>Qf $ f f f Â“ f DQG [f Â‘ Â‘ Â‘ 9 fÂ§ f 7KH DPSOLWXGH IXQFWLRQV IRU WKH WKUHH GLKHGUDOV DUH [E^Qf VLQF>U fÂ§ f@ IW Â‘ f f Â“ fÂ§ f DQG [Nf VLQF>U fÂ§ f@ N IW Â‘ f f Â“ fÂ§ f ZKHUH VLQF[f VLQ[f[ )LJXUH Df VKRZV WKH PRGXOXV RI WKH WUXH 6$5 LPDJH 1RWH WKDW WZR RI WKH GLKHGUDOV DUH ORFDWHG LQ WKH VDPH FURVVUDQJH DQG DUH FORVHO\ VSDFHG LQ UDQJH DQG WZR RI WKHP DUH ORFDWHG LQ WKH VDPH UDQJH 2I WKH IRXU WULKHn GUDOV WZR RI WKHP DUH FORVHO\ VSDFHG LQ UDQJH DQG WKH RWKHU WZR DUH ORFDWHG LQ WKH
PAGE 67
VDPH UDQJH ,Q WKLV H[DPSOH +\EULG DQG 63$5 KDYH WKH VDPH HVWLPDWHG QXPEHU RI VFDWWHUHUV fÂ§ )LJXUH Ef VKRZV WKH ZLQGRZHG ))7 6$5 LPDJH REWDLQHG E\ DSSO\LQJ WKH QRUPDOL]HG ))7 WR WKH ZLQGRZHG GDWD PDWUL[ 6$5 LPDJHV IRUPHG YLD +\EULG 63$5 5(/$;1/6 DQG WKH PRGLILHG 5(/$;1/6 DOJRULWKPV DUH VKRZQ LQ )LJXUHV Ff WKURXJK If UHVSHFWLYHO\ :H QRWH WKDW DW KLJK 615 WKH +\EULG LPDJH LV VLPLODU WR WKH 63$5 LPDJH %RWK RI WKH SDUDPHWULF 5(/$;1/6 DQG WKH PRGLILHG 5(/$;1/6 DOJRULWKPV RXWSHUIRUP WKHLU VHPLSDUDPHWULF FRXQn WHUSDUWV 63$5 DQG +\EULG VLQFH WKH GDWD PRGHO XVHG E\ WKH SDUDPHWULF PHWKRGV LV FRUUHFW UDWKHU WKDQ DSSUR[LPDWH )RU WKLV H[DPSOH RXU VLPXODWLRQV VKRZ WKDW WKH UDWLRV EHWZHHQ WKH 0$7/$% IORSV QHHGHG E\ +\EULG 63$5 WKH PRGLILHG 5(/$; 1/6 DQG 5(/$;1/6 RYHU WKH IORSV QHHGHG E\ WKH ZLQGRZHG ))7 PHWKRG DUH DQG UHVSHFWLYHO\ 1RWH WKDW ERWK 63$5 DQG +\EULG DUH FRPSXWDn WLRQDOO\ PRUH HIILFLHQW WKDQ 5(/$;1/6 DQG WKH PRGLILHG 5(/$;1/6 ZLWK WKH PRGLILHG 5(/$;1/6 EHLQJ PRUH HIILFLHQW WKDQ 5(/$;1/6 &RQVLGHU QH[W D VLPXODWHG H[DPSOH ZLWK ORZ 615 7KH 6$5 SKDVH KLVWRU\ GDWD LV WKH VDPH DV LQ WKH DERYH H[DPSOH H[FHSW WKDW WKH QRLVH YDULDQFH LV LQFUHDVHG WR 6$5 LPDJHV REWDLQHG E\ XVLQJ WKH ZLQGRZHG ))7 +\EULG 63$5 5(/$;1/6 DQG WKH PRGLILHG 5(/$;1/6 DUH VKRZQ LQ )LJXUHV Df WKURXJK Hf UHVSHFWLYHO\ 1RWH WKDW WKH DUWLIDFW SUREOHP VWDUWV WR VKRZ XS LQ WKH +\EULG LPDJH LQ )LJXUH Ef GXH WR ODUJH SDUDPHWHU HVWLPDWLRQ HUURUV %\ FRPSDULQJ )LJXUHV Ff ZLWK Ef LW FDQ EH VHHQ WKDW 63$5 FDQ HIIHFWLYHO\ PLWLJDWH WKH DUWLIDFW SUREOHP )LQDOO\ FRQVLGHU DQ H[SHULPHQWDO H[DPSOH RI 6$5 LPDJH IRUPDWLRQ E\ XVLQJ WKH 067$5 6OLF\ GDWD FROOHFWHG E\ LPDJLQJ DQ REMHFW FRQVLVWLQJ RI ERWK WULKHGUDO DQG GLKHGUDO FRUQHU UHIOHFWRUV ZKLFK LV VKRZQ LQ )LJXUH 7KH GDWD ZDV FROOHFWHG E\ WKH 6DQGLD 1DWLRQDO /DERUDWRU\ 61/f XVLQJ WKH 67$5/26 VHQVRU 7KH ILHOG GDWD ZDV FROOHFWHG E\ D VSRWOLJKWPRGH 6$5 ZLWK D FDUULHU IUHTXHQF\ RI *+]
PAGE 68
DQG EDQGZLGWK RI *+] 7KH UDGDU ZDV DERXW NLORPHWHUV DZD\ IURP WKH JURXQG REMHFW 7KH GDWD ZDV FROOHFWHG ZKHQ WKH REMHFW ZDV LOOXPLQDWHG E\ WKH UDGDU IURP DSSUR[LPDWHO\ WKH D]LPXWK DQJOH r DQG HOHYDWLRQ DQJOH r 7R FURVVYDOLGDWH WKH H[SHULPHQWDO UHVXOWV JLYHQ EHORZ ;3$7&+ >@ D KLJK IUHTXHQF\ HOHFWURPDJQHWLF VFDWWHULQJ SUHGLFWLRQ FRGH IRU FRPSOH[ REMHFWV ZDV XVHG WR JHQHUDWH YHU\ KLJK UHVROXWLRQ SKDVH KLVWRU\ GDWD IRU WKH REMHFW VKRZQ LQ )LJXUH 7KH GDWD JHQHUDWHG E\ ;3$7&+ KDV D UHVROXWLRQ RI PHWHUV LQ ERWK UDQJH DQG FURVVUDQJH DQG WKH FRUUHVSRQGLQJ ZLQGRZHG ))7 6$5 LPDJH LV VKRZQ LQ )LJXUH Df :H KDYH XVHG WKH ORJ VFDOH IRU DOO RI WKH LPDJHV VKRZQ LQ )LJXUH f 7KH RULJLQDO H[SHULPHQWDO 6OLF\ GDWD KDV D UHVROXWLRQ RI PHWHUV LQ UDQJH DQG PHWHUV LQ FURVVUDQJH 7KH [ GDWD PDWUL[ ZH XVHG WR GHPRQVWUDWH WKH SHUIRUPDQFH RI RXU DOJRULWKPV KDV D VSRLOHG UHVROXWLRQ RI PHWHUV LQ UDQJH DQG PHWHUV LQ FURVVUDQJH 7KH ZLQGRZHG ))7 6$5 LPDJH RI WKLV GDWD PDWUL[ LV VKRZQ LQ )LJXUH Ef )LJXUH Ff VKRZV WKH +\EULG LPDJH ZLWK REWDLQHG YLD *$,&f )LJXUH Gf VKRZV WKH 63$5 LPDJH ZLWK fÂ§ REWDLQHG YLD *$,&f )LJXUHV Hf DQG If VKRZ WKH 6$5 LPDJHV REWDLQHG YLD 5(/$;1/6 DQG WKH PRGLILHG 5(/$;1/6 ZLWK UHVSHFWLYHO\ 1RWH WKDW +\EULG KDV D PRUH VHYHUH DUWLIDFW SUREOHP WKDQ 63$5 1RWH DOVR WKDW WKH 63$5 LPDJH VKRZQ LQ )LJXUH Gf DSSHDUV WR ILW )LJXUH Df DQG WKH FKDUDFWHULVWLFV RI WKH REMHFW LQ )LJXUH ZHOO +RZHYHU WKH SDUDPHWULF DOJRULWKPV DUH QRW DV UREXVW DV 63$5 VLQFH WKH SDUDPHWULF LPDJHV VKRZQ LQ )LJXUHV Hf DQG If GR QRW ILW )LJXUH Df DV ZHOO ZLWK RQH RI WKH VFDWWHUHUV PLVLGHQWLILHG DQG PLVORFDWHG )RU WKLV H[SHULPHQWDO H[DPSOH WKH UDWLRV EHWZHHQ WKH 0$7/$% IORSV QHHGHG E\ +\EULG 63$5 WKH PRGLILHG 5(/$;1/6 DQG 5(/$; 1/6 RYHU WKH IORSV QHHGHG E\ WKH ZLQGRZHG ))7 PHWKRG DUH DQG UHVSHFWLYHO\ 1RWH WKDW 63$5 FDQ VRPHWLPHV EH IDVWHU WKDQ +\EULG
PAGE 69
Df Ef Ff Gf )LJXUH $PELJXLW\ HIIHFW RQ WKH 6$5 LPDJH IRUPDWLRQ LQ WKH DEVHQFH RI UDQJH HVWLPDWLRQ HUURUV Df 7UXH ZLQGRZHG ))7 6$5 LPDJH Ef :LQGRZHG ))7 LPDJH RI WKH ILUVW fVFDWWHUHUf Ff :LQGRZHG ))7 LPDJH RI WKH VHFRQG fVFDWWHUHUf Gf &RPELQHG ZLQGRZHG ))7 LPDJH RI WKH WZR fVFDWWHUHUVf 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f
PAGE 70
Df Ef Ff Gf )LJXUH $PELJXLW\ HIIHFW RQ WKH 6$5 LPDJH IRUPDWLRQ LQ WKH SUHVHQFH RI UDQJH HVWLPDWLRQ HUURUV Df :LQGRZHG ))7 LPDJH RI WKH ILUVW fVFDWWHUHUf Ef :LQGRZHG ))7 LPDJH RI WKH VHFRQG fVFDWWHUHUf Ff &RPELQHG ZLQGRZHG ))7 LPDJH RI WKH WZR fVFDWWHUHUVf ZLWK & ZLWKRXW H[WUDSRODWLRQf Gf &RPELQHG ZLQGRZHG ))7 LPDJH ZLWK e 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f
PAGE 71
Df Ef Ff Hf If )LJXUH &RPSDULVRQ RI 6$5 LPDJHV IRUPHG XVLQJ GLIIHUHQW DOJRULWKPV IRU VLPXn ODWHG GDWD DW KLJK 615 Db f Df 7UXH 6$5 LPDJH Ef :LQGRZHG ))7 6$5 LPDJH Ff +\EULG 6$5 LPDJH Gf 63$5 6$5 LPDJH Hf 5(/$;1/6 6$5 LPDJH If 0RGLILHG 5(/$;1/6 6$5 LPDJH 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f
PAGE 72
Hf )LJXUH &RPSDULVRQ RI 6$5 LPDJHV IRUPHG XVLQJ GLIIHUHQW DOJRULWKPV IRU VLPXn ODWHG GDWD DW ORZ 615 UA f Df :LQGRZHG ))7 6$5 LPDJH Ef +\EULG 6$5 LPDJH Ff 63$5 6$5 LPDJH Gf 5(/$;1/6 6$5 LPDJH Hf 0RGLILHG 5(/$;1/6 6$5 LPDJH 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVn UDQJH UHVSHFWLYHO\f
PAGE 73
Hf If )LJXUH &RPSDULVRQ RI 6$5 LPDJHV REWDLQHG YLD GLIIHUHQW DOJRULWKPV IRU WKH 6OLF\ GDWD KEO r D]LPXWK DQG r HOHYDWLRQ DQJOHVf Df :LQGRZHG ))7 6$5 LPDJH IURP WKH ;3$7&+ GDWD Ef :LQGRZHG ))7 6$5 LPDJH IURP WKH 6OLF\ GDWD Ff +\EULG 6$5 LPDJH Gf 63$5 6$5 LPDJH Hf 5(/$;1/6 6$5 LPDJH If 0RGLILHG 5(/$;1/6 6$5 LPDJH Ff a If DUH DOO REWDLQHG IURP WKH GDWD XVHG LQ Ef 7KH YHUWLFDO DQG KRUL]RQWDO D[HV DUH IRU UDQJH DQG FURVVUDQJH UHVSHFWLYHO\f
PAGE 74
&+$37(5 7$5*(7 )($785( (;75$&7,21 86,1* &/6$5 &/6$5 'DWD 0RGHO :H GHVFULEH KRZ WR REWDLQ WDUJHW IHDWXUHV YLD WKH IXOO V\QWKHWLF DSHUWXUH VKRZQ LQ )LJXUH Df ZKLFK PDNHV D SUHSDUDWLRQ IRU WKH GLVFXVVLRQ RQ &/6$5 :H VWDUW E\ HVWDEOLVKLQJ RQHGLPHQVLRQDO 'f SRLQW VFDWWHUHU GDWD PRGHO RI KLJK UDQJH UHVROXWLRQ UDGDU DV D SUHSDUDWLRQ IRU WKH HVWDEOLVKPHQW RI GDWD PRGHO KHUHLQ 'DWD 0RGHO IRU +LJK 5DQJH 5HVROXWLRQ 5DGDU 7KH UDQJH UHVROXWLRQ RI D UDGDU LV GHWHUPLQHG E\ WKH UDGDU EDQGZLGWK 7R DFKLHYH KLJK UHVROXWLRQ LQ UDQJH WKH UDGDU PXVW WUDQVPLW ZLGH EDQG SXOVHV ZKLFK DUH RIWHQ OLQHDU )0 FKLUS SXOVHV > @ $ QRUPDOL]HG FKLUS SXOVH FDQ EH ZULWWHQ DV V^Wf H$rRWWf _W_ 7R f ZKHUH I GHQRWHV WKH FDUULHU IUHTXHQF\ T GHQRWHV WKH )0 UDWH DQG 7 GHQRWHV WKH ZLGWK RI WKH SXOVH :H DVVXPH WKDW DQG 7 DUH NQRZQ 7KH VLJQDO UHWXUQHG E\ D VFDWWHUHU RI D WDUJHW KDV WKH IRUP UWf f ZKHUH 67 LV GHWHUPLQHG E\ WKH 5&6 RI WKH VFDWWHUHU DQG W GHQRWHV WKH URXQGWULS WLPH GHOD\ 7KH GHPRGXODWHG VLJQDO GWf LV REWDLQHG E\ PL[LQJ UIf ZLWK VrW fÂ§ Uf IRU VRPH JLYHQ U 5F ZKHUH ffr GHQRWHV WKH FRPSOH[ FRQMXJDWH F GHQRWHV WKH SURSDJDWLRQ VSHHG RI WKH WUDQVPLWWHG HOHFWURPDJQHWLF ZDYH DQG 5 GHQRWHV WKH UDQJH GLVWDQFH IURP UDGDU WR WKH FHQWHU RI WKH LOOXPLQDWHG SDWFK B 7HWUR7Rf77RfHM77RfHMUURfL f
PAGE 75
7KH WHUP HBLA7B7A LQ f LV XVXDOO\ FORVH WR D FRQVWDQW IRU DOO UPcQ U U UPD[ ZKHUH UPD[ DQG UPcQ FRUUHVSRQG WR WKH PD[LPXP DQG PLQLPXP YDOXHV UHVSHFWLYHO\ RI WKH URXQGWULS WLPH GHOD\V EHWZHHQ WKH VFDWWHUHUV RI D WDUJHW DQG WKH UDGDU 7KLV WHUP FDQ DOVR EH SDUWLDOO\ UHPRYHG >@ /HW 'Xf GHQRWH WKH )RXULHU WUDQVIRUP RI a a X GWf 7KHQ WKH LQYHUVH )RXULHU WUDQVIRUP RI 'XMfHLL ZLOO KDYH WKH WHUP HBA7B7RA UHPRYHG Wf 7Hn^UrURf^7nRfHM^77RfL f ZKLFK LV D FRPSOH[ VLQXVRLG ZLWK IUHTXHQF\ fÂ§Uf DQG DPSOLWXGH 7H"UI7rf7a7E :H NQRZ UPD[ DQG UPMQ DSSUR[LPDWHO\ VLQFH ZH DVVXPH WKDW WKH DOWLWXGH DQWHQQD EHDPZLGWK DQG JUD]LQJ DQJOH RI WKH UDGDU DUH NQRZQ :H DOVR DVVXPH WKDW UPD[ fÂ§ UPLQf 7R 7KHQ IRU fÂ§7 UPD[ W 7 UPLQ WKH VFDWWHUHUV RI WKH WDUJHW DW GLIIHUHQW UDQJHV FRUUHVSRQG WR GLIIHUHQW IUHTXHQFLHV RI WKH VLJQDO GWf ZKLOH WKH 5&6fV RI WKH VFDWWHUHUV DUH SURSRUWLRQDO WR WKH DPSOLWXGHV RI WKH FRUUHVSRQGLQJ VLQXVRLGV 7KH UDQJHV DQG 5&6fV RI WKH WDUJHW VFDWWHUHUV DUH WKH WDUJHW IHDWXUHV )XOO 6\QWKHWLF $SHUWXUH 5DGDU $ EURDGVLGH GDWD FROOHFWLRQ JHRPHWU\ LQ D VSRWOLJKWPRGH 6$5 LV VKRZQ LQ )LJXUH >@ 7KH ;< = FRRUGLQDWH V\VWHP LV FHQWHUHG RQ D VPDOO SDWFK RI JURXQG ZKHUH D WDUJHW LV ORFDWHG 7KH JURXQG LV LOOXPLQDWHG E\ D QDUURZ UDGLR IUHTXHQF\ 5)f EHDP IURP WKH PRYLQJ UDGDU WKDW URWDWHV ZLWK UDGLXV 5Tf DURXQG WKH FRRUGLQDWH RULJLQ ,Q )LJXUH 5 GHQRWHV WKH GLVWDQFH EHWZHHQ WKH UDGDU DQG D VFDWWHUHU DW WKH SRVLWLRQ [\]f DQG DQG Mf DUH WKH D]LPXWK DQG HOHYDWLRQ DQJOHV RI WKH UDGDU UHODWLYH WR WKH ;<= FRRUGLQDWH V\VWHP :H DVVXPH WKDW DQG 5 DUH NQRZQ 7KH UDQJH 5 RI WKH VFDWWHUHU ORFDWHG DW [ \ ]f FDQ EH ZULWWHQ DV 5 >5R FRV FRV fÂ§ [f ^5R VLQ FRV I! fÂ§ \f 5T VLQ IL fÂ§ ]fa?A f
PAGE 76
5DGDU )LJXUH %URDGVLGH VSRWOLJKWPRGH 6$5 GDWD FROOHFWLRQ JHRPHWU\ ,W KDV EHHQ VKRZQ LQ $SSHQGL[ % WKDW 5 FDQ EH DSSUR[LPDWHG DV 5T fÂ§ [ FRV FRV fÂ§ \ VLQ FRV fÂ§ ] VLQ ZKHUH R fÂ§ R >[ fÂ§ ]f VLQR FRV 2 fÂ§ [] FRV 2@ fÂ§ Rf ; fÂ§ ; \ \ R \ ]f FRV R D]VLQ 2 R [\ FRV R \] VLQ 2 R DQG ] ] fÂ§ [ \f VLQ IW f f f f f 1RWH WKDW WKH VHFRQG WHUPV RI WKH ULJKW VLGHV RI f f DQG f DUH GXH WR WKH UDQJH DQG HOHYDWLRQ FXUYDWXUH HIIHFWV DQG FDQ EH QHJOHFWHG IRU ODUJH /HW U 6LQFH U fÂ§ WKHQ IURP f ZH KDYH GW f 6;W\] H[S MWTfÂ§74Âf R fÂ§ R fÂ§ [ FRV FRV fÂ§ \ VLQ FRV fÂ§]VLQ f@ f
PAGE 77
ZKHUH 6[\] LV SURSRUWLRQDO WR WKH 5&6 RI WKH VFDWWHUHU ORFDWHG DW [\]f )RU WWR fÂ§ 7T ÂfR fÂ§ 5RfF & U ZKHUH ?W U_ 7 ZH FDQ ZULWH f DV G>W f a 6[\] SMLLW[\W\]W]f F f f ZKHUH W\ fÂ§ WW fÂ§ 7 Âf FRV FRV F WWR fÂ§ 7 Âf VLQ FRV DQG W WW A Âf VLQ f f f 1RWH WKDW GW f LV D FRPSOH[ VLQXVRLG 7KH IUHTXHQFLHV RI WKH VLQXVRLG FRUUHVSRQG WR WKH ORFDWLRQ [ \ ]f RI WKH VFDWWHUHU ZKLOH WKH DPSOLWXGH LV SURSRUn WLRQDO WR LWV 5&6 1RWH WKDW [ \ ]f LV QRW WKH WUXH ORFDWLRQ D \ ]f RI WKH VFDWWHUHU EXW LV FORVH WR [ \ ]f IRU ODUJH 5 6HH WKH $SSHQGL[ & RQ KRZ WR FDOFXODWH [ \ ]f IURP [ \ ]ff :KHQ D WDUJHW KDV PXOWLSOH VFDWWHUHUV GW f LQ f ZLOO EH D VXP RI VLQXVRLGV 7KH ORFDWLRQV DQG 5&6fV RI WKH WDUJHW VFDWWHUHUV DUH WKH WDUJHW IHDWXUHV 6LQFH XVXDOO\ WKH VDPSOHV RQ WKH W DQG D[HV DUH XQLIRUPO\ VSDFHG WKH VDPSOHV RI W[ W\ DQG W] RFFXU DW WKH SRLQWV RI D SRODU JULG +HQFH 3RODUWR&DUWHVLDQ LQWHUSRODWLRQ PD\ EH QHHGHG IRU WKH GDWD VDPSOHV WR RFFXU DW UHFWDQJXODU JULG SRLQWV 6HH 6HFWLRQ IRU DQ DOWHUQDWLYH DSSURDFKf $IWHU 3RODUWR&DUWHVLDQ LQWHUSRODWLRQ DQG VDPSOLQJ WKH VLJQDO REWDLQHG E\ WKH IXOO DSHUWXUH 6$5 LQ WKH SUHVHQFH RI QRLVH FDQ EH ZULWWHQ DV \Q Qf A DNHZIFQ2IFZA7Lf HQ Qf f IF O ZKHUH Q f f f 1 fÂ§ f f f Â“ fÂ§ DQG f f f Â“ fÂ§ ZLWK 1 1 DQG Â“ GHQRWLQJ WKH QXPEHUV RI DYDLODEOH GDWD VDPSOHV LQ WKH WKUHH GLPHQn VLRQV GHQRWHV WKH QXPEHU RI VLQXVRLGV DMW DQG ^ZMW Â Â` GHQRWH WKH XQNQRZQ
PAGE 78
FRPSOH[ DPSOLWXGH DQG IUHTXHQFLHV RI WKH NWK VLQXVRLG N Â‘ Â‘ Â‘ ILn QDOO\ ^HQ f` GHQRWHV WKH XQNQRZQ QRLVH VHTXHQFH 7KH VLQXVRLGDO IUHTXHQFLHV ^XMN QINXN 7ING!N LUIN` FRUUHVSRQG WR WKH ORFDWLRQ RI WKH NWK VFDWWHUHU RI D UDGDU WDUJHW DQG FWN LV GHWHUPLQHG E\ LWV 5&6 /HW \ f f \1 f f \ f fff \1 f Â“ fÂ§ Â“ fÂ§ f Â“ fÂ§ 79 fÂ§ f 7 \1 fÂ§ Â“ fÂ§ 9 fÂ§ f f /HW H EH GHILQHG VLPLODUO\ IURP HQ f DV \ IURP \Q f 7KHQ WKH GDWD PRGHO LQ f FDQ EH UHZULWWHQ DV \ $FW H f ZKHUH D 7 2W? 2W n f 2W[ f DQG $ DL Â£L p DL Â‘ f f O. p DN f ZKHUH p GHQRWHV WKH .URQHFNHU SURGXFW >@ DQG L DIF HAN fÂ§ HMN D N fÂ§ HM4N HM1OfXLN HMÂ“OfeMN HMÂ“OfLN &XUYLOLQHDU 6$5 f f f $VVXPH WKDW WKH FXUYLOLQHDU DSHUWXUH XVHG LQ &/6$5 FRQVLVWV RI 0 GLIIHUHQW UDGDU YLHZLQJ DQJOHV 7KHQ WKH UHFHLYHG GDWD YHFWRU \F LQ &/6$5 LV DQ 01f [ VXEYHFWRU RI \ LQ f /HW ,F GHQRWH DQ 0 [ Â“Â“f PDWUL[ ZLWK HDFK FROXPQ DQG
PAGE 79
URZ FRQWDLQLQJ RQO\ RQH QRQ]HUR XQLW HOHPHQW FRUUHVSRQGLQJ WR WKH ORFDWLRQV RI WKH DYDLODEOH GDWD VDPSOHV 7KHQ \F $FD H f ZKHUH $F ^,FÂ£L J! Â£Lf` J! DL fff ^,FD. J! DAf` J! D. 7KH XQNQRZQ VLQXVRLGDO SDUDPHWHUV ^FRN N N FWN `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r__? f ZKHUH ____ GHQRWHV WKH (XFOLGHDQ QRUP :KHQ WKH QRLVH HQ f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f
PAGE 80
SDUDPHWHU HVWLPDWHV IRU WKH ILUVW fÂ§ VFDWWHUHUV FDQ EH XVHG WR LQLWLDOL]H WKH VHDUFK IRU WKH SDUDPHWHUV RI WKH .WK RQH :H ILUVW EULHIO\ PDNH D SUHSDUDWLRQ IRU WKH 5(/$; DSSURDFK /HW 5 \F \ DmÂ£L J! Â£Af f $ $ $ ZKHUH Â£Â Â£Â DQG Â£Â DUH IRUPHG VLPLODUO\ WR WKRVH LQ (TXDWLRQ f IURP DQG UHVSHFWLYHO\ DQG ^Â£fMÂcfMXM DUH DVVXPHG WR KDYH EHHQ HVWLPDWHG 7KHQ PLQLPL]LQJ & LQ f ZLWK UHVSHFW WR DN \LHOGV WKH HVWLPDWH Â£N RI DN 2LN >Dr f DIF Jf DIF@+ \IF 111 8N fÂ§8N WAN8N r!N 8N f DQG ^Â$ 2fIF &8IF` DUJ PD[ ODr p DW p Dr@+\IF f +HQFH ^IF Â'IF ` FDQ EH REWDLQHG DV WKH ORFDWLRQ RI WKH GRPLQDQW SHDN RI WKH SHULRGRJUDP K >Dr p DIH Dr@ \N 1Â“Â“ f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r p NN J! DN@+ \N1Â“Â“f :LWK WKH DERYH SUHSDUDWLRQV ZH QRZ SURFHHG WR SUHVHQW WKH VWHSV RI WKH 5(/$; DOJRULWKP IRU WDUJHW IHDWXUH H[WUDFWLRQ ZLWK WKH IXOO DSHUWXUH 6$5 6WHS f $VVXPH 2EWDLQ ^!N N *fN DN`N L IURP \ E\ XVLQJ f DQG f
PAGE 81
6WHS f $VVXPH &RPSXWH \ ZLWK f E\ XVLQJ ^FRN&MNFXNDN`N L REWDLQHG LQ 6WHS f 2EWDLQ ^EN eRN X!N GIH`IH IURP \ 1H[W FRPSXWH \L E\ XVLQJ ^KNX!NÂ£N`N DQG WKHQ UHGHWHUPLQH ^e!IF Â ÂIF GIF`IF IURP \L ,WHUDWH WKH SUHYLRXV WZR VXEVWHSV XQWLO fSUDFWLFDO FRQYHUJHQFHf LV DFKLHYHG WR EH GLVFXVVHG ODWHU RQf 6WHS f $VVXPH &RPSXWH \ E\ XVLQJ ^FMIH Â Â DN`N REWDLQHG LQ 6WHS f 2EWDLQ ^ &RNXfN tN`N IURP \ 1H[W FRPSXWH \L E\ XVLQJ ^Â Â N Â£N`N DQG UHGHWHUPLQH ^ÂfN&RN&RNDN`N L IURP \A 7KHQ FRPSXWH \ E\ XVLQJ ^N/fNXfNÂ£N`N LÂ DQG UHGHWHUPLQH ^XN ZN X!N DLN`N IURP \ ,WHUDWH WKH SUHYLRXV WKUHH VXEVWHSV XQWLO fSUDFWLFDO FRQYHUJHQFHf 5HPDLQLQJ 6WHSV &RQWLQXH VLPLODUO\ XQWLO LV HTXDO WR WKH GHVLUHG RU HVWLPDWHG QXPEHU RI VLQXVRLGV :KHQHYHU LV XQNQRZQ LW FDQ EH HVWLPDWHG IURP WKH DYDLODEOH GDWD IRU LQVWDQFH E\ XVLQJ WKH JHQHUDOL]HG $NDLNH LQIRUPDWLRQ FULWHULRQ *$,&f UXOHV ZKLFK DUH SDUWLFXODUO\ WDLORUHG WR WKH 5(/$; PHWKRG RI SDUDPHWHU HVWLPDWLRQ 6HH HJ >@f 7KH fSUDFWLFDO FRQYHUJHQFHf LQ WKH LWHUDWLRQV RI WKH DERYH 5(/$; PHWKRG PD\ EH GHWHUPLQHG E\ FKHFNLQJ WKH UHODWLYH FKDQJH H RI WKH FRVW IXQFWLRQ &? A^ÂcfcIF XfIF fN Â£N`e ALQ f EHWZHHQ WZR FRQVHFXWLYH LWHUDWLRQV 2XU QXPHULFDO H[DPSOHV VKRZ WKDW WKH LWHUDWLRQV XVXDOO\ FRQYHUJH LQ D IHZ VWHSV &XUYLOLQHDU $SHUWXUH 7KH 5(/$; DOJRULWKP IRU WKH FXUYLOLQHDU DSHUWXUH LV VLPLODU WR WKDW RI WKH IXOO DSHUWXUH H[FHSW WKDW (TXDWLRQV f DQG f DERYH DUH UHSODFHG E\ >^,FDIF !Â£IFf` !DIF@+\&IF 2LN 01 DQG 8N fÂ§8N 8N 8N 8N 8N ^XfNNXMN` DUJ PD[ >^,FDr p DIHf` p DN@+ \&N 8N8N8N f f
PAGE 82
UHVSHFWLYHO\ ZKHUH \&E \ F fL OLMN ^,FÂ£Â p Â£Âf` Â£Â f /HW \&IF EH VLPLODU WR \r H[FHSW WKDW WKH HOHPHQWV LQ \r WKDW DUH PLVVLQJ LQ \&N DUH UHSODFHG ZLWK ]HURV LQ \&N 7KHQ WKH ULJKW KDQG VLGH RI f FDQ DOVR EH FRPSXWHG E\ DSSO\LQJ ))7 WR \&N 3HUIRUPDQFH $QDO\VLV RI 3DUDPHWHU (VWLPDWLRQ YLD &5% &RQVLGHU ILUVW WKH FDVH RI WKH IXOO DSHUWXUH /HW 4 (MHHA` EH WKH QRLVH FRYDULDQFH PDWUL[ ZKLFK LV DUELWUDU\ DQG XQNQRZQ 7KH H[WHQGHG 6OHSLDQ%DQJV IRUPXOD IRU WKH LMWK HOHPHQW RI WKH )LVKHU LQIRUPDWLRQ PDWUL[ ),0f KDV WKH IRUP > @ ^),0` WU FUA4nTf 5H >r $fn 4an $Dfn@ f ZKHUH ;f GHQRWHV WKH GHULYDWLYH RI ; ZLWK UHVSHFW WR WKH ]WK XQNQRZQ SDUDPHWHU WU;f GHQRWHV WKH WUDFH RI ; DQG 5H;f GHQRWHV WKH UHDO SDUW RI ; 1RWH WKDW ),0 LV EORFN GLDJRQDO VLQFH 4 GRHV QRW GHSHQG RQ WKH SDUDPHWHUV LQ $Df DQG $Df GRHV QRW GHSHQG RQ WKH HOHPHQWV RI 4 +HQFH WKH &5% PDWUL[ IRU WKH WDUJHW IHDWXUHV RI LQWHUHVW FDQ EH FDOFXODWHG IURP WKH VHFRQG WHUP RQ WKH ULJKW VLGH RI f /HW 9 5HDf ,Q) Df DU RMW X!7 ZKHUH D L 7 &2 O f f Â‘ /2IF &2 O7 &2 L 2 f f f /O!N DQG FR L7 &2 & f Â‘ f &2N f f f f
PAGE 83
/HW ) $ Mn$ 'Xp 'Rp 'Rp f ZKHUH $ LV GHILQHG LQ f WKH $WK FROXPQV RI 'A 'Â DQG 'T DUH G>DN J! J! DIF@ZIF >DIF Â£IF J! DIF@GfIF DQG G>DIF J! DIF J! DIH@GXfIF UHVSHFWLYHO\ DQG f 7KHQ WKH &5% PDWUL[ IRU WKH SDUDPHWHU YHFWRU UM LV JLYHQ E\ &5%UMf >5H)+4B)f@B f )RU WKH FDVH RI FXUYLOLQHDU DSHUWXUH WKH &5% PDWUL[ IRU WKH WDUJHW SDUDPHWHUV LV VLPLODU WR WKH RQH LQ f H[FHSW WKDW WKH $ LQ f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fÂ§ DQG X fÂ§ &R 7KH DGGLWLYH QRLVH LV DVVXPHG WR EH ]HURPHDQ ZKLWH *DXVVLDQ ZLWK YDULDQFH D 7KH IXOO DSHUWXUH GDWD LV JHQHUDWHG DFFRUGLQJ WR f ZLWK 1 Â“ Â“ &RQVLGHU ILUVW WKH H[DPSOH RI 0 7KH FXUYHG DSHUWXUHV ZH FRQVLGHU LQFOXGH WKH SDUDEROLF RQH LQ )LJXUH Ef WKH /VKDSHG RQH LQ )LJXUH Ff DQG WKH RQHV LQ )LJXUH )RU WKH DUF DSHUWXUHV WKH $UF DSHUWXUH LV D TXDUWHU RI D
PAGE 84
Df Ef Ff )LJXUH &XUYLOLQHDU DSHUWXUHV Df 7KH FLUFXODU DSHUWXUH Ef 7KH $UF DSHUWXUH Ff 7KH $UF DSHUWXUH $SHUWXUH &5%Df &5%Rf &5%FMf &5%R!f )XOO &LUFXODU 3DUDEROLF /VKDSHG $UF $FU 7DEOH &RPSDULVRQ RI WKH &5%V LQ G%f RI WKH WDUJHW IHDWXUHV IRU WKH FDVHV RI WKH IXOO FLUFXODU SDUDEROLF /VKDSHG $UF DQG $UF DSHUWXUHV ZKHQ 0 fÂ§ fÂ§ FX f D DQG FU FLUFOH ZKRVH FHQWHU LV DW WKH XSSHU ULJKW FRUQHU RI WKH IXOO DSHUWXUH VKRZQ LQ )LJXUH Df DQG ZKRVH UDGLXV LV DV VKRZQ LQ )LJXUH Ef DQG WKH $UF DSHUWXUH LV RQH KDOI RI D SDUDEROLF DSHUWXUH ZKRVH YHUWH[ LV DW WKH ORZHU ULJKW FRUQHU RI WKH IXOO DSHUWXUH DQG ZKR VWDUWV DW WKH XSSHU OHIW FRUQHU RI WKH IXOO DSHUWXUH VKRZQ LQ )LJXUH Df DV VKRZQ LQ )LJXUH Ff 7KHVH FXUYLOLQHDU DSHUWXUHV DUH VXEVHWV RI WKH IXOO DSHUWXUH LQ )LJXUH Df DQG DUH PDGH WR EH DV ODUJH DV SRVVLEOH )RU H[DPSOH WKH UDGLXV RI WKH FLUFXODU DSHUWXUH LV LQ WKLV H[DPSOH 7DEOH VKRZV WKH &5%V RI WKH WDUJHW SDUDPHWHUV $V H[SHFWHG WKH &5% IRU OX LV WKH VDPH IRU
PAGE 85
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f ZKHQ XL fÂ§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Â“ UHFWDQJXODU JULG DQG PDSSHG WKH GDWD VDPSOH DW HDFK W[W\W]f WR WKH QHDUHVW JULG SRLQW ZKHUH W[ W\ DQG W] DUH IXQFWLRQV RI W DQG VHH (TXDWLRQV f WR ff /HW 1 GHQRWH WKH QXPEHU RI DYDLODEOH GDWD VDPSOHV 7KHQ IRU ERWK IXOO DQG FXUYLOLQHDU DSHUWXUHV WKH GDWD PRGHO FDQ EH ZULWWHQ DV \U ,U $D H f
PAGE 86
ZKHUH ,U GHQRWHV DQ Â“ [ 1Â“Â“f PDWUL[ ZLWK HDFK FROXPQ DQG URZ FRQWDLQLQJ RQO\ RQH QRQ]HUR XQLW HOHPHQW FRUUHVSRQGLQJ WR WKH ORFDWLRQV RI WKH DYDLODEOH GDWD VDPSOHV LQ WKH UHFWDQJXODU JULG 1RWH WKDW WKH ODUJHU WKH GLPHQVLRQV RI WKH UHFWDQJXODU JULG WKH PRUH DFFXUDWH WKH PDSSLQJ DSSUR[LPDWLRQ )RU ODUJH HQRXJK GLPHQVLRQV KRZHYHU WKH QRLVH ZLOO EHFRPH WKH GRPLQDQW VRXUFH RI HUURU )RU RXU H[DPSOHV ZH FKRVH L9 1 Â“ $V FRPSDUHG WR 3RODUWR&DUWHVLDQ LQWHUSRODWLRQ WKH JULG PDSSLQJ DSSURDFK DYRLGV WKH LQWHUSRODWLRQ VWHS QHHGHG E\ WKH IRUPHU ZKLFK FRXOG EH FRPSOLFDWHG IRU DQ DUELWUDU\ FXUYLOLQHDU DSHUWXUH 7KH IRUPHU DSSURDFK KRZHYHU LV FRPSXWDWLRQDOO\ PRUH HIILFLHQW GXH WR LWV VPDOOHU GDWD GLPHQVLRQV 7KH VWHSV RI DSSO\LQJ 5(/$; WR \U DUH VLPLODU WR WKRVH IRU WKH FDVH RI IXOO DSHUWXUH H[FHSW WKDW (TXDWLRQV f DQG f DERYH DUH UHSODFHG E\ >,UDIF Dr p DIFf@nA \UN 2LN 1 DQG 8N 8N 8N 8N :N 8N ^FIF XMN` DUJ PD[ >,UDIF p DIF J! Drf@ \7N f f UHVSHFWLYHO\ ZKHUH Q [ fÂ§f WW [ @f 1RWH WKDW WKH FRPSXWDWLRQDO DGYDQWDJHV RI ))7 RYHU ')7 GLVFUHWH )RXULHU WUDQVIRUPf DQG EDFNSURMHFWLRQ GLPLQLVK DV WKH GLPHQVLRQV RI WKH UHFWDQJXODU JULG LQFUHDVHV HVSHFLDOO\ IRU FXUYLOLQHDU DSHUWXUHV
PAGE 87
:H QRZ SUHVHQW DQ H[SHULPHQWDO H[DPSOH WR GHPRQVWUDWH WKH SHUIRUPDQFHV RI WKH 5(/$; DOJRULWKP IRU WDUJHW IHDWXUH H[WUDFWLRQ ZLWK D &/6$5 :H XVH H WR WHVW WKH FRQYHUJHQFH RI WKH 5(/$; DOJRULWKP LQ WKLV H[DPSOH 7KH LQGRRU GDWD XVHG KHUHLQ ZDV REWDLQHG E\ WKH 5DGDU 6LJQDWXUH %UDQFK 1DYDO $LU :DUIDUH &HQWHU 0XJX &DOLIRUQLD 7KH UDGDU FDUULHU IUHTXHQF\ LV *+] DQG WKH EDQGZLGWK *+] 7KH [ [ GDWD VHW ZDV REWDLQHG ZLWK WKH IXOO DSHUWXUH VKRZQ LQ )LJXUH Df ZLWK VDPSOHV LQ HDFK GLPHQVLRQ DQG WKH DQJXODU LQFUHPHQWV ZHUH r 7KH WDUJHW FRQVLVWV RI FRUQHU UHIOHFWRUV ZLWK D FXELF FRQILJXUDWLRQ DQG ZDV DERXW PHWHUV DZD\ IURP WKH UDGDU 7KH VFDWWHUHUV ZHUH DERXW PHWHUV DSDUW )LJXUH VKRZV WKH H[WUDFWHG VFDWWHUHUV ZKHQ LV XVHG ZLWK 5(/$; )LJXUH Df LV REWDLQHG ZKHQ WKH IXOO DSHUWXUH VKRZQ LQ )LJXUH Df LV XVHG WR H[WUDFW WKH WDUJHW IHDWXUHV DQG )LJXUH Ef LV REWDLQHG ZKHQ WKH /VKDSHG DSHUWXUH VKRZQ LQ )LJXUH Ff LV XVHG 7KH FHQWHUV RI WKH FLUFOHV GHQRWH WKH ORFDWLRQV RI WKH H[WUDFWHG VFDWWHUHUV LQ VSDFH DQG WKH UDGLXV RI HDFK FLUFOH LV SURSRUWLRQDO WR WKH PRGXOXV RI WKH 5&6 RI WKH FRUUHVSRQGLQJ VFDWWHUHU 7KH WULDQJOHV VKRZ WKH SURMHFWLRQV RI WKH VFDWWHUHU ORFDWLRQV RQWR WKH KRUL]RQWDO SODQH DQG WKHLU VL]HV DUH DOVR VFDOHG DFFRUGLQJ WR WKH 5&6fV RI WKH VFDWWHUHUV 1RWH WKDW )LJXUH Ef LV VLPLODU WR )LJXUH Df HYHQ WKRXJK WKH IRUPHU LV REWDLQHG E\ XVLQJ RQO\ b DPRXQW RI GDWD XVHG E\ WKH ODWWHU
PAGE 88
fÂ§ )XOO $SHUWXUH 0 B 0 r rb f r 0 rf f 5DGLXV Df P RF 2 7 7 )XOO $SHUWXUH 0 fÂ§ fr 0 fÂ§ 0 5DGLXV L U B fÂ§ )XOO $SHUWXUH 0 0 0 f /B 5DGLXV Ef 7 7 fÂ§ )XOO $SHUWXUH 0 Â‘ f f 0 0 5DGLXV Ff Gf )LJXUH 7KH &5%V RI WKH WDUJHW SDUDPHWHUV DV D IXQFWLRQ RI WKH FLUFXODU DSHUWXUH UDGLXV ZKHQ D WR fÂ§ fÂ§ DQG FU Df 7KH &5% RI WKH FRPSOH[ DPSOLWXGH D Ef 7KH &5% RI WR Ff 7KH &5% RI &R Gf 7KH &5% RI &R
PAGE 89
Df Ef )LJXUH SORWV RI VFDWWHUHUV H[WUDFWHG E\ XVLQJ 5(/$; ZLWK WKH LQGRRU H[SHULPHQWDO GDWD Df 2EWDLQHG ZLWK IXOO DSHUWXUH DV VKRZQ LQ )LJXUH Df Ef 2EWDLQHG ZLWK FXUYHG DSHUWXUH DV VKRZQ LQ )LJXUH Ff
PAGE 90
&+$37(5 $872)2&86 ,1 &/6$5 7R DQDO\]H WKH HIIHFW RI &/6$5 DSHUWXUH HUURUV ZH ILUVW UHZULWH WKH &/6$5 GDWD PRGHO LQ f ZLWK UHVSHFW WR UDGDU YLHZLQJ DQJOHV $VVXPH WKDW D FXUYLOLQn HDU DSHUWXUH FRQVLVWV RI 0 GLIIHUHQW YLHZLQJ DQJOHV DQG OHW GHQRWH WKH HOHYDWLRQ DQG D]LPXWK DQJOH SDLUV RI WKH 0 ORRN DQJOHV RI WKH UDGDU /HW \QPf Q f f f $7 fÂ§ GHQRWH WKH RQHGLPHQVLRQDO 'f GDWD VDPSOHV REWDLQHG DIWHU GHFKLUSLQJ IURP WKH PWK YLHZLQJ DQJOH RI WKH UDGDU /HW WR 7R WQf WQ Q f f f 1 fÂ§ f ZKHUH WQ GHQRWHV WKH WLPH VDPSOHV /HW W;MQ Qf W9P Qf DQG W=P QfUHVSHFWLYHO\ GHQRWH WKH WLPH VDPSOHV RI WKH PWK ORRN DQJOH ZKHUH DQG W;UQ Qf WQFRVPf FRVPf f W\P^Qf LQ VLQPf &26Pf f Af fÂ§ ÂQ VLQPf f ZLWK P f f f 0 DQG Q r Â‘ $7 fÂ§ 7KHQ \Q Pf KDV WKH IRUP \Q Pf A DNHM$INW[P:INW\PQfINW]PQf? $ HQ Pf P fÂ§ f f f 0 E O f ZKHUH HQ Pf GHQRWHV WKH QRLVH DQG FOXWWHU DQG LV WKH QXPEHU RI VFDWWHUHUV :H KDYH DVVXPHG WKDW ^PP`P L DQG 5R^Pf 5R P f f f 0 DUH NQRZQ H[DFWO\ LQ WKH &/6$5 GDWD f )RU D SUDFWLFDO FXUYLOLQHDU 6$5 V\VWHP KRZn HYHU WKH UDGDU SRVLWLRQV UHODWLYH WR WKH ;<= FRRUGLQDWH V\VWHP PD\ QRW EH NQRZQ
PAGE 91
H[DFWO\ GXH WR DWPRVSKHULF WXUEXOHQFH DQG SODWIRUP SRVLWLRQ XQFHUWDLQW\ ,Q 6$5 LPDJLQJ LW LV JHQHUDOO\ DVVXPHG WKDW WKH HUURUV LQ ^P P`:B DUH QHJOLJLEOH DQG WKH HUURUV LQ ^5Pf`A FDXVH SKDVH HUURUV DORQJ WKH V\QWKHWLF DSHUWXUH ,Q &/6$5 KRZHYHU WKH HUURUV LQ ^P P`eI PD\ QR ORQJHU EH QHJOLJLEOH VLQFH WKH DSHUWXUH VKDSH LV FULWLFDO IRU WDUJHW IHDWXUH H[WUDFWLRQ 2XU SUREOHP RI LQWHUHVW KHUHLQ LV WR FRPSHQVDWH IRU WKH FXUYLOLQHDU DSHUWXUH HUURUV LQ ^5Pf`A DQG ^P IfP`A DQG H[WUDFW WKH WDUJHW SDUDPHWHUV ^RÂN r ID IN`N L IURP ^\Q Pf`Q f Â‘ f 1fÂ§ P Â‘ f f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f P DQG FRVPf 7KHVH DSSUR[LPDWLRQV DOVR KROG IRU WKH WUXH ORRN DQJOHV RI WKH UDGDU 7KHQ DFFRUGLQJ WR f f DQG f UHVSHFWLYHO\ ZH KDYH Â[Pa ÂFRV Pf Â&26 rP DQG ÂVLQ Pf R R ZKHUH ^P P`P L GHQRWH WKH WUXH ORRN DQJOHV RI WKH UDGDU DQG WWR U f f f f W fÂ§ &RQVLGHU ILUVW WKH HUURUV LQ ^FMfP`A ZKHQ ^P APL"QAf 5R`P L /HW P fÂ§P $P f
PAGE 92
/HW WQ fÂ§ WWR f ,W FDQ EH VKRZQ WKDW IRU YHU\ VPDOO $P DQG U W fÂ§ Uf ZKHUH ?W fÂ§ UR_ 7R ZH KDYH DQG LFRV P WRVLQ f$P f Z 2 a WR &26 4f P f 2 2 WVLQ P LFRV Rf$P f ZKHUH 2 LV WKH DYHUDJH RI DOO P P Â‘ f Â‘ 0 DQG KHQFH LV D FRQVWDQW 7KHQ A H[S ^MWWINW;P INW\P $2` m H[S M M"U IFW &26 IfUQ INA2 &26 c!TWL INW VLQ f H[S MU INK VLQ !R IFLR FRV 2 $f f (TXDWLRQ f VKRZV WKDW LI $P LV D FRQVWDQW WKHQ WKH SKDVH HUURU GXH WR $P DQG WKH SKDVH RI DUH DPELJXRXV +HQFH WKH SKDVH RI FDQ QHYHU EH GHWHUPLQHG 2 a H[DFWO\ LQ WKH SUHVHQFH RI $P ,I $P LV D OLQHDU IXQFWLRQ RI P WKHQ IN DQG WKH OLQHDU SKDVH HUURU GXH WR $P DUH DPELJXRXV DQG FDQQRW EH GHWHUPLQHG H[DFWO\ R 6LPLODUO\ ZH FDQ DQDO\]H WKH HUURUV LQ ^P`P L ZKHQ ^P P L"Pf 5R`P Y /HW P $P f R )RU YHU\ VPDOO P DQG $P DQG IRU U } L fÂ§ Uf ZKHUH _L fÂ§ U_ 7 ZH KDYH 2 W;P a W &26 P f r R a WR &26 R P 7$Pf f f DQG R } L VLQ P f
PAGE 93
7KHQ DK H[S ^MYINW;P INW9P IFW=Pf` a DN H[S O MLU R R INW FRV P INW&26 2U INW VLQ f H[S M.INW FRV R $} f ZKLFK VKRZV WKDW D FRQVWDQW $P DOVR UHVXOWV LQ WKH DPELJXLW\ EHWZHHQ WKH SKDVH HUURU GXH WR $P DQG WKH SKDVH RI DN +HQFH WKH SKDVH RI DN FDQ QHYHU EH GHWHUPLQHG 2 a H[DFWO\ LQ WKH SUHVHQFH RI $P $OVR LI $P LV D OLQHDU IXQFWLRQ RI P WKHQ IN DQG WKH OLQHDU SKDVH HUURU GXH WR $P DUH DPELJXRXV DQG FDQQRW EH GHWHUPLQHG H[DFWO\ R 2 )LQDOO\ FRQVLGHU WKH HUURUV LQ 5RPf ZKHQ P fÂ§P DQG P P P fÂ§ fÂ§ $I /HW 5Pf fÂ§ 5 $5Pf f 5HSODFLQJ 5 LQ f ZLWK 5RPf DQG IRU VPDOO $L"Pf ODUJH 5Pf DQG U L fÂ§ Uf ZKHUH _L fÂ§ W_ 7 ZH KDYH IURP f f X[\]W M ^[W[ \WY ]W]f HMW $5T Pf f +HQFH WKH HUURUV LQ ^5Pf`A UHVXOW LQ SKDVH HUURUV DORQJ WKH V\QWKHWLF DSHUWXUH ZKLFK LV FRQVLVWHQW ZLWK WKH DQDO\VLV LQ > @ 7KHQ WKH ]QPf LQ f VKRXOG IRU WKLV FDVH EH UHSODFHG E\ [Q Pf ( N ? 2LN t MLU>INW[PQfINW\PQfINW]P0` O S9P f ZKHUH fÂ§ W$5Pf LV WKH SKDVH HUURU FDXVHG E\ $5Pf 1RWH WKDW WKLV SKDVH HUURU GLIIHUV IURP WKH SKDVH HUURUV LQ f DQG f VLQFH LW GRHV QRW GHSHQG RQ WKH SDUDPHWHUV RI WKH $WK VFDWWHUHU DQG LV HDVLHU WR GHDO ZLWK ,I $5Pf LV D FRQVWDQW WKHQ WKH SKDVH HUURU GXH WR $5Pf DQG WKH SKDVH RI DN DUH DPELJXRXV +HQFH WKH SKDVH RI DN FDQ QHYHU EH GHWHUPLQHG H[DFWO\ LQ WKH SUHVHQFH RI $5RPf ,I R $5Pf LV D OLQHDU IXQFWLRQ RI P WKHQ IN DQG WKH OLQHDU SKDVH HUURU GXH WR $5Pf LV DPELJXRXV DQG FDQQRW EH GHWHUPLQHG H[DFWO\
PAGE 94
7KH 5HOD[DWLRQ%DVHG $XWRIRFXV $OJRULWKP $8725(/$;f 7KH $8725(/$; DOJRULWKP REWDLQV WKH HVWLPDWHV ^P DQG $ $ n/ R ^tNININLIN`N Y UHVSHFWLYHO\ RI WKH WUXH YDOXHV ^P !P "P`eI DQG ^DN IN IN IN`N L E\ PLQLPL]LQJ WKH IROORZLQJ 1/6 FULWHULRQ ? 0 1O &V Â A A _"QPf [QPf_ f P Q ZKHUH [ f Q Pf ; DNHMr?NW[PQfINW\PQfINW=PQf@ A JMnUP ZLWK W;PQf W\PQf DQG W=PQf GHILQHG LQ f f DQG f UHVSHFWLYHO\ 1RWH WKDW DV VKRZQ EHORZ ZH ZLOO GHWHUPLQH P DQG FMfP YLD D VHDUFK PHWKRG +HQFH WKHUH LV QR QHHG WR XVH WKH DSSUR[LPDWLRQV VXFK DV WKRVH LQ f f DQG f IRU W[P Qff Â\r1 Qf DQG W]P Qf M UHVSHFWLYHO\ :H XVH DQ DOWHUQDWLQJ RSWLPL]DWLRQ DSSURDFK WR PLQLPL]H & LQ f DQG WKH $8725(/$; DOJRULWKP LV RXWOLQHG DV IROORZV 6WHS ([WUDFW WKH WDUJHW IHDWXUHV ^DN IN IN IN`N L ZLWK WKH 5(/$; DOJRULWKP IURP DQ LQLWLDO FXUYLOLQHDU DSHUWXUH E\ DVVXPLQJ WKDW ^UcP fÂ§ `eI 6WHS 8SGDWH WKH FXUYLOLQHDU DSHUWXUH ^PP`f DQG WKH SKDVH HUURU ^UP`P L ZLWK 6WHSV f fÂ§ f WR EH GLVFXVVHG LQ 6HFWLRQ f ,I ^U@P`A LV NQRZQ WR EH QHJOLJLEOH WKHQ RQO\ XSGDWH ^PfP`A ZLWK 6WHS ff 6WHS 5HGHWHUPLQH WKH WDUJHW SDUDPHWHUV ZLWK WKH 5(/$; DOJRULWKP E\ XVLQJ WKH FXUYLOLQHDU DSHUWXUH ^P I!P`A O DQG ^YP`A REWDLQHG LQ 6WHS 6WHS 5HSHDW 6WHSV DQG XQWLO fSUDFWLFDO FRQYHUJHQFHf 7KH fSUDFWLFDO FRQYHUJHQFHf RI $8725(/$; LV GHWHUPLQHG VLPLODUO\ WR WKDW RI WKH 5(/$; DOJRULWKP E\ FKHFNLQJ WKH UHODWLYH FKDQJH RI WKH FRVW IXQFWLRQ & LQ f EHWZHHQ WZR FRQVHFXWLYH LWHUDWLRQV $SHUWXUH 3DUDPHWHU (VWLPDWLRQ $VVXPH WKDW WKH WDUJHW IHDWXUH HVWLPDWHV ^tN IN I IN`N L DUH JLYHQ 7KHQ P !P DQG I`P FDQ EH GHWHUPLQHG E\ PLQLPL]LQJ WKH IROORZLQJ &P ZLWK UHVSHFW WR
PAGE 95
^2PSP9P` ZKHUH 1fÂ§O &P2POfP9Pf nA?\^QPf [AQAPf@ Q f f f 1 fÂ§ P f f f 0 f ZKHUH [?Q Pf KDV WKH VDPH IRUP DV [Q Pf LQ f H[FHSW WKDW ^DN IN IN IN`N L DUH UHSODFHG E\ ^Â£IF IN IN IN`N 7R VLPSOLI\ WKH RSWLPL]DWLRQ RI &P ZH GHWHUPLQH ^P /`P L DQG ^AP`P L LWHUDWLYHO\ DV IROORZV 6WHS f 2EWDLQ ?IMM E\ PLQLPL]LQJ WKH IROORZLQJ &P? ZLWK UHVSHFW WR ^ ZKHUH 9O FPL PIPf OAr1ff a ÂLQ!Pf_} Q ff1 P f f f 0 Q f ZKHUH ÂLQ Pf KDV WKH VDPH IRUP DV ]QPf LQ f H[FHSW WKDW ^DN IN IN IN`N L DUH UHSODFHG E\ ^Â£N IN r MWOLW/L :KHQ WKHUH DUH HUURUV LQ ERWK WKH HOHYDWLRQ DQG 2 R D]LPXWK GLUHFWLRQV ZH FDQ HVWLPDWH P DQG 4P E\ WKH DOWHUQDWLQJ PLQLPL]DWLRQ A R DSSURDFK LH E\ LWHUDWLYHO\ IL[LQJ WKH HVWLPDWH P RI P DQG PLQLPL]LQJ &PL ZLWK 9 R UHVSHFW WR P DQG WKHQ IL[LQJ WKH HVWLPDWH P RI P DQG PLQLPL]LQJ &PL ZLWK UHVSHFW WR P XQWLO fSUDFWLFDO FRQYHUJHQFHf 7KH fSUDFWLFDO FRQYHUJHQFHf LQ WKH DOWHUQDWLQJ PLQLPL]DWLRQ DSSURDFK LV GHWHUPLQHG E\ WKH UHODWLYH FKDQJH RI WKH FRVW IXQFWLRQ &PL ,Q WKH QXPHULFDO H[DPSOHV ZH WHUPLQDWH WKH LWHUDWLRQ ZKHQ WKH UHODWLYH FKDQJH RI WKH FRVW IXQFWLRQ &PL EHWZHHQ WZR FRQVHFXWLYH LWHUDWLRQV LV OHVV WKDQ :KHQ WKHUH DUH HUURUV LQ RQO\ RQH DQJXODU GLUHFWLRQ LH HLWKHU P RU !P LV WR EH GHWHUPLQHG WKHQ WKH PLQLPL]DWLRQ RI &PL LV D VLPSOH RQHGLPHQVLRQDO VHDUFK SUREOHP 6WHS f 'HWHUPLQH SKDVH HUURUV ^IMP`P E\ PLQLPL]LQJ WKH IROORZLQJ FRVW IXQFn WLRQV &P 9O &P ^9Pf ?\QLPf f ]QUQfHP? f Q
PAGE 96
ZKHUH Â Pf A ANH5$INL[P^QfnLNL\PQfNL=P^Qf`A Q 79 fÂ§ P f f f 0 N f DQG W;PQf L\PQf DQG L=PQf DUH WKH VDPH DV L,PQf W\UULQf DQG W=PQf UHVSHFn WLYHO\ H[FHSW WKDW ^P cfP`A DUH UHSODFHG E\ REWDLQHG LQ 6WHS f 7KLV VWHS LV VLPLODU WR >@ DQG ZH KDYH 9P DQJOH]A\Pf P Â‘ f f 0 f ZKHUH DQG \P ? 7 Pf Pf \1 Pf =R ÂQ 7 f f Â Pf ]OPf Â‘Â‘Â‘ ]1OPf 1RWH IURP f WKDW ZH GR QRW QHHG WKH VHDUFK RYHU SDUDPHWHU VSDFH WR GHWHUPLQH ^9P`P DQG KHQFH WKH HUURUV LQ ^5Pf`A O DUH HDVLHU WR GHDO ZLWK WKDQ WKRVH LQ ^HP$P`0 P f 6WHS f 5HSHDW 6WHS f E\ UHSODFLQJ \QPf ZLWK \QPf \QPfH r1P P f f f 0 ZKHUH ^UfP`AI DUH GHWHUPLQHG LQ 6WHS f 6WHS f 5HSHDW 6WHSV f DQG f XQWLO fSUDFWLFDO FRQYHUJHQFHf ZKLFK LV GHWHUn PLQHG E\ WKH UHODWLYH FKDQJH RI WKH FRVW IXQFWLRQ &P LQ f EHWZHHQ WZR FRQVHFn XWLYH LWHUDWLRQV ,Q WKH QXPHULFDO H[DPSOHV ZH WHUPLQDWH WKH UHSHWLWLRQ RI 6WHSV f DQG f ZKHQ WKH UHODWLYH FKDQJH RI &P LV OHVV WKDQ a EHWZHHQ WZR FRQVHFXWLYH LWHUDWLRQV 1RWH WKDW LI WKH HUURUV LQ ^5Pf`A DUH NQRZQ WR EH QHJOLJLEOH WKHQ 6WHS f DORQH LV VXIILFLHQW IRU WKH DXWRIRFXVLQJ 7DUJHW )HDWXUH ([WUDFWLRQ $VVXPH DSHUWXUH SDUDPHWHU HVWLPDWHV ^P SP P`A DUH JLYHQ 7KHQ WKH SUREOHP EHFRPHV WKH WDUJHW IHDWXUH H[WUDFWLRQ SUREOHP FRQVLGHUHG LQ &KDSWHU $V
PAGE 97
VKRZQ LQ WKH SUHYLRXV FKDSWHU WKH HVWLPDWHV ^DN IN IN IN`N RI ^DIF IN IN IN`N FDQ EH REWDLQHG ZLWK WKH 5(/$; DOJRULWKP E\ HTXLYDOHQWO\ PLQLPL]LQJ WKH IROORZLQJ 1/6 FULWHULRQ 1 0 1 I ^mIH ININ D` M A Pfa PfL f n P Q ZKHUH \QPf \QPfHaLIOP DQG ]QPf KDV WKH VDPH IRUP DV ]QPf LQ f H[FHSW WKDW DQG L=PQf DUH UHSODFHG E\ W;PQf W\PQf DQG L=PQf UHVSHFWLYHO\ ZKLFK DUH GHWHUPLQHG E\ ^P LQVWHDG RI ^P !P`A /HW 0 1O )N DN IN IN rf ;@ 6 _AQfPf N ,I $P :INL\P QfINL]P 0O f P Q ZKHUH W;PQf W\PQf DQG W=PQf DUH WKH VDPH DV W;PQf W\PQf DQG W=PQf UHVSHFn WLYHO\ H[FHSW WKDW ^PMfP`A DUH UHSODFHG E\ ^P$P`P L DQG \NQPf \QPf fÂ§ 4ÂH$fU>Â/PQfÂ/PULfÂ/PQf@A f LfÂ§OLAN 7KHQ PLQLPL]LQJ )N LQ f ZLWK UHVSHFW WR DN \LHOGV WKH HVWLPDWH RWN RI DN IN/PQINL\PQfINL]PQf@ 2LN DQG ^IN K INf DUi PD[ IN ININ 01 0 1O IN ININ ININ IN f A n \ f \NQ PfH IQANL[PQfINL\PQfINL=PQf@ JBf P Q +HUHLQ ZH RPLW WKH GHWDLOHG VWHSV XVHG IRU WKH WDUJHW IHDWXUH H[WUDFWLRQ VLQFH WKH\ DUH WKH VDPH DV WKRVH LQ WKH SUHYLRXV FKDSWHU H[FHSW WKDW \N LQ f LV UHSODFHG E\ ^\NQPf` LQ f 7KH fSUDFWLFDO FRQYHUJHQFHf LQ WKH LWHUDWLRQV RI WKH 5(/$; DOJRULWKP QRZ LV GHWHUPLQHG E\ FKHFNLQJ WKH UHODWLYH FKDQJH RI WKH FRVW IXQFWLRQ ) IN IN IN`N] AM LQ f EHWZHHQ WZR FRQVHFXWLYH LWHUDWLRQV 7R VSHHG XS WKH 5(/$; DOJRULWKP YLD XWLOL]LQJ ))7 D &DUWHVLDQ JULG PDSSLQJ DSSURDFK LV LQWURGXFHG LQ &KDSWHU WR DSSUR[LPDWHO\ PDS ^L;PQf L\UQQf L=UQQf`P L RQ D
PAGE 98
&DUWHVLDQ JULG :KHQ WKH &DUWHVLDQ JULG LV ILQH HQRXJK WKH HUURUV LQWURGXFHG E\ WKH DSSUR[LPDWLRQ DUH QHJOLJLEOH ,Q WKLV FKDSWHU ZH XVH D FRDUVH &DUWHVLDQ JULG DQG WKH ))7 PHWKRG WR REWDLQ DQ LQLWLDO HVWLPDWH RI WKH SDUDPHWHUV DQG WKHQ PLQLPL]H f YLD D PXOWLGLPHQVLRQDO VHDUFK PHWKRG VXFK DV WKH )0,19 IXQFWLRQ LQ 39:$9( 7KH ODWWHU DSSURDFK XVHV OHVV FRPSXWHU PHPRU\ ([SHULPHQWDO DQG 1XPHULFDO 5HVXOWV :H ILUVW SUHVHQW DQ H[SHULPHQWDO H[DPSOH WR VKRZ WKH SHUIRUPDQFH RI WKH $8725(/$; DOJRULWKP 7KH ILHOG GDWD ZDV REWDLQHG E\ WKH 'HSOR\DEOH 6LJQDWXUH 0HDVXUHPHQW 6\VWHP '606f &DUGHURFN 'LYLVLRQ 1DYDO 6XUIDFH :DUIDUH &HQWHU %HWKHVGD 0DU\ODQG 7KH UDGDU ZDV FDUULHG RQ ERDUG RI D KHOLFRSWHU 7KH UDGDU FDUULHU IUHTXHQF\ LV *+] DQG WKH EDQGZLGWK *+] 7KH GDWD VHW ZDV REWDLQHG ZLWK D FXUYHG DSHUWXUH QRW H[DFWO\ NQRZQ EXW LV URXJKO\ WKH VDPH DV WKH GDVK OLQH VKRZQ LQ )LJXUH Df ZKHUH WKHUH DUH ORRN DQJOHV DQG VDPSOHV SHU ORRN DQJOH 7KH UDGDU ZDV DERXW PHWHUV DZD\ IURP WKH JURXQG WDUJHW 7KH JURXQG WDUJHW FRQVLVWV RI FRUQHU UHIOHFWRUV RQ WKH JURXQG SODQH DQG FRUQHU UHIOHFWRUV PRXQWHG RQ D ZRRGHQ WULSRG WKDW LV DERXW PHWHUV WDOO 7KH WUXH GLVWULEXWLRQ RI WKH VFDWWHUHUV LV VKRZQ LQ )LJXUH Ef ZKHUH WKH FHQWHUV RI WKH VTXDUHV GHQRWH WKH ORFDWLRQV RI WKH VFDWWHUHUV LQ VSDFH DQG WKH OHQJWK RI HDFK VTXDUH LV SURSRUWLRQDO WR WKH PRGXOXV RI WKH 5&6 RI WKH FRUUHVSRQGLQJ VFDWWHUHU 7KH WULDQJOHV VKRZ WKH SURMHFWLRQV RI WKH VFDWWHUHU ORFDWLRQV RQWR WKH JURXQG SODQH DQG WKHLU VL]HV DUH DOVR VFDOHG WR EH SURSRUWLRQDO WR WKH 5&6fV RI WKH VFDWWHUHUV )LJXUHV Ff DQG Gf UHVSHFWLYHO\ VKRZ WKH VFDWWHUHU GLVWULEXWLRQ REWDLQHG ZLWK WKH 5(/$; DOJRULWKP IURP WKH LQLWLDO DQG DQ PDQXDOO\ DGMXVWHG DSHUWXUHV VKRZQ LQ )LJXUH Df :H QRWH WKDW DV FRPSDUHG WR )LJXUH Ef WKH UHVXOWV LQ )LJXUH Gf DUH REYLRXVO\ EHWWHU WKDQ WKRVH LQ )LJXUH Ff :H QRZ ILUVW FRQVLGHU DXWRIRFXVLQJ RQO\ LQ WKH HOHYDWLRQ GLUHFWLRQ E\ DVn VXPLQJ WKDW QR HUURUV H[LVW LQ ERWK WKH D]LPXWK DQJOHV ^P`P L DQG WKH GLVWDQFHV
PAGE 99
^5Pf`0 7KH $8725(/$; DOJRULWKP FRQYHUJHV DIWHU LWHUDWLRQV )LJXUH VKRZV WKH WDUJHW VFDWWHUHUV H[WUDFWHG E\ $8725(/$; ZLWK WKH VHDUFK LQWHUYDO IRU ^$P`AI EHLQJ sr &RPSDULQJ )LJXUHV Ef DQG Gf ZLWK )LJXUH Ef ZH ILQG WKDW $8725(/$; ZRUNV ZHOO DQG WKH $8725(/$; UHVXOWV DUH VOLJKWO\ EHWWHU WKDQ WKRVH REWDLQHG E\ XVLQJ WKH PDQXDOO\ DGMXVWHG DSHUWXUH )LJXUH VKRZV WKDW WKH PDQXDOO\ DGMXVWHG DQG DXWRIRFXVHG DSHUWXUHV ILW TXLWH ZHOO DIWHU DGGLQJ D OLQH WR WKH IRUPHU 6LQFH WKH UHVXOWV LQ )LJXUH Ef DUH EHWWHU WKDQ WKRVH LQ )LJXUH Gf LW DSSHDUV WKDW D OLQHDU SKDVH HUURU LQ MfP DV D IXQFWLRQ RI Pf ZDV LQWURGXFHG LQ WKH PDQXDOO\ DGMXVWHG DSHUWXUH 7KH OLQHDU SKDVH HUURU FDQ FDXVH VKLIWV LQ r DQG WKH DPRXQW RI VKLIW LV GLIIHUHQW IRU GLIIHUHQW VFDWWHUHU N DV FDQ EH VHHQ IURP f &RQVLGHU QH[W DXWRIRFXVLQJ LQ ERWK WKH HOHYDWLRQ DQG D]LPXWK GLUHFWLRQV E\ DVn VXPLQJ WKDW QR HUURUV H[LVW LQ ^5Pf`A 7KH $8725(/$; DOJRULWKP FRQYHUJHV DIWHU LWHUDWLRQV )LJXUH VKRZV WKH UHVXOWV REWDLQHG ZLWK WKH VHDUFK LQWHUYDOV IRU ^$P`eI DQG ^$P`A EHLQJ sr DQG sr UHVSHFWLYHO\ &RPSDUHG ZLWK WKH UHVXOWV REWDLQHG E\ DXWRIRFXVLQJ RQO\ LQ WKH HOHYDWLRQ GLUHFWLRQ ZH ILQG WKDW DXWn RIRFXVLQJ LQ ERWK GLUHFWLRQV SURYLGHV OLWWOH IXUWKHU LPSURYHPHQW WR WKH DFFXUDF\ RI WKH HVWLPDWHG WDUJHW SDUDPHWHUV LQ WKLV H[DPSOH +HQFH IRU WKLV H[DPSOH LW DSSHDUV WKDW WKH DSHUWXUH HUURUV PDLQO\ RFFXU LQ WKH HOHYDWLRQ GLUHFWLRQ )RU WKLV H[SHULPHQWDO H[DPSOH ZH KDYH DOVR XVHG $8725(/$; WR H[WUDFW WDUJHW IHDWXUHV ZKHQ ZH DVVXPHG WKDW HUURUV H[LVW LQ ERWK DQG ^5RPf`0 $JDLQ DV FRPSDUHG ZLWK )LJXUH Ef ZH KDYH QRWLFHG OLWWOH FKDQJH LQ WKH H[WUDFWHG WDUJHW SDUDPHWHUV ,W FRXOG EH WKDW '606 KDV DOUHDG\ GRQH D JRRG MRE FRPSHQVDWLQJ IRU WKH HUURUV LQ ^5RPf`A O ZLWK VRPH WUDGLWLRQDO PHWKRG VR WKDW WKH\ DUH QRZ QHJOLJLEOH DV FRPSDUHG ZLWK WKH HUURUV LQ ^MfP`P L )LQDOO\ ZH XVH D VLPXODWLRQ H[DPSOH WR VKRZ WKDW WKH DFFXUDF\ RI WDUJHW IHDWXUH H[WUDFWLRQ YLD &/6$5 LV YHU\ VHQVLWLYH WR WKH DFFXUDF\ RI WKH FXUYLOLQHDU DSHUWXUH :H DVVXPH WKDW WKHUH DUH VFDWWHUHUV ZLWK D VLPLODU GLVWULEXWLRQ WR WKDW
PAGE 100
(VWLPDWLRQ (UURU 0DJQLWXGH 5DQJH f &URVVUDQJH f +HLJKW f $8725(/$; 5(/$; 7DEOH (VWLPDWLRQ HUURUV IRU WKH SDUDPHWHUV RI WKH KLJKHVW VFDWWHUHU LQ WKH VLPXn ODWLRQ H[DPSOH EHIRUH DQG DIWHU DXWRIRFXVLQJ LQ WKH H[SHULPHQWDO H[DPSOH DERYH 7KH WUXH DSHUWXUH LV WKH GRW OLQH VKRZQ LQ )LJXUH Df DQG WKH VFDWWHUHU GLVWULEXWLRQ LV VKRZQ LQ )LJXUH Ef +HUH ZH VLPXODWH WKH FDVH ZKHUH WKH DSHUWXUH HUURUV H[LVW RQO\ LQ WKH HOHYDWLRQ GLUHFWLRQ )LJXUH Ff VKRZV WKH VFDWWHUHU GLVWULEXWLRQ REWDLQHG ZLWK WKH 5(/$; DOJRULWKP IURP WKH LQLWLDO DSHUWXUH GDVK OLQHf VKRZQ LQ )LJXUH Df :H QRWH WKDW WKH VFDWWHUHU GLVWULEXWLRQ LQ )LJXUH Ff LV TXLWH GLIIHUHQW IURP WKH WUXH RQH LQ )LJXUH Ef :KHQ XVHG ZLWK WKH VLPXODWLRQ GDWD $8725(/$; FRQYHUJHV DIWHU IRXU LWHUDWLRQV 7KH DXWR IRFXVHG DSHUWXUH LV WKH GRW OLQH VKRZQ LQ )LJXUH Df DQG WKH VFDWWHUHU GLVWULEXWLRQ REWDLQHG ZLWK $8725(/$; LV LOOXVWUDWHG LQ )LJXUH Gf :H VHH WKDW WKH VFDWn WHUHU GLVWULEXWLRQ REWDLQHG E\ $8725(/$; LV DOPRVW WKH VDPH DV WKH WUXH RQH :H DOVR QRWLFH WKDW WKH VKDSH RI WKH DXWRIRFXVHG DSHUWXUH LV FORVHU WR WKH WUXH RQH WKDQ WKH LQLWLDO DSHUWXUH 7KH FRQVWDQW GLIIHUHQFH EHWZHHQ WKH DXWRIRFXVHG DQG WKH WUXH DSHUWXUHV ZLOO FDXVH SKDVH HUURUV LQ ^FÂN`N Q ZKLFK FDQQRW EH HOLPLQDWHG GXH WR WKH DPELJXLW\ SUREOHPV GLVFXVVHG LQ 6HFWLRQ 7R TXDQWLWDWLYHO\ LOOXVWUDWH WKH DFFXUDF\ LPSURYHPHQW RI WKH SDUDPHWHU HVWLPDWHV YLD $8725(/$; ZH FRPSDUH WKH HVWLn PDWLRQ HUURUV RI WKH KLJKHVW VFDWWHUHU LQ )LJXUH Ef EHIRUH DQG DIWHU DXWRIRFXVLQJ 7DEOH VKRZV WKH GLIIHUHQFHV EHWZHHQ WKH WUXH YDOXHV DQG WKHLU HVWLPDWHV REWDLQHG YLD 5(/$; DQG $8725(/$; UHVSHFWLYHO\
PAGE 101
'HYRWLRQ PHWHUVnf )LJXUH Df &XUYLOLQHDU DSHUWXUHV IRU WKH H[SHULPHQWDO H[DPSOH Ef 7UXH VFDWWHUHU GLVWULEXWLRQ Ff SORW RI VFDWWHUHUV H[WUDFWHG IURP WKH LQLWLDO DSHUWXUH GDVKHG OLQHf LQ Df ZLWK 5(/$; Gf SORW RI VFDWWHUHUV H[WUDFWHG IURP WKH PDQXDOO\ DGMXVWHG DSHUWXUH VROLG OLQHf LQ Df ZLWK 5(/$;
PAGE 102
(OHYDWLRQ $QJOH GHJUHHVf (OHYDWLRQ $QJOH GHJUHHVf )LJXUH $XWRIRFXVHG FXUYLOLQHDU DSHUWXUH DQG VFDWWHUHU GLVWULEXWLRQ REWDLQHG ZLWK $8725(/$; E\ DXWRIRFXVLQJ RQO\ LQ WKH HOHYDWLRQ GLUHFWLRQ DQG XVLQJ IRU WKH H[SHULPHQWDO H[DPSOH Df $XWRIRFXVHG FXUYLOLQHDU DSHUWXUH GRWWHG OLQHf Ef 6FDWWHUHU GLVWULEXWLRQ $]LPXWK $QJOH GHJUHHVf $]LPXWK $QJOH GHJUHHVf Df Ef )LJXUH Df 0DQXDOO\ DGMXVWHG VROLG OLQHf DQG DXWRIRFXVHG GRWWHG OLQHf FXUYLn OLQHDU DSHUWXUHV IRU WKH H[SHULPHQWDO H[DPSOH Ef )LWWLQJ WKH PDQXDOO\ DGMXVWHG DSHUWXUH WR WKH DXWRIRFXVHG DSHUWXUH E\ DGGLQJ D OLQH WR WKH IRUPHU
PAGE 103
LQ &' 7 U nfÂ‘ Â‘Â‘Â‘Â‘Â‘Â‘ Â‘ Â‘ Â‘ L GHJUH M Z ? V f ? ? f ? FQ F B ? ? fÂ§ ,QLWLDO DSHUWXUH ? 61 F R B $SHUWXUH REWDLQHG E\ $8725(/$; DIWHU LWHUDWLRQV rr B ? ? R ? /OL/ LL LL LO L Â‘ L Â‘ L L L Â‘ L L L L Â‘ Â‘ Â‘ L $]LPXWK $QJOH GHJUHHVf Df Y LXHUrWLRQV )LJXUH $XWRIRFXVHG FXUYLOLQHDU DSHUWXUH DQG VFDWWHUHU GLVWULEXWLRQ REWDLQHG ZLWK $8725(/$; E\ DXWRIRFXVLQJ LQ ERWK WKH HOHYDWLRQ DQG D]LPXWK GLUHFWLRQV DQG XVLQJ IRU WKH H[SHULPHQWDO H[DPSOH Df $XWRIRFXVHG FXUYLOLQHDU DSHUWXUH GRWWHG OLQHf Ef 6FDWWHUHU GLVWULEXWLRQ
PAGE 104
(OHYDWLRQ $QJOH GHJUHHVf f LcWUD!LrQr Ff Gf )LJXUH Df &XUYLOLQHDU DSHUWXUH IRU WKH VLPXODWLRQ H[DPSOH Ef 7UXH VFDWWHUHU GLVWULEXWLRQ IRU WKH VLPXODWLRQ H[DPSOH Ff 6FDWWHUHU GLVWULEXWLRQ REWDLQHG IURP WKH LQLWLDO DSHUWXUH GDVKHG OLQHf LQ Df E\ XVLQJ 5(/$; ZLWK Gf 6FDWWHUHU GLVn WULEXWLRQ REWDLQHG YLD $825(/$; ZLWK fÂ§ DQG XVLQJ WKH DXWRIRFXVHG DSHUWXUH GRWWHG OLQHf VKRZQ LQ Df
PAGE 105
&+$37(5 &21&/86,21 6XPPDU\ ,Q WKLV VWXG\ ZH KDYH H[WHQGHG WZR SDUDPHWULF UHOD[DWLRQEDVHG DOJRULWKPV LQFOXGLQJ 5(/$; DQG 5(/$;1/6 DQG SUHVHQWHG D QRYHO VHPLSDUDPHWULF DSn SURDFK 63$5 WR IRUP HQKDQFHG 6$5 LPDJHV LQ GLIIHUHQW VFHQDULRV $Q HIIHFWLYH UHOD[DWLRQEDVHG SDUDPHWULF PHWKRG UHIHUUHG WR DV $8725(/$; KDV DOVR EHHQ SURn SRVHG IRU ERWK WDUJHW IHDWXUH H[WUDFWLRQ DQG PRWLRQ FRPSHQVDWLRQ YLD &/6$5 7KH 5(/$; DQG 5(/$;1/6 DOJRULWKPV DUH SDUDPHWULF VSHFWUDO HVWLPDn WLRQ PHWKRGV DQG SURYLGH WKH SRWHQWLDO IRU VLJQLILFDQW UHVROXWLRQ LPSURYHPHQW RI WKH IRUPHG 6$5 LPDJHV SURYLGHG WKDW WKH DVVXPHG GDWD PRGHO LV YDOLG 5(/$; LV EDVHG RQ WKH SRLQW VFDWWHUHU WULKHGUDOf GDWD PRGHO DQG 5(/$;1/6 LV GHYLVHG DFFRUGLQJ WR D PL[HG GLKHGUDO DQG WULKHGUDO GDWD PRGHO ZKLFK GHVFULEHV D GLKHGUDO FRUQHU UHn IOHFWRU DV D FRPSOH[ VLQXVRLG ZLWK DPSOLWXGH GHVFULEHG DV D VLQH IXQFWLRQ DQG D FRQVWDQW SKDVH LQ FURVVUDQJH DQG ZLWK FRQVWDQW DPSOLWXGH DQG SKDVH LQ UDQJH 6XSHU UHVROXWLRQ 6$5 LPDJHV FDQ EH JHQHUDWHG E\ ILUVW H[WUDFWLQJ WKH WDUJHW IHDWXUHV YLD 5(/$; RU 5(/$;1/6 EDVHG RQ WKH FRUUHVSRQGLQJ GDWD PRGHOV DQG WKHQ DSn SO\LQJ ))7 WR WKH V\QWKHVL]HG SKDVH KLVWRU\ GDWD ZLWK ODUJHU GLPHQVLRQV LQ ERWK UDQJH DQG FURVVUDQJH E\ XVLQJ WKH H[WUDFWHG WDUJHW IHDWXUHV ([SHULPHQWDO H[DPSOHV ZLWK (5,0 DQG 067$5 GDWD LQ &KDSWHU KDYH VKRZQ WKDW 5(/$; DQG 5(/$; 1/6 FDQ EH XVHG WR REWDLQ HLWKHU KLJKHU UHVROXWLRQ LPDJHV IURP WKH VDPH VL]H GDWD RU HTXLYDOHQWO\ HTXDOTXDOLW\ LPDJHV IURP VLJQLILFDQWO\ IHZHU QXPEHU RI GDWD VDPn SOHV 2I WKH WZR PHWKRGV 5(/$; LV PRUH VXLWDEOH IRU SRLQW VFDWWHUHU WDUJHWV DQG
PAGE 106
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n SOH[ VLQXVRLGV ZLWK DUELWUDU\ DPSOLWXGHV DQG FRQVWDQW SKDVHV LQ FURVVUDQJH DQG ZLWK FRQVWDQW DPSOLWXGHV DQG SKDVHV LQ UDQJH 7KH 6$5 LPDJHV ZLWK VXSHU UHVROXWLRQ LQ UDQJH REWDLQHG YLD 63$5 FDQ EH DFKLHYHG E\ DSSO\LQJ ))7 WR WKH SKDVH KLVWRU\ GDWD ZLWK D ODUJH GLPHQVLRQ LQ UDQJH V\QWKHVL]HG E\ XVLQJ WKH H[WUDFWHG IHDWXUHV DQG WKH SDUDPHWULF GDWD PRGHO LQ UDQJH 7KH FURVVUDQJH UHVROXWLRQ RI 6$5 LPDJHV FDQQRW EH VLJQLILFDQWO\ LPSURYHG YLD 63$5 RYHU WKH FRQYHQWLRQDO ))7 PHWKRG GXH WR WKH QRQSDUDPHWULF GDWD PRGHO LQ FURVVUDQJH
PAGE 107
IRU WDUJHW IHDWXUH H[WUDFWLRQ YLD &/6$5 DQG PRWLRQ FRPSHQVDWLRQ DXWRIRFXVf IRU &/6$5 LQ &KDSWHUV DQG UHVSHFWLYHO\ $ &/6$5 GDWD PRGHO KDV EHHQ ILUVW HVWDEOLVKHG DQG WKH DPELJXLW\ SUREOHPV LQKHUHQW LQ WKH GDWD PRGHO KDYH EHHQ LQn YHVWLJDWHG ZKHQ DSHUWXUH HUURUV DUH SUHVHQW 7R SURYLGH D JXLGH IRU WKH FKRLFH RI &/6$5 WUDMHFWRU\ WKH SHUIRUPDQFH RI GLIIHUHQW FXUYLOLQHDU DSHUWXUHV LV DQDO\]HG YLD &UDPU5DR %RXQG ZKLFK LV WKH ORZHVW ERXQG DFKLHYDEOH E\ DQ\ XQELDVHG HVWLPDn WRUV $XWRIRFXVLQJ IRU &/6$5 EHFRPHV PRUH FRPSOLFDWHG VLQFH WKH DSHUWXUH HUURUV LQ &/6$5 XQOLNH LQ WKH 6$5 V\VWHPV FDQ QR ORQJHU EH DSSUR[LPDWHG DV SKDVH HUURUV RQO\ LQ FURVVUDQJH GXH WR WKH QRQQHJOLJLEOH HUURUV LQWURGXFHG LQ WKH D]LPXWK DQG HOHYDWLRQ DQJOHV RI WKH WUDMHFWRU\ :H KDYH SUHVHQWHG D UHOD[DWLRQEDVHG DXWRIRn FXV DOJRULWKP UHIHUUHG WR DV $8725(/$; IRU WDUJHW IHDWXUH H[WUDFWLRQ XVLQJ &/6$5 LQ WKH SUHVHQFH RI DSHUWXUH HUURUV $8725(/$; FDQ EH XVHG WR VLPXOWDn QHRXVO\ H[WUDFWV WDUJHW IHDWXUHV DQG FRPSHQVDWHV PRWLRQ HUURUV LQFOXGLQJ SKDVH HUURUV DQG DSHUWXUH HUURUV E\ PLQLPL]LQJ DQ 1/6 FRVW IXQFWLRQ YLD DQ DOWHUQDWLQJ RSWLPL]DWLRQ DSSURDFK ([SHULPHQWDO H[DPSOHV KDYH EHHQ XVHG WR GHPRQVWUDWH WKH HIIHFWLYHQHVV RI $8725(/$; IRU WDUJHW IHDWXUH H[WUDFWLRQ XVLQJ &/6$5 GDWD LQ WKH SUHVHQFH RI FXUYLOLQHDU DSHUWXUH HUURUV )XWXUH :RUNV ,Q WKLV VWXG\ ZH KDYH FRQFHQWUDWHG RQ WKH VWXG\ RI 6$5 WDUJHW IHDWXUH H[WUDFn WLRQ VXSHU UHVROXWLRQ LPDJH IRUPDWLRQ DQG DXWRIRFXV YLD UHOD[DWLRQEDVHG SDUDPHWn ULF DQG VHPLSDUDPHWULF PHWKRGV +RZHYHU WKH SUREOHPV LQYROYHG LQ LQYHUVH 6$5 ,6$5f EHFRPH PRUH FRPSOLFDWHG DQG GLIILFXOW &RPSDUHG ZLWK WKH VWDWLRQDU\ WDUJHWV RI 6$5 WDUJHWV EHLQJ LPDJHG LQ ,6$5 DUH RIWHQ HQJDJHG LQ FRPSOLFDWHG PDQHXYHUV WKDW FRPELQH WUDQVODWLRQDO DQG URWDWLRQDO PRWLRQV DQG IXUWKHUPRUH WKH PRWLRQV EHn WZHHQ WKH UDGDU DQG WKH PRYLQJ WDUJHWV LQ ,6$5 DUH QRQFRRSHUDWLYH +HQFH WKH
PAGE 108
PRYLQJ WDUJHW LPDJHV REWDLQHG YLD FRQYHQWLRQDO 6$5 SURFHVVLQJ DOJRULWKPV DUH W\SLn FDOO\ PLVORFDWHG DQG VPHDUHG LQ FURVVUDQJH GXH WR SKDVH HUURUV DQG UDQJH PLJUDWLRQV LQGXFHG E\ WKH PRWLRQV 7KH DPRXQW RI VPHDULQJ JHQHUDOO\ GHSHQGV RQ WKH FKDUDFn WHU DQG PDJQLWXGH RI WKH WDUJHW PRWLRQ GXULQJ WKH GDWD FRKHUHQW FROOHFWLRQ LQWHUYDO &7,f 7KH IXWXUH ZRUN WR EH FRQGXFWHG LQFOXGHV WKH HVWDEOLVKPHQW RI DQ HIIHFWLYH ,6$5 GDWD PRGHO FRQVLGHULQJ UDQJH PLJUDWLRQV DQG SKDVH HUURUV DQG KRZ WR DSSO\ WKH UHOD[DWLRQEDVHG RSWLPL]DWLRQ DSSURDFK WR VLPXOWDQHRXV ,6$5 DXWRIRFXVLQJ DQG LPDJH IRUPDWLRQ
PAGE 109
$33(1',; $ $ 6,1*/( 6&$77(5(5 )($785( (;75$&7,21 9,$ 63$5 7KH HVWLPDWHV RI ^! [ ` FDQ EH REWDLQHG E\ PLQLPL]LQJ WKH IROORZLQJ 1/6 FRVW IXQFWLRQ e &LI[f __ Hr\fXVf@ $f ,QVHUWLQJ $f LQWR $f ZH REWDLQ WKH 1/6 HVWLPDWHV ^I!` RI ^` E\ HTXLYDOHQWO\ PD[LPL]LQJ WKH IROORZLQJ FRVW IXQFWLRQ Â“ Q ? [ ^_\!mfO 5r >\IfAff H>n!r@` f $f
PAGE 110
/HW \V^Q If GHQRWH WKH QWK HOHPHQW RI \Vf Q f f Â‘ 1 fÂ§ 7KHQ FMf LV JLYHQ E\ $f ,QVHUWLQJ $f LQWR $f DQG LJQRULQJ WKH VFDOLQJ IDFWRU ZH FDQ VLPSOLI\ $f WR L MDUJMA >\!0f@ 1 &0f 1fÂ§O ( >\O1If@ Hr1n} OO\VfLU 1 MLUIfQ Q $f 7KHQ WKH 1/6 HVWLPDWHV ^ ` RI ^ ` DUH GHWHUPLQHG E\ ^ ` DUJPD[ & f $f 1RWH WKDW \Vf LQ f FDQ EH REWDLQHG E\ DSSO\LQJ ))7 WR HDFK FROXPQ RI
PAGE 111
$33(1',; % $3352;,0$7,21 2) 7+( 5$1*( 5 7KH UDQJH 5 LQ f FDQ EH DSSUR[LPDWHG DV IROORZV 8QGHU WKH FRQGLWLRQV Â‘N A 5IW On DQG aN ZH KDYH 5 5T 5T L L a? [ \ ] [ r \ r aU ] r fÂ§ fÂ§ FRV FRV IW fÂ§ A VLQ FRV IW fÂ§ fÂ§ VLQ IW + fÂ§a 5R 5T 5T 5IW L [ D M \ Â‘ D M ] [ \ ] fÂ§ fÂ§ FRV FRV IW fÂ§ fÂ§ VLQ FRV IW fÂ§ fÂ§ VLQ IW G fÂ§fÂ§ L"R 1R W M5T ^[ FRV FRV IW \ VLQ FRV IW ] VLQ IWfn i %Of /HW LS IW fÂ§ R ZKHUH LV WKH DYHUDJH RI DOO IW XVHG WR IRUP WKH V\QWKHWLF DSHUWXUH )RU YHU\ VPDOO LS ZH KDYH FRV IW ID FRV IWR fÂ§ VLQ IW4fLS %f VLQ IW ID VLQ IWR FRV IWRfLS %f )RU YHU\ VPDOO ZH KDYH FRV ID VLQ a 7KHQ NHHSLQJ ILUVWRUGHU WHUPV \LHOGV [ VLQ IW [ VLQ FRV IW \ VLQ IW \ FRV FRV IW ] FRV IW ID [ VLQ IW>VP IW FRV IWfLS? \ VLQ IW>VLQ IW FRV IWRfLS@ \ FRV >FRV IW fÂ§ VLQ IWfLS@ ] FRV >FRV IW fÂ§ VLQ IWfLS@ ID >[ \f VLQ IW@ VLQ IW >\ ]f FRV IW? FRV FRV IW >[ fÂ§ ]f VLQRFRVIWR?LS %f :H DOVR KDYH [\ FRV VLQ FRV IW \] VLQ FRV VLQ S [] FRV FRV IW VLQ IW
PAGE 112
m [\ VLQ FRV c!?FRV 2 VLQ 2f@ \] VLQ FRV >VLQ 2 FRV 2fLS? [] FRV FRV >VLQ 2 FRV 2fLS@ LQ FRV 2 \] VLQ Rf VLQ FRV [] VLQ 2f FRV FRV D]FRV Rf +HQFH 5 LV VLPSOLILHG WR f %f
PAGE 113
$33(1',; & &$/&8/$7,1* [\]f )520 [\]f *LYHQ 5 2 DQG [\]f [\]f FDQ EH GHWHUPLQHG IURP f f DQG f DV IROORZV YOY 9L5R >YL O Utf 5 W@]f "L5R UMc[@ \H 5 >Ub O f \ L f Â 5"""" UIHf ] A O UIHf [] 5 "L f [ 5OUM> f@ \ A A Uf?[f \\ L"R > f 5R"@ ULOf WAL@ WW @O5\\ &Of ] ^YLYK 5R >5R9L 9f =YLU@]O YOf@ \ \\[5O\ 5AAA\` >L"A \Ycf\? &f DQG ? ^Y": 5R"L >LÂ f 5R Âf@ \ UM5\\ 5Rb\O Af` >5RWL6 9Y'Y@ &f ZKHUH FRVR DQG VLQ2 1RWH WKDW WKH OHIW VLGH RI &Of LV DQ HLJKWKRUGHU SRO\QRPLDO DQG KHQFH KDV HLJKW ]HURV :H SLFN WKH URRW WKDW LV FORVHVW WR \ DV WKH VROXWLRQ IRU \ :H QRZ SUHVHQW WZR H[DPSOHV )LUVW OHW 5T W DQG [\]f f :H REWDLQ [\]f f :H UHPDUN WKDW WKH RWKHU VHYHQ SRVVLEOH VROXWLRQV IRU \ DUH HLWKHU FRPSOH[ RU IDU DZD\ IURP \f 1H[W OHW 5 fÂ§ 2 W DQG [\]f f :H REWDLQ
PAGE 114
[\]f f 1RWH WKDW LQ ERWK H[DPSOHV [\]f DUH YHU\ FORVH WR [\]f DQG KHQFH FDQ EH XVHG WR DSSUR[LPDWH [\]f
PAGE 115
5()(5(1&(6 >@ $QGHUVK 0 +D]OHWW 6 : /HH ' 5HHYHV 3 6XOOLYDQ DQG < &KX ;3$7&+ D KLJKIUHTXHQF\ HOHFWURPDJQHWLF VFDWWHULQJ SUHGLFWLRQ FRGH DQG HQn YLURQPHQW IRU FRPSOH[ WKUHHGLPHQVLRQDO REMHFWV ,((( $QWHQQDV DQG 3URSDn JDWLRQ 0DJD]LQH ffÂ§ )HEUXDU\ >@ $ $XVKHUPDQ $ .R]PD / :DONHU + 0 RQHV DQG ( & 3RJJLR 'HYHORSPHQWV LQ UDGDU LPDJLQJ ,((( 7UDQVDFWLRQV RQ $HURVSDFH DQG (OHFWURQLF 6\VWHPV f XO\ >@ : %DQJV $UUD\ SURFHVVLQJ ZLWK JHQHUDOL]HG EHDPIRUPHUV 3K' GLVVHUWDWLRQ @ 6HUJLR %DUEDURVVD /XLJL 0DUVLOL DQG *DHWDQR 0XQJDUL 6$5 VXSHUUHVROXWLRQ LPDJLQJ E\ VLJQDO VXEVSDFH SURMHFWLRQ WHFKQLTXHV $(8 ,QWHUQDWLRQDO RXUQDO RI (OHFWURQLFV DQG &RPPXQLFDWLRQV f fÂ§ 0DUFK >@ 5 %HQLW] $GDSWLYH KLJKGHILQLWLRQ LPDJLQJ 63,( 3URFHHGLQJV RQ 2SWLFDO (QJLQHHULQJ LQ $HURVSDFH 6HQVLQJ 2UODQGR )/ SDJHV $SULO >@ 5 %HQLW] +LJK GHILQLWLRQ YHFWRU LPDJLQJ IRU V\QWKHWLF DSHUWXUH UDGDU 3URn FHHGLQJV RI WKH VW $VLORPDU &RQIHUHQFH RQ 6LJQDOV 6\VWHPV DQG &RPSXWHUV 3DFLILF *URYH &$ 1RYHPEHU >@ =KDRTLDQJ %L LDQ /L DQG =KHQJ6KH /LX 6XSHU UHVROXWLRQ 6$5 LPDJH IRUn PDWLRQ YLD SDUDPHWULF VSHFWUDO HVWLPDWLRQ PHWKRGV ,((( 7UDQVDFWLRQV RQ $HURVSDFH DQG (OHFWURQLF 6\VWHPV f fÂ§ DQXDU\ >@ % %XQGD\ %DVLF 2SWLPL]DWLRQ 0HWKRGV (GZDUG $UQROG /WG /RQGRQ >@ 3 %XUJ 0D[LPXP HQWURS\ VSHFWUDO DQDO\VLV 3URFHHGLQJV RI WKH WK 0HHWLQJ 6RFLHW\ RI ([SORUDWLRQ *HRSK\VLFLVWV 2NODKRPD &LW\ 2. 2FW >@ &DSRQ +LJK UHVROXWLRQ IUHTXHQF\ZDYHQXPEHU VSHFWUXP DQDO\VLV 3URFHHGLQJV RI WKH ,((( $XJXVW >@ :DOWHU &DUUDUD 5RQ 6 *RRGPDQ DQG 5RQDOG 0 0DMHZVNL 6SRWOLJKW 6\Qn WKHWLF $SHUWXUH 5DGDU 6LJQDO 3URFHVVLQJ $OJRULWKPV $UWHFK +RXVH ,QF 1RUn ZRRG 0$ >@ &ODUNH 6XSHUYLVHG ,QWHUSUHWDWLRQ RI 6DPSOHG 'DWD 8VLQJ (IILFLHQW ,PSOHn PHQWDWLRQV RI +LJKHU5DQN 6SHFWUXP (VWLPDWLRQ 9RO ,, 3UHQWLFH +DOO (QJOHn ZRRG &OLII 1>@ 6 5 'H*UDDI 6$5 LPDJLQJ YLD PRGHUQ VSHFWUDO HVWLPDWLRQ PHWKRGV 63,( 3URFHHGLQJV RQ 2SWLFDO (QJLQHHULQJ LQ $HURVSDFH 6HQVLQJ 2UODQGR )/ $SULO
PAGE 116
>@ 6 5 'H*UDDI 6LGHOREH UHGXFWLRQ YLD DGDSWLYH ),5 ILOWHULQJ LQ 6$5 LPDJHU\ ,((( 7UDQVDFWLRQV RQ ,PDJH 3URFHVVLQJ f fÂ§ 0D\ >@ 6 5 'H*UDDI 6$5 LPDJLQJ YLD PRGHUQ VSHFWUDO HVWLPDWLRQ PHWKRGV ,((( 7UDQVDFWLRQV RQ ,PDJH 3URFHVVLQJ f 0D\ >@ ( 'XGJHRQ DQG 5 7 /DFRVV $Q RYHUYLHZ RI DXWRPDWLF WDUJHW UHFRJQLWLRQ 7KH /LQFROQ /DERUDWRU\ RXUQDO ffÂ§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f DQXDU\ >@ 0 7 )HQQHOO DQG 5 3 :LVKQHU %DWWOHILHOG DZDUHQHVV YLD V\QHUJLVWLF 6$5 DQG 07, H[SORLWDWLRQ ,((( $HURVSDFH DQG (OHFWURQLF 6\VWHPV 0DJD]LQH f )HEUXDU\ >@ + *ROXE DQG & ) 9DQ /RDQ 0DWUL[ &RPSXWDWLRQV RKQV +RSNLQV 8QLn YHUVLW\ 3UHVV %DOWLPRUH 0' >@ 3 7 *RXJK $ IDVW VSHFWUDO HVWLPDWLRQ DOJRULWKP EDVHG RQ WKH ))7 ,((( 7UDQVDFWLRQV RQ 6LJQDO 3URFHVVLQJ XQH >@ / & *UDKDP 6\QWKHWLF LQWHUIHURPHWHU UDGDU IRU WRSRJUDSKLF PDSSLQJ 3URn FHHGLQJV RI WKH ,((( XQH >@ 5 2 +DUJHU 6\QWKHWLF $SHUWXUH 5DGDU 6\VWHP 7KHRU\ DQG 'HVLJQ &KDSWHU 9, f3KDVH DQG 0RWLRQ (UURUVf $FDGHPLF 3UHVV ,QF 1HZ @ 5 +HLGEUHGHU 0D[LPXP HQWURS\ PHWKRGV LQ FRKHUHQW UDGDU LPDJLQJ ,Q WHUQDWLRQO RXUQDO RI ,PDJLQJ 6\VWHPV DQG 7HFKQRORJ\ >@ 1 +HOG DQG 2f%ULHQ 1RUGHQ V\VWHPVf GLPHQVLRQDO LQWHUIHURPHWULF V\QWKHWLF DSHUWXUH UDGDU IO\LQJ WHVWEHG 3URFHHGLQJV RI WKH ,((( ,QWHUn QDWLRQDO *HRVFLHQFH DQG 5HPRWH 6HQVLQJ 6\QRLVXP ,*$566f +RXVWRQ 7; 0D\ >@ $ +EJERP $SHUWXUH V\QWKHVLV ZLWK D QRQUHJXODU GLVWULEXWLRQ RI LQWHUIHUn RPHWHU EDVHOLQHV $VWURQRP\ DQG $VWURSK\VLFV 6XSSOHPHQWV >@ & 9 DNRZDW] U ( :DKO 3 + (LFKHO & *KLJOLD DQG 3 $ 7KRPSn VRQ 6SRWOLJKW0RGH 6\QWKHWLF $SHUWXUH 5DGDU $ 6LJQDO 3URFHVVLQJ $SSURDFK .OXZHU $FDGHPLF 3XEOLVKHUV 1RUZHOO 0$
PAGE 117
>@ 6 0 .D\ 0RGHUQ 6SHFWUDO (VWLPDWLRQ 7KHRU\ DQG $SSOLFDWLRQ 3UHQWLFH+DOO (QJOHZRRG &OLIIV 1>@ 6 0 .D\ DQG 6 / 0DUSOH U 6SHFWUXP DQDO\VLV D PRGHUQ SHUVSHFWLYH 3URFHHGLQJV RI ,((( f fÂ§ 1RYHPEHU >@ .ODXGHU U $& 3ULFH 6 'DUOLQJWRQ DQG : $OEHUVKHLP 7KH WKHRU\ DQG GHVLJQ RI FKLUS UDGDUV 7KH %HOO 6\VWHP 7HFKQLFDO RXUQDO f XO\ >@ .HQQHWK .QDHOO 7KUHHGLPHQVLRQDO 6$5 IURP FXUYLOLQHDU DSHUWXUHV 63,( 3URn FHHGLQJV RQ 2SWLFDO (QJLQHHULQJ LQ $HURVSDFH 6HQVLQJ 2UODQGR )/ $SULO >@ .HQQHWK .QDHOO 7KUHHGLPHQVLRQDO 6$5 IURP SUDFWLFDO DSHUWXUHV 63,( 3URn FHHGLQJV RQ 2SWLFDO (QJLQHHULQJ LQ $HURVSDFH 6HQVLQJ 6DQ 'LHJR >@ .HQQHWK .QDHOO 7KUHHGLPHQVLRQDO 6$5 IURP FXUYLOLQHDU DSHUWXUHV 3URFHHGLQJV RI WKH ,((( 1DWLRQDO 5DGDU &RQIHUHQFH SDJHV $QQ $UERU 0, 0D\ >@ 9 /DUVRQ DQG / 1RYDN 3RODULPHWULF VXEVSDFH WDUJHW GHWHFWRU IRU 6$5 GDWD EDVHG RQ WKH +X\QHQ GLKHGUDO PRGHO 3URFHHGLQJV RI 63,( 2Un ODQGR )ORULGD $SULO >@ +RQJELQ /L LDQ /L DQG 3HWUH 6WRLFD 3HUIRUPDQFH DQDO\VLV RI IRUZDUGEDFNZDUG PDWFKHGILOWHUEDQN VSHFWUDO HVWLPDWRUV ,((( 7UDQVDFWLRQV RQ 6LJQDO 3URFHVVLQJ f XO\ >@ /L DQG 3 6WRLFD $Q DGDSWLYH ILOWHULQJ DSSURDFK WR VSHFWUDO HVWLPDWLRQ DQG 6$5 LPDJLQJ ,((( 7UDQVDFWLRQV RQ 6LJQDO 3URFHVVLQJ f XQH >@ /L DQG 3 6WRLFD (IILFLHQW PL[HGVSHFWUXP HVWLPDWLRQ ZLWK DSSOLFDWLRQV WR WDUJHW IHDWXUH H[WUDFWLRQ ,((( 7UDQVDFWLRQV RQ 6LJQDO 3URFHVVLQJ )HEUXDU\ >@ LDQ /L =KHQJ6KH /LX DQG 3HWUH 6WRLFD WDUJHW IHDWXUH H[WUDFWLRQ YLD LQWHUIHURPHWULF 6$5 ,(( 3URFHHGLQJV RQ 5DGDU 6RQDU DQG 1DYLJDWLRQ f $SULO >@ =KHQJ6KH /LX +RQJELQ /L DQG LDQ /L (IILFLHQW LPSOHPHQWDWLRQ RI &DSRQ DQG $3(6 IRU VSHFWUDO HVWLPDWLRQ ,((( 7UDQVDFWLRQV RQ $HURVSDFH DQG (OHFWURQLF 6\VWHPV f 2FWREHU >@ =KHQJ6KH /LX DQG LDQ /L )HDWXUH H[WUDFWLRQ RI 6$5 WDUJHWV FRQVLVWLQJ RI WULKHGUDO DQG GLKHGUDO FRUQHU UHIOHFWRUV ,(( 3URFHHGLQJV RQ 5DGDU 6RQDU DQG 1DYLJDWLRQ 0DUFK >@ =KHQJ6KH /LX DQG LDQ /L 6$5 PRWLRQ FRPSHQVDWLRQ DQG IHDWXUH H[WUDFWLRQ YLD 0&5(/$; RXUQDO RI WKH 2SWLFDO 6RFLHW\ RI $PHULFD $ f 0DUFK >@ & ( 0DQFLOO (QKDQFHPHQW RI UDGDU LPDJHU\ E\ PD[LPXP HQWURS\ SURFHVVLQJ 3URFHHGLQJ RI WKH +XPDQ )DFWRUV 6RFLHW\ VW $QQXDO 0HHWLQJ SDJHV
PAGE 118
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f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f $SULO >@ 5 5R\ DQG 7 .DLODWK (635,7 (VWLPDWLRQ RI VLJQDO SDUDPHWHUV YLD URn WDWLRQDO LQYDULDQFH WHFKQLTXHV ,((( 7UDQVDFWLRQV RQ $FRXVWLFV 6SHHFK DQG 6LJQDO 3URFHVVLQJ $663f XO\ >@ 7 5XFN ( %DUULFN : 6WXDUW DQG & .ULFKEDXP 5DGDU &URVV 6HFWLRQ +DQG %RRN 3OHQXP 3UHVV 1HZ @ 5 2 6FKPLGW 0XOWLSOH HPLWWHU ORFDWLRQ DQG VLJQDO SDUDPHWHU HVWLPDWLRQ ,((( 7UDQVDFWLRQV RQ $QWHQQDV DQG 3URSDJDWLRQ $3f 0DUFK >@ & : 6KHUZLQ 3 5XLQD DQG 5 5DZFOLIIH 6RPH HDUO\ GHYHORSPHQWV LQ V\QWKHWLF DSHUWXUH UDGDU V\VWHP ,5( 7UDQVDFWLRQV RQ 0LOLWDU\ (OHFWURQLFV 0,/f $SULO >@ 3DXO / 6PLWK U .HQQHWK 5 +DUG\ DQG .HQQHWK 0 *ORYHU $SSOLFDWLRQV RI UDGDU WR PHWHRURORJLFDO RSHUDWLRQV DQG UHVHDUFK 3URFHHGLQJV RI WKH ,((( f XQH
PAGE 119
>@ 7 6GHUVWUP DQG 3 6WRLFD 6\VWHP ,GHQWLILFDWLRQ 3UHQWLFH+DOO ,QWHUQDWLRQDO /RQGRQ 8. >@ % 6WHLQEHUJ DQG % .DQJ 5DGDU GHWHFWLRQ VHQVLWLYLW\ DV D IXQFWLRQ RI WDUJHW GLPHQVLRQDOLW\ ,((( ,QWHUQDWLRQDO 5DGDU &RQIHUHQFH SDJHV >@ 3 6WRLFD DQG 5 / 0RVHV ,QWURGXFWLRQ WR 6SHFWUDO $QDO\VLV 3UHQWLFH+DOO (QJOHZRRG &OLIIV 1>@ 3 6WRLFD DQG $ 1HKRUDL 3HUIRUPDQFH VWXG\ RI FRQGLWLRQDO DQG XQFRQGLWLRQDO GLUHFWLRQRIDUULYDO HVWLPDWLRQ ,((( 7UDQVDFWLRQV RQ $FRXVWLFV 6SHHFK DQG 6LJQDO 3URFHVVLQJ $663f 2FWREHU >@ .L\R 7RPL\DVX 7XWRULDO UHYLHZ RI V\QWKHWLFDSHUWXUH UDGDU 6$5f ZLWK DSSOLn FDWLRQV WR LPDJLQJ RI WKH RFHDQ VXUIDFH 3URFHHGLQJV RI WKH ,((( f 0D\ >@ 7VDR DQG % 6WHLQEHUJ 5HGXFWLRQ RI VLGHOREH DQG VSHFNOH DUWLIDFWV LQ PLFURZDYH LPDJLQJ 7KH &/($1 WHFKQLTXH ,((( 7UDQVDFWLRQV RQ $QWHQQDV DQG 3URSDJDWLRQ $SULO >@ ( :DKO 3 + (LFKHO & *KLJOLD DQG & 9 DNRZDW] U 3KDVH JUDGLn HQW DXWRIRFXV D UREXVW WRRO IRU KLJK UHVROXWLRQ 6$5 SKDVH FRUUHFWLRQ ,((( 7UDQVDFWLRQV RQ $HURVSDFH DQG (OHFWURQLF 6\VWHPV XO\ >@ 1 @ + =HENHU DQG 5 0 *ROGVWHLQ 7RSRJUDSKLF PDSSLQJ IURP LQWHUIHURPHWULF V\QWKHWLF DSHUWXUH UDGDU REVHUYDWLRQV RXUQDO RI *HRSK\VLFDO 5HVHDUFK $SULO >@ + $ =HENHU 6 1 0DGVHQ 0DUWLQ % :KHHOHU 7 0LOOHU < /RX $On EHUWL 6 9HWUHOOD DQG $ &XFFL 7KH 7236$5 LQWHUIHURPHWULF UDGDU WRSRJUDSKLF PDSSLQJ LQVWUXPHQW ,((( 7UDQVDFWLRQV RQ *HRVFLHQFH DQG 5HPRWH 6HQVLQJ f 6HSWHPEHU
PAGE 120
%,2*5$3+,&$/ 6.(7&+ =KDRTLDQJ %L ZDV ERUQ LQ +HLORQJMLDQJ 3URYLQFH &KLQD RQ 2FWREHU +H DWWHQGHG +XD]KRQJ 8QLYHUVLW\ RI 6FLHQFH DQG 7HFKQRORJ\ :XKDQ &KLQD LQ 6HSWHPEHU ZKHUH KH HDUQHG KLV %DFKHORU RI (QJLQHHULQJ GHJUHH LQ HOHFWULFDO HQJLQHHULQJ LQ XO\ +H HDUQHG KLV 0DVWHUfV GHJUHH LQ FRPPXQLFDWLRQ DQG HOHFn WURQLF V\VWHPV IURP &RPPXQLFDWLRQ 7HOHPHWHULQJ DQG 7HOHFRQWURO 5HVHDUFK ,QVWLn WXWH &7,f LQ 6KLMLD]KXDQJ &KLQD LQ 0DUFK +H ZRUNHG DV DQ ,& GHVLJQHU IRU GLJLWDO FRPPXQLFDWLRQ V\VWHPV ZLWK DGYDQFHG ('$ HOHFWURQLF GHVLJQ DLGHGf WRROV LQ &7, IURP WR +H KDV EHHQ D JUDGXDWH UHVHDUFK DVVLVWDQW LQ WKH 6SHFWUDO $QDO\VLV /DE 6$/f RI WKH 'HSDUWPHQW RI (OHFWULFDO DQG &RPSXWHU (QJLQHHULQJ DW WKH 8QLYHUVLW\ RI )ORULGD VLQFH $XJXVW GXULQJ ZKLFK WLPH KH KDV FRQGXFWHG UHVHDUFK LQ WKH DUHD RI UDGDU VLJQDO SURFHVVLQJ LQFOXGLQJ WDUJHW IHDWXUH H[WUDFWLRQ DQG V\QWKHWLF DSHUWXUH UDGDU LPDJH IRUPDWLRQ
PAGE 121
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
PAGE 122
7KLV GLVVHUWDWLRQ ZDV VXEPLWWHG WR WKH *UDGXDWH )DFXOW\ RI WKH &ROOHJH RI (QJLQHHULQJ DQG WR WKH *UDGXDWH 6FKRRO DQG ZDV DFFHSWHG DV SDUWLDO IXOILOOPHQW RI WKH UHTXLUHPHQWV IRU WKH GHJUHH RI 'RFWRU RI 3KLORVRSK\ 0D\ :LQIUHG 0 3KLOOLSV 'HDQ &ROOHJH RI (QJLQHHULQJ 0LKUDQ 2KDQLDQ 'HDQ *UDGXDWH 6FKRRO
4.6 CramrRao Bound (CRB) of the SemiParametric Data Model
Let Y be an N x matrix with its (n, )th element being y(n,fi) with
52
K
y(n, n) = ^ xk{n)eik + e(n, n), n 0,1, , N 1, n = 0,1, , N 1,
k=1
(4.51)
where K denotes the number of scatterers; {xk()}%=Q denotes the realvalued am
plitude function of for the /cth scatterer; q*)k and {fk,fk}, respectively, are the
constant phase and the frequency pair of the Â£;th scatterer; finally, {e(n, )} denotes
the unknown 2D noise and clutter sequence. Then we can rewrite (4.51) as:
K
Y = Â£e*Wjv(/*)xÂ£D(/fc) + E,
(4.52)
k=1
where E denotes an N x N matrix with e(n, n) being its (n, n)th element,
n t
>N{fk)
wv(/fc)
l e^k . e^Â¡k{N\)
1 gj27r/fc . ej2irfh(Nl)
(4.53)
1 T
(4.54)
with ()r denoting the transpose, D(fk) denoting the following diagonal matrix:
D(/*)=diag l, ..., e^Ui) \
and
X/c =
*1 T
xk(0) xk(l) xk(N 1)
(4.55)
(4.56)
Let y = vec(Y) = [ yf y?
yL ]T with yn{n 1,2, , N) being the nth
column of Y. Then the data model can be rewritten as follows:
y = A/i. + e,
(4.57)
where
A 
[w(/l) OXl] ww(/i) (w/i(/jt)XK]WW(/i)
(4.58)
CHAPTER 5
3D TARGET FEATURE EXTRACTION USING CLSAR
5.1 CLSAR Data Model
We describe how to obtain 3D target features via the full synthetic aperture
shown in Figure 1.2(a), which makes a preparation for the discussion on CLSAR. We
start by establishing onedimensional (1D) point scatterer data model of high range
resolution radar as a preparation for the establishment of 3D data model herein.
5.1.1 1D Data Model for High Range Resolution Radar
The range resolution of a radar is determined by the radar bandwidth. To
achieve high resolution in range, the radar must transmit wide band pulses, which
are often linear FM chirp pulses [29, 49]. A normalized chirp pulse can be written as
s{t) = eA2*/ot+7t2), t < To/2, (5.1)
where f0 denotes the carrier frequency, 2q denotes the FM rate, and T0 denotes the
width of the pulse. We assume that /0, 7, and T0 are known. The signal returned by
a scatterer of a target has the form
r(t) = (5.2)
where ST is determined by the RCS of the scatterer and t denotes the roundtrip
time delay. The demodulated signal d(t) is obtained by mixing r(f) with s*(t r0)
for some given r0 = 2R0/c, where ()* denotes the complex conjugate, c denotes the
propagation speed of the transmitted electromagnetic wave, and R0 denotes the range
distance from radar to the center of the illuminated patch.
_ (5Tet2(7r/o7To)(TTo)ej7(TTo)2ej27(rro)i (53)
63
CHAPTER 7
CONCLUSION
7.1 Summary
In this study, we have extended two parametric relaxationbased algorithms,
including RELAX and RELAXNLS, and presented a novel semiparametric ap
proach SPAR to form enhanced 2D SAR images in different scenarios. An effective
relaxationbased parametric method, referred to as AUTORELAX, has also been pro
posed for both 3D target feature extraction and motion compensation via CLSAR.
The RELAX and RELAXNLS algorithms are parametric spectral estima
tion methods and provide the potential for significant resolution improvement of the
formed SAR images provided that the assumed data model is valid. RELAX is based
on the point scatterer (trihedral) data model and RELAXNLS is devised according
to a mixed dihedral and trihedral data model, which describes a dihedral corner re
flector as a 2D complex sinusoid with amplitude described as a sine function and
a constant phase in crossrange and with constant amplitude and phase in range.
Super resolution SAR images can be generated by first extracting the target features
via RELAX or RELAXNLS based on the corresponding data models, and then ap
plying FFT to the synthesized phase history data with larger dimensions in both
range and crossrange by using the extracted target features. Experimental examples
with ERIM and MSTAR data in Chapter 3 have shown that RELAX and RELAX
NLS can be used to obtain either higher resolution images from the same size data,
or, equivalently, equalquality images from significantly fewer number of data sam
ples. Of the two methods, RELAX is more suitable for point scatterer targets and
94
86
shown in the previous chapter, the estimates {ak, fk, fk, fk}k=1 of {afc, fk, fk, fk}k=1
can be obtained with the RELAX algorithm by equivalently minimizing the following
NLS criterion:
/ K N M N1
f {fe, fk,fk, a} j = ^2 m)~ m)i2 > (632)
' m=1 n=0
where y(n,m) = y(n,m)e~:iflm and z3(n,m) has the same form as z(n,m) in (6.5)
except that and iZm(n) are replaced by tXm(n), tym(n), and iZm(n),
respectively, which are determined by {9m, instead of {9m, 4>m}^=1. Let
M Nl
Fk (ak, fk, fk, /*.) = X] S ^(nm)
J 2k If Am W+fkiym (n)+fkizm Ml
 (6.33)
m=1 n=0
where tXm(n), tym(n), and tZm(n) are the same as tXm(n), tym(n), and tZm(n), respec
tively, except that {9m,(j)m}^=1 are replaced by {9mAm}m=i, and
K
yk(n,m) = y(n,m) Q,eAr[/Lm(n)+/Lm(ri)+/Lm(n)]^ (6.34)
il,i^k
Then minimizing Fk in (6.33) with respect to ak yields the estimate otk of ak:
Yfml En0 yk(n> m)e~j2K[fkLmn+fkiym(n)+fkizm(n)]
Oik =
and
{fk, h, fk) = arÂ§ max
fk ,fk,fk
MN
M Nl
fk=fk,fk=fk,fk=fk
^ y ) yk(n, m)e f2n^kixm(n)+fkiym(n)+fkiZm(n)] (g_36)
m=1 n=0
Herein, we omit the detailed steps used for the 3D target feature extraction
since they are the same as those in the previous chapter except that yk in (5.25)
is replaced by {yk(n,m)} in (6.34). The practical convergence in the iterations
of the RELAX algorithm now is determined by checking the relative change of the
cost function F fk, fk, fk}kz=^j in (6.32) between two consecutive iterations. To
speed up the RELAX algorithm via utilizing FFT, a Cartesian grid mapping approach
is introduced in Chapter 5 to approximately map {iXm(n), iyrn(n), iZrn(n)}m=i on a
This work is dedicated to my wife, Jing Wang, for her love and sharing my
happiness and depression in the past years; to my parents, who continue supporting
and encouraging me for my education.
88
{R0(m)}M=1. The AUTORELAX algorithm converges after 6 iterations. Figure 6.2
shows the target scatterers extracted by AUTORELAX with the search interval for
{A0m}^f=1 being 0.34. Comparing Figures 6.2(b) and 6.1(d) with Figure 6.1(b),
we find that AUTORELAX works well and the AUTORELAX results are slightly
better than those obtained by using the manually adjusted aperture. Figure 6.3 shows
that the manually adjusted and autofocused apertures fit quite well after adding a
line to the former. Since the results in Figure 6.2(b) are better than those in Figure
6.1(d), it appears that a linear phase error (in (j)m as a function of 6m) was introduced
in the manually adjusted aperture. The linear phase error can cause shifts in /*, and
the amount of shift is different for different scatterer k, as can be seen from (6.14).
Consider next autofocusing in both the elevation and azimuth directions by as
suming that no errors exist in {R0(m)}^=1. The AUTORELAX algorithm converges
after 7 iterations. Figure 6.4 shows the results obtained with the search intervals for
{A0m}Â£f=1 and {A9m}^=1 being 0.34 and 0.006, respectively. Compared with
the results obtained by autofocusing only in the elevation direction, we find that aut
ofocusing in both directions provides little further improvement to the accuracy of
the estimated target parameters in this example. Hence for this example, it appears
that the aperture errors mainly occur in the elevation direction.
For this experimental example, we have also used AUTORELAX to extract
target features when we assumed that errors exist in both and
{Ro(m)}M=1. Again, as compared with Figure 6.2(b), we have noticed little change
in the extracted target parameters. It could be that DSMS has already done a good
job compensating for the errors in {Ro(m)}^=l with some traditional method so that
they are now negligible as compared with the errors in {
Finally, we use a simulation example to show that the accuracy of target
feature extraction via CLSAR is very sensitive to the accuracy of the curvilinear
aperture. We assume that there are 20 scatterers with a similar distribution to that
62
(e) (f)
Figure 4.5: Comparison of SAR images obtained via different algorithms for the
Slicy data hbl5533.015 (0 azimuth and 30 elevation angles), (a) Windowed 2D
FFT SAR image from the XPATCH data, (b) Windowed 2D FFT SAR image from
the Slicy data, (c) Hybrid SAR image, (d) SPAR SAR image, (e) RELAXNLS
SAR image, (f) Modified RELAXNLS SAR image, (c) ~ (f) are all obtained from
the data used in (b). (The vertical and horizontal axes are for range and crossrange,
respectively.)
15
and parametric methods for feature extraction and image formation of SAR targets
has not been addressed in the literature before.
2.2 3D SAR Feature Extraction
One relatively mature technology for 3D SAR imaging and feature extrac
tion is IFSAR [24, 55, 69, 27, 70] which exploits two different apertures. A simple
implementation of a IFSAR is to use two antenna displaced in the crosstrack plane
to obtain two coherent and parallel measurements. The relative height information
can be estimated from the phase difference between the two measurements. The use
of interferometric techniques with SAR imagery to provide height information makes
IFSAR suitable for 3D imaging since it inherits the finer resolutions (in range and
crossrange) associated with the spotlightmode SAR and also has reasonable height
resolution. Many 2D modern spectral estimation algorithms including Capon [10],
APES [38], and MUSIC [58] have been generalized for 3D imaging and target feature
extraction using IFSAR [40, 47]. An NLS method that outperforms MUSIC in speed
and estimation accuracy was also reported in [40]. However, such a system cannot
resolve more than one target scatterer at the same projected range and crossrange
but at different heights [40].
Another relatively new technology for 3D imaging is curvilinear SAR. CLSAR
traverses a curvilinear aperture path and avoids the height ambiguity problem inher
ent in IFSAR systems. CLSAR images are useless in practice due to high sidelobes.
However, CLSAR can be used in conjunction with modern spectral estimation meth
ods to accurately extract 3D features of small targets, which consist of a finite number
of isolated point scatterers. CLSAR is still at its exploratory stage of development
and only a few studies were reported in the literature. In [33], a coherent CLEAN
algorithm was presented by Knaell to effectively eliminate sidelobes in CLSAR im
ages. The CLEAN algorithm was originally proposed in radio astronomy [28] and
64
The term e_i7^T_T0^2 in (5.3) is usually close to a constant for all rmÂ¡n < r, r0 < rmax,
where rmax and rmÂ¡n correspond to the maximum and minimum values, respectively,
of the roundtrip time delays between the scatterers of a target and the radar. This
term can also be partially removed [49]. Let D(u) denote the Fourier transform of
~ ~ u2
d(t). Then the inverse Fourier transform of D(uj)eJi:i will have the term e_J7^T_To^
removed. Yet this removal can only be approximate since d(t) is not known for all
t and hence D(lo) is not known exactly. The closer e_J7(TT)2 is to a constant for
rmin < t < rmax, the better its removal. With this removal, we have
d[t) = (5TeJ'2{7r/7ro){T7'o)ej27{TTo)i, (5.4)
which is a complex sinusoid with frequency 27(7r0) and amplitude <5Te?2(7rf07T)(T~T0b
We know rmax and rmjn approximately since we assume that the altitude, antenna
beamwidth, and grazing angle of the radar are known. We also assume that (rmax
rmin) < To Then for T0/2 + rmax < t < T0/2 + rmin, the scatterers of the target at
different ranges correspond to different frequencies of the signal d(t), while the RCSs
of the scatterers are proportional to the amplitudes of the corresponding sinusoids.
The ranges and RCSs of the target scatterers are the 1D target features.
5.1.2 Full Synthetic Aperture Radar
A broadside data collection geometry in a spotlightmode SAR is shown in
Figure 5.1 [29], The XY Z coordinate system is centered on a small patch of ground,
where a target is located. The ground is illuminated by a narrow radio frequency
(RF) beam from the moving radar that rotates (with radius Rq) around the coordinate
origin. In Figure 5.1, R denotes the distance between the radar and a scatterer at
the position (x,y,z), and 9 and (j) are the azimuth and elevation angles of the radar
relative to the XYZ coordinate system. We assume that 9, 4>, and R0 are known.
The range R of the scatterer located at (x, y, z) can be written as
R = [(Ro cos 9 cos 4> x)2 + {Ro sin 9 cos y)2 + (Rq sin
75
where Ir denotes an x (N) matrix with each column and row containing
only one nonzero unit element corresponding to the locations of the available data
samples in the rectangular grid. Note that the larger the dimensions of the rectangular
grid, the more accurate the mapping approximation. For large enough dimensions,
however, the noise will become the dominant source of error. For our examples, we
chose iV = N = = 128. As compared to PolartoCartesian interpolation, the grid
mapping approach avoids the interpolation step needed by the former, which could
be complicated for an arbitrary curvilinear aperture. The former approach, however,
is computationally more efficient due to its smaller data dimensions.
The steps of applying RELAX to yr are similar to those for the case of full
aperture except that Equations (5.26) and (5.27) above are replaced by
[Ir(afc <8> a* afc)]'^ yrk
Oik =
N
and
Uk =Uk ,Uk =Uk ,Wk =Uk
{c5fc, ujk} = arg max [Ir(afc afc a*)] yTk
(5.40)
(5.41)
respectively, where
Yrk = Yr
K
Y <*dr
(5.42)
Let yric be similar to y except that the elements in y that are missing in yTk are
replaced with zeros in yTk. Then the right hand side of (5.41) can also be computed
by applying 3D FFT to yTk. Since the dimensions of the rectangular grid are larger
than the dimensions of the original data due to zerofilling, the 3D FFT of yTk is
approximately periodic and we should limit our attentions to only one period. For our
examples, the former is four times as large as the latter and hence the peak searching
in (5.41) is limited to the frequency intervals [ 2n x (0.125) 2tt x 0.125 ] Note
that the computational advantages of FFT over DFT (discrete Fourier transform)
and backprojection diminish as the dimensions of the rectangular grid increases,
especially for curvilinear apertures.
APPENDIX C
CALCULATING (x,y,z) FROM (x,y,z)
Given R0, 0O, and (x,y,z), (x,y,z) can be determined from (5.8), (5.9), and
(5.10) as follows:
vlv+ 2ViRo [vi (l + r&) (R0 + 2t]2z) + 7?iRo 2rjÂ¡x] ye + R20 [2r% (l + 772) y2
+47i + 1 )2 2 + + 4R0???7?2 (5 + rfe) z 877^771 (l + rfe) xz
+4R0 (2?7i 3772) x + Rlrj[ (8 + 772)] y4 877^ (771^0 + 2r)\x) yy3
+2772i?o [25 (1 + 77!) +Ro?]2 (3 + ril) + 2t7^77i] t/2t/2 87]lR40y3y +
= 0, (C.l)
z = {vivh4 + Ro [RoVi(2 + V2) + Zvir]2z(l + vl)] y2
2yyxRly R^^^y2} / [i?0(^2/ + yvÂ¡)y\ (C.2)
and
2 = \ {v?W + Ro?7i [277i(1 + 772) + 772(77!Ro 2772)] y2
2rj2R20yy + Ro%y2(l + ^2)} / [Ro(t7iS/ + VvDv] > (C.3)
where 771 = cos0o and 772 = sin0O Note that the left side of (C.l) is an eighthorder
polynomial, and hence has eight zeros. We pick the root that is closest to y as the
solution for y.
We now present two examples. First, let Rq = 15, 00 = 7t/5, and (x,y,z) =
(0.5,0.4,0.3). We obtain (x,y,z) = (0.5005,0.3849,0.3078). (We remark that the
other seven possible solutions for y are either complex or far away from y.) Next, let
R0 300, 0O = 7t/4, and (x,y,z) = (5,4,2). We obtain
102
13
In recent years, many sophisticated and high resolution methods based on
point scatterer data model, such as eigendecomposition based methods [4], including
MUSIC (Multiple Signal Classification) [58] and ESPRIT (Estimation of Signal Pa
rameters via Rotational Invariance Technology) [56], and relaxationbased nonlinear
least squares (NLS) method such as RELAX [39], have also been applied to super
resolution SAR image formation. The eigendecomposition based methods are signal
or noise subspace based approaches and generally are not suitable for SAR image
formation since they require that the number of data samples be much larger than
that of the dominant point scatterers and the process of whitening the noise sub
space (used in MUSIC) will destroy the spatial inhomogeneities of the terrain clutter
[15]. The RELAX algorithm [39] was originally proposed for sinusoidal parameter
estimation in colored noise. RELAX obtains the parameter estimates by minimizing
the NLS cost function via a sequence of FFTs, which is conceptually and computa
tionally simple and hence robust. RELAX coincides with the maximum likelihood
(ML) method when the noise is white and is also asymptotically (for large numbers
of data samples) statistical efficient when the noise is an AR or autoregressive mov
ing average (ARMA) process that can be described by a finite number of unknowns
[63]. Due to the iterations performed in each step of RELAX, a set of more accurate
initial estimates are available, which make RELAX converge to the global minimum
with a high probability. RELAX can also be referred to as SUPER CLEAN when
compared with two other approximate relaxation methods including CLEAN [28] and
MCLEAN (more clean) [23]. CLEAN was first proposed in radio astronomy and later
used in microwave imaging [66]. If no iteration is performed in each step of RELAX,
then it is equivalent to CLEAN. RELAX becomes MCLEAN if only one iteration
is performed in each step of RELAX. Hence RELAX is a computationally efficient
and robust algorithm with good statistical performance and is suitable for super
resolution SAR image formation of targets consisting of point scatterers (trihedrals).
83
6.2 The RelaxationBased Autofocus Algorithm (AUTORELAX)
The AUTORELAX algorithm obtains the estimates {9m, and
A A 'L o 0
{&k,fk,fkifk}k=v respectively, of the true values {9m, m, 7?m}Â£f=1 and
{ak, fk, fk, fk}k=i by minimizing the following NLS criterion:
/ K \ M Nl
Cs ^ ^ ?/(n,m) x(n,m)2, (6.23)
m=1 n=0
where
x(
(6.24)
;(n, m) = (X akej2*\Jktxm(n)+fktym(n)+fktZm(n)] ^ gj'r/m ,
with tXm(n), tym(n), and tZm(n) defined in (6.2), (6.3), and (6.4), respectively. Note
that as shown below, we will determine 9m and cj)m via a search method. Hence there
is no need to use the approximations such as those in (6.11), (6.12), and (6.13) for
txm (n)) y (n) > and tzm (n) j respectively. We use an alternating optimization approach
to minimize C5 in (6.23) and the AUTORELAX algorithm is outlined as follows:
Step 1: Extract the target features {ak, fk, fk, fk}k=i with the RELAX algorithm
from an initial curvilinear aperture by assuming that {rÂ¡m 0}Â£f=1.
Step 2: Update the curvilinear aperture {0m,0m}=1 and the phase error {r7m}m=i
with Steps (1) (4) (to be discussed in Section 6.2.1). (If {r]m}^=1 is known to be
negligible, then only update {9m,(/)m}^=1 with Step (1).)
Step 3: Redetermine the target parameters with the RELAX algorithm by using
the curvilinear aperture {9m, (f>m}^=l and {vm}^=1 obtained in Step 2.
Step 4: Repeat Steps 2 and 3 until practical convergence.
The practical convergence of AUTORELAX is determined similarly to that
of the RELAX algorithm by checking the relative change of the cost function C5 in
(6.23) between two consecutive iterations.
6.2.1 Aperture Parameter Estimation
Assume that the target feature estimates {&k, fk, f fk}k=i are given Then
6m, 4>m, and f}m can be determined by minimizing the following Cm with respect to
29
without the background clutter are illustrated in Figures 3.1 (c) ~ (f) with K = 59.
Note that the RELAX images have a higher resolution than the FFT images for the
dominant scatterers. Utilizing only 25% of the 40 x 40 phase history data, i.e., using
a 20 x 20 phase history data, we form the 2D FFT SAR images shown in Figures
3.2(a) and (b). Figures 3.2(c) ~ (f) show the RELAX images with K 41. Compar
ing Figures 3.2(f) and 3.2(b), we note that the two images are quite similar although
the former uses only 25% of the data used by the latter.
Consider next an example of the MSTAR Slicy data consisting of both trihe
dral and dihedral corner reflectors collected by the SNL using the STARLOS sensor.
The field data was collected by a spotlightmode SAR with a carrier frequency 9.559
GHz and bandwidth 0.591 GHz. The radar was about 5 kilometers away from the
ground target shown in Figure 3.3. The SAR images are obtained when the tar
get is illuminated by the radar from the azimuth angle 0 and elevation angle 30.
The range and crossrange resolutions of the original data (54 x 54) are 0.3 meters
and 0.32 meters, respectively, and the windowed 2D FFT SAR image (obtained
via Taylor window) is shown in Figure 3.4(a). To demonstrate the super resolution
property of RELAX and RELAXNLS, we use part of the original data, whose di
mension is 32 x 32. Figure 3.4(b) shows the windowed 2D FFT images with the
spoiled resolution 0.51 x 0.54 meters obtained from this reduced size data. RELAX
and RELAXNLS are also applied to this 32 x 32 phase history data matrix. Figures
3.4(c) and (d), respectively, show the windowed RELAX SAR images with and with
out estimated background clutter with K = 36. We note that RELAX images have a
better resolution than the FFT images for the dominant trihedrals and resemble the
FFT images for dihedrals even though the data model in the crossrange dimension
used by RELAX is not correct for this example. The windowed RELAXNLS SAR
images with and without estimated background clutter with K 9 are illustrated
in Figures 3.4(e) and (f), respectively. We note again that the RELAXNLS images
Elevation Angle (degrees)
93
iÂ¡t.ra>in*
(c) (d)
Figure 6.5: (a) Curvilinear aperture for the simulation example, (b) True scatterer
distribution for the simulation example, (c) Scatterer distribution obtained from the
initial aperture (dashed line) in (a) by using RELAX with K = 20. (d) Scatterer dis
tribution obtained via AUORELAX with K 20 and using the autofocused aperture
(dotted line) shown in (a).
101
2xy sin 9 cos <Â¡>\cos 0O (sin 0O)0] + 2yz sin 9 cos 0[sin 0O + (cos 0O)ip\
+2xz cos 9 cos 0[sin 0O + (cos 0O)ip]
2(.in/ cos 0O + yz sin 0o) sin 9 cos 0 + (2xz sin 0O) cos 9 cos 0
+ (2a:zcos2 0o)0
Hence R is simplified to (5.6).
(B.5)
31
(e) (f)
Figure 3.1: RELAX SAR Images by using the 40 x 40 ERIM data with K = 59.
(a) Unwindowed 2D FFT image, (b) Windowed 2D FFT image, (c) Unwindowed
RELAX image without background clutter, (d) Windowed RELAX image without
background clutter, (e) Unwindowed RELAX image with background clutter, (f)
Windowed RELAX image with background clutter.
21
We consider using GAIC (see [39] and the references therein for details) to
determine K, the number of sinusoids, by assuming the unknown noise and clutter
being white. The estimate K of K can be determined as an integer that minimizes
the following GAIC cost function:
/ Nl 1 \
GAIC^ = N\n EE e(n, n)2 j + j\n[ln(NN)](4:K + l), (3.8)
where
K
^ 710 71=0
(n, n) = y(n, n) ^ Ajte7^'cn+Wfcn\ n = 0,1, , N 1, n = 0,1, , N 1,
k=1
(3.9)
4K + 1 denotes the total number of unknown realvalued parameters (of which 4K
are for the sinusoids and 1 is for the white noise), and 7 is a parameter of user choice.
3.1.2 Image Formation
We can form SAR images by using the point scatterer data model and the
target features extracted via RELAX. SAR images obtained via RELAX are referred
to as RELAX SAR images or simply RELAX images. The steps of using RELAX
for SAR image formation are outlined as follows.
Step 1: Obtain the parameter estimates of {o^, uk, 4}^ yia RELAX by using the
measured phase history data (see Section 3.1.1 for details).
Step 2: Generate ys{ns,s), the simulated phase history data of large dimensions,
from the estimated parameters and based on the data model in (3.1):
K
, ns = 0,1, , CAT 1, s = 0,1, , ( 1,(3.10)
y.
,{ns,ns) = ^ ake
j(uikns+ujkns)
fc=1
where Â£ (Â£ > 1) denotes the extrapolation factor, which is a parameter of user
choice, and 0>fc, and u>k, respectively, denote the estimates of ak, u)k, and wfc,
k = 1,2, , iÂ£.
Step 3: Form RELAX SAR images containing only the dominant target features by
10
to establish. In Chapter 4, we also present a modified RELAXNLS algorithm by
using the initial conditions provided by SPAR to save the amount of computations
needed by the original RELAXNLS algorithm. In Chapter 5, a CLSAR data model
is established and RELAX is introduced for 3D target feature extraction via CLSAR.
Chapter 6 investigates the ambiguity problems inherent in the proposed CLSAR data
model when aperture errors are present. A relaxationbased autofocus algorithm for
3D target feature extraction in the presence of aperture errors is also presented in
Chapter 6. Finally, Chapter 7 give the summary and conclusions of this research and
outlines the future work.
LIST OF FIGURES
1.1 Illustration of 2D radar imaging 2
1.2 Illustration of SAR imaging modes, (a) Spotlightmode SAR. (b)
Stripmap SAR 3
1.3 Possible apertures for a 3D SAR system, a) Full aperture, b) Parabolic
aperture, (c) Orthogonal subaperture. ( 9 and denote the azimuth
and elevation angles, respectively.) 6
1.4 Mesh plots of the modulus of the RCS obtained by using 2D FFT with
different apertures (range information suppressed for the illustration
purpose only), (a) Full aperture as shown in Figure protect 1.2(a). (b)
Curvilinear aperture as shown in Figure 1.2(c) 7
3.1 RELAX SAR Images by using the 40 x 40 ERIM data with K 59. (a)
Unwindowed 2D FFT image, (b) Windowed 2D FFT image, (c) Un
windowed RELAX image without background clutter, (d) Windowed
RELAX image without background clutter, (e) Unwindowed RELAX
image with background clutter, (f) Windowed RELAX image with
background clutter 31
3.2 RELAX Images by using the 20 x 20 ERIM data with K 41. (a)
Unwindowed 2D FFT image, (b) Windowed 2D FFT image, (c) Un
windowed RELAX image without background clutter, (d) Windowed
RELAX image without background clutter, (e) Unwindowed RELAX
image with background clutter, (f) Windowed RELAX image with
background clutter 32
3.3 Target photo taken at 45 azimuth angle 33
3.4 SAR images of the Slicy data (0 azimuth and 30 elevation angles). (a)
Original windowed 2D FFT image, (b) Windoed 2D FFT image with
spoiled resolution, (c) ~ (d): Windowed RELAX image without and
with background clutter with K = 36. (e) ~ (f): Windowed RELAX
NLS image without and with background clutter with K = 9. (c) ~
(f) are obatined using the data in (b). (The vertical and horizontal
axes are for range and crossrange, respectively.)
vii
34
12
[37]. Empirically, it was observed in [38] that Capon appears to give biased spec
tral estimates whereas APES is unbiased. This observation was later proved in [37]
by using the secondorder Taylor expansion technique. Also the statistical perfor
mance of the Capon estimator degrades rapidly as the filter length increases, while
that of APES remains unaffected. The study in [37] gives a compelling reason for
preferring APES to Capon in most spectral estimation applications including SAR
imaging. In cases where an appropriate parametric data model of targets is not
available, APES is a good candidate for SAR imaging which provides much lower
sidelobes and narrower spectral peaks than Capon and the conventional FFT based
methods. However, significant resolution improvement cannot be obtained by using
nonparametric methods since they do not fully exploit the characteristics of targets,
especially when such information is available.
Parametric Spectral Estimators
When parametric data models can be used to describe radar targets, paramet
ric spectral estimators provide attractive alternatives for super resolution SAR image
formation. The Burgs maximum entropy (ME) method [9] was applied to radar im
age enhancement by Mancill in 1977 [44]. Unlike FFT methods which assume that
the signal beyond the observations is zero, ME methods make linear prediction of the
signal beyond the observation window so that they can achieve better resolution than
FFT methods. Identical to ME methods in principles, the autoregressive (AR) model
based methods [15, 13, 20, 19, 26] have attracted a lot of research attention. However,
the high resolution property of the AR based methods are highly dependent on the
presence of observation noise [31]. As the signaltonoise ratio (SNR) decreases, so
does the resolution. On the other hand, the AR based methods are more appropriate
for SAR image formation of distributed scatterers rather than point scatterers since
the AR spectral estimator by nature is a continuous spectral estimator.
6
(a)
0
$
(c)
e
Figure 1.3: Possible apertures for a 3D SAR system, a) Full aperture, b) Parabolic
aperture, (c) Orthogonal subaperture. ( 9 and <Â¡> denote the azimuth and elevation
angles, respectively.)
no accurate parametric data model is available in crossrange. Based on the flexible
data model, we present a robust and computationally simple semiparametric (SPAR)
algorithm for target feature extraction and SAR image formation by exploiting a
hybrid of nonparametric and parametric spectral estimation methodes.
It has been shown in [62, 16] that threedimensional (3D) target features
including the height information as well as the 2D target features provide very useful
information in certain applications, such as automatic target recognition (ATR). 3D
SAR image formation and feature extraction are becoming active research areas. It
is possible to obtain 3D information of an illuminated scene with Spotlightmode
SAR when a nonstraight line synthetic aperture is exploited, which is determined by
the azimuth angle 9 and elevation angle . In our work, we will also study 3D target
feature extraction and motion compensation (autofocus) using curvilinear SAR, in
which the trajectory of a spotlightmode SAR is a suitable curvilinear aperture.
Among existing 3D SAR imaging techniques, interferometric SAR (IFSAR)
is now a welldeveloped technique that plays an important role in 3D (as well as 2D)
imaging. However, IFSAR suffers from ambiguity problems [11, 40]. Theoretically,
3D target features and images can also be obtained when the trajectory of SAR is a
full aperture as shown in Figure 1.2(a) even though such an aperture is not feasible in
108
[61] T. Sderstrm and P. Stoica. System Identification. PrenticeHall International,
London, U.K., 1989.
[62] B. D. Steinberg and B. Kang. Radar detection sensitivity as a function of target
dimensionality. IEEE International Radar Conference, pages 106110, 1990.
[63] P. Stoica and R. L. Moses. Introduction to Spectral Analysis. PrenticeHall,
Englewood Cliffs, NJ, 1997.
[64] P. Stoica and A. Nehorai. Performance study of conditional and unconditional
directionofarrival estimation. IEEE Transactions on Acoustics, Speech, and
Signal Processing, ASSP38(10):17831795, October 1990.
[65] Kiyo Tomiyasu. Tutorial review of syntheticaperture radar (SAR) with appli
cations to imaging of the ocean surface. Proceedings of the IEEE, 66(5):563583,
May 1978.
[66] J. Tsao and B. D. Steinberg. Reduction of sidelobe and speckle artifacts in
microwave imaging: The CLEAN technique. IEEE Transactions on Antennas
and Propagation, 36:543556, April 1988.
[67] D. E. Wahl, P. H. Eichel, D. C. Ghiglia, and C. V. Jakowatz, Jr. Phase gradi
ent autofocus a robust tool for high resolution SAR phase correction. IEEE
Transactions on Aerospace and Electronic Systems, 30:827835, July 1994.
[68] G. N. Yoji. Phase Difference Auto Focusing for Synthetic Aperture Radar Imag
ing. U.S. Patent 4,999,635, 1991.
[69] H. Zebker and R. M. Goldstein. Topographic mapping from interferometric
synthetic aperture radar observations. Journal of Geophysical Research, 91:4993
4999, April 1986.
[70] H. A. Zebker, S. N. Madsen, J. Martin, K. B. Wheeler, T. Miller, Y. Lou, G. Al
berti, S. Vetrella, and A. Cucci. The TOPSAR interferometric radar topographic
mapping instrument. IEEE Transactions on Geoscience and Remote Sensing,
30(5):933940, September 1992.
71
respectively, where
yCJb = y c
K
i
= l,ijk
{Ic( )} 0
(5.30)
Let yCfc be similar to y*, except that the elements in y*. that are missing in yCk are
replaced with zeros in yCk. Then the right hand side of (5.29) can also be computed
by applying 3D FFT to yCk.
5.3 Performance Analysis of Parameter Estimation via CRB
Consider first the case of the full aperture. Let Q = Ejee^} be the noise
covariance matrix, which is arbitrary and unknown. The extended SlepianBangs
formula for the ijth element of the Fisher information matrix (FIM) has the form
[3, 64]:
{FIM},, = tr (cr^Q'q;) + 2Re [(<*" A")'. Q~' (Aa)'] (5.31)
where X) denotes the derivative of X with respect to the zth unknown parameter,
tr(X) denotes the trace of X, and Re(X) denotes the real part of X. Note that FIM
is block diagonal since Q does not depend on the parameters in (Aa), and (Aa)
does not depend on the elements of Q. Hence the CRB matrix for the target features
of interest can be calculated from the second term on the right side of (5.31). Let
V =
ReJ (a) InF (a) a/r ojt u>T
where
a;
i T
CO l 0J2 LOfc
CO =
lT
CO i O2 Ll>k
and
co
iT
CO 1 C2 COk
(5.32)
(5.33)
(5.34)
(5.35)
84
{Om,(pm,Vm}, where
Nl
Cm(Om,(l)m,Vm) = '^2\y{n,m) x^n^m)]2, n = 0, , N 1, m = 1,2, , M,
71=0
(6.25)
where x\(n, m) has the same form as x(n, m) in (6.24) except that {ak, fk, fk, fk}k=i
are replaced by {fc, fk, fk, fk}k=1. To simplify the optimization of Cm, we determine
{0m,
Step (1): Obtain \f9jj1, > by minimizing the following Cm\ with respect to {$777, 5
where
JVl
cmi (9m,fm) = l^771) ~ i(n>m)2 n = 0, ,N 1, m = 1, M,
n=0
(6.26)
where i(n, m) has the same form as z(n,m) in (6.5) except that {ak, fk, fk, fk}k=i
are replaced by {k, fk, /*,, /jtlitLi When there are errors in both the elevation and
O o
azimuth directions, we can estimate 0m and Qm by the alternating minimization
^ o
approach, i.e., by iteratively fixing the estimate 0m of 0m and minimizing Cmi with
/V o
respect to 9m, and then fixing the estimate 9m of 0m and minimizing Cmi with respect
to 0m until practical convergence. The practical convergence in the alternating
minimization approach is determined by the relative change of the cost function Cmi.
In the numerical examples, we terminate the iteration when the relative change of the
cost function Cmi between two consecutive iterations is less than 103. When there
are errors in only one angular direction, i.e., either 6m or 4>m is to be determined,
then the minimization of Cmi is a simple onedimensional search problem.
Step (2): Determine phase errors {fjm}m=1 by minimizing the following cost func
tions Cm2:
JVl
Cm2 {Vm) = \y(nim) z2(n,rn)e:,11m\2 (6.27)
n=0
70
Step (2): Assume K = 2. Compute y2 with (5.25) by using {cok,Cjk,cuk,ak}k=i
obtained in Step (1). Obtain {bk, Â£ok, u>k, dfe}fe=2 from y2. Next, compute yi by using
{h,k,u>k,k}k=2 and then redetermine {Â£>fc, 4, 5fc, dfc}fc=1 from yi.
Iterate the previous two substeps until practical convergence is achieved (to
be discussed later on).
Step (3): Assume K = 3. Compute y3 by using {cjfe, 4, 4, ak}2k==1 obtained in Step
(2). Obtain {04, Cok,u)k, &k}k=3 from y3. Next, compute yi by using {4, 4, k, k}k=2
and redetermine {)k,Cok,Cok,ak}k=i from y^ Then compute y2 by using
{k,L)k,u)k,k}k=i and redetermine {uk, wk, u>k, aik}k=2 from y2.
Iterate the previous three substeps until practical convergence.
Remaining Steps: Continue similarly until K is equal to the desired or estimated
number of sinusoids. (Whenever K is unknown, it can be estimated from the available
data, for instance, by using the generalized Akaike information criterion (GAIC) rules
which are particularly tailored to the RELAX method of parameter estimation. See,
e.g., [39].)
The practical convergence in the iterations of the above RELAX method
may be determined by checking the relative change e of the cost function
C\ ^{Â¡)Â¡fc, u)fc, )k, k}Â£=1^J in (5.24) between two consecutive iterations. Our numerical
examples show that the iterations usually converge in a few steps.
5.2.2 Curvilinear Aperture
The RELAX algorithm for the curvilinear aperture is similar to that of the
full aperture except that Equations (5.26) and (5.27) above are replaced by
[{Ic(afc <8>fc)} <8>afc]HyCfc
Oik
MN
and
Uk Uk !Uk =Uk ,Uk=Uk
{u)k,k,ujk} = arg max [{Ic(a* afe)} a.k]H yCk
Uk,Uk,Uk
(5.28)
(5.29)
APPENDIX B
APPROXIMATION OF THE RANGE R
The range R in (5.5) can be approximated as follows. Under the conditions
k ^ Rft < l' and ~k < we have
R
Rq
Rq
2 i 2 i 2~\ 1/2
x y z x y ~r z *
1 2 cos 9 cos (ft 2^ sin 9 cos (ft 2 sin (ft H ~
Ro Rq Rq Rft
i x a j. y a j. z x2 + y2 + z2
1 cos 9 cos (ft  sin 9 cos (ft  sin (ft d 5
i?o No /t0 2jRq
{x cos 9 cos (ft + y sin 9 cos (ft + z sin (ft)'
2/Â§
(B.l)
Let ip = (ft 0o, where is the average of all (ft used to form the synthetic aperture.
For very small ip, we have
cos (ft fa cos (fto (sin
(B.2)
sin (ft fa sin (fto + (cos (fto)ip.
(B.3)
For very small 9, we have cos9 fa 1, sin# ~ 9. Then keeping firstorder terms yields
x2 sin2 (ft + x2 sin2 9 cos2 (ft + y2 sin2 (ft + y2 cos2 9 cos2 (ft + z2 cos2 (ft
fa x2 sin (ft[sm (ft0 + (cos (ft0)ip\ + y2 sin (ft[sin (ft0 + (cos (fto)ip]
+y2 cos 0[cos (ft0 (sin (ft0)ip] + z2 cos 0[cos (ft0 (sin (ft0)ip]
fa [(x2 + y2) sin (ft0] sin (ft + [(y2 + z2) cos (ft0\ cos 9 cos
+[(x2 z2) sinocos(fto\ip. (B.4)
We also have
2xy cos 9 sin 9 cos2 (ft + 2yz sin 9 cos 0 sin 0 p 2xz cos 9 cos (ft sin (ft
100
4.1 Ambiguity effect on the SAR image formation in the absence of range
estimation errors, (a) True windowed FFT SAR image, (b) Windowed
FFT image of the first scatterer. (c) Windowed FFT image of the
second scatterer. (d) Combined windowed FFT image of the two
scatterers. (The vertical and horizontal axes are for range and cross
range, respectively.) 58
4.2 Ambiguity effect on the SAR image formation in the presence of range
estimation errors, (a) Windowed FFT image of the first scatterer.
(b) Windowed FFT image of the second scatterer. (c) Combined
windowed FFT image of the two scatterers with Â£ = 1 (without
extrapolation), (d) Combined windowed FFT image with Â£ = 2.
(The vertical and horizontal axes are for range and crossrange, re
spectively.) 59
4.3 Comparison of SAR images formed using different algorithms for sim
ulated data at high SNR (a2 = 0.6). (a) True SAR image, (b) Win
dowed 2D FFT SAR image, (c) Hybrid SAR image, (d) SPAR SAR
image, (e) RELAXNLS SAR image, (f) Modified RELAXNLS SAR
image. (The vertical and horizontal axes are for range and crossrange,
respectively.) 60
4.4 Comparison of SAR images formed using different algorithms for sim
ulated data at low SNR (a2 = 6). (a) Windowed 2D FFT SAR image.
(b) Hybrid SAR image, (c) SPAR SAR image, (d) RELAXNLS SAR
image, (e) Modified RELAXNLS SAR image. (The vertical and hor
izontal axes are for range and crossrange, respectively.) 61
4.5 Comparison of SAR images obtained via different algorithms for the
Slicy data hbl5533.015 (0 azimuth and 30 elevation angles), (a)
Windowed 2D FFT SAR image from the XPATCH data, (b) Win
dowed 2D FFT SAR image from the Slicy data, (c) Hybrid SAR
image, (d) SPAR SAR image, (e) RELAXNLS SAR image, (f) Mod
ified RELAXNLS SAR image, (c) ~ (f) are all obtained from the
data used in (b). (The vertical and horizontal axes are for range and
crossrange, respectively.) 62
5.1 Broadside spotlightmode SAR data collection geometry 65
5.2 Curvilinear apertures, a) The circular aperture, b) The Arc1 aper
ture. c) The Arc2 aperture 73
5.3 The CRBs of the target parameters as a function of the circular aper
ture radius when K = 1, a = 1, u) = = = 0, and o2 = 40. a) The
CRB of the complex amplitude a. b) The CRB of u. c) The CRB of
Co. d) The CRB of Go 77
5.4 3D plots of K = 8 scatterers extracted by using RELAX with the
indoor experimental data, (a) Obtained with full aperture as shown in
Figure 1.2(a). (b) Obtained with curved aperture as shown in Figure
1.2(c)
viii
78
92
in
CD
45.4
1 11 1 1 1 1 1 1 1 T r ' 1 1 11 i
(degre
45.2
45.0
:
 j
\
s . \
0
44.8
\
. \
cn
c
<
44.6
_ \
\
: Initial aperture \ SN
c
o
44.4
_ .. : Aperture obtained by AUTORELAX after 7 iterations **. _
\
\
+1
o
>
0
44.2
44.0
\
LlJ
J iL i.i. ii il i i i i i i I i i i i
3210 1 2 3
Azimuth Angle (degrees)
(a)
v iuertions
Figure 6.4: Autofocused curvilinear aperture and scatterer distribution obtained with
AUTORELAX by autofocusing in both the elevation and azimuth directions and
using K = 20 for the experimental example, (a) Autofocused curvilinear aperture
(dotted line), (b) Scatterer distribution.
99
Let ys{n, f) denote the nth element of ys(/), n = 0,1, , N 1. Then cj) is given by
(A.6)
Inserting (A.6) into (A.5) and ignoring the scaling factor, we can simplify (A.5) to:
i = jargj^1 [yJ>M/)] .
N1
CMJ) = +
71=0
Nl
E [yl"N(f)]2 e'
71=0
llys(/)ir +
N1
j2ir(2f)n
n=0
(A.7)
Then the NLS estimates {/, /} of {/, /} are determined by
{/, /} = argmax C4(/, /). (A.8)
/,/
Note that ys(/) in (4.37) can be obtained by applying 1D FFT to each column of Ys
and the second term in (A.7) can also be readily obtained by applying 1D FFT to
the sequence {y2s{fi, f)}n=o w^h 2/ as the frequency variable. Hence {/,/} can be
obtained via a 2D search for the location corresponding to the peak of Cif, /), which
can be computed efficiently via 1D FFTs. Note also that padding with zeros for the
1D FFTs is necessary to achieve high accuracy for the frequency estimates. An
alternative approach is to find an approximate location corresponding to the global
maximum with 1D FFT without much zeropadding and then use the approximate
location as the initial condition to find a more accurate position via, for example,
alternatively using the FMIN function in MATLAB.
37
4.2 Review of the APES Algorithm
The APES [38] algorithm belongs to the adaptive filterbank class, in which
the adaptive finite impulse response (FIR) filter is designed based on not only the data
to be processed, but also the interference at the frequency of interest. APES mimics
the maximum likelihood estimator of a complex sinusoid in circularly symmetric
zeromean Gaussian noise. It has been shown in [37] that APES has better statistical
performance than Capon [10], another popularly used adaptive filterbank approach
for SAR imagery.
Let {zn}^I{g1 denote the available 1D data sequence and have the form:
zn = a(co)exp{jnu>} + n = 0,1, , N 1, (4.3)
where a(u) denotes the complex amplitude of a sinusoid with frequence lo and en(u)
denotes the unmodeled noise and interference at u. Then APES can be used to obtain
the complex amplitude estimate a(co) of a(u) from {zu}^Zq. It has been shown in
[38] that the leastsquares estimate of a{u) using the forward and backward FIR
filter output has the form:
&(lu)
h//(a;)Z(w) I /3(co)ZH (uj)h(uj) ,
(4.4)
where h(u>) denotes the impulse response of an Mtap FIR filter and has the following
form:
h(w) =
l T
hi(u) h2(co)
hM(uj)
(4.5)
Z(u>) and Z(cu), respectively, denote the normalized Fourier transforms of {zi}^M
and {zi.e., the Fourier transforms of {zj}^QM and {z}^qM divided by (N
M + 1), where z and zrespectively, denote the overlapping subvectors of the data
vector z =
z0 z i
data sequence z =
zni
ZN1 ZN2
and the complex conjugated and backwardordered
and have the form
Z=
T
Zi+1
Zi+M1
i = 0,1, , N M,
(4.6)
25
and
{t,tt} = argmax
u%(ujt)Ytv*(ujt)\2
N
(3.21)
Thus {ujt,u>t} corresponds to the peak location of aj^(a>)Ytuj*(jt)\2, which can be
calculated efficiently via 2D FFT, and at corresponds to the complex height of the
normalized 2D FFT of Yt at {jt,Ct}.
Feature Extraction of a Single Dihedral Corner Reflector
Let Yd denote the data of a single dihedral in the presence of unknown noise
Erf. Then
Yrf = adUJN(u>d)gr I Erf, (3.22)
where {uid, ud ad, b, r) are the target features of a dihedral corner reflector, and
r \T
g =
9(0) 0(1)
g(1)
(3.23)
with g() sinc[f)7r(r )]eJ'w
{rf,cÂ¡)rf,rf, b, f} of {ad,tod, Cob, r} are determined by minimizing the following cost
function:
Crf (ad, ud, id, 6, r) =  Yrf adu>N(ud)gr\\2F (3.24)
It can be shown that ad is determined by
. ^N^d)Yrfg*
ad Yg2
and {uod,u)d,b, f} can be obtained by maximizing the cost function
r, l \ \^N^d)Ydg*\2
c3 (ujd,ud,b,T) ^g2 (3.26)
RELAXNLS maximizes C3 in (3.26) by utilizing an alternating maximization ap
proach [8], i.e., alternatively updating {ujd, u>d, b, f} while fixing other parameter esti
mates at their recently determined values. Note that fourdimensional search over the
parameter space is performed to estimate the dihedral corner reflector parameters.
96
for 3D target feature extraction via CLSAR and motion compensation (autofocus)
for CLSAR in Chapters 5 and 6, respectively. A CLSAR data model has been first
established and the ambiguity problems inherent in the data model have been in
vestigated when aperture errors are present. To provide a guide for the choice of
CLSAR trajectory, the performance of different curvilinear apertures is analyzed via
CramrRao Bound, which is the lowest bound achievable by any unbiased estima
tors. Autofocusing for CLSAR becomes more complicated since the aperture errors
in CLSAR, unlike in the 2D SAR systems, can no longer be approximated as phase
errors only in crossrange due to the nonnegligible errors introduced in the azimuth
and elevation angles of the trajectory. We have presented a relaxationbased autofo
cus algorithm, referred to as AUTORELAX, for 3D target feature extraction using
CLSAR in the presence of aperture errors. AUTORELAX can be used to simulta
neously extracts 3D target features and compensates motion errors including phase
errors and aperture errors by minimizing an NLS cost function via an alternating
optimization approach. Experimental examples have been used to demonstrate the
effectiveness of AUTORELAX for 3D target feature extraction using CLSAR data
in the presence of curvilinear aperture errors.
7.2 Future Works
In this study, we have concentrated on the study of SAR target feature extrac
tion, super resolution image formation, and autofocus via relaxationbased paramet
ric and semiparametric methods. However, the problems involved in inverse SAR
(ISAR) become more complicated and difficult. Compared with the stationary targets
of SAR, targets being imaged in ISAR are often engaged in complicated maneuvers
that combine translational and rotational motions and furthermore, the motions be
tween the radar and the moving targets in ISAR are noncooperative. Hence, the
89
Estimation Error
Magnitude
Range (/)
Crossrange (/)
Height (/)
AUTORELAX
0.015
0.00288
0.00179
0.00348
RELAX
0.296
0.31544
0.01858
0.24444
Table 6.1: Estimation errors for the parameters of the highest scatterer in the simu
lation example before and after autofocusing.
in the experimental example above. The true aperture is the dot line shown in Figure
6.5(a) and the scatterer distribution is shown in Figure 6.5(b). Here we simulate the
case where the aperture errors exist only in the elevation direction. Figure 6.5(c)
shows the scatterer distribution obtained with the RELAX algorithm from the initial
aperture (dash line) shown in Figure 6.5(a). We note that the scatterer distribution
in Figure 6.5(c) is quite different from the true one in Figure 6.5(b). When used
with the simulation data, AUTORELAX converges after four iterations. The auto
focused aperture is the dot line shown in Figure 6.5(a) and the scatterer distribution
obtained with AUTORELAX is illustrated in Figure 6.5(d). We see that the scat
terer distribution obtained by AUTORELAX is almost the same as the true one. We
also notice that the shape of the autofocused aperture is closer to the true one than
the initial aperture. The constant difference between the autofocused and the true
apertures will cause phase errors in {ck}k=n which cannot be eliminated due to the
ambiguity problems discussed in Section 6.1.To quantitatively illustrate the accuracy
improvement of the parameter estimates via AUTORELAX, we compare the esti
mation errors of the highest scatterer in Figure 6.5(b) before and after autofocusing.
Table 6.3 shows the differences between the true values and their estimates obtained
via RELAX and AUTORELAX, respectively.
ACKNOWLEDGEMENTS
I am sincerely grateful to my advisor and committee chairman, Dr. Jian
Li, not only for her academic guidance and insightful suggestions throughout the
development of this dissertation and providing an invaluable environment for the
research, but also for her financial support in the past years.
I would like to thank my graduate committee members, Dr. Jos C. Principe,
Dr. Fred J. Taylor, Dr. William W. Edmonson, and Dr. David C. Wilson, for their
time and interest in serving on my supervisory committee.
I would also like to thank Dr. ZhengShe Liu, Dr. Renbiao Wu, and the fellow
students, past and present, of the Spectral Analysis Laboratory (SAL) for their great
help.
Finally, I would like to gratefully acknowledge all the people who helped me
during my Ph.D program.
7
(a) (b)
Figure 1.4: Mesh plots of the modulus of the RCS obtained by using 2D FFT with
different apertures (range information suppressed for the illustration purpose only),
(a) Full aperture as shown in Figure protect 1.2(a). (b) Curvilinear aperture as shown
in Figure 1.2(c).
practice. Nevertheless, a suitable subset of the full aperture can be used as a trajectory
of SAR to obtain 3D information. This technique is known as curvilinear SAR
(CLSAR). CLSAR traverses a curvilinear aperture path, which is a subset of the full
aperture (see, for example, Figures 1.2(b) and (c)), and collects a sparsely sampled
dataset, from which 3D information can be obtained. CLSAR avoids ambiguity
problems inherent in IFSAR systems, but suffers from severe high sidelobes when
used with FFT to form SAR images due to the limited size of measurements. This
observation is illustrated in Figure 1.2. (Figures 1.2(a) and (b) show the modulus
of the RCS of a single point scatterer obtained by applying 2D FFT to the data
collected via the full aperture as shown in Figure 1.2(a) and the curvilinear aperture
as shown in Figure 1.2(c).) Due to its high sidelobes, CLSAR is of little interest in
SAR imaging. Yet CLSAR can be used with parametric spectral estimation methods
to extract 3D target features of small targets consisting of a finite number of isolated
point scatterers. In Chapter 5, we will establish a CLSAR data model and describe
how the RELAX algorithm [39] can be used for 3D target feature extraction using
CLSAR with different curvilinear apertures.
54
j2neJ(t>k {[uft(fk) Xfc] (uN(fk) dN)}
(4.64)
with
and
1N
i T
0 1 TV 1
j2ire^k {[w^(/jfc) Xfc Ody^] wjv(/fc)}
(4.65)
(4.66)
with
lJV
n T
0 1 TV 1 (4.67)
Let F4 denote an TVTV x K matrix whose [(k 1)TV + ]th column is determined
by
eJ(^k {[wjj d>yv(/fc)] u?yv(/fe)} ) k = 1,2,... ,K, = 1, 2, , TV, (4.68)
where wn is an TV x 1 vector with the th element being unit and all other elements
being zeros. Let
F =
Fi F2 F3 F4
(4.69)
Then the CRB matrix for the parameter vector 77 is given by:
CRB(ij) = [2Re(FffQ1F)]_1.
(4.70)
4.7 Numerical and Experimental Results
We demonstrate and compare the SAR image formation performances of
SPAR, and the modified RELAXNLS algorithm with both simulated and experi
mental examples. The algorithms are also compared with Hybrid and RELAXNLS.
In the following examples, the dimensions of the original SAR phase history data
matrix are TV = TV = 32 and GAIC with 7 = 5.5 is used to determine K for the
relaxationbased feature extraction algorithms of SPAR and Hybrid. The threshold
Ti used in the isolation process of SPAR is 10% of the peak value. The maximization
of C4(/,/) in (4.41) is done in two steps. First, initial frequency estimates / and /
73
(a) (b) (c)
Figure 5.2: Curvilinear apertures, a) The circular aperture, b) The Arc1 aperture,
c) The Arc2 aperture.
Aperture
CRB(a)
CRB(o;)
CRB(cj) CRB(o>)
Full
5.9875
51.4511
51.4511 51.4511
Circular
5.3663
39.3415
40.8469 40.8788
Parabolic
5.8693
39.3415
40.4915 39.6521
Lshaped
4.8584
39.3415
38.2414 38.2414
Arc1
11.7482
39.3415
31.7272 31.7272
Acr2
17.2473
39.3415
26.8015 27.4636
Table 5.1: Comparison of the CRBs (in dB) of the target features for the cases of the
full, circular, parabolic, Lshaped, Arc1, and Arc2 apertures when M 63, K 1,
cu = (2> = ) = 0, a = 1, and cr2 = 40.
circle whose center is at the upper right corner of the full aperture shown in Figure
1.2(a) and whose radius is 31, as shown in Figure 5.2(b), and the Arc2 aperture
is one half of a parabolic aperture whose vertex is at the lower right corner of the
full aperture and who starts at the upper left corner of the full aperture shown in
Figure 1.2(a), as shown in Figure 5.2(c). These curvilinear apertures are subsets
of the full aperture in Figure 1.2(a) and are made to be as large as possible. For
example, the radius of the circular aperture is 15.5 in this example. Table 1 shows
the CRBs of the target parameters. As expected, the CRB for lu is the same for
RELAXATIONBASED METHODS FOR SAR TARGET
FEATURE EXTRACTION AND IMAGE FORMATION
By
ZHAOQIANG BI
A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY.
UNIVERSITY OF FLORIDA
This dissertation was submitted to the Graduate Faculty of the College of
Engineering and to the Graduate School and was accepted as partial fulfillment of
the requirements for the degree of Doctor of Philosophy.
May 1999
Winfred M. Phillips
Dean, College of Engineering
Mihran J Ohanian
Dean, Graduate School
80
exactly due to atmospheric turbulence and platform position uncertainty. In 2D SAR
imaging, it is generally assumed that the errors in {9m, 0m}W_1 are negligible and the
errors in {R0(m)}^=1 cause phase errors along the synthetic aperture. In CLSAR,
however, the errors in {9m, 0m}Â£f=1 may no longer be negligible since the aperture
shape is critical for 3D target feature extraction. Our problem of interest herein is to
compensate for the curvilinear aperture errors in {R0(m)}^=1 and {9m, (f)m}^=1 and
extract the 3D target parameters {ok, /*,, fa, fk}k=i from {y(n, m)},n = 0,1, , N
1, m = 1, 2, , M. We first analyze aperture errors in CLSAR and then present an
autofocus algorithm, referred to as AUTORELAX, for 3D target feature extraction
in the presence of aperture errors.
6.1 Aperture Error Effects
We will consider the approximations and the ambiguity problems in our data
model in the presence of aperture errors. We omit the dependence on n in the
following analysis For the sake of notation convenience. For the broadside data
collection geometry shown in Figure 5.1, 9m is very small. For very small 9m, we have
sin(0m) 9m and cos(9m) 1. These approximations also hold for the true look
angles of the radar. Then according to (6.2), (6.3), and (6.4), respectively, we have
xm~ (cos 0m),
(COS 0J 9m,
and
(sin 0m),
o o
where {9m, 0m}m=i denote the true look angles of the radar and
4(tt/o 7r0 + 71)
(6.6)
(6.7)
(6.8)
t =
Consider first the errors in {cj)m}^=1 when {9m =^m,i?0(n^) = Ro}m=i Let
0m 0m +A0m
(6.9)
76
We now present an experimental example to demonstrate the performances
of the RELAX algorithm for 3D target feature extraction with a CLSAR. We use
e = 0.01 to test the convergence of the RELAX algorithm in this example. The
indoor data used herein was obtained by the Radar Signature Branch, Naval Air
Warfare Center, Mugu, California. The radar carrier frequency is 9.968 GHz and the
bandwidth 1.524 GHz. The 32 x 32 x 32 data set was obtained with the full aperture
shown in Figure 1.2(a) with 32 samples in each dimension and the angular increments
were 0.28. The target consists of K = 8 corner reflectors with a cubic configuration
and was about 15 meters away from the radar. The scatterers were about 0.5 meters
apart.
Figure 5.4.2 shows the extracted scatterers when K = 8 is used with RELAX.
Figure 5.4.2(a) is obtained when the full aperture shown in Figure 1.2(a) is used to
extract the 3D target features and Figure 5.4.2(b) is obtained when the Lshaped
aperture shown in Figure 1.2(c) is used. The centers of the circles denote the locations
of the extracted scatterers in 3D space and the radius of each circle is proportional
to the modulus of the RCS of the corresponding scatterer. The triangles show the
projections of the scatterer locations onto the horizontal plane and their sizes are also
scaled according to the RCSs of the scatterers. Note that Figure 5.4.2(b) is similar
to Figure 5.4.2(a) even though the former is obtained by using only 6.15% amount
of data used by the latter.
6.1 (a) Curvilinear apertures for the experimental example, (b) True scat
terer distribution, (c) 3D plot of K = 20 scatterers extracted from
the initial aperture (dashed line) in (a) with RELAX, (d) 3D plot
of K = 20 scatterers extracted from the manually adjusted aperture
(solid line) in (a) with RELAX 90
6.2 Autofocused curvilinear aperture and scatterer distribution obtained
with AUTORELAX by autofocusing only in the elevation direction
and using K = 20 for the experimental example, (a) Autofocused
curvilinear aperture (dotted line), (b) Scatterer distribution 91
6.3 (a) Manually adjusted (solid line) and autofocused (dotted line) curvi
linear apertures for the experimental example, (b) Fitting the manu
ally adjusted aperture to the autofocused aperture by adding a line to
the former 91
6.4 Autofocused curvilinear aperture and scatterer distribution obtained
with AUTORELAX by autofocusing in both the elevation and azimuth
directions and using K 20 for the experimental example, (a) Auto
focused curvilinear aperture (dotted line), (b) Scatterer distribution. 92
6.5 (a) Curvilinear aperture for the simulation example, (b) True scatterer
distribution for the simulation example, (c) Scatterer distribution ob
tained from the initial aperture (dashed line) in (a) by using RELAX
with K = 20. (d) Scatterer distribution obtained via AUORELAX
with K = 20 and using the autofocused aperture (dotted line) shown
in (a) 93
IX
3
Figure 1.2: Illustration of SAR imaging modes, (a) Spotlightmode SAR. (b) Strip
map SAR.
processing radar returns from a sequence of locations when a nominallysized antenna,
which is carried on board of an aircraft, sequentially moves along the flight path
(crossrange direction). Consequently, a synthetic antenna with an effective antenna
length proportional to the synthetic aperture can be realized even with a small real
antenna. The longer the synthetic aperture, the finer the crossrange resolution
achieved. It has been demonstrated in [29] that the crossrange resolution of SAR
depends only on the wavelength A and the synthetic aperture interval. To achieve
such a desirable crossrange resolution, many techniques have been developed to
record data, compensate motion errors, and construct high quality SAR images ever
since the 1960s. (Detailed descriptions of these techniques are well documented in
many books, e.g. [29, 11]). SAR has grown so dramatically through advances in
digital signal processing, computing power, and microwave technology that it is one
of the most important techniques for high resolution radar image formation.
Spotlightmode and stripmap SAR are two commonly used SAR imaging
modes. In a spotlightmode SAR, antenna beams continuously steer to the same
terrain patch (scene) illuminated by a radar when the aircraft traverses the flight
path, while the antenna steering of a stripmap SAR is fixed relative to the flight
path when the aircraft moves along the flight path (see Figure 1.2 for the difference).
42
where Â£ denotes an extrapolation factor (Â£ > 1) and is a parameter of user choice.
Note that the super resolution property of the soformed SAR images is determined
by the feature extraction algorithm and Â£ > 1 is only used to demonstrate the super
resolution property of the target feature extraction algorithm. The estimated noise
and clutter data matrix is
(n, n) = y(n, ) ss(n, ), n = 0,1, , N 1, = 0,1, , 1, (4.24)
which is also important in many SAR applications since, for example, important
target information such as the target shadow information is contained in (n, ). We
cannot extrapolate (n, ) in either range or crossrange since no parametric data
model is available for (n, ).
To obtain SAR images with low sidelobes in range via 1D FFT, we apply 1D
windows to ss(n, ) and (n, ) in range. We obtain a new (Â£Y) x data matrix Y
as follows:
{y(n, ) = ss(n, n)ws(n) + Â£(n, )we(n), n = 0,1, , N 1, = 0,1, , 1,
y(n, ) = ss(n, )ws(n), n = N, N + 1, , Â£Y 1, = 0,1, , 1,
(4.25)
where y(n, fi) denotes the (n, )th element of Y and ws{n) and we(n) are 1D windows
of lengths Â£Y and N, respectively, satisfying
Civi
53 Ws(n) = CW. (4.26)
n=0
and
N1
^2we(n) = N. (4.27)
n0
Note that scaling (n,) in y(n,) by a factor of Â£ is necessary since the range
dimension of ss(n,) is Â£ times of that of (n, ). The steps needed for SAR image
formation are as follows:
Step (1): Form Y from y(n, ) by using (4.23), (4.24), and (4.25).
44
k
Oik
4>k
fk
Ik
bk
Tk
k=l
9.6
0
0.1
0.3
0.3
18.6
k=2
6.4
0
0.1
0.1
0.2
18.6
Table 4.1: True parameters of the two dihedrals used in Figures 1 and 2.
to simulate the dihedrals, where and brespectively, are proportional to the
maximal RCS and the length of the kth dihedral corner reflector and rdenotes the
peak location of the data sequence and is determined by the orientation of the kth.
dihedral. The size of the simulated data matrix is 32 x 32 (i.e., N = = 32).
The parameters for the two dihedrals are given in Table 1. An ambiguous set of
target features can be obtained by choosing fi /2 = 0.1, f1 = f2 = f = 0.2,
4>i = 0, and (Â¡>2 7t/2 in (4.19) and (4.20). The windowed FFT SAR images of
the two scatterers are shown in Figures 4.1(b) and (c), respectively, which differ
considerably from the two dihedral scatterers in Figure 4.1(a). The combined SAR
image of the two scatterers is given in Figure 4.1(d), which is exactly the same as
the true image shown in Figure 4.1(a). However, due to the presence of noise and
clutter, parameter estimation errors are inevitable. The errors in range are the main
cause of the artifact problem in the high resolution SAR image formation. In Figure
4.2, we assume that all parameters are accurate except that /2 = /2 + 0.01 = 0.11.
Figures 4.2(a) and (b), respectively, show the windowed FFT images of the two
aforementioned scatterers in the presence of estimation errors and Figure 4.2(c)
shows the combined SAR image. By comparing Figures 4.1(a) and 4.2(c), we note
that an extra line (artifact) shows up next to the short dihedral. The reason is that
due to the estimation errors, ^ f2. Hence the two scatterers in Figures 4.2(a)
and (b) are not exactly in the same range and cannot be combined perfectly to
obtain the two dihedral lines in Figure 4.1(a). This problem becomes even worse
when super resolution SAR images are formed via data extrapolation in range. The
105
[14] S. R. DeGraaf. Sidelobe reduction via adaptive FIR filtering in SAR imagery.
IEEE Transactions on Image Processing, 3(3) :292301, May 1994.
[15] S. R. DeGraaf. SAR imaging via modern 2D spectral estimation methods. IEEE
Transactions on Image Processing, 7(5):729761, May 1998.
[16] D. E. Dudgeon and R. T. Lacoss. An overview of automatic target recognition.
The Lincoln Laboratory Journal, 6(1):310, 1993.
[17] P. H. Eichel, D. C. Ghiglia, and C. V. Jakowatz, Jr. Speckle processing methods
for syntheticaperture radar phase correction. Optics Letters, 14:13, January
1989.
[18] P. H. Eichel and C. V. Jakowatz, Jr. Phasegradient algorithm as an optimal
estimator of the phase derivative. Optics Letters, 14:11011103, October 1989.
[19] A. Farina, A. Forte, F. Prodi, and F. Vinelli. Superresolution capabilities in
2D direct and inverse SAR processing. International Symposium on Noise and
Clutter Rejection in Radars and Imaging Sensors, pages 151156, Kanagawa
Science Park, Japan, November 1994.
[20] A. Farina, F. Prodi, and F. Vinelli. Application of superresolution techinques to
radar imaging. Chinese Journal of Systems Engineering and Electronics, 5(1):1
14, January 1994.
[21] M. T. Fennell and R. P. Wishner. Battlefield awareness via synergistic SAR and
MTI exploitation. IEEE Aerospace and Electronic Systems Magazine, 13(2):39
45, February 1998.
[22] G. H. Golub and C. F. Van Loan. Matrix Computations. Johns Hopkins Uni
versity Press, Baltimore, MD, 1984.
[23] P. T. Gough. A fast spectral estimation algorithm based on the FFT. IEEE
Transactions on Signal Processing, 42:13171322, June 1994.
[24] L. C. Graham. Synthetic interferometer radar for topographic mapping. Pro
ceedings of the IEEE, 62:763768, June 1974.
[25] R. O. Harger. Synthetic Aperture Radar System: Theory and Design. Chapter
VI, Phase and Motion Errors. Academic Press Inc., New York, 1970.
[26] G. R. Heidbreder. Maximum entropy methods in coherent radar imaging. In
ternationl Journal of Imaging Systems and Technology, 2:239247, 1990.
[27] D. N. Held and J. D. OBrien. Norden systems 2dimensional interferometric
synthetic aperture radar flying testbed. Proceedings of the 1992 IEEE Inter
national Geoscience and Remote Sensing Synoisum (IGARSS), Houston, TX,
2:646648, May 1992.
[28] J. A. Hbgbom. Aperture synthesis with a nonregular distribution of interfer
ometer baselines. Astronomy and Astrophysics Supplements, 15:417426, 1974.
[29] C. V. Jakowatz, Jr., D. E. Wahl, P. H. Eichel, D. C. Ghiglia, and P. A. Thomp
son. SpotlightMode Synthetic Aperture Radar: A Signal Processing Approach.
Kluwer Academic Publishers, Norwell, MA, 1996.
66
where Sx,y,z is proportional to the RCS of the scatterer located at (x,y,z). For
4(tt/o 7Tq + 7)(o Ro)/c
d[t, 9, 0) ~ Sx,y,z
pjiitx+yty+ztz)
c )
(5.12)
where
ty
4(tt/0 7T0 + 7) cos 9 cos 0
5
c
4(tt/o 7T0 + 7) sin 0 cos 0
and
t, = 
4(tt/0 7^0 + 7) sin0
(5.13)
(5.14)
(5.15)
Note that d(t, 9,0) is a 3D complex sinusoid. The frequencies of the 3D sinusoid
correspond to the 3D location (x, y, z) of the scatterer, while the amplitude is propor
tional to its RCS. Note that (x, y, z) is not the true location (a;, y, z) of the scatterer,
but is close to (x, y, z) for large R0. (See the Appendix 3C on how to calculate (x, y, z)
from (x, y, z).) When a target has multiple scatterers, d(t, 9, 0) in (5.12) will be a sum
of sinusoids. The 3D locations and RCSs of the target scatterers are the 3D target
features. Since usually the samples on the t, 9, and 0 axes are uniformly spaced, the
samples of tx, ty, and tz occur at the points of a polar grid. Hence PolartoCartesian
interpolation may be needed for the data samples to occur at rectangular grid points.
(See Section 5.4.2 for an alternative approach.)
After PolartoCartesian interpolation and sampling, the signal obtained by
the 3D full aperture SAR in the presence of noise can be written as:
K
y(n, , n) = ^ akeJ
(wfcn+Ofc+w^Ti)
e(n, , n),
(5.16)
fc=l
where n = 0,1, , N 1, = 0,1, , 1, and = 0,1, , 1, with
N, N, and denoting the numbers of available data samples in the three dimen
sions; K denotes the number of sinusoids; ajt and {wjt, 4, 4} denote the unknown
CHAPTER 3
PARAMETRIC METHODS FOR SAR IMAGING
When an appropriate data model is available for radar targets, parametric
spectral estimation algorithms, which are less sensitive to the model mismatch, offer
the potential to significantly improve the resolution of the formed SAR images. In
this chapter, we consider forming enhanced SAR images with robust relaxationbased
methods including RELAX and RELAXNLS.
3.1 RELAX
When the dominant features of targets can be approximated as point scatter
ed, enhanced SAR images can be formed by using the target features extracted via
RELAX and the point scatterer data model. We first briefly introduce the RELAX
algorithm [39] and then discuss how to form SAR images using RELAX.
3.1.1 Review of the RELAX Algorithm
Assume that there are K dominant point scatterers or trihedral corner re
flectors in a target. An ideal trihedral acts as a triplebounce reflector and can be
modeled as a strong point scatterer [53] since its radar cross section (RCS) can be
considered as a constant due to the very small angle variation of the radar beam in
SAR. Then the target data model in the presence of unknown noise has the form:
K
y(n, ) ake^ulkn+Jkn'> + e(n, ), n = 0,1, , N 1, fi = 0,1, , 1,
k=1
(3.1)
where N and denote the numbers of data samples in range and crossrange, respec
tively; ok and {ujk, 4}, respectively, denote the unknown complex amplitude and 2D
18
40
modulated by 2cos[7r(/a fb)n + (a 4>b)/2]. Thus the data model in (4.1) cannot
be used to describe each of the two corner reflectors in this case.
Type 3: Two different scatterers located in the same range
Assume that two different corner reflectors with parameters {^()}^1 f, /}?=1
are located in the same range. Then the target model in the absence of noise has the
form:
r(n, n) = Xi{n)e^ieP2ir^n+^in\ n = 0,1, , N 1, n = 0,1, , N 1.
=1
(4.17)
With straightforward calculations, we can rewrite (4.17) as
2
r(n,) = Xj () e3^' eJ'27r(A n+/"); (4.18)
2=1
where 2 = i + 7t/2 with denoting an arbitrary phase, f1=f2 = f,f1 = f2 = f
with / denoting an arbitrary crossrange location,
and
i(n) = ^2xi(n) cos[2vr(/i f)n + (<^ 1)],
=i
2(n) = sin[27r(/ /)+(& ~ 0i)]
2=1
(4.19)
(4.20)
Note from (4.18) that {0, {(n)}^r=01,/,/,}=1 are the ambiguous features of
{cj)U {xi()}%~,f,fi}2i=v
Type 4: Multiple scatterers located in the same range
When more than two scatterers are located in the same range, the data model in the
absence of noise can still be written as (4.18) except that
L
i() = 5^z()cos[27r(/j f)+ ( &)], (4.21)
2=1
and
X2(n) = Y^xi{n) sin[27r(/ f)n+ (<& 0i)],
2 = 1
(4.22)
33
Figure 3.3: Target photo taken at 45 azimuth angle.
38
and
zNil zNi2
, i = 0,1, , N M,
(4.7)
"NiM
respectively; finally, (3(u) exp{j(N l)u>}. It has been shown in Appendix C
of [38] that by using a certain conjugate symmetric condition, a(u) in (4.4) can be
reduced to
a(uj) = hH (lo)Z(uj). (4.8)
The adaptive FIR filter h.APEs{u) corresponding to APES is derived in Ap
pendix A of [38] and has the form:
Q 1(c<;)a(a;)
Â¡APES
M =
where the steering vector a(o>) is given by
(4.9)
a(u) =
1 T
l gjw . ej(Ml)t
(4.10)
and Q(oi) denotes the estimate of the covariance matrix of the noise and is given by:
Q(w) = R Z(w)Zff(u>),
(4.11)
with
ZM = ^=[ Z(w) ZH] (4.12)
and R = (R + R)/2 denoting the average of the forward and backward sample
covariance matrices R and R. The forward sample covariance matrix R is obtained
from Zi as follow:
NM
R =
N
 y
M + l z
i=0
Zjzf,
(4.13)
and the backward sample covariance matrix R can be obtained similarly from z
4.3 Data Model Ambiguities and Their Effects on SAR Image Formation
Target feature extraction methods devised based on the data model in (4.2)
are robust against data mismodeling errors due to the model flexibility. However,
85
where
K
2(12, m) = ^ ^keRAfkixm{n)+'ikiym(,n)+JkiZm{n)}^ n = 0,1,  , TV 1, m = 1, 2, , M,
k=1
(6.28)
and tXm(n), iym(n), and iZm(n) are the same as iIm(n), tyrri(n), and tZm(n), respec
tively, except that {9m, (Â¡)m}!^=1 are replaced by obtained in Step (1).
This step is similar to [43] and we have
Vm = angle(z^ym), m = 1, 2, , M,
(6.29)
where
and
ym
\ T
2/(0, m) 2/(1, m)
y(N 1, m)
Zo
.n
1 T
(6.30)
(6.31)
2(0, m) z2(l,m) z2(Nl,m)
Note from (6.29) that we do not need the search over parameter space to determine
{Vm}m=1 and hence the errors in {R0(m)}^=l are easier to deal with than those in
{emAm}M
m= 1
Step (3): Repeat Step (1) by replacing y(n,m) with y(n,m) = y(n,m)e m,
m = 1,2, , M, where {r)m}^f=1 are determined in Step (2).
Step (4): Repeat Steps (2) and (3) until practical convergence, which is deter
mined by the relative change of the cost function Cm in (6.25) between two consec
utive iterations. In the numerical examples, we terminate the repetition of Steps (2)
and (3) when the relative change of Cm is less than 10~3 between two consecutive
iterations.
Note that if the errors in {R0(m)}^=1 are known to be negligible, then Step
(1) alone is sufficient for the autofocusing.
6.2.2 Target Feature Extraction
Assume aperture parameter estimates {9m,
problem becomes the target feature extraction problem considered in Chapter 5. As
78
(a)
(b)
Figure 5.4: 3D plots of K = 8 scatterers extracted by using RELAX with the indoor
experimental data, (a) Obtained with full aperture as shown in Figure 1.2(a). (b)
Obtained with curved aperture as shown in Figure 1.2(c).
32
(a)
(b)
# #
(e) (f)
Figure 3.2: RELAX Images by using the 20 x 20 ERIM data with K = 41. (a)
Unwindowed 2D FFT image, (b) Windowed 2D FFT image, (c) Unwindowed
RELAX image without background clutter, (d) Windowed RELAX image without
background clutter, (e) Unwindowed RELAX image with background clutter, (f)
Windowed RELAX image with background clutter.
43
Step (2): Apply the normalized 1D FFT to each column of Y to obtain
an intermediate matrix and then apply 1D APES to each row of the intermediate
matrix. (See [41] for the efficient implementation of APES.) Note that the normalized
1D FFT has the form
rN1
(4.28)
71=0
4.3.3 Model Ambiguity Effects on SAR Image Formation
All of the aforementioned types of ambiguities will have no effect on the SAR
image formation if no parameter estimation errors exist since the scatterers will be
perfectly reconstructed by using any of the possible ambiguous data models. For
example, when there are two identical scatterers located in the same range, the data
model in (4.2) can still be used for SAR image formation since the two scatterers are
now described as one scatterer with (4.16), which still fits the data model of (4.2)
with K = 1. Hence the original SAR image can still be reconstructed by using the
parameters of the one scatterer described by the right side of (4.16).
In the presence of parameter estimation errors due to the presence of noise
and clutter, however, Types 1 and 2 ambiguities discussed in Section 3.1 will have
little effect on SAR image formation, whereas Types 3 and 4 ambiguities can result in
artifact problems for the high resolution SAR image formation. Generally speaking,
the higher the SNR, the more accurate the parameter estimates and hence the less
significant the artifact problem. The effect of Type 3 ambiguity on SAR image
formation in the presence of estimation errors is illustrated in Figures 4.1 and 4.2.
(The effects are similar for Type 4 ambiguity.) Figure 4.1 is obtained by assuming
no parameter estimation errors. Figure 4.1(a) shows the FFT image of a target
consisting of two dihedrals of different lengths located in the same range. We use
xk(n) = aksmc[bkn(n rk)], k = 1,2, n = 0,1, 31,
(4.29)
81
Let
tn
4tt/o
(6.10)
It can be shown that for very small A0m and 7r/0 7(t r0), where \t ro < To/2,
we have
and
icos 0m to(sin 00)A0m,
(6.11)
w 0 O
~ to (COS 0Q) 0m,
(6.12)
O O
tsin 0m +i0(cos 0o)A0m,
(6.13)
where 0O is the average of all 0m, m = 1, 2, , M, and hence is a constant. Then
^ exp {j2tt(fktXm + fktym + AO} exp j j2?r
/fct COS (f)rn + fk^O COS (Â¡>q9ti
+fkt sin 0
exp < j27r
fkh sin o +/fcio cos 0O
A0
(6.14)
Equation (6.14) shows that if A0m is a constant, then the phase error due to A0m
and the phase of are ambiguous. Hence the phase of can never be determined
O ~
exactly in the presence of A0m. If A0m is a linear function of 9m, then fk and the
linear phase error due to A0m are ambiguous and cannot be determined exactly.
o
Similarly, we can analyze the errors in {0m}m=i when {0m =0m, i?0(m) =
Ro}m=v Let
=#m +A0m. (6.15)
o
For very small 9m and A9m and for 7r/0 7(i r0), where i r0 < T0/2, we have
O
tXm ~ t COS 0m,
o
~ to COS 0o (9m TA9m) 1
(6.16)
(6.17)
and
o
i sin 0m .
(6.18)
53
with and (g>, respectively, denoting the elementwise product and Kronecker prod
uct of two vectors.
H =
(4.59)
and e = vec(E).
Next, we derive the CRB matrix for the data model in (4.57) where the noise
covariance matrix is assumed to be arbitrary and unknown. Let Q E{eeH} be
the noise covariance matrix with ()H denoting the conjugate transpose. Then the
extended SlepianBangs formula for the ijth element of the Fisher information matrix
(FIM) has the form [3, 64]:
{FIM},, = tr (q'Q'CT'Q;) + 2Re [(/A1)! Q_1 (A/i)j] (4.60)
where X) denotes the derivative of X with respect to the th unknown parameter,
tr(X) denotes the trace of X, and Re(X) denotes the real part of X. Note that FIM
is block diagonal since Q does not depend on the parameters in (A/x), and (A/x)
does not depend on the elements of Q. Hence the CRB matrix for the target features
of interest can be calculated from the second term on the right side of (5.31). Let
(4.61)
where
and
V =
fT
fT xe
(f) =
4> 2
f =
fi
2
K
f 
h
h
Ik
xe =
T
T
X2
T
XK
1 t
i T
(4.62)
Let Fi, F2, and F3 denote NN x K matrices whose A;th columns are determined,
respectively, by
j*k {[u(/i)0xJ8m(/i)),
(4.63)
14
Recently, RELAXNLS [42] was introduced for the feature extraction of targets
consisting of both trihedrals and dihedrals based on a mixed data model rather than
the point scatterer data model. An ideal trihedral acts as a triplebounce reflector
and can be modeled as a strong point scatterer [53]. The point scatterer data model,
however, is no longer valid in crossrange for an ideal dihedral corner reflector, which
acts as a doublebounce reflector, since its radar cross section (RCS) is described by
a sine function, i.e., sinc(0) = sin(0)/0, where 0 denotes the angle between the radar
beam and the line perpendicular to the dihedral.
Super resolution SAR images obtained via modern spectral estimation meth
ods provide more details required for the detection and automatic target recognition
(ATR) than the conventional FFT based methods. Quantitative results showing the
advantages of using these sophisticated spectral estimation methods in SAR appli
cations begin to appear in the literature. For example, it was shown in [51, 50] that
a significant improvement in ATR performance can be achieved by using SAR im
ages with enhanced resolution obtained via the reducedrank variations of the Capon
method.
Compared with nonparametric spectral estimators, high resolution paramet
ric spectral estimators provide great potential for SAR imaging with better resolution
and lower sidelobes when the assumed data model is valid. On the other hand, para
metric methods are more sensitive to data model mismatch than nonparametric
methods. Hence, practically only robust super resolution parametric methods can be
used for SAR image formation. We will use robust relaxationbased algorithms in
cluding RELAX and RELAXNLS to form SAR images of different manmade targets
in our work. In cases where an appropriate data model of manmade targets is not
available in crossrange, robust nonparametric spectral estimators with good perfor
mances such as APES and high resolution parametric methods may work together to
obtain satisfactory results. To the best of our knowledge, a hybrid of nonparametric
Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy.
RELAXATIONBASED METHODS FOR SAR TARGET
FEATURE EXTRACTION AND IMAGE FORMATION
By
Zhaoqiang Bi
May 1999
Chairman: Dr. Jian Li
Major Department: Electrical and Computer Engineering
Synthetic aperture radar (SAR) has been a mature but actively researched
technology due to its dayandnight and allweather capability of offering high res
olution imaging for both military and civilian applications. As the foundation of
automatic target detection and recognition, SAR imaging and autofocusing continue
to attract more research interest. Nonparametric spectral estimation methods are
robust methods for SAR image formation. However, nonparametric methods cannot
be used to significantly improve the resolution of the formed SAR images since they
generally do not fully exploit the characteristics of radar targets of interests even
when such information is available.
In this dissertation, efforts have been made to form super resolution two
dimensional (2D) SAR images via relaxationbased parametric methods. The
relaxationbased optimization methods have been proved to be quite useful in several
x
97
moving target images obtained via conventional SAR processing algorithms are typi
cally mislocated and smeared in crossrange due to phase errors and range migrations
induced by the motions. The amount of smearing generally depends on the charac
ter and magnitude of the target motion during the data coherent collection interval
(CTI). The future work to be conducted includes the establishment of an effective
ISAR data model considering range migrations and phase errors and how to apply
the relaxationbased optimization approach to simultaneous ISAR autofocusing and
image formation.
107
[45] C. E. Mandil and J. M. Swiger. A map drift autofocus technique for correcting
higher order sar phase errors. 27th Annual TriService Radar Symposium Record,
Monterey, CA, pages 391400, June 1981.
[46] S. L. Marple, Jr. Digital Spectral Analysis with Applications. Upper Saddle
River, Prentice Hall NJ, 1987.
[47] R. P. Marzban. Interferometric Synthetic Aperture Radar Image Formation and
Segmentation. M.S Thesis, University of Florida, Ganiesville, FL, 1996.
[48] Dean L. Mensa. High Resolution RADAR Imaging. Artech House, Inc., 610
Washington St., Dedham, MA, 1981.
[49] D. C. Munson, Jr., J. D. OBrien, and W. K. Jenkins. A tomographic formulation
of spotlightmode synthetic aperture radar. Proceedings of the IEEE, 71:917 
925, August 1983.
[50] L. M. Novak, G. R. Benitz, and G. J. Owirka. Classifier performance using
enhanced resolution SAR data. Proceedings of the IEE International Conference
on Radar, pages 634638, Edinburgh, UK, October 1997.
[51] L. M. Novak, G. R. Benitz, G. J. Owirka, and L. A. Bessette. ATR performance
using enhanced resolution SAR. SPIE Proceedings on Optical Engineering in
Aerospace Sensing, Orlando, FL, pages 332337, April 1996.
[52] F. Prodi, F. Vinelli, and F. Dente. An autofocus techniques of SAR data based
on superresolution. International Geoscience and Remote Sensing Symposium,
pages 15861588, August 1993.
[53] August W. Rihaczek and Hershkowitz Steven J. Radar Resolution and Complex
Image Analysis. Artech House Inc., Norwood, MA, 1996.
[54] E. A. Robinson. A historical perspective of spectrum estimation. Proceedings of
the IEEE, 70:885907, September 1982.
[55] E. Rodriguez and J. M. Martin. Theory and design of interferometric synthetic
aperture radars. IEE Proceedings F, 139(2):147159, April 1992.
[56] R. Roy and T. Kailath. ESPRIT Estimation of signal parameters via ro
tational invariance techniques. IEEE Transactions on Acoustics, Speech, and
Signal Processing, ASSP37(7):984995, July 1989.
[57] G. T. Ruck, D. E. Barrick, W. D. Stuart, and C. K. Krichbaum. Radar Cross
Section Hand Book. Plenum Press, New York, 1970.
[58] R. O. Schmidt. Multiple emitter location and signal parameter estimation. IEEE
Transactions on Antennas and Propagation, AP34(3):276280, March 1986.
[59] C. W. Sherwin, J. P. Ruina, and R. D. Rawcliffe. Some early developments
in synthetic aperture radar system. IRE Transactions on Military Electronics,
MIL6(2):111115, April 1962.
[60] Paul L. Smith, Jr., Kenneth R. Hardy, and Kenneth M. Glover. Applications
of radar to meteorological operations and research. Proceedings of the IEEE,
62(6):724745, June 1974.
CHAPTER 6
AUTOFOCUS IN CLSAR
To analyze the effect of CLSAR aperture errors, we first rewrite the CLSAR
data model in (5.12) with respect to radar viewing angles. Assume that a curvilin
ear aperture consists of M different viewing angles and let denote the
elevation and azimuth angle pairs of the M look angles of the radar. Let y(n,m),
n = 0,1, , AT 1, denote the onedimensional (1D) data samples obtained after
dechirping from the mth viewing angle of the radar. Let
4(7t/o 7To + 7tn)
t"n.
n = 0,1, , N 1,
(6.1)
where tn denotes the time samples. Let tXjn (n), tVm (n), and tZm (n),respectively, denote
the time samples of the mth look angle, where
and
tXrn (n) = tncos(9m) cos(0m),
(6.2)
tym{n) = in sin(9m) COS(0m),
(6.3)
(^) n sin(0m),
(6.4)
with m = 1, 2, , M and n = 0,1, , AT 1. Then y(n, m) has the form
K
y(n, m) = ^ akej2AfktxmW+fktym(n)+fktzm(n)\ A + e(n, m), m 1, 2, , M,
Jb=l
(6.5)
where e(n, m) denotes the noise and clutter and K is the number of scatterers.
We have assumed that {0m,0m}m=i and Ro{m) = Ro, m = 1,2, , M, are
known exactly in the CLSAR data (6.5). For a practical curvilinear SAR system, how
ever, the radar positions relative to the XYZ coordinate system may not be known
79
23
3.2 RELAXNLS
When targets consist of both trihedral (point scatterers) and dihedral corner
reflectors, we can form enhanced SAR images by using the RELAXNLS algorithm
[42], We first give a brief review of the RELAXNLS algorithm and then discuss SAR
image formation via RELAXNLS.
3.2.1 Review of the RELAXNLS Algorithm
RELAXNLS [42] is devised to identify and obtain target features of multiple
corner reflectors (trihedral and dihedral) based on a mixed parametric data model.
An ideal dihedral corner reflector acts as a doublebounce reflector and the RCS
achieves the maximum when the radar beam is perpendicular to the dihedral and
falls off as a function of sinc(0) = sin(0)/9 when the angle between the radar beam
and the line perpendicular to the dihedral increases [49, 57]. Then dihedrals can be
modeled as
Kd
sd(n,n) = ^ adksmc{nbk(nTk)]e3{Udkn+Uikn\n = 0,1, , Nl, = 0,1, , N1,
k=1
(3.14)
where N and denote the numbers of the available data samples in range and
crossrange, respectively; adk, {ujdk,dk}, and bk, k = 1, 2, , Kd, are, respectively,
proportional to the maximal RCS, the central location, and the length of the A:th
dihedral corner reflector; rk, k 1, 2, ,Kd, denotes the peak location of the data
sequence in crossrange and is determined by the orientation of the fcth dihedral
corner reflector; finally, Kd is the number of the dihedral corners. Assume that a
target consists of Kd dihedrals and Kt trihedrals and K = Kd + Kt Then the target
data model in the presence of noise and clutter has the form:
y(n, ft) = sd(n, ) + s(n, ) + e(n, ), n = 0,1, , N 1, = 0,1, , 1,
(3.15)
36
The received signal reflected from a target scatterer herein is modeled as:
s(n, ) = x()e?2*Un+fn\ n = 0,1, , N 1, = 0,1, , 1, (4.1)
where N and denote the dimensions of the available data samples in range and
crossrange, respectively; x() is an arbitrary unknown realvalued function of
determined by the radar cross section (RCS) of the scatterer; 0 is a constant phase;
finally, {/, /} is the frequency pair proportional to the 2D location (range and cross
range) of the scatterer. This data model is essentially semiparametric since little
parameterization is done in crossrange.
Assume that a target consists of K scatterers. Then the target data model in
the presence of noise has the form:
K
y(n, ) = ^ Xk{)e^k2R^knJrk1^ + e(n, ), n = 0,1, , N 1, = 0,1, , 1,
k=1
(4.2)
where {^()}^1 denotes the realvalued amplitude function of for the kth scat
terer; cf)k and {.fk,fk}, respectively, are the constant phase and the frequency pair
of the A;th scatterer; finally, {e(n,)} denotes the unknown 2D noise and clutter
sequence.
Since SAR images are often used in SAR applications, our problem of interest
herein is to estimate the target parameters {fik, {xk()}%~Q, fk, fk}k=1 from the 2
D data sequence {y(n, )} and then to form high resolution SAR images with the
estimated target parameters. Before presenting the effects of the semiparametric
data model in (4.2) on SAR image formation and the SPAR (SemiPARametric)
algorithm for feature extraction and image formation, we first briefly introduce the
APES algorithm.
45
larger the extrapolation factor Â£, the more significant the artifact problem since the
difference between fx and f2 is exaggerated Â£ times. Figures 4.2(d) shows the SAR
image obtained with Â£ = 2. By comparing Figures 4.2(d) and (c) (here Â£ = 1,
without extrapolation), we note that the artifact next to the short dihedral becomes
more significant. Severe artifacts may exist at low SNR since the accuracy of the
parameter estimates is poor. The SPAR algorithm we present below attempts to
avoid this problem by using windows to isolate the multiple scatterers located in the
same range.
4.4 The SPAR Algorithm
SPAR can be summarized by the following two steps:
Step 1: Scatterer Isolation based Target Feature Extraction: See Chapter
4.4.1 below for details.
Step 2: SAR Image Formation: See Chapter 4.3.2 for details, where the
estimated target features are obtained by using Step 1.
4.4.1 Target Feature Extraction
The basic idea behind SPAR is to extract the features of each scatterer sepa
rately. Before we present the target feature extraction algorithm, we first summarize
the steps needed for the feature extraction of a single scatterer as a preparation. The
generalized Akaike information criterion (GAIC) is also introduced to estimate K,
the number of scatterers, at the end of this subsection.
Feature Extraction of a Single Scatterer
The data model of a single scatterer in the presence of noise has the form:
ys(n, ) = s(n, ) + es(n, ), n = 0,1, , IV 1, = 0,1, , 1, (4.30)
106
[30] S. M. Kay. Modern Spectral Estimation: Theory and Application. PrenticeHall,
Englewood Cliffs, N.J., 1988.
[31] S. M. Kay and S. L. Marple, Jr. Spectrum analysis a modern perspective.
Proceedings of IEEE, 69(11) :13801418, November 1981.
[32] Klauder, Jr., A.C. Price, S. Darlington, and W. J. Albersheim. The theory and
design of chirp radars. The Bell System Technical Journal, 39(4):745808, July
1960.
[33] Kenneth Knaell. Threedimensional SAR from curvilinear apertures. SPIE Pro
ceedings on Optical Engineering in Aerospace Sensing, 2230:120134, Orlando,
FL, April 1994.
[34] Kenneth Knaell. Threedimensional SAR from practical apertures. SPIE Pro
ceedings on Optical Engineering in Aerospace Sensing, San Diego, 2562:3141,
1995.
[35] Kenneth Knaell. Threedimensional SAR from curvilinear apertures. Proceedings
of the 1996 IEEE National Radar Conference, pages 220225, Ann Arbor, MI,
May 1996.
[36] V. Larson and L. Novak. Polarimetric subspace target detector for SAR data
based on the Huynen dihedral model. Proceedings of SPIE, 2487:235250, Or
lando, Florida, April 1995.
[37] Hongbin Li, Jian Li, and Petre Stoica. Performance analysis of forwardbackward
matchedfilterbank spectral estimators. IEEE Transactions on Signal Processing,
46(7): 19541966, July 1998.
[38] J. Li and P. Stoica. An adaptive filtering approach to spectral estimation and
SAR imaging. IEEE Transactions on Signal Processing, 44(6):14691484, June
1996.
[39] J. Li and P. Stoica. Efficient mixedspectrum estimation with applications to
target feature extraction. IEEE Transactions on Signal Processing, 44:281295,
February 1996.
[40] Jian Li, ZhengShe Liu, and Petre Stoica. 3D target feature extraction via
interferometric SAR. IEE Proceedings on Radar, Sonar and Navigation, 144(2),
April 1997.
[41] ZhengShe Liu, Hongbin Li, and Jian Li. Efficient implementation of Capon and
APES for spectral estimation. IEEE Transactions on Aerospace and Electronic
Systems, 34(4):13141319, October 1998.
[42] ZhengShe Liu and Jian Li. Feature extraction of SAR targets consisting of
trihedral and dihedral corner reflectors. IEE Proceedings on Radar, Sonar and
Navigation, 145:161172, March 1998.
[43] ZhengShe Liu and Jian Li. SAR motion compensation and feature extraction
via MCRELAX. Journal of the Optical Society of America A, 15(3):599610,
March 1998.
[44] C. E. Mancill. Enhancement of radar imagery by maximum entropy processing.
Proceeding of the Human Factors Society 21 st Annual Meeting, pages 241243,
1977.
72
Let
F= A j'A Du Do Do
(5.36)
where A is defined in (5.20), the Ath columns of D^, D, and Dq are d[a.k
afc]/3wfc, 3[afc <8> fc (g> afc]/<9d)fc, and d[afc afc afe]/du)fc, respectively, and
(5.37)
Then the CRB matrix for the parameter vector rj is given by:
CRB(rj) = [2Re(FHQ_1F)]_1.
(5.38)
For the case of curvilinear aperture, the CRB matrix for the target parameters
is similar to the one in (5.38) except that the A in (5.38) is now replaced by Ac.
5.4 Numerical and Experimental Results
We first use CRBs to investigate the performances of different curved apertures
for target feature extraction and then demonstrate with an experimental example the
performance of the RELAX algorithm for 3D target feature extraction using CLSAR.
5.4.1 Performance Analysis of Different Curvilinear Apertures via CRBs
Since CRBs are the best unbiased performance any asymptotic estimator can
achieve, we not only compare CRBs of target features when different curved apertures
are used for 3D target feature extraction, but compare them with the CRB of the
full aperture. Without loss of generality, let us consider the case of a single scatter
with a. 1 and u Co = = 0. The additive noise is assumed to be zeromean
white Gaussian with variance a2 = 40. The full aperture data is generated according
to (5.16) with N = = = 32.
Consider first the example of M = 63. The curved apertures we consider
include the parabolic one in Figure 1.2(b), the Lshaped one in Figure 1.2(c), and
the ones in Figure 5.2. For the arc apertures, the Arc1 aperture is a quarter of a
59
(a)
(b)
(c) (d)
Figure 4.2: Ambiguity effect on the SAR image formation in the presence of range
estimation errors, (a) Windowed FFT image of the first scatterer. (b) Windowed
FFT image of the second scatterer. (c) Combined windowed FFT image of the
two scatterers with C = 1 (without extrapolation), (d) Combined windowed FFT
image with Â£ = 2. (The vertical and horizontal axes are for range and crossrange,
respectively.)
Devotion (meters')
90
Figure 6.1: (a) Curvilinear apertures for the experimental example, (b) True scatterer
distribution, (c) 3D plot of K = 20 scatterers extracted from the initial aperture
(dashed line) in (a) with RELAX, (d) 3D plot of K = 20 scatterers extracted from
the manually adjusted aperture (solid line) in (a) with RELAX.
26
Identification of Crner Reflector
It has been shown in [42] that Cd is much less than Ct when the true corner
reflector is a dihedral and Ct is almost the same as Cd when the true corner reflector
is a trihedral. Also b 26f when a trihedral is mistakenly considered as a dihedral,
where Sf is the smallest positive solution to
sm(N5fir)
N
~2~'
(3.27)
sin(5/7r)
Then a corner reflector can be identified by checking the estimate b and the cost
functions Ct in (3.19) (with {at,u)t, ojt} replaced by {at,u)t,u)t}) and Cd in (3.24)
(with {cOdi Cbd, Od, b, r} replaced by {j, uid, b,f}. A corner reflector is identified as
a dihedral if
cd
>0.1 and b > 25f.
Steps of RELAXNLS are outlined as follows:
Step (1): Assume K 1. Identify and obtain the parameters of a single
corner reflector.
Step (2): Assume K = 2. Subtract the signal obtained in Step (1) from
{y(n,)} and then identify and obtain parameters of the second corner reflector.
Subtract the second signal from {y(n,n)} and reidentify and redetermine the pa
rameters of the first signal.
Iterate the previous two substeps until practical convergence.
Step (3): Assume K = 3. Subtract the two signals obtained in Step (2) from
{y(n,)} and then identify and obtain the parameters of the third signal. Subtract
the second and third signals from {y(n, )} and reidentify and redetermine the first
signal. Subtract the first signal (most recently determined) and the third signal from
{y(n,n)} and then reidentify and redetermine the second signal.
Iterate the previous three substeps until practical convergence.
Remaining Steps: Continue similarly until K K (K can be determined
by using GAIC to be discussed later on).
I certify that I have read this study and that in my opinion it conforms to
acceptable standards of scholarly presentation and is fully adequate, in scope and
quality, as a dissertation for the degree of Doctor of Philosophy.
Jian Li Chairman
Associate Professor of Electrical and
Computer Engineering
I certify that I have read this study and that in my opinion it conforms to
acceptable standards of scholarly presentation and is fully adequate, in scope and
quality, as a dissertation for the degree of Doctor of Philosophy.
Jos C. Principe
Professor of Electrical and
Computer Engineering
I certify that I have read this study and that in my opinion it conforms to
acceptable standards of scholarly presentation and is fully adequate, in scope and
quality, as a dissertation for the degree of Doctor of Philosophy.
Frederick J. Taylor
Professor of Electrical and
Computer Engineering
I certify that I have read this study and that in my opinion it conforms to
acceptable standards of scholarly presentation and is fully adequate, in scope and
quality, as a dissertation for the degree of Doctor of Philosophy.
William W. Edmonson
Assistant Professor of Electrical and
Computer Engineering
I certify that I have read this study and that in my opinion it conforms to
acceptable standards of scholarly presentation and is fully adequate, in scope and
quality, as a dissertation for the degree of Doctor of Philosophy.
David C. Wilson
Professor of Mathematics
CHAPTER 2
LITERATURE SURVEY
In this chapter, we give a brief review of the topics related to the work, namely
super resolution 2D SAR image formation, 3D SAR target feature extraction, and
SAR motion compensation.
2.1 2D Super Resolution SAR Image Formation
Since SAR image formation can be considered as a spectral estimation prob
lem, modern spectral estimation methods find extensive applications in improving
the resolution and accuracy of the formed SAR images. Excellent texts on spectral
estimation are available [63, 30, 46] and historical and modern perspectives on the
general topics of spectral analysis can be found in [54, 31]. A brief review of super res
olution SAR image formation via nonparametric and parametric spectral estimation
methods is given below.
Nonparametric Spectral Estimators
An important class of nonparametric spectral estimators are the adaptive
filterbank approaches which can be used to form SAR images with better resolu
tion and lower sidelobes than the FFT based methods [15]. The adaptive sidelobe
reduction (ASR) approach [14] and the widely used Capon method [10] as well as
its reduced rank variations [5, 6] are members of this class. A recently introduced
matchedfilter bank based complex spectral estimator that is of interest is the APES
(Amplitude and Phase Estimation of a Sinusoid) method [38]. Despite their different
derivations, it was found that both Capon and APES are members of the general
class of the matchedfilterbank (MAFI) approaches to complex spectral estimation
11
51
scatterers. However, the Hybrid SAR images may suffer from more severe artifact
problem than SPAR, especially at low SNR. When no multiple scatterers are located
in the same range, SPAR and Hybrid perform similarly.
4.5 Modified RELAXNLS Algorithm
RELAXNLS [42] is a parametric approach for the feature extraction of targets
consisting of both trihedrals and dihedrals. It is based on a mixed data model in
which x(n) is modeled as a realvalued constant for a trihedral or a sine function of
for a dihedral. Like SPAR, RELAXNLS extracts the target features by minimizing
an NLS cost function via a relaxationbased approach. However, RELAXNLS is
computationally expensive since a 4D search over the parameter space is required
for dihedral corner reflectors. Since SPAR is more robust and computationally more
efficient than RELAXNLS, the former can be used to provide the initial conditions
needed by the latter.
Let {fa, fa, fk, fk}k=i denote the parameter estimates obtained via SPAR
according to the data model in (4.2), where K is the estimated number of scatterers
obtained via GAIC. The SPAR estimates {fa, fa, fk, fk}k=1 cannot be used directly
as initial conditions for RELAXNLS. The initial conditions are obtained by applying
the first step of RELAXNLS [42] to each Ufc, k 1,2, , K, where
k
Ufc = Y Â£ e**G ifrjiji), k = 1,2, , K. (4.50)
=l^zk
Once we have the initial conditions, we can use the last step of RELAXNLS [42] to
obtain the dihedral and trihedral parameter estimates, which are then used for SAR
image formation [7]. We refer to this approach as the modified RELAXNLS.
69
parameter estimates for the first K 1 scatterers can be used to initialize the search
for the parameters of the Kth one. We first briefly make a preparation for the RELAX
approach. Let
R
y/c = y a(i <8> ^), (5.25)
A A A
where , , and are formed, similarly to those in Equation (5.21), from
and respectively, and {)j,Â¡)j,u;j, are assumed to have been estimated.
Then minimizing C4 in (5.24) with respect to ak yields the estimate k of ak:
Oik
[a* <8) afc (g) afc]H yfc
NNN
Uk Uk t^kUk !<*>k =Uk
(5.26)
and
{A, O)fc, CUfc}
arg max
la* at a*]Hyfc
(5.27)
Hence {0fc, Dfc, } can be obtained as the location of the dominant peak of the 3D
periodogram,
h 2
[a* afe 0 a*] yk
N
which can be computed efficiently with 3D FFT. Note that padding with zeros
for the 3D FFT is necessary to achieve high accuracy. An alternative approach is
to find an approximate location corresponding to the global maximum with the 3D
FFT without much zeropadding and then use the approximate location as the initial
condition to find a more accurate position via a multidimensional search method, such
as the FMINV function in PVWAVE. We used the latter approach in our examples
presented in Section 5.4.2. Note that ak is easily computed from the complex height
of the peak of [a*, kk ak]H yk/(N).
With the above preparations, we now proceed to present the steps of the
RELAX algorithm for 3D target feature extraction with the full aperture SAR.
Step (1): Assume K = 1. Obtain {(2>k, k, G)k, ak}k=i from y by using (5.27) and
(5.26).
CHAPTER 4
SEMIPARAMETRIC METHODS FOR SAR IMAGING
4.1 Problem Formulation and Data Model
Compared with parametric spectral estimation methods, nonparametric spec
tral estimation methods are robust, but cannot be used to significantly improve the
resolution of the formed SAR images. As can be seen from the previous chapter that
relaxationbased parametric methods can be used to generate super resolution SAR
images when the data model for targets of interest is appropriate. However, the super
resolution property of parametric methods will be compromised when the appropriate
data model is not available. In this chapter, we consider the case where it is difficult
to establish a good parametric data model in crossrange for target scatterers due
to a variety of physical phenomena including glints, resonances, and motioninduced
phase errors, etc. Instead of modeling the target of interest as 2D sinusoid (point
scatterer) in both range and crossrange, we apply herein a semiparametric model
and simply assume that the data model in crossrange due to one or more corner re
flectors is a complex sinusoid with an arbitrary amplitude and a constant phase. Due
to the flexibility of this data model, feature extraction and image formation meth
ods based on this data model should be more robust compared with those based on
the approximated dihedral and trihedral data models in Chaper 3. The robustness,
however, comes at the cost of small resolution improvement in crossrange over the
conventional fast Fourier transform methods due to assuming little in crossrange. To
preserve the robustness, we use a refined nonparametric spectral estimation method
such as APES (Amplitude and Phase Estimation of a Sinusoid) algorithm [38] to
improve the crossrange resolution.
35
(a)
(b)
(c)
(e) (f)
Figure 4.3: Comparison of SAR images formed using different algorithms for simu
lated data at high SNR (a% = 0.6). (a) True SAR image, (b) Windowed 2D FFT
SAR image, (c) Hybrid SAR image, (d) SPAR SAR image, (e) RELAXNLS SAR
image, (f) Modified RELAXNLS SAR image. (The vertical and horizontal axes are
for range and crossrange, respectively.)
30
not only have a better resolution than the 2D FFT images, but also more closely
reflects the characteristics of the photo shown in Figure 3.3.
50
{^fc, 5c*Â¡, /fc, /fc}*=i,3, calculate Y2 from Y2, and then redetermine {0fc, xfe, fk, /fc}fc=2
from Y2.
Iterate the previous three substeps until practical convergence.
Remaining Steps: Continue similarly until K is equal to the desired or
estimated number of scatterers.
The practical convergence in the iterations of the above relaxationbased
NLS algorithm may be determined by checking the relative change e of the cost
function C5 f{0fc,Xfc,/fc,/fc}in (4.44) between two consecutive iterations. Our
numerical and experimental examples show that the algorithm usually converges in
a few iterations.
We can determine K, the number of scatterers in (4.2), by extending the
generalized Akaike information criterion (see [61] for details). By assuming that the
noise is white, the estimate iY of iY is determined as an integer that minimizes the
following extended GAIC cost function:
(N11 \
EE \pc(n,n)\2\ + y\n[\ny(NN)][K(N + 3) + 1], (4.48)
n=0 =0 /
where 7 is a constant of user choice,
k
fcin, ) = y(n, )Xk{n)e^ke:2'K^kn+^kn\ n = 0,1, , TV1, = 0,1, , TV1,
fc=i
(4.49)
and Tf(TV + 3) + 1 is the total number of realvalued unknown parameters (of which
K(TV + 3) is for the scatterers and 1 is for the white noise variance).
Note that the NLS estimates of {(j)k, xfc, fk, fk}k=i can also be estimated from
determined in (4.45), rather than {Yfc}^, via the above relaxationbased
optimization algorithm. We refer to this approach as the Hybrid method. When
multiple scatterers are located in the same range, Hybrid may be computationally
more efficient than SPAR since Hybrid does not isolate the scatterers so that multiple
scatterers located in the same range can be more efficiently described as at most two
28
where ns = 0,1, , Â£N 1 and fis = 0,1, , ( 1 with Â£ being an extrapolation
factor (Â£ > 1); dk, Codk, u>dk, h, and fk, respectively, denote the estimates of adk,
Ujdk, Qdk, bk, and Tk, k = 1, 2, , Kd\ finally, Jfe, wifc, and utk, respectively, denote
the estimates of atk, cotk, and Qtk, k = 1, 2, , Kt. The clutter estimate (n, )
used in Step 3 is now determined by (3.29) with Kt and Kd replaced by Kt and Kd,
respectively. Note that scaling the simulated data for the dihedrals by a factor of Â£
in (3.30) is needed since the sine function goes to zero as ns increases or decreases
away from rk. Note also that Â£ is a parameter of user choice and Â£ > 1 is just
used to demonstrate the super resolution property of RELAXNLS. Finally, since the
Fourier transforms of the sine functions of sufficient lengths do not result in sidelobes,
when ys(ns,s) in (3.30) is windowed, the second term of ys(jis,fis) is multiplied by
ws(ns,s) and the first term is multiplied by wi(ns), where w\ins) is a 1D window
sequence satisfying
CAT1
w{na) = (N. (3.31)
7ls=0
3.3 Experimental Results
In this section, we demonstrate the image formation performances of RELAX
and RELAXNLS with the experimental data. In the following examples, the extrap
olation factor Â£ = 2 is used and GAIC with 7 = 4 and 7 = 18 are used to determine
K for RELAX and RELAXNLS, respectively. Kaiser windows with shape parameter
(3 6 are used whenever needed.
First consider SAR image formation via 2D FFT and RELAX by using a
portion of the 2D data corresponding to some roof rims collected by one of the two
apertures of the ERIMs (Environment Research Institute of Michigans) DCS inter
ferometric synthetic aperture radar (IFSAR). Figures 3.1(a) and (b), respectively,
show the unwindowed and windowed 2D FFT images obtained by zeropadding the
40 x 40 phase history data. The unwindowed and windowed RELAX images with and
46
where s(n,n) is given in (4.1) and {es(n,n)} denotes the unknown 2D noise and
clutter sequence. Let
UnU)
1 ej2irf
T
ej2nf(Nl)
(4.31)
and
Wff(/) =
\T
1 ej2nf . ej2Trf(Nl)
(4.32)
where ()T denotes the transpose. Let D(/) denote the following diagonal matrix:
D(/) = diagl, ..., eK*Wi)J.
(4.33)
Define
T 1T
x x(0) x(l) x( 1) (434)
Let Ys be an N x matrix with its (n, )th element being ys(n,). Then we can
rewrite (4.30) as:
Ya = e*G(x,/,/)+Es, (4.35)
where
G(x,/,/) = w(/)xTD(/), (4.36)
and Es denotes an N x TV matrix with es(n,n) being its (n,)th element. Let yS,
= 0, 1, ,iV 1, denote the th column of Ys and define
Mf) = Y WM), (4.37)
where ()* denotes the complex conjugate. Then the NLS estimates jx, , /,/j of
{x, , f, /} are (see Appendix A for the detailed derivations):
x = ^Re [eJ0ys(/) 0w^(/)] (4.38)
where Re (a:) denotes the real part of x and denotes the Hadamard matrix product
or the elementwise product of two matrices;
i = arg ( [y]>M/)]V22,2/,4 (4.39)
xml version 1.0 encoding UTF8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchemainstance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID EU7XFUN0J_IVOJ1T INGEST_TIME 20141223T18:19:36Z PACKAGE AA00024493_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
4
The crossrange resolution of stripmap SAR is limited by the azimuth beamwidth,
while that of spotlightmode SAR is independent of the azimuth beamwidth since the
antenna steers its beam to continuously illuminate the same scene. Hence Spotlight
mode SAR is more suitable for high resolution SAR image formation of a small scene
than stripmap SAR.
1.2 Scope of the Work
Spotlightmode SAR has been extensively applied to 2D high resolution im
age formation and target feature extraction of small scenes due to its finer resolutions
and readily controllable flight path, which is a straight line. The 2D target features
include the radar cross section (RCS) and the 2D location (range and crossrange) of
each scatterer. One topic of our work is super resolution 2D image formation from
phase history data collected by spotlightmode SAR. The super resolution herein is
defined as the resolution that is better than the Rayleigh limits determined by the
transmission wavelength and the length of the synthetic aperture of SAR. A conven
tional approach for 2D image formation from SAR phase history data is the fast
Fourier transform (FFT), which is computationally efficient and robust. However,
SAR images obtained via FFT suffer from poor resolution and high sidelobes. SAR
image formation can be considered as a spectral estimation problem since SAR images
are the reconstruction of scene reflectivity intensities versus space locations. Mod
ern spectral estimation methods including parametric and nonparametric methods
have found extensive applications in improving the resolution and accuracy of the
formed SAR images. In particular, parametric spectral estimation methods are more
suitable to form super resolution SAR images when parametric data models can be
used to model the target. The super resolution of the formed SAR images can be
achieved by extrapolating the phase history data beyond observations using an appro
priate data model and the parameter estimates. We will discuss how to form super
other applications, such as radio astronomy, microwave imaging, and spectral estima
tion. The relaxationbased methods are extended for super resolution SAR imaging
of radar targets consisting of only trihedrals or both trihedrals and dihedrals. We
have also devised a robust and computationally simple SPAR (SemiPARametric) al
gorithm for 2D SAR imaging based on a flexible semiparametric data model when
it is difficult to establish an accurate target data model in crossrange. Hence SPAR
takes advantages of both parametric and nonparametric spectral estimation meth
ods to form enhanced SAR images. Numerical and experimental examples have been
used to demonstrate the performances of the proposed algorithms. We have observed
that the relaxationbased parametric methods provide super resolution SAR images
when the assumed data model is valid; otherwise SPAR performs better.
Threedimensional (3D) target features, including the height information,
radar cross section (RCS), and 2D location (range and crossrange), provide quite
useful information for such applications as automatic target recognition. Thus, ef
forts have also been made in this dissertation to devise an effective relaxationbased
algorithm, referred to as AUTORELAX, for both 3D target feature extraction and
motion compensation via curvilinear SAR (CLSAR), a novel technology which is
still at its developing stage. The proposed AUTORELAX algorithm is shown to be
promising when evaluated by using both the experimental and simulated examples.
xi
24
where
Kt
st(n, n) = y^jatke^utkn+mkn\ n = 0,1, , N 1, = 0,1, , 1, (3.16)
k=1
with and {aHk,tk}k=v respectively, denoting the unknown complex am
plitudes and 2D unknown frequencies of the Kt trihedrals. The RELAXNLS al
gorithm proposed in [42] can be used to identify and effectively estimate the tar
get features{adk, bk, Lodk, Odk, Tk}kii and {atk, utk, tk by minimizing the following
NLS criterion:
C2 (^{atk > fc=l> iadk > bk,Udk dk rA:}fc=i^  Y St Sd\\F (317)
where [Y]y = y(i,j), [St]*j = st(i,j), and [Sd]ij = sd(i,j). RELAXNLS converts
the problem of multiple corner reflector identification and feature extraction to that
of a sequence of single reflector identification and feature extraction by using the
relaxation based optimization approach. The identification and feature extraction of
a single corner reflector in RELAXNLS is described as follows:
Feature Extraction of a Single Trihedral Corner Reflector
Let Yt denote the data of a single trihedral in the presence of unknown noise
Et. Then
Y t = cx.tu n + E, (3.18)
where at and respectively, are proportional to the RCS and 2D location
of the trihedral. The NLS estimates {at,ujt,tit} of {at, ut, Cot} can be obtained by
minimizing the following cost function
Ct {auuuQt) = Yj atuN(y
and respectively, are given by
v%(ut)Ytu;*()t)
NN
(3.20)
BIOGRAPHICAL SKETCH
Zhaoqiang Bi was born in Heilongjiang Province, China on October 15, 1968.
He attended Huazhong University of Science and Technology, Wuhan, China in
September 1986, where he earned his Bachelor of Engineering degree in electrical
engineering in July 1990. He earned his Masters degree in communication and elec
tronic systems from Communication, Telemetering, and Telecontrol Research Insti
tute (CTI) in Shijiazhuang, China in March 1993. He worked as an IC designer for
digital communication systems with advanced EDA (electronic design aided) tools in
CTI from 1993 to 1996. He has been a graduate research assistant in the Spectral
Analysis Lab (SAL) of the Department of Electrical and Computer Engineering at
the University of Florida since August 1996, during which time he has conducted
research in the area of radar signal processing, including target feature extraction
and synthetic aperture radar image formation.
109
82
Then
ah exp {j2v(fktXm + fktVm + /fctZm)} ~ ak exp l j2ir
o o
fkt cos 0m +fkt0COS 00Or
+fkt sin 0
exp j2Kfkt0 cos 0o A6,,
(6.19)
which shows that a constant A9m also results in the ambiguity between the phase
error due to A9m and the phase of ak. Hence the phase of ak can never be determined
O ~
exactly in the presence of A9m. Also, if A9m is a linear function of 9m, then fk and
the linear phase error due to A9m are ambiguous and cannot be determined exactly.
o O
Finally, consider the errors in Ro(m) when 9m 9m and 0m =0m, m
1,2, , Af. Let
R0(m) R0 + AR0(m). (6.20)
Replacing R0 in (5.7) with Ro(m) and for small Ai?0(m), large R0(m), and 7r/0
7(i r0), where i t0 < T0/2, we have from (5.11)
9, 0)
ux,y,zt
j {xtx +ytv +ztz) ejt0 ARq (m)
(6.21)
Hence the errors in {R0(m)}!^=1 result in phase errors along the synthetic aperture,
which is consistent with the analysis in [29, 43]. Then the z(n,m) in (6.5) should,
for this case, be replaced by
x(n, m)
K
E
. k=\
Oik &
j2ir[fktxm(n)+fktym(,n)+fktzmM} l p3Vm
(6.22)
where t0AR0(m) is the phase error caused by AR0(m). Note that this phase
error differs from the phase errors in (6.14) and (6.19) since it does not depend on the
parameters of the A;th scatterer and is easier to deal with. If AR0(m) is a constant,
then the phase error due to AR0(m) and the phase of ak are ambiguous. Hence
the phase of ak can never be determined exactly in the presence of ARo(m). If
o
AR0(m) is a linear function of 9m, then fk and the linear phase error due to AR0(m)
is ambiguous and cannot be determined exactly.
68
row containing only one nonzero unit element corresponding to the locations of the
available data samples. Then
yc = Aca + e, (5.22)
where
Ac = {Ic(i i)} ai {Ic(aK a^)} aK
The unknown sinusoidal parameters {cok, k, k, ctk, }(L1 are the features of our
interest and are to be estimated from the yc collected by CLSAR. In the following
sections, we will consider parametric 3D target feature extraction using CLSAR.
5.2 The RELAX Algorithm
The RELAX algorithm [39] can be extended to extract 3D target features. We
first consider using RELAX with the full aperture for 3D target feature extraction
and then extend this approach to the case of curvilinear apertures.
5.2.1 Full Aperture
The RELAX algorithm [39] minimizes the following NLS cost function:
C4 = y Ac*\ (5.24)
where  denotes the Euclidean norm. When the noise e(n, , ) is the zeromean
white Gaussian random process, the NLS estimates obtained with RELAX coincide
with the ML estimates of the target features. When the noise is colored, the NLS
estimates are no longer the ML estimates, but they still possess excellent statistical
performance [39].
We present below the relaxationbased minimization approach that leads to a
conceptually and computationally simple method. For each fixed K, the intermediate
number of scatterers, we perform a complete relaxationbased search by letting only
the parameters of one scatterer vary and freezing all others at their most recently
determined values. In this way, we will also take advantage of the fact that the
(5.23)
27
The practical convergence in the iterations can also be determined by check
ing the relative change of the cost function C2 in (3.17) between two consecutive
iterations.
We can also determine K, the total number of trihedrals and dihedrals, by
extending the GAIC discussed in Section 3.1 and assuming white noise and clutter.
The estimate K of K can be determined as an integer that minimizes the following
extended GAIC cost function:
(N11 \
EE Ie(n, n)2 I + 7ln[ln(AGV)](4iG + 6Kd + 1), (3.28)
TiO =0 /
where
Kd
e(n, n) = y(n, n) dd,sinc[7r6fe(n fk)\e
j{^dkn+ihdkn)
Kt
J2tkej{Cjt*n+Â£lt*), (3.29)
k=l
k=1
with n = 0,1, , N 1 and = 0,1, , 1; 7, as before, is a parameter of user
choice; K = Kt + KÂ¡i with Kt and Kd denoting the numbers of trihedral and dihedral
corners, respectively, determined by RELAXNLS given K\ finally, 4Kt + 6Kd + 1
is the total number of unknown realvalued parameters (of which 4Kt and 6Kdi
respectively, are for the trihedral and dihedral corners and 1 is for the white noise).
3.2.2 Image Formation
We can form enhanced SAR images of targets consisting of both trihedrals
and dihedrals by using the extracted target features via RELAXNLS and the mixed
data model in (3.15). The SAR images obtained via RELAXNLS are referred to
as RELAXNLS SAR images or simply RELAXNLS images. To form RELAXNLS
images, the dominant target features are first extracted by using the RELAXNLS
algorithm (see Section 3.2.1 for details) and then Steps 2 and 3 discussed in Section
3.1.2 are used except that the simulated phase history data of large dimensions used
in Step 2 is now determined by
Kd Kt
ys(ns, s) = CX] dfcsincMfc(s fk)}eUdkns+dknA + ^ (3 30)
k=1 fc=l
74
all of the curvilinear apertures since u is in the range direction, which is normal
to the plane of the curvilinear apertures. Comparing the CRBs for the Lshaped,
Arc1, and Arc2 apertures, we note that as expected, the more curved the aperture
is, the lower the CRBs for the target features. Since the Arc2 aperture is the least
curved, it has the largest CRBs. When the aperture becomes one straight line and
hence no longer curved, the CRBs go to infinity since we can no longer extract 3D
target features. Note also that from the top to bottom of Table 5.4.1, the CRBs
for J and Co increase since the aperture length decreases. The CRB for a, however,
does not always increase. Consider next the circular aperture in Figure 5.2(a) when
ui = Q = 0. Figure 5.4.2 shows the CRBs of the target parameters as a function
of the radius of the circular aperture for different M. Note that as expected, the
larger the radius and/or M, the lower CRBs of the target parameters. Note also
that, as expected, the CRB for co does not change with the radius for the fixed
M since again to is in the range direction, which is orthogonal to the plane of the
curvilinear aperture.
5.4.2 Experimental Examples
Before we present an experimental example to demonstrate the performance
of the RELAX algorithm for 3D target feature extraction with a CLSAR, we first
introduce an alternative to the PolartoCartesian interpolation. Instead of Polar
toCartesian interpolation, we created an N x N x rectangular grid and mapped
the data sample at each (tx,ty,tz) to the nearest grid point, where tx, ty, and tz are
functions of t, 9, and 4> (see Equations (5.13) to (5.15)). Let N denote the number of
available data samples. Then for both full and curvilinear apertures, the data model
can be written as
yr = Ir Aa + e,
(5.39)
LIST OF TABLES
4.1 True parameters of the two dihedrals used in Figures 1 and 2 44
5.1 Comparison of the CRBs (in dB) of the target features for the cases
of the full, circular, parabolic, Lshaped, Arc1, and Arc2 apertures
when M = 63, K = 1, uj = Q = Q = 0, a = 1, and a2 = 40 73
6.1 Estimation errors for the parameters of the highest scatterer in the
simulation example before and after autofocusing 89
vi
19
frequencies of the fcth sinusoid or point scatterer, k = 1,2, K] {e(n, )} denotes
the unknown noise sequence. Note that {cok,cok} corresponds to the 2D location of
the kth scatterer and ak is determined by the RCS and range of the fcth scatterer.
The NLS estimates {ak,ujk,LOk}k=1 f {ak, ^k)k=i can be obtained by minimizing
the following NLS criterion [39]:
Ci ({ak,ujk,u>k}^=1)
K
Y Y cvN(uk)uj:fr(ujl.)ak
k=1
(3.2)
where ()T denotes the transpose;  \\p denotes the Frobenius norm; Y is an N x N
matrix with y(n, ) being its (n)th element;
^N^k)
1 ejuJk
T
ejuk(Nl)
(3.3)
n T
^N^k) 1 e^k e^k^N b (34)
RELAX minimizes the NLS cost function C\ in (3.2) by using a complete relaxation
based search, i.e., for each K, the intermediate number of scatterers, only the param
eters of one scatterer vary while all others are fixed at their most recently determined
values. Let
K
~ (3.5)
Yt = Y ^2 QÂ¡W/(wÂ¡)w5(iij),
iz=\,izjLk
where {, >, are assumed to have been estimated. Then {o;*, Â¡>*, are
obtained by
Oik =
NN
\Uk=Uk,Uk=Uk
(3.6)
where ()H and ()*, respectively, denote the conjugate transpose and complex con
jugate, and
{ujk,uk} = argmax \u>%(ujk)YkuJ(k) 2 (3.7)
Uk ,Uk
which indicates that {tuk,uk} can be obtained as the location of the dominant peak
of the periodogram a;^(w*;)Yfea;^.(;fc)2 /N, which can be efficiently calculated
via FFT. Then, ak can be calculated from the complex amplitude of the peak of
20
\uj%(u>k)Ykuj*()k)^ /N. The RELAX algorithm is summarized as follows:
Step (1): Assume K 1. Obtain {Ck,u>k,ak}k=i from Y by using (3.7) and (3.6).
Step (2): Assume K = 2. Compute Y2 with (3.5) by using {uok, uk, k}k=i ob
tained in Step (1). Obtain {uk,)k,k}k~2 from Y2. Next, compute Yi by using
{vk,u>k, ak}k=2 and then redetermine {ujk, uk, k}k=\ from YL.
Iterate the previous two substeps until practical convergence is achieved.
Step (3): Assume K = 3. Compute Y3 by using {k,k,k}k=i obtained in Step
(2). Obtain {u)k,k,k}k=3 from Y3. Next, compute Yi by using {A, k, k}\=2 and
redetermine {k,u)k, k}k=i from Yx. Then compute Y2 by using {tbk,uk, k}k=1
and redetermine {u)k, Ok, oÂ¡k}k=2 from Y2.
Iterate the previous three substeps until practical convergence.
Remaining Steps: Continue similarly until K is equal to K, which can be deter
mined by using generalized Akaike information criterion (GAIC) to be discussed later
on.
The practical convergence in the iterations may be determined by checking
the relative change e of the cost function C\ ({)k,u)k,&k}k=i) in (3.2) between two
consecutive iterations.
It has been demonstrated in [39] that RELAX is related to CLEAN [28], which
is only an approximate relaxationbased method that minimizes the NLS criterion.
RELAX turns out to be CLEAN when no iteration is performed in each step of
RELAX. Due to the iterations performed in each step of RELAX, a more accurate
initial can be obtained, which makes RELAX converge to the global minimum with a
high probability. Meanwhile, RELAX is asymptotically (for large number of samples)
efficient, while CLEAN gives biased estimates, especially when two scatterers are
closely spaced. The RELAX algorithm, thereby, can also be referred to as SUPER
CLEAN.
56
same range. In this example, Hybrid and SPAR have the same estimated number of
scatterers, K 7. Figure 4.3(b) shows the windowed 2D FFT SAR image obtained
by applying the normalized 2D FFT to the windowed data matrix. SAR images
formed via Hybrid, SPAR, RELAXNLS and the modified RELAXNLS algorithms
are shown in Figures 4.3(c) through 4.3(f), respectively. We note that at high SNR,
the Hybrid image is similar to the SPAR image. Both of the parametric RELAXNLS
and the modified RELAXNLS algorithms outperform their semiparametric coun
terparts SPAR and Hybrid since the data model used by the parametric methods is
correct rather than approximate. For this example, our simulations show that the
ratios between the MATLAB flops needed by Hybrid, SPAR, the modified RELAX
NLS, and RELAXNLS over the flops needed by the windowed FFT method are 27.4,
28.4, 50.1, and 70.8, respectively. Note that both SPAR and Hybrid are computa
tionally more efficient than RELAXNLS and the modified RELAXNLS, with the
modified RELAXNLS being more efficient than RELAXNLS.
Consider next a simulated example with low SNR. The SAR phase history
data is the same as in the above example except that the noise variance is increased
to = 6. SAR images obtained by using the windowed 2D FFT, Hybrid, SPAR,
RELAXNLS, and the modified RELAXNLS are shown in Figures 4.4(a) through
4.4(e), respectively. Note that the artifact problem starts to show up in the Hybrid
image in Figure 4.4(b) due to large parameter estimation errors. By comparing
Figures 4.4(c) with (b), it can be seen that SPAR can effectively mitigate the artifact
problem.
Finally, consider an experimental example of SAR image formation by using
the MSTAR Slicy data collected by imaging an object consisting of both trihedral
and dihedral corner reflectors, which is shown in Figure 3.3. The data was collected
by the Sandia National Laboratory (SNL) using the STARLOS sensor. The field
data was collected by a spotlightmode SAR with a carrier frequency of 9.559 GHz
49
the interval n\ < n+ < n2 so that the magnitude of Vfe is above a certain threshold,
say within the interval. Similarly, we can fix n to n+ and search for the interval
i < + < 2. Then the N x rectangular window w(n,fi) has unit value for
ni < n < n2 and \ < < 2 and zero value elsewhere. The threshold T we use in
our numerical and experimental examples is 10% of the peak value of the magnitude
of Vfc.
Step (iii): Determine Yk by applying 2D inverse FFT (IFFT) to Vfc W,
where the (n, )th element of W is w(n,n).
Instead of minimizing Ce{4>k, Xfc, fk, fk), we now minimize
2
5
F
C7((j)k, xjt, A, A) == Y*, eJ^Gfc(xfe, A, /fe)
(4.47)
where Yk is used to replace Yk in C'6(^fc,Xfe, fk, fk), by using the method presented
in Section 4.1.1.
With the above preparations, now we provide the steps of the scatterer isola
tion and relaxation based NLS algorithm, which are the substeps of Step 1 of SPAR.
Step I: Assume K = 1. Calculate Y from Y by using the isolation process.
Obtain {/>fc, xfc, fk, fk}k=i from Y.
Step II: Assume K = 2. Compute Y2 with (4.45) by using {(frki'Xk, fk, f k}k=\
obtained in Step I. Calculate Y2 from Y2. Obtain {(Â¡>k,xk, fk, fk}k=2 from Y2. Next
compute Yx with (4.45) by using {k,xk, fk, fk}k=2, calculate Y\ from Yi, and then
redetermine {
Iterate the previous two substeps until practical convergence is achieved (to
be discussed later on).
Step III: Assume K = 3. Compute Y3 with (4.45) by using {(/fc, x.k, fk, fk}k=i
obtained in Step II. Calculate Y3 from Y3. Obtain {
compute Yi with (4.45) by using {(j>k, xfc> fk, Jk}\=2^ calculate Yx from Y3, and then
redetermine {(j)k,x.k, fk, fk}k=1 from Yx. Then compute Y2 with (4.45) by using
55
are obtained via 1D FFT with zeropadding to a total length of 128 in range and to a
total length of 64 in crossrange. Next, these initial estimates are refined by using the
FMIN function in MATLAB alternately, i.e., by updating / while fixing / at its most
recently determined value and vice versa, until practical convergence, which is de
termined by checking the relative change of the cost function C^(f,f). We have used
103 to determine the convergence of this fine search as well as the relaxationbased
algorithm. The extrapolation factor Â£ = 8 is used in range for SPAR and Hybrid
and in both range and crossrange for RELAXNLS and the modified RELAXNLS.
Both 1D and 2D Kaiser windows with shape parameter Â¡3 = 6 are used whenever
needed.
First consider a simulated example with high SNR. The SAR phase history
data matrix is simulated by assuming that there are four trihedrals and three di
hedrals in the presence of zeromean white complex Gaussian noise with variance
o2n = 0.6. The amplitude functions for the four trihedrals are generated as follows:
xk[n) = 1, A: = 1,2,3, = 0,1, , 1, (4.71)
and
x4() = 2, = 0,1, , JV 1. (4.72)
The amplitude functions for the three dihedrals are
xb{n) = 9.6sinc[0.37r( 18.6)], ft = 0,1, , 1, (473)
and
xk() = 6.4sinc[0.27r( 18.6)], k = 6,7, ft = 0,1, , 1, (474)
where sinc(x) = sin(x)/x. Figure 4.3(a) shows the modulus of the true SAR image.
Note that two of the dihedrals are located in the same crossrange and are closely
spaced in range and two of them are located in the same range. Of the four trihe
drals, two of them are closely spaced in range and the other two are located in the
16
later was used in microwave imaging [66]. Encouraged by the initial results in [33],
Knaell later presented in [34] a correlation technique which works in conjunction with
the coherent CLEAN algorithm to reduce sidelobe interference. Furthermore, Knaell
in [35] compared the performances of the coherent CLEAN method and two other
methods including a maximum likelihood (ML) method and a sidelobe leakage reduc
tion algorithm [12] in extracting 3D target features by using CLSAR datasets. We
will discuss how RELAX (SUPER CLEAN) can be extended for 3D target feature
extraction using CLSAR in Chapter 5.
2.3 SAR Motion Compensation
Aperture errors exist due to atmospheric turbulence and platform uncertainty.
These undesirable errors may significantly degrade the quality of SAR images in sev
eral ways. For example, aperture errors may cause image geometric distortion, spu
rious targets, loss of resolution, and decrease in image contrast [32, 25] depending on
their natures. Therefore, effective motion compensation methods are indispensable
for SAR imaging and target feature extraction in practice. In 2D SAR systems, mo
tion errors mainly cause phase errors in crossrange [29]. Autofocus techniques that
compensate for phase errors using radar phase history data have been extensively
exploited by 2D SAR systems. Of the existing autofocus algorithms, mapdrift [45]
and phase difference [68] algorithms are used to compensate for phase errors asso
ciated with certain order polynomials, whereas the phasegradient autofocus (PGA)
algorithm [17, 18, 67] is independent of the phase error model, thereby is more ro
bust. Autofocus algorithms based on super resolution spectral estimation methods
also begin to appear in the literature [43, 52], All the aforementioned 2D autofocus
algorithms assume that phase errors change in crossrange only and are constant for
all range bins. This assumption is generally valid for the 2D SAR systems where
57
and bandwidth of 0.591 GHz. The radar was about 5 kilometers away from the
ground object. The data was collected when the object was illuminated by the radar
from approximately the azimuth angle 0 and elevation angle 30. To crossvalidate
the experimental results given below, XPATCH [1], a high frequency electromagnetic
scattering prediction code for complex 3D objects, was used to generate very high
resolution phase history data for the object shown in Figure 3.3. The data generated
by XPATCH has a resolution of 0.038 meters in both range and crossrange, and the
corresponding windowed FFT SAR image is shown in Figure 4.5(a). (We have used
the log scale for all of the images shown in Figure 4.5.) The original experimental
Slicy data has a resolution of 0.3 meters in range and 0.32 meters in crossrange.
The 32 x 32 data matrix we used to demonstrate the performance of our algorithms
has a spoiled resolution of 0.51 meters in range and 0.54 meters in crossrange. The
windowed 2D FFT SAR image of this data matrix is shown in Figure 4.5(b). Figure
4.5(c) shows the Hybrid image with K = 7 (obtained via GAIC). Figure 4.5(d)
shows the SPAR image with K 7 (obtained via GAIC). Figures 4.5(e) and (f)
show the SAR images obtained via RELAXNLS and the modified RELAXNLS
with K = 7, respectively. Note that Hybrid has a more severe artifact problem
than SPAR. Note also that the SPAR image shown in Figure 4.5(d) appears to fit
Figure 4.5(a) and the characteristics of the object in Figure 3.3 well. However, the
parametric algorithms are not as robust as SPAR since the parametric images shown
in Figures 4.5(e) and (f) do not fit Figure 4.5(a) as well with one of the scatterers
misidentified and mislocated. For this experimental example, the ratios between the
MATLAB flops needed by Hybrid, SPAR, the modified RELAXNLS, and RELAX
NLS over the flops needed by the windowed FFT method are 29.2, 16.7, 32.8, and
43.1, respectively. Note that SPAR can sometimes be faster than Hybrid!
61
(e)
Figure 4.4: Comparison of SAR images formed using different algorithms for simu
lated data at low SNR (
SAR image, (c) SPAR SAR image, (d) RELAXNLS SAR image, (e) Modified
RELAXNLS SAR image. (The vertical and horizontal axes are for range and cross
range, respectively.)
5
resolution SAR images by using the target features extracted via relaxationbased
methods including RELAX [39] and RELAXNLS [42] in Chapter 3. The relaxation
based optimization approach is one of the effective ways to minimize a nonlinear least
squares (NLS) cost function and has been proved to be quite useful in many applica
tions, such as radio astronomy [28], microwave imaging [66], spectral estimation [23],
and sinusoidal parameter estimation and target feature extraction [39, 42], RELAX
was originally proposed for the parameter estimation of complex sinusoids in colored
noise. Compared with other NLS algorithms, RELAX is relatively robust against
data mismodeling errors [39]. We will show that RELAX can be used to enhance
SAR images of targets consisting of point scatterers. However, the point scatterer
data model may not be valid in practice since, for example, for many manmade
targets, such as vehicles and buildings, much of the returned energy is caused not
only by trihedral corner reflectors, but also by dihedral corner reflectors of a target
[36]. We also describe how to form super resolution SAR images of targets consisting
of both trihedrals and dihedrals using RELAXNLS [42]. RELAXNLS is based on
a mixed data model, which models a trihedral as a point scatterer and a dihedral as
a complex sinusoid with a constant phase and a variable amplitude described by a
sine function in crossrange. We will demonstrate the super resolution property of
the RELAX and RELAXNLS algorithms with experimental data including MSTAR
and ERIM (Environmental Research Institute of Michigan) data.
Parametric methods may outperform nonparametric methods provided that
the assumed data models are valid. In the more likely cases where it is difficult to
establish a good parametric data model in crossrange due to a variety of physical
phenomena including glints, resonances, and motioninduced phase errors [15], a more
flexible data model is desired and nonparametric and parametric methods may be
combined together for 2D SAR target feature extraction and image formation. In
our work, we also establish a more flexible data model for manmade targets when
34
(a)
(b)
v.
"1. W
(C)
(d)
(e) (f)
Figure 3.4: SAR images of the Slicy data (0 azimuth and 30 elevation angles),
(a) Original windowed 2D FFT image, (b) Windoed 2D FFT image with spoiled
resolution, (c) ~ (d): Windowed RELAX image without and with background clutter
with K = 36. (e) ~ (f): Windowed RELAXNLS image without and with background
clutter with K = 9. (c) ~ (f) are obatined using the data in (b). (The vertical and
horizontal axes are for range and crossrange, respectively.)
17
significant phase errors are caused by motion measurement errors, such as alongtrack
velocity errors or lineofsight acceleration errors [11].
As in 2D SAR systems, motion error compensation is also indispensable for
3D SAR systems. Autofocus in CLSAR, however, becomes more complicated since
the aperture errors can no longer be approximated as phase errors only in crossrange
since the errors introduced in 6 (azimuth angle) and <Â¡> (elevation angle) may not be
negligible and the extracted 3D target features are sensitive to the shape of the
curvilinear aperture (see the discussion in Chapter 6). To the best of our knowledge,
autofocus for CLSAR has not been reported in the literature, partly because CLSAR
is a relatively new technology.
77
1 1 1
1
Full Aperture
M=32
_
M=64
* *% *..
 M=80
*

1 1 1
1
10
Radius
(a)
<3,
m 40
oc
O
"T"
T
Full Aperture
M=32
* M=64
M=80
10
Radius
36
1 1 1
i r
_
38

40

42

44
Full Aperture
M=32

46
 M=64
 M=80

48
50

52
1 1 1
1 L_
10
Radius
(b)
T"
T
Full Aperture
M=32
M=64
 M=80
10
Radius
(c) (d)
Figure 5.3: The CRBs of the target parameters as a function of the circular aperture
radius when K = 1, a = 1, to = 0, and cr2 = 40. a) The CRB of the
complex amplitude a. b) The CRB of to. c) The CRB of Co. d) The CRB of Co.
4.4.1 Target Feature Extraction 45
4.5 Modified RELAXNLS Algorithm 51
4.6 CramrRao Bound (CRB) of the SemiParametric Data Model 52
4.7 Numerical and Experimental Results 54
5 3D TARGET FEATURE EXTRACTION USING CLSAR 63
5.1 CLSAR Data Model 63
5.1.1 1D Data Model for High Range Resolution Radar .... 63
5.1.2 Full Synthetic Aperture Radar 64
5.1.3 Curvilinear SAR 67
5.2 The RELAX Algorithm 68
5.2.1 Full Aperture 68
5.2.2 Curvilinear Aperture 70
5.3 Performance Analysis of Parameter Estimation via CRB 71
5.4 Numerical and Experimental Results 72
5.4.1 Performance Analysis of Different Curvilinear Apertures
via CRBs 72
5.4.2 Experimental Examples 74
6 AUTOFOCUS IN CLSAR 79
6.1 Aperture Error Effects 80
6.2 The RelaxationBased Autofocus Algorithm (AUTORELAX) . 83
6.2.1 Aperture Parameter Estimation 83
6.2.2 Target Feature Extraction 85
6.3 Experimental and Numerical Results 87
7 CONCLUSION 94
7.1 Summary 94
7.2 Future Works 96
APPENDIXES
A A SINGLE SCATTERER FEATURE EXTRACTION VIA SPAR . 98
B APPROXIMATION OF THE RANGE R 100
C CALCULATING (x,y,z) FROM (x,y,z) 102
REFERENCES 104
BIOGRAPHICAL SKETCH l0g
v
47
where arg(a;) denotes the argument of a complex variable x; finally,
{/,/} = argmax C4(/, /),
fif
(4.40)
where
The steps needed to obtain the NLS estimates of a single scatterer are summarized
as follows:
Step (I): Use (4.37) to obtain ys(/) and obtain the cost function C^f, f)
according to (4.41). Determine {/, /} by maximizing C4(/, /) using the method given
in Appendix A.
Step (II): Calculate 4> according to (4.39) with {/,/} replaced by {/,/}
obtained in Step (I).
Step (III): Calculate x via (4.38) with {0,/,/} replaced by {0,/,/} deter
mined in Steps (I) and (II), respectively.
Feature Extraction of Multiple Scatterers
When a target consists of multiple scatterers, we can obtain the NLS esti
mates of the target features based on (4.2) by using a relaxationbased optimization
approach. Let
T
Xfc Xfc(0) xk(l) xk(N 1)
(4.42)
and let Y and E be N x N matrices with their (n,n)th elements being y(n,n) and
e(n, ), respectively. Then we can rewrite (4.2) as
K
(4.43)
k=1
where Gk(xk, fk, fk) has the same form as the G(x,/,/) in (4.36) except that x, /,
and / are replaced by xfc, fk, and fk, respectively. Let y, = 0,1, , 1, be the
41
where L > 2 denotes the number of scatterers located in the same range. Thus the
L scatterers located in the same range are considered as two scatterers when using
the data model in (4.2).
Before we discuss the impact of the model ambiguities on SAR image for
mation, we first describe how the image formation is done if we have the estimated
model parameters.
4.3.2 Image Formation
Assume for now that we have extracted the target features based on any of
the ambiguous data models. For notational convenience, we will use the notation
used in (4.2). Since the target data model in range is a sum of several complex
sinusoids with constant amplitudes and phases, we can use the estimated sinusoidal
parameters to simulate a data matrix with a larger dimension in range and then use
FFT to demonstrate the super resolution property of the feature extraction algorithm
we shall present. Yet we cannot extrapolate the estimate {()} of {^()} since it is
assumed to be an arbitrary unknown realvalued function of and hence FFT cannot
be used to obtain SAR images with enhanced resolution in crossrange. Instead, we
use 1D APES [37, 38, 41] in crossrange when forming SAR images since APES
belongs to the class of matched interbank spectral estimators and provides lower
sidelobes, narrower spectral peaks, and more accurate spectral estimates than FFT.
Let {ss(n, )} denote the simulated data sequence with a larger dimension in
range based on the estimated target features {(j)k, {xk()}^~Q fk, fk}^=l of
{(/)k, {xk()}^ZQ, /fc, fk}k=i where K denotes the estimate of the scatterer number
K. Then
K
ss(n, n) Xk()e^kkU\ n = 0,1, , Â£N 1, = 0,1, , 1,
fc=i
(4.23)
Elevation Angle (degrees) Elevation Angle (degrees)
91
Figure 6.2: Autofocused curvilinear aperture and scatterer distribution obtained with
AUTORELAX by autofocusing only in the elevation direction and using K = 20 for
the experimental example, (a) Autofocused curvilinear aperture (dotted line), (b)
Scatterer distribution.
Azimuth Angle (degrees) Azimuth Angle (degrees)
(a) (b)
Figure 6.3: (a) Manually adjusted (solid line) and autofocused (dotted line) curvi
linear apertures for the experimental example, (b) Fitting the manually adjusted
aperture to the autofocused aperture by adding a line to the former.
67
complex amplitude and 3D frequencies of the kth sinusoid, k = 1,2, , K; fi
nally, {e(n, , )} denotes the unknown noise sequence. The sinusoidal frequencies
{ujk = 2nfk,uJk = 27Tfk,,d>k = 2irfk} correspond to the 3D location of the kth
scatterer of a radar target and ctk is determined by its RCS.
Let
y =
2/(0,0,0) 2/(1,0,0) y(N 1,0,0) 2/(0,1,0) y( 1,1,0)
y(N 1,1,0) 2/(0, 1, 1) 2/(1, 1, TV 1)
T
y(N 1, 1,7V 1)
(5.17)
Let e be defined similarly from e(n, , ) as y from y(n, , ). Then the data model
in (5.16) can be rewritten as:
y = Act + e,
(5.18)
where
a =
T
Ot\ Ot2 Otx )
and
A =
ai <8> i ai 3lK
ak )
where denotes the Kronecker product [22] and
1
i
1
afc =
e3^k
>
ejk
, a k
ejQk
ej(Nl)uik
ej(l)Â£jk
ej(l)ik
5.1.3 Curvilinear SAR
(5.19)
(5.20)
(5.21)
Assume that the curvilinear aperture used in CLSAR consists of M different
radar viewing angles. Then the received data vector yc in CLSAR, is an (MN) x 1
subvector of y in (5.18). Let Ic denote an M x () matrix with each column and
2
Flight Path
CrossRange
Figure 1.1: Illustration of 2D radar imaging.
bandwidth. The crossrange resolution 5C of a conventional radar or real aperture
radar (RAR) is rangedependent and is determined by
(1.1)
where R represents the range of the target; A and L, respectively, denote the wave
length and the size of the antenna aperture. To increase crossrange resolution (given
R), an obvious way is either to reduce the wavelength, or to increase the size of the
antenna. Using shorter wavelength requires more transmission power since electro
magnetic signals with smaller wavelengths are more likely absorbed by the atmo
sphere. In the meantime, it is not feasible to increase the antenna size too much,
especially when the radar is carried on board of an airplane or spacecraft. Hence
a conventional radar or RAR inherently has a limited crossrange resolution. It is
desired to achieve high crossrange resolution without reducing the wavelength and
exploiting large physical antennas. Synthetic aperture radar (SAR) is a solution and
can achieve high crossrange resolution independent of range.
The SAR technology was first proposed in the 1960s [59]. Instead of exploiting
large antennas and reducing wavelength, SAR makes use of a synthetic aperture to
achieve high crossrange resolution. A synthetic aperture is formed by coherently
95
RELAXNLS is more effective for SAR image formation of targets consisting of both
trihedral and dihedral corner reflectors.
A more robust semiparametric algorithm, referred to as SPAR, was proposed
in Chapter 4 based on a more flexible and robust data model for manmade targets
when it is difficult to establish an accurate parametric data model in crossrange
for targets of interest. Instead of using the aforementioned approximate dihedral
and trihedral data model, all scatterers of targets in SPAR are modeled as 2D com
plex sinusoids with arbitrary amplitudes and constant phases in crossrange and with
constant amplitudes and phases in range. The SAR images with super resolution in
range obtained via SPAR can be achieved by applying FFT to the phase history data
with a large dimension in range synthesized by using the extracted features and the
parametric data model in range. The crossrange resolution of SAR images cannot
be significantly improved via SPAR over the conventional FFT method due to the
nonparametric data model in crossrange. Yet it can be improved slightly by using
APES, which is a matched filterbank based nonparametric method with better sta
tistical performance than Capon. It has been shown in Chapter 4 with numerical and
experimental examples that SPAR can not only form accurate and super resolution
SAR images by taking advantages of both parametric and nonparametric spectral
estimation methods, but also effectively mitigate the artifacts in SAR images due
to the flexible data model by first isolating out the most dominant scatterer via a
rectangular window. Another advantage of SPAR is that it can be used to provide
good initial conditions needed by other parametric target feature extraction methods
to reduce the total amount of computations needed. Numerical and experimental
examples have shown that SPAR is more robust and computationally simple than
RELAXNLS and its variation discussed in Chapter 4.
CLSAR is a relatively new technology for 3D target feature extraction and is
still at its developing stage. In the study, we have presented parametric algorithms
