and image processing for moving target indication and side lobe suppression

Synthetic Aperture Radar signal

ABSTRACT
This thesis summarizes a selection of my re- algorithm on an image. The results from LORA
search contributions within Synthetic Aperture data are satisfactory especially with regard to
Radar (SAR). The research is mainly aimed at forest clutter suppression. These methods can
Synthetic Aperture Radar signal and
developing signal processing methods and app- also be applied in other fields. The ability to image processing for moving target
lying them on single channel and multi channel detect and focus images of ships allows for sur-
wideband SAR systems. veilling coastal areas to help rescuing persons
indication and side lobe suppression
SAR systems can generate images similar to at sea. Detecting and locating of cars and trucks
optical pictures, i.e. photos, and for satellite bor- allow for traffic monitoring to obtain statistics
ne systems, sometimes with much finer resolu- of the number of cars traveling on the roads and
tion compared to optical systems. SAR has also their speed.
been used for instance to obtain fine resolution
images of the moon, Venus and the satellites of In the thesis, two more important aspects for
Saturn. Other applications of SAR are detecting SAR processing are presented. One paper pre-
e.g. changes in ice sheets and deforestation. sents windowing of UWB SAR images. A stron- Thomas Sjögren
In this thesis, SAR systems capable of ima- gly reflecting object such as a power line in a
ging with high relative resolution are conside- SAR image causes strong side lobes, i.e. ringing
red. Data from such systems, namely the VHF on its both sides. This ringing can cause a small
system CARABAS-II and the UHF system house to be covered by the side lobes. Applying
LORA, are used. The term high relative resolu- a window can suppress these side lobes in the
tion imaging is used with regard to high relative image. However, the power line will broaden in
bandwidth. the image and may cover the small house if an
Two of the topics in this thesis are related unsuitable window is chosen. The last topic in
to detecting and estimating parameters of mo- the thesis is on the theoretical limits on accura-
ving targets in SAR, with CARABAS-II data cy in measuring parameters of a moving target
in the first topic and LORA data in the second. in a SAR image. The parameters are position,
A speed estimating and refocusing method is velocity, radar cross section and phase shift in
introduced for the CARABAS-II data. The re- the reflected waves. The theoretical expressions
sults show good estimation accuracy as well as of the limits on target position and relative
good ability to focus the moving targets and to speed are verified using simulation of a single
suppress forest clutter by applying a refocusing channel system.

Thomas Sjögren

Blekinge Institute of Technology

Doctoral Dissertation Series No. 2012:14
2012:14

ISSN 1653-2090 School of Engineering

2012:14 ISBN: 978-91-7295-243-0
 Synthetic Aperture Radar Signal and
Image Processing for Moving Target
Indication and Side Lobe Suppression

Thomas Sjögren
Blekinge Institute of Technology doctoral dissertation series
No 2012:14

Synthetic Aperture Radar Signal and

Image Processing for Moving Target
Indication and Side Lobe Suppression

Thomas Sjögren

Doctoral Dissertation in
Applied Signal Processing

School of Engineering
Blekinge Institute of Technology
SWEDEN
2012 Thomas Sjögren
School of Engineering
Publisher: Blekinge Institute of Technology,
SE-371 79 Karlskrona, Sweden
Printed by Printfabriken, Karlskrona, Sweden 2012
ISBN: 978-91-7295-243-0
ISSN 1653-2090
urn:nbn:se:bth-00542
Abstract

The thesis summarizes a selection of my research contributions within Synthetic

Aperture Radar (SAR). The research is mainly aimed at developing signal
processing methods and applying them on single channel and multi channel
wideband SAR systems. SAR systems can generate images similar to optical
pictures, i.e. photos, and for satellite borne systems, sometimes with much
finer resolution compared to optical systems. SAR has also been used for
instance to obtain fine resolution images of the moon, Venus and the satellites
of Saturn. Other applications of SAR are detecting e.g. changes in ice sheets
and deforestation. In this thesis, SAR systems capable of imaging with high
relative resolution are considered. Data from such systems, namely the VHF
system CARABAS-II and the UHF system LORA, are used. The term high
relative resolution imaging is used with regard to high relative bandwidth. Two
of the topics in this thesis are related to detecting and estimating parameters
of moving targets in SAR, with CARABAS-II data in the first topic and LORA
data in the second. A speed estimating and refocusing method is introduced
for the CARABAS-II data. The results show good estimation accuracy as well
as good ability to focus the moving targets and to suppress forest clutter by
applying a refocusing algorithm on an image. The results from LORA data
are satisfactory especially with regard to forest clutter suppression. These
methods can also be applied in other fields. The ability to detect and focus
images of ships allows for surveilling coastal areas to help rescuing persons at
sea. Detecting and locating of cars and trucks allow for traffic monitoring to
obtain statistics of the number of cars traveling on the roads and their speed.
In the thesis, two more important aspects for SAR processing are presented.
One paper presents windowing of UWB SAR images. A strongly reflecting
object such as a power line in a SAR image causes strong side lobes, i.e. ringing
on its both sides. This ringing can cause a small house to be covered by the
side lobes. Applying a window can suppress these side lobes in the image.
However, the power line will broaden in the image and may cover the small
house if an unsuitable window is chosen. The last topic in the thesis is on the
theoretical limits on accuracy in measuring parameters of a moving target in
a SAR image. The parameters are position, velocity, radar cross section and

vii
viii Abstract

phase shift in the reflected waves. The theoretical expressions of the limits
on target position and relative speed are verified using simulation of a single
channel system.
Preface

This doctoral thesis summarizes part of my work within the field of radar
signal processing, specifically Synthetic Aperture Radar. The work has been
performed at the Department of Electrical Engineering, Blekinge Institute of
Technology.
The thesis consists of four selected parts:

Part
I Moving Target Relative Speed Estimation And Refocusing In Synthetic
Aperture Radar Images.
II Suppression of Clutter in Multichannel SAR GMTI.
III Linear Apodization for UWB SAR Images.
IV Performance of Target Parameter
Estimation Using SAR.

ix
Acknowledgments

I have many people to thank for the help, support, friendship and encourage-
ment over the last 5 years. First i would like to thank Dr. Sven Johansson,
which accepted me to join the Department of Signal Processing back in 2006.
Then i would like to thank my supervisor Prof. Mats Pettersson which has
with his rich experience and deep knowledge in SAR system design and signal
processing helped a lot. It was Mats that initially gave me the idea of starting
this work after having a nice talk in the spring of 2006, as this position was
announced on the web of BTH. Furthermore I would like to thank my super-
visor Prof. Magnus Herberthson which with the help of his deep knowledge
in mathematics, radar and SAR given very valuable feedback for which I am
much thankful.
In the everyday work at the department I would like to thank all colleagues
in the department for the positive attitude and the open and very friendly style
in coffeebreaks, lunches as well as all nice parties. Especially I would like to
thank Dr. Mikael Swartling for all nice walks in the forest and sessions of board
game play, Josef for the nice company and for being a great personal trainer
at the gym, and Feng Wang for all nice talks, dinners as well as guidance and
help in mathematics. Also I must give a special personal thank for Dr. Viet
Vu which has been a great close colleague in research, coursework as well as
very friendly and social in work and outside of work.
Most importantly, I would like to give my warm wishes to my parents, Lars
and Elisabeth, my sister Erika, my godmother Karin and specially Grosspapi
Karl, which has all been a great support throughout my life and given me
strength to keep fighting in tough situations. What must not be forgotten as
well is all joy that all Tetrinet blocks have given me throughout the years,
especially the almighty stick and the wonderful L-block, being the base for
all offensive play in Tetrinet. Also I would like to thank Carlshamn Chamber
Choir for all joyful singing moments brought to me during my time in Blekinge
so far. It has been challenging, but a weekly boost of joy and energy coming
from great friendship and great music. Putting it all together, with family,
music and the great things in life, I have been so fortunate to have met my
fiancé Aubrey during my study time. In the period of very tough work the last

xi
xii Acknowledgments

year it has been specially hard strain for you to bare with me under stress. I
thank you deeply for sticking out with me even though I am sure at times I
have not been an easy person.
Karlskrona, Autumn 2012
Thomas
List of selected publications

T.K.Sjögren, V.T.Vu, M.I.Pettersson, A. Gustavsson, L.M.H. Ulander, Moving

Target Relative Speed Estimation and Refocusing in Synthetic Aperture Radar
Images, IEEE Transactions on Aerospace and Electronic Systems, vol. 48,
no. 3, pp. 2426-2436, 2012.

T.K.Sjögren, V.T.Vu, M.I.Pettersson, F. Wang, D. Murdin, A. Gustavsson and

L.M.H. Ulander, Suppression of Clutter in Multichannel SAR GMTI, IEEE
Transactions on Geoscience and Remote Sensing, In Review.

T.K.Sjögren, V.T.Vu and M.I.Pettersson, Linear Apodization for UWB SAR

Images, IEEE Transactions on Geoscience and Remote Sensing, In Review.

M.I.Pettersson, T.K.Sjögren and V.T.Vu, Performance of Target Parameter

Estimation Using SAR, IEEE Transactions on Aerospace and Electronic Sys-
tems, In Review.

xiii
Contents

Abstract vii

Preface ix

Acknowledgments xi

List of selected publications xiii

Contents xv

Introduction 1
1 Radar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Synthetic Aperture Radar . . . . . . . . . . . . . . . . . . . . . 4
3 Earth Observation, Surveillance and Security . . . . . . . . . . 5
4 SAR Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
5 Moving targets in SAR . . . . . . . . . . . . . . . . . . . . . . . 9

References 21

I Moving Target Relative Speed Estimation and Refo-

cusing in Synthetic Aperture Radar Images 25
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2 The Radar model and Moving target focusing . . . . . . . . . . 30
3 Phase of a moving target in a SAR image . . . . . . . . . . . . 32
4 Estimation procedure . . . . . . . . . . . . . . . . . . . . . . . 34
5 Refocusing in a SAR image . . . . . . . . . . . . . . . . . . . . 35
6 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . 37
7 Real data results . . . . . . . . . . . . . . . . . . . . . . . . . . 39
8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

References 47

xv
xvi CONTENTS

II Suppression of Clutter in Multichannel SAR GMTI 51

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2 Fast Backprojection . . . . . . . . . . . . . . . . . . . . . . . . 55
3 Moving target focusing and relocation . . . . . . . . . . . . . . 56
4 Clutter Suppression for GMTI . . . . . . . . . . . . . . . . . . 59
5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

References 75

IIILinear Apodization for UWB SAR Images 79

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2 Linear Apodization in SAR . . . . . . . . . . . . . . . . . . . . 86
3 Optimum Linear Apodization Technique Adapted For UWB SAR 89
4 Sidelobe suppression results . . . . . . . . . . . . . . . . . . . . 94
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

References 103

IVPerformance of Target Parameter

Estimation Using SAR 105
Introduction

The wish for humans to explore nature is the core for the development and
survival of the human race. Since the space race began in the 1950s with
the launch of Sputnik, we have used spacecrafts to explore our planet’s clos-
est surroundings. These explorations have given us a better understanding of
how dependent we are on the surrounding atmosphere, for instance the ozone
layer, and how important it is to supervise it[1]. Radar and especially Syn-
thetic Aperture Radar (SAR) systems are very important for the continuous
supervision of the Earth climate[2],[3],[4] and they are also crucial as sensors
capable of high-resolution mapping of space bodies[5],[6]. This thesis consists
of an introduction part, whose aim is to introduce radar, SAR, and the most
basic SAR processing algorithms among other algorithms, which are used in
the research presented in this thesis. In the final part of the introduction,
ground moving targets in SAR and signal processing for focusing and detect-
ing moving targets are first explained. After this, side lobe suppression in SAR
is introduced. The aspects covered in this thesis with regard to moving targets
are detection, parameter estimation, focusing and measurement accuracy.

1 Radar
The main sensors used in remote sensing that create 2-D mappings of the
ground are active sensors such as radar and LIDAR or passive sensors such
as IR sensors, optical cameras and radiometers. Out of these sensors, the
radiometer and the radar are least sensitive to bad weather conditions such as
cloud and rain. In order to use a camera, not only good weather is needed, but
also sunlight to illuminate the ground.
The abbreviation radar comes from RAdio Detection And Ranging and
the basic principle of radar is to illuminate an object with the help of elec-
tromagnetic (EM) radiation and analyze the received response. In order to
radiate EM waves, a transmitter and an antenna are needed. The transmit-
ter generates and amplifies a signal. The signal goes from the transmitter to
an antenna which radiates EM waves that contain the signal from the trans-
mitter. In order to later receive an echo, another antenna is needed which is

1
2 Introduction

(a) Bistatic setup (b) Monostatic setup

Figure 1: Two different configurations of a basic radar system. For the bistatic
system, there are two antennas and they are separated by a distance B, here
called a baseline.

connected to a receiver. The receiver uses a detector to detect the incoming

signal by means of e.g. a matched filter. Such a system with one transmitter,
one transmitting antenna, one receiver and one receiving antenna is called a
bistatic radar system. It is possible, however, that the transmitter and receiver
share one common antenna; such a system is called a monostatic system. In
Fig. 1a and 1b, the bistatic and monostatic radar configurations are shown
respectively. Another radar configuration is the multistatic setup that make
use of many transmitters and receivers that together form a radar network. A
kind of multi static radar systems are passive bistatic or passive multistatic
systems. In passive systems, the echoed signals from one or many third part
transmitters are used. This can, for instance, be GPS or TV signals.
Radar systems were introduced in the beginning of the 20th century, how-
ever, it was not used widely until the 1930s, when the military realized the use-
fulness of this technology. Most of the continued development of radar has also
been conducted by the military and in Sweden some famous radar systems have
been developed, including ERIEYE, PS-05/A, GIRAFFE, ARTHUR, LORA
and CARABAS-II. Data from CARABAS-II and LORA has been used in the
work presented in this thesis.
Fully coherent radar needs very precise oscillators. Therefore, the system
has precise knowledge of the transmitted signal and with the use of GPS in-
formation it can get good information about the position of the transmitter
and receiver. Since the wave speed in air is well known, and also the position
of the transmitter and receiver, the flight time, and therefore also the distance
 3

the wave has traveled from transmitter to receiver can be precisely measured.
The most common type of radar is probably the one that incoherently
transmits short pulses. This type of system uses a magnetron to generate a
pulse which has a large power over short time. For long-distance surveillance
high-performance systems, it is more common to transmit a wide-band pulse
over a long time and afterwards perform pulse compression. In these radars, it
is common to use chirp waveforms. In order to obtain a good range resolution,
the transmitted signal is saved and later convolved with the received signals,
something referred to as pulse compression. A standard chirp signal can be
written as
 
2 Tp Tp
st (τ ) ∝ e−i2π(φ0 +fc τ +kτ ) , τ ∈ − , (1)
2 2

where τ is fast-time, φ0 an arbitrary phase, k is the chirp rate, fc is the

center frequency and Tp is the pulse length.
For a monostatic systems, the received echo signal from an object at the
distance R will be

 
Tp Tp
 
2R 2
c )+k (τ − c )
−i2π φ0 +fc (τ − 2R
sr (τ ) ∝ e ,τ ∈ − + τd , + τd (2)
2 2

∗
After convolving sr (τ ) with its matched filter namely (st (−τ )) , where ∗
denotes complex conjugation, the pulse compressed signal spc (τ ) can be deter-
mined. This signal can be shown to be
 2R !
 − τ
ei2π( c −τ )fc
2R
spc (τ ) ∝ Tp 1 − c
Tp
 2R  !!

2R
  − τ 
2R 2R

c
· sinc 2πkTp −τ 1− ,τ ∈ − Tp , + Tp
c Tp c c
(3)

The windowed sinc function has a much better resolution in range compared
to sr (τ ). We say that we have compressed the signal energy to a short time
duration. The way to obtain good resolution in elevation and azimuth for a
radar is by creating very narrow beams.
4 Introduction

2 Synthetic Aperture Radar

In SAR systems as well as in traditional radar, both monostatic and bistatic
configurations are possible. In this section, a monostatic SAR system is as-
sumed.
The basic principle of SAR is the collection of radar echoes from a hypothet-
ical stationary ground scene as the SAR system moves along a straight flight
track, called the synthetic aperture. The longer the synthetic aperture is, the
larger the integration angle will be. For a point in the middle of the SAR scene,
the synthetic aperture covers a certain viewing angle, and this angle is called
the integration angle. The resolution in a final SAR image is a function of
the integration angle, signal bandwidth and centre frequency only. This makes
the resolution independent of the distance from the scene. That means that a
satellite imaging a scene on the ground from long distance can obtain the same
resolution as an aircraft imaging the same area from very short distance. This
can be contrasted with optical sensors that are limited by the resolution in the
viewing angle.
In the same way as for each radar echo, pulse compression is used, but this
time to compress the signal along the synthetic aperture. As the SAR platform
moves along a hypothetical straight flight track, the radar emits and receives
pulses. Considering one target of interest, the difference between pulses will be
determined by the attenuation of the signal from the path loss and the antenna
pattern as well as by the change in distance to the target. Based on this, the
pulse-compressed signal before SAR processing is

  
 2R(t)
 c − τ

2R(t)
spc (τ, t) ∝ Tp 1 −  ei2π( c −τ )fc
Tp
    
 2R(t)
− τ
  
2R (t)  c 
· sinc 2πkTp
 −τ  1− 
c Tp
 
2R (t) 2R (t)
τ∈ − Tp , + Tp (4)
c c

where the
q range to a target is a function of t, so-called slowtime, and is given by
2
R (t) = (X0 − vp t) + Y02 . Here, X0 and Y0 are the azimuth and slant-range
coordinates for the target. Slowtime is the azimuth time vector, which relates
 5

to the movement of the SAR platform during illumination of the scene of in-
terest. Based on this pulse-compressed data, many algorithms that are able
to obtain a SAR image have been proposed. Among algorithms performing
well even for large integration angles, Range Migration(RMA)[7] and Global
Backprojection(GBP)[8] should be mentioned. Consequently, their basic func-
tions are presented later in the introduction.

3 Earth Observation, Surveillance and Security

Radars are able to detect targets very far away, as well as able to create images
from a large distance, such as radar images of the moon and of the interna-
tional space station from Earth or images of planet surfaces from interplanetary
space probes. Radars have also been developed to have the abilities to track
satellites, detect changes over time such as change in ice sheets or deforesta-
tion of rain forest and generate topographic maps. These different abilities
to gather important knowledge on the health of Earth has been used during
many years already. Important European Space Agency (ESA) Earth observa-
tion satellites include ERS-I, ERS-II and ENVISAT, all having a SAR system
on board. Recently launched and highly important satellites are TerraSAR-X
and Tandem-X, which are two German SAR-satellites, as well as the Italian
Cosmos-Skymed which will have a system of 4 SAR-satellites. The last men-
tioned projects are combined civilian and military satellites which also have
modes of operation able to detect moving targets. It should also be mentioned
that the first of the coming set of 5 different ESA SENTINEL satellite missions
is a SAR satellite mission. Furthermore, a P-band SAR for Earth biomass mea-
surements is a possible future satellite SAR project which is evaluated by ESA.
All these former, current and future SAR satellite systems show how greatly
ESA values SAR systems in space.
As has already been mentioned, SAR has some specific advantages in regard
to operation in darkness and in bad weather conditions. Important is also the
high precision oscillator that forms pulses that are almost exactly known in time
and the fact that the position of the SAR platform for each pulse is known with
high accuracy. These properties makes the sensitivity for change between two
SAR images over the same area, but separated in time, high. The decorrelation
between measurements is mainly driven by clutter statistics. This fact is the
base for interferometric measurements that has been used to measure small
movement such as land raise from earthquakes and land surface movements
and also the base for the possibility of topographic measurements by using
6 Introduction

multiple passes. The mentioned decorrelation property mentioned above has

been used in so-called coherent change detection [9]. Change detection refers
to that the same ground scene is imaged two times during two flights with an
aircraft or during two passes by a satellite. If instead two or more receivers are
mounted on the same platform, displaced along the flight direction, a small time
difference between the two SAR images is obtained. This creates the ability to
detect movement in an image. In this case, the decorrelation of the clutter will
be very low between the measurements. In surveillance of, for instance, land
or sea, the use of formation flying satellites, using a high-pass filter in time -
also known as MTI filtering - makes it possible to detect moving targets.

4 SAR Processing
As with traditional radar, the first step in many SAR processing algorithms is
pulse compression. After this, the SAR image is formed using either processing
in the 2D frequency domain or in the 2D spatial domain. In connection to
this thesis, many algorithms have been implemented [10]-[13]. However, only
two basic algorithms will be explained here. In this section, a spatial-domain
algorithm, namely GBP will first be presented, followed by a frequency-domain
algorithm named RMA. The basic SAR geometry is shown in Fig. 2.

4.1 Global Backprojection

In GBP [8], the SAR image is formed by evaluating the GBP integral, which
is given by
Z ∞  
R (Xp , Yp )
h(Xp , Yp ) = spc 2 , t dt (5)
−∞ c
For each pixel position, (Xp , Yp ) in the SAR image, a coherent summation of
the radar echoes over all slowtime positions is carried out. In Fig. 3a, the
procedure of GBP is illustrated. In Fig. 3b - 3d , SAR images generated using
GBP with an increasing part of the aperture are shown.

4.2 Range Migration Algorithm

When using the Range Migration Algorithm (RMA)[7], the idea is to do the
processing in the 2D frequency domain, also known as the 2D wave-domain.
The main step in RMA is a change of variables in the 2D wave-domain,
 7

η0 η
h
R(t0)=Y0

α
ξ0=X0

Figure 2: The general SAR imaging geometry consists of a platform moving

along a straight path on an altitude h. The integration angle α is the angle
formed from the endpoints of the aperture and the centre of the image area.
The cartesian coordinates spanning the ground plane are denoted by ξ and
η. Also, t0 is the slow time when the distance from the airplane to the scene
centre is shortest.
8 Introduction

(a) Illustration of Global Backprojection. To the left, the SAR

data after pulse compression is shown. To the right, the process
of backprojecting this data onto the image is illustrated.
Local Azimuth [m]
Local Azimuth [m]

50 50

100 100

150 150

200 200
50 100 150 200 50 100 150 200
Local Range [m] Local Range [m]
(b) GBP image using one pulse (c) GBP image using one third of the
for a scene with 5 point-like targets. aperture for a scene with 5 point-like
targets.
Local Azimuth [m]

50

100

150

200
50 100 150 200
Local Range [m]
(d) GBP image using the complete aperture
for a scene with 5 point-like targets.

Figure 3: SAR image formation using Global Backprojection(GBP) consists of

the backprojection of the pulse-compressed data onto the image grid in a pulse-
by-pulse manner. In the above example, the increase in SAR image resolution
as more radar echoes are backprojected, is shown. In this simulation, 5 point-
like targets were chosen.
 9

(a) The range wavenumber and the azimuth (b) The slantrange wavenumber and the az-
wavenumber space. imuth wavenumber space.

Figure 4: SAR image formation using the Range Migration Algorithm(RMA)

consists essentially of the transformation from a to b.

p
kρ = kR 2 − k 2 , where k is the wavenumber in slant range, k is the az-
x ρ x
imuth wavenumber and kR is the wavenumber in pulse-compressed raw data.
After this, the signal is multiplied with two functions and transformed back to
the spatial domain, giving a SAR image. The first function, which is a pure
magnitude function, is
q
S1 = kρ / kx2 + kρ2 (6)
while the second one, which is a pure phase function, is
√ 2 2
S2 = e−j (rc kx +kρ −kρ ) (7)

5 Moving targets in SAR

In a SAR image, the image of a moving target is typically effected in two ways,
namely through defocusing and by shifting its position. For a moving target,
its phase information in the SAR image is modified, compared to that of a
stationary target, in a way determined by target motion parameters. This fact
is utilized in this thesis. An illustrative example, produced by moving targets in
SAR imagery, is the one produced by trains when imaged by spaceborne SAR
systems. In [14] the train appears to move parallelly to the train track. In very
10 Introduction

high resolution SAR, such as UWB SAR, that makes use of long integration
times, the moving target will have its energy spread out in a curved shape,
either as a parabola or a hyperbola [15]. The papers in this thesis focus on
moving targets and their effects on SAR images.

5.1 Model for radar echo signal

For detectability and low false alarm rate, it is important to model a received
radar echo and a SAR image with and without a moving target present. This
means that we have to have an appropriate model for the target, the clutter
and the noise. In the first two papers of this thesis, it is for simplicity assumed
that the radar echo is described by (8)

x=s+c+n (8)
where, as presented in Paper IV, x is the radar echo, s the moving target
signal, c the clutter signal and n is the noise. Clutter in this case refers to
everything in the echo that is not noise and not a moving target. This can
be anything that reflects the radar signal, such as grass, water, trees, roads,
houses, power lines etc.
Furthermore, it is generally assumed that the SAR processing is a linear
system, i.e. if the radar echo is the sum of moving target signal, clutter signal
and noise, then the SAR image is the sum of the SAR image of the moving
target, the SAR image of the clutter and the SAR image of the noise as in (9)

hx (Xp , Yp ) = hs (Xp , Yp ) + hc (Xp , Yp ) + hn (Xp , Yp ) (9)

where h (Xp , Yp ) is the SAR image at the azimuth image coordinate Xp and
slant range image coordinate Yp .
This model assumes that the return echo from the moving target is indepen-
dent of the clutter, i.e. anything surrounding the moving target. This would
mean that as the target moves e.g. along a road, where for certain angles the
moving target is obstructed by trees or houses, the radar echo must not be
affected by the trees of the houses. As this cannot be true, the return echo
cannot be considered independent of the clutter and the model is not true.
However, within signal processing, most models are of the form given in (8),
mainly due to simplicity when implementing filtering which in our case has
the purpose to filter out the noise and clutter from the radar echo and only
keep the moving target. In the following subsection, several methods for such
filtering is presented.
 11

5.2 Moving Target dynamics model

In general, movement in a 2-D surface such as flat ground can be described by
a polynomial function in slow time, t.

aξ 2
ξ (t) = ξ0 + vξ (t − t0 ) +(t − t0 ) + . . .
2
aη 2
η (t) = η0 + vη (t − t0 ) + (t − t0 ) + . . .
2
ζ (t) = 0 (10)

where v is speed, a is acceleration, ξ and η speed components in azimuth and

slant range and t0 the time instant where the distance between the target and
the platform is minimal. As the time for SAR data collation is usually short,
it is throughout the research in this thesis assumed that there are no target
accelerations giving (11). There are cases in this thesis where the collection
times are not short, in which one can argue about the validity of the model,
as for VHF SAR data using CARABAS-II [16]. Having wavelengths between
3 and 15 meter causes only very large targets to give reflections. A large
target is heavy and thus cannot accelerate very quickly. Also, the vibrations
for the moving vehicles do not not cause severe signal loss, since the vibration
amplitude is small in comparison to signal wavelength. This may explain why
the approximation of no accelerations still give the good results in Part II.

ξ (t) = ξ0 + vξ (t − t0 )
η (t) = η0 + vη (t − t0 )
ζ (t) = 0 (11)

5.3 Increasing Detectability of Moving Targets

In order to extract parameters and image moving targets in SAR images, one
must detect and localize the target. This is done by applying signal processing
algorithms that increase the detectability. Such algorithms can be divided into
different categories, for instance based on the number of channels used:

1. Single-channel

2. Dual-channel
12 Introduction

Figure 5: Standard setup for a 4-channel along track SAR system used for
moving target detection, with the baseline of length B.

3. Multi-channel

In Fig. 5, a standard SAR setup for the detection of moving targets is shown.
The two antennas alternate as transmitters and both antennas are connected
to the receiver. In multi-channel systems, more antennas are mounted in along
track direction.

Single-channel methods
One antenna on a SAR platform is most common and therefore much effort has
been put into being able to detect moving targets for single-channel SAR,[14]-
[22]. Firstly, as mentioned at the start of the section, moving targets appear
smeared in a SAR image, for low bandwidth systems it appears just as a line,
similar to taking a photo of a car in the night, where the lights become lines.
Knowing the speed and direction of movement for the moving target it can be
focused, by reprocessing the data according to the target dynamic properties.
This is called moving target focusing. By focusing the target, it becomes more
prominent in the image in comparison to the original SAR image, thus increas-
ing the detectability. Secondly, the phase information of a smeared moving
 13

target in a SAR image is in azimuth direction a chirp, i.e. the phase informa-
tion increases or decreases quadratically. This is only true for moving targets.
Trying to find such chirps in a SAR image is a method to detect moving tar-
gets. Such methods can employ, for instance, time-frequency analysis. Thirdly,
as the platform and the target move, several medium resolution images can be
generated. For each of these images, the target has moved a short distance
in between. This effect can be used by applying change detection using, most
easily, only subtraction of two images. Another method used for detecting tar-
gets involves on a patch-by-patch basis processing the SAR image [22]. For
each patch the image is multiplied with its own conjugate. If a moving target
is present, it will focus as its so-called residual phase or chirp will be compen-
sated by multiplying with the conjugate. This is based on the same idea as for
detecting a chirp. Finally, the total opposite of trying to detect the moving
target by finding it in the image is trying to find its shadow. As was mentioned
in the beginning of this section, a moving target is smeared and displaced in
the SAR image. However, on the original place of the target, a shadow can be
seen, since the target blocked the background behind it. Thus, trying to detect
these shadows is also a way to detect moving targets.

Dual and multi-channel methods

Adding a second channel strongly increases the ability to suppress the clut-
ter. There are a couple of standard methods in which two channels can be
used, including for instance the Displaced Phase Centre Antenna (DPCA) [14],
Along-track Interferometry (ATI) [14], Adaptive DPCA (ADPCA) and Space
Time Adaptive Processing (STAP) [20]. In all of these methods, two or more
antennas are places on the platform along the flight direction. Having more
than one antenna, the model given in (8) must be extended. First of all, it is
assumed that the antennas are placed equidistantly in the flight direction. For
simplicity it is also assumed that each antenna acts as transmitter and receiver
separately and that all antennas and receivers are identical. Given these con-
ditions, the signal of the k’th antenna at the azimuth position x0 is given by
(12).

 
(k − 1) d
xk (t) = x t−
vp
     
(k − 1) d (k − 1) d (k − 1) d
= s t− +c t− +n t− (12)
vp vp vp
14 Introduction

where t is slow time.

Now we make the assumption that the clutter does not change during the
small time differences d(k−1)
vp , which is typically much less than 0.1 seconds,
and we assume that noise is a stationary random process. This allows us to
rewrite (12) to (13).
   
(k − 1) d (k − 1) d
xk (t) = x t − =s t− + c (t) + n (13)
vp vp
The DPCA method assumes that as the platform moves, the pulse repetition
Nv
frequency is chosen by P RF = d p , where N is any positive integer, vp the
platform speed and d the distance between the antennas.
Using the same sub-aperture for both antennas to generate an image, the
two antenna signals can be subtracted for each aperture position, then applying
SAR processing to generate an image. Alternatively, two images are formed
and then the two images subtracted. If some target moves, and the antenna
channels are identical, then the difference of the channels is decided only by the
movement of the target. If the two signals are then subtracted from each other,
the moving target signal is the only information left in the signal. In reality
there are always imperfections which reduce the effectiveness of DPCA, such
as e.g. antennas not having the same antenna pattern, receiver characteristics,
etc. Moreover, in reality there is always noise present. The noise will not be
suppressed in the DPCA, rather the SNR worsens. The DPCA method can be
written mathematically as

x1 − x2 1 · x1 + (−1) · x2
y= = (14)
x1 + x2 x1 + x2
Combining (13) with (14) gives
 
s (t) − s t − d
vp + n0
y= (15)
x1 + x2
where n0 is a random process. This shows that even though SNR was lost,
the Signal to Clutter Ratio (SCR) gain is very significant.
Similarly as for DPCA, subtraction of two channels is the central part for
ATI. But instead of calculating the difference of the measurements or the im-
ages, the difference of the phase of the signals are calculated. This is conducted
by multiplying the first signal with the conjugate of the second signal as pre-
sented in (16).
 15

y = x1 ◦ x∗2 (16)
In this equation, ◦ refers to the Hadamard product, or element-wise multi-
plication.
Adding adaptivity to the filtering is usually made using either ADPCA or
STAP. ADPCA is an extension of DPCA in which the coefficients for combining
the two channels are adaptively updated, instead of always set to 1 and -1. The
coefficients are calculated based on the covariance matrix of the clutter between
the two channels. How often the coefficients are updated is an implementation
choice.
STAP takes ADPCA much further, allowing for K number of channels and
each fast time sample of the measurements vectors to have a filter coefficient,
i.e. in the case of two channels, instead of 2 coefficients as with DPCA, it
allows for 2M coefficients if M is the number of fast time samples. STAP
has the ability just as DPCA to suppress clutter, but it can also be used for
suppressing for instance jammers. Also, in STAP in contrast to DPCA, the
SNR is not degraded as much. STAP filtering can be described by (17)
X
y= wk ◦ x k (17)
k

where xk is a vector of range samples from antenna k and wk are filter

coefficients for antenna k. This means that the filtered signal is a weighted
sum of the k channels in which each sample also has a different filter coefficient.
Thus, the filter has many more coefficients than the DPCA. We can choose
 
w1
w =  w2 
 
..
.
and  
x1
x =  x2 
 
..
.
and then the optimal coefficients for w are found to be
w = C−1 u (18)
16 Introduction

in which C is the clutter and noise covariance matrix, and u is the steering
vector which depends on moving target dynamics parameters. The adaptive
part of STAP refers to that the optimal coefficients changes depending on the
clutter and noise statistics. These statistics changes depending on, for instance,
time, range and angle. In the thesis in Part II, STAP filtering is used. However,
x,w,u and C are defined a little differently.

5.4 Imaging of Moving Targets

As explained for the single-channel methods, focusing of the target according
to target dynamic properties is possible. Since imaging of a moving target and
estimation of moving target parameters are very closely linked, the reader is
here introduced to the concept of focusing of moving targets and the important
parameter NRS, which also will help in understanding the last section in the
introduction.
If it is possible to exactly know the movement of the target during the SAR
data collection time, then it is possible to focus the target to obtain an image
of the target appearing as if it would be stationary, thus producing a normal
SAR image of the target.
Based on the model for a moving target (11), and the model for the platform
movement given by

ξp (t) = ξ0 + vp (t − t0 ) (19)
ηp (t) = 0 (20)
ζp (t) = h (21)

where p refers to platform and h to platform altitude, it is possible to

formulate the distance between platform and target as
q
2
R (t) = γ 2 (vp t − Xo ) + Yo2 (22)

where γ is the normalized relative speed (NRS), and Xo and Yo the azimuth
and slant range image coordinates of the target.
As presented in paper I, γ is found to be
s
2
(vp − vξ ) + vη2
γ= (23)
vp2
 17

The most basic time-domain SAR algorithm is GBP, described by (5). The
algorithm consist of a coherent summation of radar echoes along the synthetic
aperture for each image pixel. In the summation, the distance for each azimuth
position to the image pixel is calculated. Since there is a moving target, the
range-history will not be the same as for a stationary image pixel. Instead we
can consider that the image pixel grid moves with the same NRS as that of the
moving target. This will make the moving target focused in the SAR image.
This simple principle of considering that the image grid moves according to a
certain NRS generalizes to all SAR algorithms, and allows for the focusing of
moving targets.
However, if instead we want to focus a moving target based on a SAR image
only, not having the radar echo data, we have to modify the focusing method.
Such a method is presented in part I.

5.5 Estimation of Moving Target Dynamics Parameters

Estimation and detection are very closely linked together. With correct target
dynamics parameters in STAP and SAR processing, the target becomes focused
and best chance for detection is obtained. In contrary, when trying to estimate
target parameters we have detected a moving target, but we do not know its
speed or moving direction, also we do not know the position of the target.
What we want is to determine are these parameters. In order to accurately
estimate the position, we need to focus the target as much as possible, i.e. we
have to find the speed and moving direction of the target first.
As mentioned in 5.3, a moving target can be detected by for instance trying
to detect a chirp in the SAR image. This chirp also contains information about
the NRS allowing it to be estimated. This information can be extracted by for
instance phase unwrapping or by using time-frequency techniques such as the
Wigner-Ville transform combined with the Hough transform [21]. Another way
to find NRS is by attempting to refocus the target locally in the SAR image
by varying the NRS as the focusing parameter and using the peak magnitude
value of the moving target in the SAR image as the objective function. This
method is presented in part I of this thesis. NRS represents the magnitude
of the difference vector between the target and platform velocity vectors. In
order to also find the direction of the difference vector and thus determining the
target speed parameters the steering vector in STAP is used. Further, it is also
possible, if the moving target is very strong in its surrounding, to calculate a
2D FFT and from the SAR spectrum determine both NRS and target bearing.
18 Introduction

The position where the target appears in the image, and the velocity vector
is enough to determine the azimuth shift in the SAR image, and to reposition
the target. Thus, when the moving target velocity vector has been estimated
and the target is focused in the SAR image, the target position can be esti-
mated.

5.6 Paper I
In paper I, a moving target estimation and SAR focusing technique is presented.
The method is based on single channel SAR images. The main contributions
are the derivation of an expression for the phase in a SAR image as a function of
target NRS and the development of a refocusing algorithm, and its integration
with the estimation method. This paper also proposes an iterative method
for estimating the NRS of a moving target in a SAR image and it presents
successful results from a simulation as well as from using CARABAS-II data.

5.7 Paper II
In paper II, a filtering method for multichannel SAR data is used to filter
real radar data from the multichannel UHF SAR system LORA. The filtering
method is composed of combining target focusing using NRS with STAP and
integrating it into the SAR processing algorithm denoted Fast Backprojection
(FBP). In the measurement, a corner reflector was mounted on the back of a
truck and the truck drove along a road. To evaluate the result, the Signal-To-
Clutter-Ratio before and after filtering was measured. The results show a gain
of around 19dB and also a good ability to suppress strongly reflecting objects.

5.8 Paper III

In Paper III, a study on linear apodization is presented. The aim of the study
is to evaluate suitable windows for use in UWB SAR imaging. The Cosine-
On-Pedestal is a very flexible window due to its trade-off parameter, and it
is therefore chosen for the study. Limiting the window to belonging only to
Cosine-On-Pedestal was needed in order to focus on the effect of choosing
suitable parameters for SAR systems with small to large integration angles
and small to large bandwidths. An evaluation was made with regard to main-
lobe broadening and ISLR. It was found that polar windows were better than
rectangular shaped windows even for systems with small integration angles and
small bandwidths. Furthermore, it was found that for systems with matched
 19

azimuth and range resolution it was best to use the same parameter value in
the azimuth and range direction. It was also found that trying to suppress
ISLR below -25dB leads to a very large increase in the main lobe per dB of
suppressed ISLR.

5.9 Paper IV
In Paper IV, the main contributions are the derivation of bounds for how
good parameters can be estimated for targets in SAR. These parameters in-
clude position, speed in azimuth and range, target radar cross section (RCS)
and phase of target. The derived bounds are validated by simulations using
single channel SAR for the position parameters in azimuth, slant range and
the speed parameter normalized relative speed (NRS). The paper presents a
method showing how to numerically evaluate the derived equation to obtain
the estimation bounds.
References

[1] J.C. Farman, B.G. Gardiner and J.D. Shanklin, ”Large losses of total ozone
in Antarctica reveal seasonal ClOx/NOx interaction,” Nature, vol. 315,
no. 6016, pp. 207-210, 1985.

[2] J.F. Vesecky, R.H. Stewart, ”The observation of ocean surface phenomena
using imagery from the SEASAT synthetic aperture radar: An assessment,”
Journal of Geophysical Research, vol. 87-5, pp. 3397-3430, 1982.

[3] E. Rignot, J.L. Bamber, M.R. van den Broeke, C. Davis, Y. Li, W.J. van
den Berg, E. van Meijgaard, ”Recent Antarctic ice mass loss from radar
interferometry and regional climate modelling,” Nature Geoscience, vol.
1, pp. 106-108, 2008.

[4] D. Lu, ”Recent Antarctic ice mass loss from radar interferometry and re-
gional climate modelling,” International Journal of Remote Sensing, vol.
27, pp. 1297-1328, 2006.

[5] E. R. Stofan et al, ”The lakes of Titan,” Nature, vol. 445, no. 7123,
pp.61-64, 2007.

[6] V.A. Kotelnikov, A.F. Bogornolov, O.N. Rzhiga, ”Radar study of Venus
surface by Venera-15 and -16 spacecraft,” Advances in Space Research,
vol. 5, no. 8, pp.5-16, 1985.

[7] C. Cafforio, C. Prati, and F. Rocca, ”SAR Data Focusing Using Seismic
Migration Techniques,” IEEE Transactions on Aerospace and Electronic
Systems, vol. 27, no. 2, pp. 194-207, 1991.

[8] L.E.Andersson, On the Determination of a Function From Spherical Aver-

ages SIAM Journal on Mathematical Analysis, vol. 19, no. 1, pp. 214-232,
1988.

[9] L. M. H. Ulander, M. Lundberg, W. Pierson, and A. Gustavsson, Change

detection for low-frequency SAR ground surveillance IEE Proceedings on
Radar Sonar and Navigation, vol. 152, no. 6, pp. 413-420, 2005.

21
22 REFERENCES

[10] T.K. Sjogren, V.T. Vu, M.I. Pettersson, A comparative study of the po-
lar version with the subimage version of Fast Factorized Backprojection in
UWB SAR International Radar Symposium 2008, Wroclaw, Poland, May
2008.

[11] V.T. Vu, T.K. Sjogren, M.I. Pettersson, Ultrawideband Chirp Scaling Al-
gorithm IEEE Geoscience and Remote Sensing Letters, vol. 7, no. 2, pp.
281-285, 2010.

[12] T.K. Sjogren, V.T. Vu, M.I. Pettersson, Moving target refocusing algo-
rithm for synthetic aperture radar images IEEE Gescience and Remote
Sensing Symposium 2010, Honolulu, HI, July 2010.

[13] V.T. Vu, T.K. Sjogren, M.I. Pettersson, Fast factorized backprojection al-
gorithm for UWB SAR image reconstruction IEEE Gescience and Remote
Sensing Symposium 2011, Vancouver, Canada, July 2011.

[14] S.Chiu, MV.Dragos̆ević, Moving Target Indication via RADARSAT-2

Multichannel Synthetic Aperture Radar Processing EURASIP Journal on
Advances in Signal Processing, 2010.

[15] J.K.Jao, Theory of synthetic aperture radar imaging of a moving target.

IEEE Transactions on Geoscience and Remote Sensing, vol. 39, no. 9, pp.
1984-1992, 2001.

[16] H.Hellsten, L.M.Ulander, A. Gustavsson, B. Larsson, Development of

VHF CARABAS II SAR. In Proceedings of SPIE, vol. 2747, Orlando,
FL, Apr. 1996, pp. 48-60.

[17] H.Hellsten, L.M.H.Ulander, Airborne array aperture UWB UHF

radar–Motivation and system considerations. IEEE Aerospace and Elec-
tronic Systems Magazine, vol. 15, no. 5, pp. 35-45, 2000.

[18] C.H., Gierull , Ground moving target parameter estimation for two-channel
SAR. IEE Proceedings–Radar Sonar and Navigation, vol. 153, no. 5, pp.
224-233, 2006.

[19] V.T. Vu, T.K. Sjögren, M.I. Pettersson, A. Gustavsson, L.M.H Ulander.,
Detection of Moving Targets by Focusing in UWB SAR—Theory and Ex-
perimental Results. IEEE Transactions on Geoscience and Remote Sensing,
vol. 48, no. 10, pp. 3799-3815, 2010.
REFERENCES 23

[20] M.I. Pettersson, Detection of moving targets in wideband SAR. IEEE

Transactions on Aerospace and Electronic Systems, vol. 40, no. 3, pp. 780-
796, 2004.
[21] S. Barbarossa, A. Farina, Detection and imaging of moving targets with
synthetic aperture radar. 2. Joint time-frequency analysis by Wigner-Ville
distribution. IEE Proceedings–Radar and Signal Processing, vol. 139, no.
1, pp. 89-97, 1992.
[22] J.R. Fienup, Detecting moving targets in SAR imagery by focusing. IEEE
Transactions on Aerospace and Electronic Systems, vol. 37, no. 3, pp. 794-
809, 2001.
 Part I

Moving Target Relative Speed

Estimation and Refocusing in Synthetic
Aperture Radar Images
 Moving Target Relative Speed Estimation and
Refocusing in Synthetic Aperture Radar Images

Thomas K. Sjögren, Viet T. Vu, Mats I. Pettersson,

Anders Gustavsson and Lars M. H. Ulander.

Abstract

In this paper, a method for moving target relative speed estima-

tion and refocusing based on Synthetic Aperture Radar (SAR) images is
derived and tested in simulation and on real data with good results. Fur-
thermore, an approach on how to combine the estimation method with
the refocusing method is introduced. The estimation is based on a chirp
estimator that operates in the SAR image and the refocusing of the mov-
ing target is performed locally using subimages. Focusing of the moving
target is achieved in the frequency domain by phase compensation, and
therefore makes it even possible to handle large range cell migration in
the SAR subimages. The proposed approach is tested in a simulation
and also by on real Ultra Wide Band (UWB) SAR data with very good
results. The estimation method works especially well in connection with
low frequency UWB SAR, where the clutter is well focused and the phase
of the smeared moving target signal becomes less distorted. The main
limitation of the approach is target accelerations where the distortion
increases with the integration time.

1 Introduction
When observing the Earth from spaceborne or airborne platforms with the
use of imaging sensors such as Synthetic Aperture Radar (SAR) or electro-
optical cameras, it is possible to e.g. observe climate change, classify forest
and crop types, and monitor effects of earthquakes. In the case of an optical
sensor, the image is obtained during an extremely short time which is decided

27
28 Part I

by the shutter opening time. In the case of a SAR system, however, the data
is gathered over seconds and even minutes. On further comparison of the two
types of sensors, SAR is able to operate in bad weather and during the night.
SAR is also capable of detecting change and movement. It is possible to detect
e.g. moving ships or other vehicles on the ground. Such information can be
used for many different applications, e.g. surveillance of ships and traffic [1]-
[4]. In order to obtain good results for the purpose of e.g. traffic monitoring,
the SAR system should be equipped with at least two antennas physically
separated in the flight direction. The reason is that combining the antennas
allows for strong suppression of stationary scatterers, i.e. clutter [5], [6].
In Narrow Band (NB) SAR, which also can be referred to as conventional
SAR, the bandwidth of the signal is normally small compared to the center fre-
quency, often less than 10 percent. Two examples of conventional SAR systems
are the spaceborne Radarsat-2 [7] -[9] and TerraSAR-X [3],[10]. Such systems
often have integration angles of less than 10 degrees. For these conditions SAR
processing algorithms such as Chirp Scaling (CS) [11] are very efficient to form
the images. NB SAR systems often operate at high center frequencies. This
allows for a high resolution in the SAR image in terms of both range and az-
imuth, even when only a low integration angle and a low relative bandwidth
are available. However, for each frequency band, the higher the resolution, the
better the ability to detect and separate targets. If very high resolution is of
interest, larger relative bandwidths and larger integration angles are needed.
In these cases, Ultra Wide signal bandwidth and Ultra Wide beamwidth are
needed. We denote such systems UWB-SAR systems.
If there is an interest in imaging the ground in forested areas, the Foliage
Penetrating (FOPEN) capability characteristic of Low Frequency (LF) SAR
can be utilized. LF SAR systems must at the same time make use of UWB
SAR in order to obtain good resolution. Furthermore, having wavelength res-
olution such as CARABAS-II [12] allows for resolution cell often smaller than
the scatterer. This gives unique capabilities to separate e.g. target from trees.
To clarify further, in this paper we use the terminology that stationary scat-
terers are clutter. This should be understood by the fact that with a UWB
system at low frequencies, the number of scatterers per resolution cell is less
than one. This is a unique property that leads to very stable clutter signa-
ture between illuminations and the high performance, for example in change
detection [13]. LF SAR in combination with UWB is related to extreme con-
ditions such as very long integration times as well as a hyperbolic target range
history. These conditions demand that either algorithms such as Range Migra-
tion Algorithm (RMA) [14] or any of the family of Back Projection algorithms,
Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 29
e.g. Global Backprojection (GBP) [15], are used if good image quality is to be
obtained. When combining foliage penetration with the previously mentioned
SAR abilities, it is possible to detect changes under foliage [13].
Usually in SAR images, the signature of a moving object is contained in
one range cell. In this case, a one-dimensional matched filter can be applied
locally in order to focus the moving target. However, the signature of moving
objects in UWB-SAR images spreads over multiple range cells. In these cases,
a one-dimensional matched filter will be insufficient. Therefore, the refocusing
method needed to obtain a high resolution SAR image of the moving target
based on a SAR image depends strongly on the system parameters.
Many methods have been presented that estimate the speed of moving tar-
gets in SAR imagery [16] -[24]. In [22], a method for estimating Normalized
Relative Speed (NRS) is given using subapertures. There are also methods
which estimate target acceleration, as presented, e.g., in [17]. When using sin-
gle channel SAR, only one degree of freedom is usually available for parameter
estimation. Because of this, only the speed parameter in the along-track or
the across-track direction can be measured, while the use of more antennas
allows for more degrees of freedom. Some recent publications show solutions
on how to estimate more than one parameter using single channel SAR data
[18]. Once the parameters have been estimated, it is also important to obtain
a focused SAR image. Different methods have been suggested for this process.
For instance, based on an estimate of the speed parameter, GBP can be used
to form a focused, very small SAR image that surrounds the moving target[23].
Other methods apply a matched filter to the final SAR image or use an inter-
polation in the 2D frequency domain such as in [24], allowing for high-quality
refocusing of SAR images with long integration times, such as those produced
by UWB SAR systems. It must be noted again that this assumes constant
relative speed. In the refocusing methods given in [24], the basic assumption
is that the original SAR image was produced to image stationary objects on
the ground.
This paper presents a method to estimate the speed of a moving object
and to refocus it. Here we consider estimation and assume that the target
has been detected. The presented method can be used for detection, however
this is left out in this presentation. The estimator is operating on the phase
information of a SAR image. The model of the phase information is a more
general version compared to the one developed in [23]. The refocusing equation
is derived for a general case when the original SAR image was formed with an
arbitrary focusing speed, making it more general compared to [24] where only
stationary speed was considered. Further benefits of the method are that the
30 Part I

size of image for refocusing can be any, as long as it is rectangular. To our

knowledge, both the estimation method and the refocusing method are new and
can be used independently of each other. Moreover, in this paper an approach
how to combine the methods is presented.
The paper is divided into eight sections. In section II, the radar model used
and the theory of SAR moving target focusing employing NRS are presented.
In section III, an analytical expression is given for the phase of a moving target
in a SAR image. The procedure for obtaining the NRS estimate is presented
in section IV and in section V the refocusing method is derived. Section VI
shows results based on simulations. In Section VII, results using a real SAR
image with a simulated target superimposed is presented. Finally, in section
VIII and IX the findings are discussed and conclusions drawn.

2 The Radar model and Moving target focusing

To estimate speed of and to refocus a moving target in the SAR image, a model
is needed. The model used in this paper is based on a set of assumptions. We
assume a point target, a non-accelerating flying platform above a flat earth
and that the start-stop approximation is valid. However, as discussed later,
the platform motion compensation is already done in the image formation.
Thus, a signal model where the received signal is a time-shifted and attenuated
version of the transmitted signal is attained. Given any radar signal, the pulse-
compressed signal before SAR processing for a target is given by (1)
 
2Rt (t)
−j2π −τ fc
spc (τ, t) = p (τ, t) e c
(1)

where τ is fast time, p (τ, t) is the magnitude of the pulse compressed signal, t
is slow time, Rt (t) is the range history to target, c is the speed of light, and fc
is the center frequency of the radar signal.
If a target is moving on the ground during the SAR illumination time at
a constant speed, i.e. with no acceleration, we can parametrize the position
according to

xtc (t) = vx (t − t0 ) + x0
ytc (t) = vy (t − t0 ) + y0
ztc (t) = 0 (2)

where t0 is the time where the distance is at the minimum between the target
and the platform and (x0 , y0 ) is the position of the target at minimum range.
Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 31
In the same coordinate system we can also write the movements of the SAR
platform with constant speed vp , moving parallel with the x-axis. The platform
movements can be written as
xpc (t) = vp t
ypc (t) = 0
zpc (t) = h (3)
and the range between the target and the platform in Cartesian coordinates
can be described according to
R (vx , vy , vp , t) =
q
2 2
(vp t − vx (t − t0 ) − x0 ) + (vy (t − t0 ) + y0 ) + h2
(4)
Now, if we want to make a SAR image in which we focus a target with
the parameters x0 , y0 , vx , vy , we can rewrite (4) using image coordinates. They
take on the following form
q
2
Rt (t) = γt2 (vp t − Xt ) + Yt2 (5)
Identifying the t2 coefficients in (4) and (5), we get the NRS γt as well as
expressions for the image coordinates Xt and Yt of target positions according
to [25] s
2
(vp − vx ) + vy2
γt = (6)
vp2
vy
Xt = x0 − y0 (7)
(vp − vx )
and v
u !
u vy2
Yt = th2 + y02 1+ 2 (8)
(vp − vx )
where t0 = Xt /vp from (5) was used [25]. Based on this expression for a moving
target range history, we are able to process the SAR scene using
q
2
Rp (Xp , Yp , γp , t) = γp2 (vp t − Xp ) + Yp2 (9)
where γp is the processing NRS, Xp is the azimuth image coordinate and Yp
the slant range image coordinate in order to focus all targets with γt = γp .
The geometries are illustrated in Fig. 1 and Fig. 2.
32 Part I

Figure 1: SAR image formation geometry with respect to an arbitrary station-

ary ground point (Xp , Yp ). The integration time is here denoted by tn .

Figure 2: SAR image formation geometry with respect to a moving target. In

the figure, Xt is the aircraft position, and Yt is the range between the aircraft
and the target, at t = t0 . The integration time is here denoted by tn .

3 Phase of a moving target in a SAR image

Starting with the expression for the pulse-compressed signal given in (1), we
know that we can focus a moving target if γp = γt . At the same time as we
focus a moving target, all scatterers with other NRSs will become defocused
and therefore all stationary targets will defocus or smear.
In SAR processing, phase is fundamental. When looking at a defocused
target, information about its phase will be essential in order to be able to
estimate the speed of the target. Therefore, finding the phase from a target
with γt in a SAR image processed at γp is important.
Using the GBP integral [15] combined with NRS [19], the image response
can be represented by (10), where Xp and Yp are the processed SAR images
pixels and spc is the pulse-compressed target radar echo [26]. To evaluate the
Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 33

Z∞  
2Rp (Xp , Yp , γp , t)
h (Xp , Yp , γp ) = spc , t Rp (Xp , Yp , γp , t) dt
c
−∞
Z∞  
2Rp (Xp , Yp , γp , t) −j4πfc
= p , t e c [Rt (t)−Rp (Xp ,Yp ,γp ,t)] Rp (Xp , Yp , γp , t) dt
c
−∞
(10)

integral in (10) and derive the phase component of the moving target impulse
response, i.e.
−j4πfc
[Rt (t) − Rp (Xp , Yp , γp , t)] (11)
c
we use the approximation of stationary phase[27]. In order to find the time of
the stationary phase tst , we take the first derivative with respect to time in the
phase component of (10) and set it to zero. The time of stationary phase is
shown to be
Rt (t) γp2 Xp − Rp (Xp , Yp , γp , t) γt2 Xt
tst =   (12)
vp Rt (t) γp2 − Rp (Xp , Yp , γp , t) γt2
In the phase component of (10), we leave out j4πfc /c in the coming deriva-
tions and try to simplify it after inserting t = tst from (12). To start with, we
use the difference of two squares on (13) and obtain (14)
δR (tst ) = Rt (tst ) − Rp (Xp , Yp , γp , tst ) (13)
2 2
γt2 (vp tst − Xt ) + Yt2 − γp2 (vp tst − Xp ) − Yp2
R (tst ) = (14)
Rt (tst ) + Rp (Xp , Yp , γp , tst )
Equation (12) can be rewritten in the following form
γt2 Rp (Xp , Yp , γp , tst )
vp tst − Xp = (vp tst − Xt ) (15)
γp2 Rt (tst )
Substituting (15) into (14), the following expression is obtained
δR (tst ) =
γp2 γt2 (γp2 −γt2 ) 2

Rp (Xp ,Yp ,γp ,tst )
2 (Xp − Xt ) + Yt2 − Yp2
γp2 −γt2 R (t )
t st
(16)
Rt (tst ) + Rp (Xp , Yp , γp , tst )
34 Part I

Thus, we can write the final expression for the phase in the SAR image by
multiplying with 4π/λc in (16) as

θ (Xp , Yp ) =
γp2 γt2 (γp2 −γt2 ) 2
4π  Rp (Xp ,Yp ,γp ,tst )
2 (Xp − Xt ) + Yt2 − Yp2
γp2 −γt2 Rt (tst )
(17)
λc (Rt (tst ) + Rp (Xp , Yp , γp , tst ))

Since Rp (Xp , Yp , γp , tst ) and Rt (tst ) are very large in relation to the exten-
sion of the object after smearing in the SAR image, we can make the approxi-
mation
Rp (Xp , Yp , γp , tst )
≈1 (18)
Rt (tst )
and (17) can therefore be approximated by
γp2 γt2 2
γp2 −γt2
(Xp − Xt ) + Yt2 − Yp2
θ (Xp , Yp ) = 4π (19)
2λc Rt (tst )

4 Estimation procedure
The phase in (19) indicates that the phase in azimuth can be described as a
quadratic function. By estimating this quadratic term we can find γˆt . In our
model, we assume one target, clutter and noise. These three components of
the signal are assumed to be added together. In [28] it is shown that high
SNR allows for approximating the phase as a sum of the phase from the signal
and phase noise. Using this approximation, and that we assume that the slant
range of the target is known, we can write the phase in the SAR image in
azimuth direction according to [28] as
a0 2
θ (Xp , Yp ) |Yp =Yt = X + a1 Xp + a2 + n (Xp ) (20)
2 p
where n (Xp ) is a phase noise component from clutter, radio frequency inter-
ference, receiver and SAR algorithm. We will now use the method presented
in [28]. We sample the phase of a range bin, Θp,s , in the given SAR image.
This range bin in the SAR image is denoted by S (Xp ).
 
= [S (Xp )]
Θp,s = atan (21)
< [S (Xp )]
Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 35
The number of pixels used is experimentally decided. In this paper all pixels
within the 3 dB width were used. Having extracted these pixels, we can find
the second derivative of this sequence by

Θ̈p,s = ∆2Xp Θp,s (22)

where ∆2Xp is the finite difference operator [28]. This second derivative is of
interest as it directly corresponds to a0 in (20). We suggest that the same
estimator for the chirp rate as proposed in [28] and [23] is used, namely a
BLUE estimator. The estimate is then
1T C−1 Θ̈p,s
â0 = (23)
1T C−1 1
where C is the covariance matrix of the second derivative of the phase from
n (Xp ).
By obtaining an estimate of a0 , we can estimate γt by identifying the
quadrature term of (19) together with (20). Using the fact that for pixels
close to the image point (Xt , Yt ), Rt (tst ) can be assumed to be constant and
approximately equal to minimum range to target, the relation between a0 and
γt found from (19) is
 − 21
8π −2
γ̂t = + γp (24)
2λc Yt â0
In [29], a moving target and a stationary scatterer were placed at the same
position in the SAR image and the estimation accuracy was investigated by
comparing if C was based on the contributions from n or on assuming n to
be white noise. The results did not indicate any difference between the two
cases unless in a special case when the image was formed for γp = 1. For this
reason as well as for simplicity, C will in this paper be used in its simplest
form, namely based on a white noise approximation.

5 Refocusing in a SAR image

The idea is to refocus the moving target based on a processed SAR image. In
comparison to [24] however, the image does not need to have been formed for
γp = 1, but could be created using arbitrary relative speed within the limits
0 < |vrel | < 2 |vp |. In this paper, we propose a refocusing algorithm. First, one
unfocused subimage of any size is selected. This image is transformed to wave
domain by a two-dimensional FFT. Wave domain is here understood to be the
36 Part I

domain the SAR image is in after applying two-dimensional Fourier transform.

In this domain, the refocusing is applied using a reinterpolation of the wave
domain image. Then this signal is transformed back to the image domain and
if the correct γp is used the target will be focused. Range migration can be
described by the change of variable
q
kρstat = kR 2 − k2 (25)
x

where kR is the range wave number, kρstat is the slant range wavenumber when
focusing a stationary ground, i.e. γp = 1, and kx is the azimuth wavenumber.
If instead a moving target is to be focused, the azimuth wavenumber needs to
be rescaled with the NRS of that moving target, corresponding to [30]
q
kρM T = kR 2 − γ 2 k2 (26)
t x

However, if the original image is processed at an arbitrary NRS γp , the slant

range wavenumber for that speed will be [30]
q
kρstat = kR 2 − γ 2 k2 (27)
p x

and all moving targets with the NRS γp will be focused. If there is a moving
target with γt 6= γp , that target will not be focused. To focus the moving target
at γt from an image focused at γp , the following procedure should be carried
out.
• Selecting an area surrounding the moving target in the image processed
at γp .
• Transforming subimage to the wave domain.
• Applying the coordinate transform given by (28) which is found by in-
serting (27) in (26).
q

kρM T = kρ2stat + kx2 γp2 − γt2 (28)

• Multiplying with (29)

q  
−jr0  (γp2 −γt2 )kx
2 +k 2
 
ρstat −kρstat
kρstat e
q
(29)
kρ2stat + kx2 γp2 − γt2

• Applying 2D inverse Fourier transform.

Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 37

SAR Extraction of For i = 1 to

Subimage i
image subimages #targets

Moving
target
Sampling of
detector
phase in
azimuth
Refocus
SAR
NO subimage
processor
using t
Unwrap
phase and
raw data estimate t

NO
Final Insert subimages
Target
SAR into original YES i=#targets? YES
focused?
image SAR image

Figure 3: Schematic sketch illustrating how the estimation and refocusing ap-
proach is applied. The three gray blocks are outside the scope of this paper.

Table 1: Simulated System and Real CARABAS-II Parameters

Parameter Simulation CARABAS-II

Center frequency fc 350 MHz 52 MHz
Integration time tn 20 s 100 s
Bandwidth B 300 MHz 60 MHz
Minimum range r0 1412 m 9922 m
Platform speed vp 100 m/s 129 m/s
γt 1.0155
γˆt 1.0150

6 Simulation results
In this section, we present simulation results that illustrate how speed can be
estimated and a moving target refocused using the approach outlined above.
The simulations are based on parameters chosen for a low-frequency SAR sys-
tem on board an airborne platform at a distance close to the imaged ground
38 Part I

Table 2: Simulated Target Parameters

Target vx vy γt x0 y0
[m/s] [m/s] [m] [m]
A 4 0 0.9689 1288 925
B 1 0 0.9922 1288 975
C 5 −2 0.9613 1288 1000
D 2 0 0.9845 1288 1000
E −4 0 1.0311 1288 1025
F −2 0 1.0155 1288 1050

scene, leading to a large integration angle even at short integration times. In

Table 1, the sensor system parameters for the simulation are given.
The procedure in which the refocusing and estimation were made is illus-
trated in Fig. 3. First, subimages are extracted from the image and then each
subimage is focused by iteratively estimating and refocusing until the target
is focused. The estimated value of γt for the subimage is the estimate of the
target NRS, γt . After all subimages are refocused, they are combined together
to form the complete SAR image.
For the simulation, six targets with different speeds were chosen. The speeds
of the targets are found in Table 2. Based on the originally obtained image,
Fig. 4, the signatures for each target are extracted, thus producing six smaller
images. For each of these images, the estimation procedure described in section
IV was used to obtain an NRS estimate. The images were then refocused
according to the procedure presented in section V. In these refocused images,
new estimates were obtained and the images are once again refocused. For each
of the targets, up to three iterations are made. After this, the small focused
images of the targets produced by the process are combined and the final result
is shown in Fig. 5. The azimuth displacement of target C in the image seen
in Fig. 5 can be understood by the nonzero vy , causing Xt 6= x0 according to
(7). In Table 3, the NRS estimates obtained for all targets are found, as well
as the errors denoted δγt .
The errors are larger for the targets having larger |1 − γt |, thus leading to
Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 39
longer convergence times.

7 Real data results

7.1 Measurement scene

To investigate the performance of the method when using real data, a simulated
moving target was inserted into a real SAR image obtained by the CARABAS-
II system [12]. The SAR measurement was conducted on the Island of Visingsˆ
on the lake Vättern in Sweden. Also in Table 1, the CARABAS-II parameters
for this specific measurement and the moving target parameters are given. The
target has been inserted into a forested area of the original SAR image which
can be seen in Fig. 6. The weak signature of the unfocused moving target can
hardly be seen in Fig. 7, i.e. the zoomed part of Fig. 6. In order to evaluate the
method, it is important to find the signal level in relation to surrounding clutter
and noise, and therefore parameters like Signal to Clutter Ratio (SCR), Signal
to Noise Ratio (SNR), Clutter to Noise Ratio (CNR) and Dynamic Range of
SAR scene are of interest.
In this work we have chosen the integral method defined in [31] to evaluate

1 Dynamic Range of SAR Scene is defined as the ratio of the energy of strongest reflecting
object in the scene to the noise energy. The scene is in this case shown in Fig. 6.

Table 3: Simulation Results

Target γˆt γt δγt

A 0.9673 0.9689 0.0016
B 0.9922 0.9922 0.0000
C 0.9586 0.9613 0.0027
D 0.9841 0.9845 0.0004
E 1.0290 1.0311 0.0021
F 1.0150 1.0155 0.0005
40 Part I

0
A
20
Local Range[m]

40 B C
60 D

80 E

100 F

0 50 100 150 200

Local Azimuth[m]

Figure 4: Original simulated SAR image.

the properties. According to this definition, SCR in dB scale is given by

   
 ZZ   ZZ 
   
SCR = 10log  S dA − 10log  m (RCC ) dA (30)
   
   
AS,3dB  AR 
CC,3dB

In (30) S is the impulse response for the moving target after focusing and
AS,3dB its resolution area. RCC is the autocorrelation of the clutter, ARCC ,3dB
its correlation area and m (RCC ) the mean of the autocorrelation.
As a simulated target was inserted, the SCR was chosen and calculated
easily. The energy of the target was found from an image without clutter
surrounding. The SCR was in this case selected to be 5.7 dB. Similar to SCR,
the CNR and SNR were calculated. The clutter energy was found from the
imaged forest in Fig. 7 and the noise energy was found from an open water
area near the forest. From this and the measured CNR, the SNR was obtained
Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 41

0
A
20
Local Range[m]

40
B C
60
D
80 E

100 F

120
0 50 100 150 200
Local Azimuth[m]
Figure 5: Final image after iterative focusing of all targets seen in Fig. 4 and
composition of respective subimages.

as shown in Table 4. In Table 4, the Dynamic Range of the SAR Scene is

also shown. The Dynamic Range of the SAR Scene was calculated as the ratio
between the energy in the strongest reflecting object in the scene of Fig.6 to
the noise level. This Dynamic Range should therefore not be confused with the
total Dynamic Range of the system which is higher. A very notable number
is the CNR of only 10 dB. This relates to the very low backscatter from trees
at these low frequencies. In fact, at low frequency both forest attenuation and
backscatter are low and therefore at low frequency targets hidden even in dense
forest can be detected [13].

7.2 Measurement Results

The scene with the moving target superimposed is shown in Fig. 6 and a
zoomed-in version is depicted in Fig. 7. The pixels are extracted and the
42 Part I

Table 4: Measurement Parameters

Parameter Value
Signal to Clutter Ratio (SCR) 5.7 dB
Signal to Noise Ratio (SNR) 15.7 dB
Clutter to Noise Ratio (CNR) 10.0 dB
1
Dynamic Range of SAR Scene 38.9 dB

processing scheme illustrated in Fig. 3, using (24) for estimation, is applied.

The number of focusing iterations for the moving target was set to three.
In Fig. 8 and Fig. 9 the subimage processed at γp = γt is shown in
the surroundings of other subimages processed at γp = 1, i.e. for stationary
scatters. The target subimage can easily be identified to the other parts of
the image with a focused target in the middle and all other parts appearing as
white. The moving target is well focused with the method and the brightness
of the target is at the same level as the strongest clutter scatterers in the
surrounding subimages. This is expected in this case with a SCR of only 5.7
dB. However, the most amazing result is the strong suppression of the clutter
in the target subimage. The reasons for the suppression are mainly two: The
radar system operates at low frequency and that the system is an UWB system.
Normally for microwave SAR systems the backscattering from forest is caused
by volume scattering in the foliage. At low frequency, however, the scattering
occurs mainly by the trunk ground interaction, and the scattering can therefore
mainly be considered as a one scatter scattering mechanism. In UWB systems
the scatterers size is commonly larger than the resolution cell. Therefore, with
the used system there will be a low density of scatterers per resolution cell, even
in forested area, and in fact there will be many resolution cells per scatter. The
low density of scatters will not lead to a normal speckle process but rather to
the situation that each scatter will smear, resulting in unfocused scatters that
will have much lower peak values. The effect will be lower levels on peaks from
the clutter areas. However the total energy is the same in both cases. For
large integration angles this defocusing effect is large as seen, and it is also
illustrated in [19].
In this experiment, the target had an NRS of 1.0155 and the estimate
Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 43

200

400

600
Local Range[m]

800

1000

1200

1400

1600

1800

2000
0 500 1000 1500 2000
Local Azimuth[m]

Figure 6: SAR image of Visingsö where a simulated moving target has been
synthetized and placed in the forest. White corresponds to weakest intensity
and black to strongest intensity.

obtained after the third and last iteration was 1.0150, giving an error of 0.0005.
For a target moving only in azimuth, this corresponds to a speed of 2 m/s, and
an estimate of 1.93 m/s which gives a relative error of 3.5%. On the other
hand, for a target moving only in range, this corresponds to 22.76 m/s and an
estimate of 22.39 m/s, which gives a relative error of 1.6%. The estimate is
thus close to the true value of the target NRS, showing that even in the case
where the simulated target is superimposed into clutter that partly covers the
signal, these proposed approach appears to work.

8 Discussion
In order to obtain good results, the platform should preferably be located close
to the imaged scene. The reason is that for shorter integration times, the
dynamics change less for the moving targets and that high resolution SAR
images are produced in short time. Another advantage is that this allows the
platform to fly at a lower speed, allowing for higher accuracy when conducting
44 Part I

900

950
Local Range[m]

1000

1050

1100

950 1000 1050 1100 1150

Local Azimuth[m]

Figure 7: A zoom-in of the simulated moving target which has been added to
the SAR scene. White corresponds to weakest intensity and black to strongest
intensity.

the speed measurement since the accuracy of the actual speed measurement is
related to the ratio between the target speed and the platform speed. It should
be noted, however, that the method does not depend on these observations,
it only performs better when they are in place. What should be remembered,
also, is that the refocusing is made based on processed SAR images in which
the stationary objects have been focused, i.e. γp = 1. This means that motion
compensation has already been performed between the stationary objects and
the moving platform, and therefore the moving target focusing can more easily
be done. However, how appropriate the model of no acceleration in (2) strongly
depends on radar system parameters and actual target motion. Anyhow, it has
been shown at low frequencies that the model in (2) works on real targets for
very long integration times [19],[23]. One should further note the ability to
apply the refocusing method for detection purposes.
Moving Target Relative Speed Estimation and Refocusing in Synthetic Aperture
Radar Images 45

200

400

600
Local Range[m]

800

1000

1200

1400

1600

1800

2000
0 500 1000 1500 2000
Local Azimuth[m]

Figure 8: Final SAR image after iterative focusing and combination with origi-
nal SAR image. White corresponds to weakest intensity and black to strongest
intensity.

9 Conclusions

This paper proposes an approach to estimate the speed of a moving target and
to refocus it. The estimation of relative speed is based on a chirp estimator
that makes use of the target phase. Focusing is made in subimages where the
SAR image is transformed to the wave domain and compensated according to
the target NRS, γt and the NRS used in SAR processing of the image. The
estimation and the focusing are then combined in an iterative process. The
approach is illustrated with one simulation and one test in which a simulated
target is superimposed onto a complex SAR image. We can conclude that even
when the peak of the moving target signal is at the same level as the peaks of
the surrounding clutter, the target can be refocused successfully. The refocused
target can also be combined with the original SAR image, thus providing the
end user with good information on where the target was located in relation
to the surroundings as well as the magnitude of its NRS. To conclude, we can
say that the simulated and real results show that the proposed approach to
46 Part I

900

950
Local Range[m]

1000

1050

1100

950 1000 1050 1100 1150

Local Azimuth[m]

Figure 9: Zoom-in of the focused moving target of the final SAR image after
iterative focusing and combination. White corresponds to weakest intensity
and black to strongest intensity.

estimate and refocus appears to be functional.

Acknowledgment
The authors would like to thank the KK-Foundation for making this research
project possible, and the support from Swedish Defence Research Agency, Saab
Bofors Dynamics, Saab Microwave Systems and RUAG Aerospace Sweden.
References

[1] D. J. Weydahl, C. Brekke, P. Selvik, ÿ. Helleren and R. Olsen, “Ship

traffic monitoring using satellite SAR images in combination with AIS
reports,” in Proc. SPIE Remote Sensing for Environmental Monitoring,
GIS Applications, and Geology VII, vol. 6749, Florence, Italy, Sept. 2007,
pp. 67490D–67496D.

[2] S. V. Baumgartner and G. Krieger, “SAR traffic monitoring using time-

frequency analysis for detection and parameter estimation,” in Proc.
IGARSS, vol. II, Boston, MA, July 2008, pp. 25–28.

[3] S. Suchandt, H. Runge, H. Breit, U. Steinbrecher, A. Kotenkov and U.

Balss, “Automatic extraction of traffic flows using TerraSAR-X along-
track interferometry,” IEEE Trans. Geosci. Remote Sensing, vol. 48, no. 2,
pp. 807–819, 2010.

[4] M. V. Dragoševic and S. Chiu, “Space-based motion estimators - Evalu-

ation with the first RADARSAT-2 MODEX data,” IEEE Geosci. Rem.
Sens. Lett., vol. 6, no. 3, pp. 438–442, 2009.

[5] H. Hellsten and L. M. H. Ulander, “Airborne array aperture UWB UHF

radar - Motivation and system considerations,” IEEE Aerosp. Electron.
Syst. Magazine, vol. 15, no. 5, pp 35–45, 2000.

[6] L. M. H. Ulander and H. Hellsten, “Low-frequency ultra-wideband array-

antenna SAR for stationary and moving target imaging,” in Proc. Radar
Sensor Technology IV, vol. 3704, Orlando, FL, Apr. 1999, pp. 149–158.

[7] P. Meisl, A. A. Thompson and P. A. Luscombe, “RADARSAT-2 mission -

Overview and development status,” in Proc. EUSAR, Munich, Germany,
May 2000, pp. 373–376.

[8] T. Toutin and R. ChÈnier, “3-D radargrammetric modeling of

RADARSAT-2 ultrafine mode: Preliminary results of the geometric cali-
bration,” IEEE Geosci. Rem. Sens. Lett., vol. 6, no. 3, pp. 611–615, 2009.

47
48 REFERENCES

[9] S. Cote, S. Muir, S. K. Srivastava and R. K. Hawkins, “SAR image quality

and calibration operations for the RADARSAT satellites at the Canadian
space agency,” in Proc. RADAR, Bordeaux, France, Oct. 2009, pp. 1–6.

[10] R. Werninghaus and S. Buckreuss, “The TerraSAR-X mission and system

design,” IEEE Trans. Geosci. Remote Sensing, vol. 48, no. 2 pp. 606–614,
2010.

[11] R. K. Raney, H. Runge, R. Bamler, I. G. Cumming and F. H. Wong, “Pre-

cision SAR processing using chirp scaling,” IEEE Trans. Geosci. Remote
Sensing, vol. 32, no. 4, pp. 786–799, 1994.

[12] H. Hellsten, L. M. H. Ulander, A. Gustavsson and B. Larsson, “Develop-

ment of VHF CARABAS II SAR,” in Proc. SPIE Radar Sensor Technol-
ogy, vol. 2747, Orlando, FL, Apr. 1996, pp. 48ñ60.

[13] L. M. H. Ulander, M. Lundberg, W. Pierson and A. Gustavsson, “Change

detection for low-frequency SAR ground surveillance,” IEE Proc. Radar
Sonar and Navigation, vol. 152, no. 6, pp. 413–420, 2005.

[14] C. Cafforio, C. Prati and F. Rocca, “SAR data focusing using seismic
migration techniques,” IEEE Trans. Aerosp. Electron. Syst., vol. 27, no.
2, pp. 194–207, 1991.

[15] L. E. Andersson, “On the determination of a function from spherical aver-

ages,” SIAM Journal on Mathematical Analysis, vol. 19, no. 1 pp. 214–232,
1988.

[16] S. Axelsson, “Position correction of moving targets in SAR-imagery,” in

Proc. SPIE - The International Society for Optical Engineering, vol. 5236,
Barcelona, Spain, Sept. 2003, pp. 80–92.

[17] C. H. Gierull, “Ground moving target parameter estimation for two-

channel SAR,” IEE Proc. Radar Sonar and Navigation, vol. 153, no. 3,
pp. 224–233, 2006.

[18] P. A. C. Marques and J. M. B. Dias, “Velocity estimation of fast moving

targets using a single SAR sensor,” IEEE Trans. Aerosp. Electron. Syst.,
vol. 41, no. 1, pp. 75–89, 2005.

[19] V. T. Vu, T. K. Sjögren, M. I. Pettersson, A. Gustavsson and L. M. H

Ulander, “Detection of moving targets by focusing in UWB SAR - Theory
REFERENCES 49

and experimental results,” IEEE Trans. Geosci. Remote Sensing, vol. 48,
no. 10, pp. 3799–3815, 2010.

[20] S. Barbarossa and A. Farina, “Detection and imaging of moving targets

with synthetic aperture radar. 2. Joint time-frequency analysis by Wigner-
Ville distribution,” IEE Proc. Radar and Signal Processing, vol. 139, no.
1 pp. 89–97, 1992.

[21] F. Zhou, R. Wu, M. Xing and Z. Bao, “Approach for single channel
SAR ground moving target imaging and parameter estimation,” IET Proc.
Radar, Sonar and Navigation, vol. 1, no. 1, pp. 59–66, 2007.

[22] H. Zhou, X. Huang, Y. Chang and Z. Zhou, “Single-channel UWB SAR

ground moving targets detection method using change detection based on
single-pass sub-aperture images,” in Proc. APSAR, Huangshan, China,
Nov. 2007, pp. 266–270.

[23] T. K. Sjögren, V. T. Vu, M. I. Pettersson, H-J. Zepernick and A. Gus-

tavsson, “Speed estimation experiments for ground moving targets in low
frequency UWB SAR,” in Proc. RADAR, Edinburgh, UK, Oct. 2007, pp.
519-523.

[24] J. K. Jao, “Theory of synthetic aperture radar imaging of a moving tar-

get,” IEEE Trans. Geosci. Remote Sensing, vol. 39, no. 9 pp. 1984–1992,
2001.

[25] M. I. Pettersson, “Detection of moving targets in wideband SAR,” IEEE

Trans. Aerosp. Electron. Syst., vol. 40, no. 3, pp. 780–796, 2004.

[26] M. I. Pettersson, “Optimum relative speed discretisation for detection of

moving objects in wide band SAR,” IET Radar, Sonar and Navigation,
vol. 1, no. 3, pp. 213–220, 2007.

[27] J. C. Curlander and R. N. McDonough, Synthetic Aperture Radar Systems

and Signal Processing, Wiley, ISBN: 978-0-471-85770-9, 1991.

[28] P. M. Djuric and S. M. Kay, “Parameter estimation of chirp signals,”

IEEE Trans. Acoustics, Speech and Signal Processing, vol. 38, no. 12, pp.
2118–2126, 1990.

[29] T. K. Sjˆgren, V. T. Vu and M. I. Pettersson, “Moving target relative

speed estimation in the presence of strong stationary surrounding using a
50 REFERENCES

single antenna UWB SAR system,” in Proc. IGARSS, vol. I, Boston, MA,
Jul. 2008, pp. 157–160.
[30] V. T. Vu, T. K. Sjögren and M. I. Pettersson, “Moving target detection by
focusing for frequency domain algorithms in UWB low frequency SAR,”
in Proc. IGARSS, vol. I, Boston, MA, July 2008, pp. 161–164.
[31] L. M. H. Ulander, “Accuracy of using point targets for SAR calibration,”
IEEE Trans. Aerosp. Electron. Syst., vol. 28, no. 1, pp. 139–148, 1991.
 Part II

Suppression of Clutter in Multichannel

SAR GMTI
 Suppression of Clutter in Multichannel SAR GMTI

Thomas K. Sjögren, Viet T. Vu, Mats I. Pettersson, Feng Wang,

Daniel Murdin, Anders Gustavsson and Lars M. H. Ulander.

Abstract

In this paper, results of moving target detection in multichannel

UHF-band Synthetic Aperture Radar (SAR) data are shown. The clutter
suppression is done using Finite Impulse Response (FIR) filtering of mul-
tichannel SAR in combination with a 2-stage Fast Backprojection (FBP)
algorithm to focus the moving target using relative speed. The FIR filter
coefficients are chosen with the use of STAP filtering. Two parameters
are used for target focusing, target speed in range and in azimuth. When
the target is focused, both speed parameters of the target are found. In
the experimental results, two channels were used in order to suppress
clutter. In the resulting SAR images it is obvious that very strong scat-
terers and the forest areas have been suppressed in comparison to the
moving target in the image scene. The gain obtained can be measured
using SCNR gain, which is about 19dB. Another way to measure signal
processing gain is the ability to suppress the strongest reflecting object
in the SAR scene. The gain of target in relation to this object is 25dB.
This shows that using UHF-band SAR GMTI for suppressing forest and
increasing the target signal can work.

1 Introduction
Target detection and position determination constituted the core mission for
radar systems from the beginning. In early times, monopulse radar was used to
help to determine the position of a target. However, to be able to perform such
tasks, the target should stand out in respect to the surrounding, i.e. it first has
to be detected and its angle found with good precision. If a target can not be
found just by scanning in angle, then knowledge of the target such as signature

53
54 Part II

or Doppler generated by target motion can be used in order to detect it. To

increase detection ability, many echoes can be stored and combined together
for noise suppression. This suppression is due to the very short correlation
time of noise and the longer correlation time for the echo of a target. However,
we are not only interested in detecting an object and estimating its speed, but
also in obtaining an image of the object.
Synthetic Aperture Radar (SAR) has the ability to obtain high resolution
images of ground scenes including targets. Thus it is a suitable sensor for
target imagery. Normally in a SAR image, it is hard to directly detect moving
targets, especially if the clutter is strong and the target smeared and displaced
due to the movement of the target during the SAR integration time. In order to
reliably detect the moving target, new signal processing algorithms are needed.
Much research has been made in algorithms to detect the moving targets, to
estimate their parameters as well as to obtain focused images of the targets.
Most SAR systems are equipped with one antenna, and getting good results
with moving target detection, parameter estimation and target focusing for
single channel SAR require much effort [1]-[9]. For instance in [4], [5], the
detection is made in the SAR raw data, by detecting the azimuth chirp of the
moving target which is one characteristic of moving targets. In [2], [3], [7]-[9],
focusing of the target in the final image is used to detect moving targets. As
suggested in [6], many images can be created along the aperture and change
detection applied.
With the use of two or more channels, the chance of detecting a moving
target can be increased. For detection of moving targets in multi-channel SAR,
the Displaced Phase Center Antenna (DPCA) and Along Track Interferometry
(ATI) techniques have be used. DPCA basically subtracts the signal of two
different antennas, as they are moved in azimuth direction [10],[11], while the
ATI algorithm uses the phase difference information of the two antennas in
order to detect targets [10],[12]. Many experiments with DPCA have been
conducted, for instance [10],[11].
Space Time Adaptive Processing (STAP) is another selection for moving
target detection in multi-channel SAR. In STAP, instead of just subtracting
two signals, statistical properties are used in order to suppress the clutter and
noise. In e.g. [13] and [14], STAP for SAR was suggested and in [13] results
were also presented. Further results were shown in e.g. [15], [16]. However,
in most of these experiments, the operating frequencies of the systems were at
X-band or C-band, in which the foliage penetration ability is not as good as in
UHF and VHF. The experimental results shown in this paper are within the
field of STAP for lower UHF band SAR [17].
Suppression of Clutter in Multichannel SAR GMTI 55

2 Fast Backprojection
This section gives an overview of the Fast Backprojection algorithm (FBP)[18].
FBP is one special case of the Fast Factorized Backprojection algorithm (FFBP)
[19],[20]. FBP consists of one pre-stage and two processing stages. The main
stages and its pseudocode are presented in Algorithm 1.

Input: Rawdata
Output: Image
k=1;
Image = 0;
for k ≤ N umberOf Subapertures do
k = k + 1;
GeneratePolarGrid;
P olarImage=GeneratePolarImage(Rawdata) ;
CartesianImage=PolarToCartesian(P olarImage);
Image = Image + CartesianImage;
end
Algorithm 1: Fast Backprojection

In the pre-stage, the complete aperture is split into a number of equally

sized subapertures. For each subaperture, a polar grid is generated for which
the step size in α is decided by ∆ cos α ≤ λmin2l , where l is sub aperture length
and λmin the shortest signal wavelength, while the step size in R is decided by
c
∆R ≤ 2B [18]. Fig. 1 illustrates the polar grid which is used in FBP.

hk (R, α) =
 q 
tk + 2vlp 2 2
Z
 2 (c (t − tk ) /2) + R + 2R (c (t − tk ) /2) cos α 
St,  dt (1)
 c 
tk − 2vlp

where c is speed of light, t slow time, tk slow time for centre for sub aperture,
R and α is range and angle in the polar grid, in which the angle α is counted
from the positive azimuth axis.
In the first stage of the algorithm, the polar image is generated by, for
each position in the polar grid, the sum of the contribution from all aperture
56 Part II

S ( t 1, τ )

l1
( α ' , R' )
S ( t 2, τ )

Δα
ΔR

x
Figure 1: The size of the sub aperture l1 is decided based on ∆α. In this figure,
α is the angle from the positive x-axis and R is the distance from the centre of
the sub aperture l1 to a position in the polar image grid.

positions in the sub-aperture is computed. The generation of the polar image

can be mathematically expressed by (1).
In the second stage, for each subaperture, the data from the polar image
is interpolated into the final Cartesian image. Then all Cartesian images are
superpositioned to generate the final SAR image.

3 Moving target focusing and relocation

This section describes moving target focusing and how moving target focusing
can be integrated with FBP. In SAR, what distinguishes a moving target from
a stationary target is the different range history over the integration time. We
assume a target with movement parameters given by
Suppression of Clutter in Multichannel SAR GMTI 57

xtc (t) = vx (t − t0 ) + x0
ytc (t) = vy (t − t0 ) + y0
ztc (t) = 0 (2)

where t0 is the time where the distance is at the minimum between the target
and the platform and (x0 , y0 ) is the position of the target at minimum range.
With a platform moving at constant speed vp in azimuth, the range history
of the target can be written as [21].
q
2
Rt (t) = γt2 (vp t − Xt ) + Yt2 (3)

where γt is normalized relative speed (NRS), Xt and Yt azimuth and slant

range image coordinates of the target.
As presented in [21], γt , Xt and Yt are found to be
s
2
(vp − vx ) + vy2
γt = (4)
vp2

vy
Xt = x0 − y0 (5)
(vp − vx )

v !
u
u vy2
Yt = th2 + y02 1+ 2 (6)
(vp − vx )

Since we now know the targets range history, we can process the SAR image
accordingly, thus focusing the target, i.e.
q
2
Rp (Xp , Yp , γp , t) = γp2 (vp t − Xp ) + Yp2 (7)

where γp is the processing NRS, Xp is the azimuth image coordinate and Yp

the slant range image coordinate in order to focus all targets with γt = γp .
This holds for all SAR processing algorithms.
More specifically, one can modify (1) to
58 Part II

hk (R, α) =
 q 
tk + 2vlp
2
Z
 2 cγp2 (t − tk ) /2 − R cos α + R2 sin2 α 
t,
S  dt (8)
c 
tk − 2vlp

The focusing can be conducted in a similar manner in the second stage of

FBP, when making the transformation from polar to Cartesian image grid.

H (Xp , Yp ) =
  
Xp − Xck
X q
2
hk  γp2 (Xp − Xck ) + Yp2 , cos−1  q  (9)
2
k γp2 (Xp − Xck ) + Yp2

where Xck is the azimuth position corresponding to t = tk , i.e. the azimuth

position for the origin of the k-th polar image.
If we are able to focus the target for a certain γp , the gain in SCNR is
usually significant. However, we do not have information about the relation
of vx and vy , and since Xt in (5) depends strongly on vy , we need to have
information about target bearing in order to find the actual x0 and y0 of the
target.
Let us have a look at (4) again. We can see that (4) can be rewritten to
the following form.
s 2
vt vt
γt = −2 cos αt + 1 (10)
vp vp

where vt is the speed of the target, and αt the bearing.

Solving (10) for vt , we obtain (11)
 q 
vt = vp cos αt + γt2 + cos2 αt − 1 (11)

Having an estimate of γt , a 1-D search in αt can be performed. Equation

(11) gives us vt from the estimate of γt and the found value of αt .
Suppression of Clutter in Multichannel SAR GMTI 59

4 Clutter Suppression for GMTI

In this section, the theory for suppressing clutter in multi-channel SAR is
presented. First of all, in order to perform suppression, we must know what
to suppress, therefore we need a model for our received signal. In this paper,
the assumption is that the received signal consists of target signal, clutter and
receiver noise, as formulated in (12).

xl = sl + cl + nl (12)

where l is antenna number, sl the reflection from target, cl the received

reflection from clutter and nl is the noise in the radar receiver.
In SAR, multiple radar pulses are transmitted and received with a certain
pulse repetition frequency (P RF ) during the time when the radar system moves
along a path, called the synthetic aperture. Between each pulse the radar
system has moved a certain distance ∆l, the noise changes rapidly, while the
clutter and the target signal do not change so fast. Such a property is the basis
for the development of the clutter suppression techniques such as DPCA, ATI
and STAP for GMTI. This paper focuses on the STAP technique for performing
clutter suppression in multi-channel SAR.
In this section,we first give some background to STAP processing, and then
we move to the implementation used in this paper and differences from earlier
publications.

4.1 STAP
This subsection explains how the STAP processing was implemented in the
research presented in this paper. As part of a larger processing scheme pre-
sented in this paper, STAP processing is an essential part. In Fig. 6 one finds
a general overview.
In order to retrieve sl from xl in (12), an FIR filter is applied which combines
range frequency samples of the channels to obtain a single filtered output signal
xf iltered as presented in (13)

L
X
xf iltered (n) = xf ilteredL (L (n − 1) + l) (13)
l=1

and
60 Part II

xf ilteredL = w ◦ x (14)

where w are filter weights and x is the measurement vector including all the
measurements of all channels and xf ilteredL is the intermediate result where all
samples for all L channels have been weighted.
In order to find an optimum w, it has been suggested in [15],[22],[23] to use
w as given in (15).

wopt = uC−1
i (15)

In the above equation, Ci refers to the covariance matrix of the combination

of all interference, in our case assumed as noise and clutter. The steering
vector which is used to maximize signal strength for certain target dynamic
parameters is denoted u. Generally, Ci is not of full rank, and thus its inverse
does not exist, therefore we choose to use the Moore-Penrose pseudo inverse,
thus we rewrite (15) to (16).

wopt = uC+
i (16)

Hence, we can reformulate (13) as

uC+
i ◦x
xf ilteredL = (17)
N
N refers to a normalization factor. In this paper we chose a normalization
factor according to (18). Through this choice, a constant false alarm rate
(CFAR) is obtained over the image, i.e. the noise level will be equalized over
the image.

q
N= uC+
i u
H (18)

In the implementation of STAP combined with SAR in this paper, each

beam in the polar image is split into a certain number of sub-beams. After
this, each sub-beam is transformed to the Fourier domain using FFT’s. In
Fig. 2 the beam from a polar image is shown and in Fig. 3 one sub-beam from
Suppression of Clutter in Multichannel SAR GMTI 61

the beam in Fig. 2. The FFT of Fig. 3 is shown in Fig. 4. As seen in Fig. 4,
almost all energy of the signal is around the frequency 310Mhz. All of these
steps are always taken for both radar channels. We denote the sub-beams and
their Fourier transforms by

Xk,l (f ) = F{xk,l (τ )} (19)

where k is sub-beam number, l channel number, f frequency and τ time.

Next step is to obtain an estimate of the covariance matrix, Ci . This is
made by calculating a weighted sum of the Fourier transforms from the sub-
beams. The signal strength varies strongly between sub-beams. In order to
compensate for this and obtain a good estimate for the covariance matrix, the
different sub-beam FFTs are equalized.

X xk H xk
Ĉi = p (20)
k xk xk H

with xk given by

 T
Xk,1,fmin
 Xk,2,fmin 
 
xk = Xk,1,fmin +df  (21)
 
Xk,2,fmin +df 
..
 
.

After having an estimate of Ci , we need to define the steering vector only.

In this paper, we use the steering vector for focusing a moving target with a
constant speed vt and moving in the bearing αt . The steering vector can be
found as

 T
1

 exp (−j2πfmin vt dant cos (α − αt ) /vp ) 

 1 
u = exp (−j2π (fmin + df ) vt dant cos (α − αt ) /vp ) (22)
 
 
 .. 
 . 
62 Part II

−3
x 10
4

3.5

Polar Image Range beam

2.5

1.5

0.5

0
2000 3000 4000 5000 6000 7000 8000 9000
Range [m]

Figure 2: This figure shows one range beam from a polar image in the middle
stage of FBP.

−3
x 10
1.8

1.6

1.4
Polar Image Range sub−beam

1.2

0.8

0.6

0.4

0.2

0
3250 3300 3350 3400 3450 3500 3550
Range [m]

Figure 3: After the beam in Fig. 2 is split, several sub-beams are obtained, in
this figure one sub-beam of Fig. 2 is seen.

where fmin refers to lower signal frequency, df the frequency step, dant the
distance between the effective phase centers of the two channels, and vp the
platform speed.
The steering vector is in this paper used as a means to improve signal-to-
interference ratio. It does so by shifting the filtered beams according to the
movement of the target during the time span between the first antenna until
second antenna is in same azimuth position. The steering vector is the key in
being able to find target bearing as it is able to maximize the target signal with
regard to target dynamics. Furthermore, the chosen steering vector assumes
target moving with constant velocity, in the same manner as for focusing using
Suppression of Clutter in Multichannel SAR GMTI 63

0.12

0.1

FFT of sub−beam 0.08

0.06

0.04

0.02

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency [Hz] 8
x 10

Figure 4: After calculating the FFT of the sub-beam in Fig. 3, one obtains
this sub-beam range spectrum. It is clear that almost all signal energy is
concentrated around 310Mhz. Therefore, only the samples surrounding these
frequencies are filtered. This helps to keep the processing burden down.

NRS.

4.2 Integration of STAP in FBP

Multiple methods are combined in order to perform the filtering as presented
in this paper. Methods in this context refer to physical setup of radar system
and also to the different signal processing techniques that are used.
First of all, with regard to the physical setup, a number of receiving an-
tennas are used. All antennas are placed on an airborne platform with the
antennas placed in along-track direction. In Fig. 5, the antenna used in the
experiment presented in this paper is shown. The specific system used in the
experiment in this paper is LORA [24], [25] which is a very high resolution
UHF SAR system with MTI capabilities.
Secondly, with regard to SAR signal processing, a two-stage back projection
algorithm is used, namely FBP which was presented in section 2.
In order to suppress clutter in the SAR images, STAP is used as presented
in section 4.1. This clutter suppression is conducted on the range beams in the
polar images in the first stage of FBP. This is performed for each subaperture
in a similar manner as in [15]. The main differences in comparison to [15] is the
normalization factor used when estimating the covariance matrix, the use of
FBP instead of the Local Backprojection (LBP) algorithm, and the fact that
no assumptions are made with regard to dependencies within the covariance
64 Part II

TX2 RX3 RX2 RX1 TX1

Figure 5: In this figure, the geometrical configuration of the LORA system

is shown, with two transmitting antennas and three receiving antennas. The
spacing between each antenna is approximately 0.7m

matrix. Both LBP and FBP are 2-stage implementations of FFBP, but the
difference is that LBP is based on splitting the full image grid into sub-images
and forming small beams over each sub-image instead of as in FBP generate
beams that cover the whole image area. Fig. 6 shows the complete processing
scheme for two channels from raw data to obtaining detection, focusing and
speed estimation of moving targets. The pseudocode of the processing scheme
is presented in Algorithm 2.
Thirdly, the hypothesis testing for target relative speed as well as the target
bearing is conducted. The bearing estimation is an integrated part of the
STAP, as it is conducted by varying target parameters in the steering vector.
The hypothesis testing for different relative speed is performed in the last stage
of FBP, as the clutter-suppressed polar images are being backprojected onto
the Cartesian image grid.

5 Results
In this section, results using the LORA system developed by Swedish Defence
Research Agency (FOI) [17] running in MTI mode with two receiving antennas
and one transmitting antenna is presented. The experiment was conducted
in 2005 on the island of Visingsö in the lake Vättern. The main experiment
parameters can be found in Table 1. In this section, coordinates in the images
will be referred to as (X, Y ), where X and Y is the local azimuth and the local
range respectively.
In the experiment, one truck was driven at a speed of approximately 50
km/h along a small road along the west coast of Visingsö island, while the
airplanes flight path was almost orthogonal to the target path. The experiment
setup is shown in Fig. 7. The target carrying a corner reflector is shown in
Fig. 8. The aircraft and the truck do not move parallelly, nor perpendicularly.
This leads to the target being displaced in azimuth and smeared out in the
image.
Suppression of Clutter in Multichannel SAR GMTI 65

Input: Rawdata
Output: Image
k=1;
Image = 0;
for k ≤ N Subapertures do
k = k + 1;
GeneratePolarGrid;
P olarImage=GeneratePolarImage(Rawdata) ;
l=1;
for l ≤ N Angles do
k = k + 1;
RangeBeam=ExtractRangeBeam(P olarImage,l) ;
SubBeams=SplitRangeBeam(RangeBeam,N Subbeams) ;
m=1;
for m ≤ N Subbeams do
m = m + 1;
F SubBeam0(m)=FFT(SubBeams(m)) ;
F SubBeam(m)=ExtractSignalFreq(F SubBeam0(m));
end
C=EstimateC(F SubBeam) ;
u=SteerVector(vt ,αt ) ;
m = 1;
for m ≤ N Subbeams do
m = m + 1;
F SubBeamF iltered(m)=STAP(C,u,F SubBeam(m));
SubBeamF iltered(m)=IFFT(F SubBeamF iltered(m));
end
RangeBeamF iltered=CombineBeams(SubBeamF iltered(m));
P olarImage(l)=RangeBeamF iltered;
end
CartesianImage=PolarToCartesian(P olarImage,γt );
Image = Image + CartesianImage;
end
Algorithm 2: Integrated SAR and STAP algorithm
66 Part II

Raw data Channel 1 Raw data Channel 2

Pulse compression Pulse compression

Aligning data Aligning data

Generating beams Generating beams

Range FFT of beams Range FFT of beams

Extraction of frequency Extraction of frequency

components components

Estimation of
covariance matrix
(vx,vy)
Ĉ
Prewhitening and
(ɣ,u)
filtering with C and u

Backprojection to final
SAR Image

Filtered SAR Image

Figure 6: In this figure, the complete processing scheme is illustrated using two
radar channels, from raw data to SAR image with a focused moving target.

Table 1: Experiment setup, Visingsö experiment site

Parameter Value
Centre Frequency 313 MHz
Bandwidth 12.8 MHz
Target Type Truck with corner reflector
Integration Time 4.6 s
Aircraft Speed 130 m/s

In Fig. 9, the target is visible close to the position(1500, 500). However, the
energy of the target is much lower compared to the strong reflecting objects
such as at (2700, 500) which are specular reflections from large houses. In this
original image, the square of peak value of the target is 2dB below the average
energy of the clutter. First, a certain target moving direction was guessed to
Suppression of Clutter in Multichannel SAR GMTI 67

be 120◦ . After this, a set of images for different NRS were created and the
gain in Signal To Clutter and Noise Ratio (SCNR) was measured according
to [26], as well as the gain of the target in relation to the strongest point in
the image. In Fig. 10a the different measurements of SCNR gain vs NRS are
shown. A peak can be seen at 1.03, corresponding to the highest SCNR gain
in the process. A closer search shown in Fig. 10b was performed around 1.03,
in which the maximum was found at 1.031. The gain is approximately 19 dB.
Next step is to, based on the NRS estimate, try to focus the target better
with regard to target bearing and thus obtain an estimate of αt . The results
are shown in Fig. 11. Here we see that a bearing of 104 degrees appears to be
a good estimate. The bearing of 104 degrees and NRS of 1.031 gives a speed
of 13.87 m/s.
The highest gain in SCNR is approximately 20dB, with an SNR loss in the
filtering at only 0.6dB. SNR loss is here defined as the ratio of the energy of the
peak pixel of the target to the mean energy of the noise. In this best case, the
gain is over 23dB with regard to the strongest object in the SAR image prior
to filtering. The image corresponding to these best values are found in Fig. 12,
in which the target is highlighted with a red rectangle, and a zoom-in of the
target signature can be seen. Please note that the target is the strongest object
found in the whole image after the filtering. Also note that the clutter energy
is on the same level as the noise all over the image. This indicates successful
filtering.

6 Conclusions

This paper presents how filtering with two-channel UHF SAR can be conducted
using STAP and relative speed focusing. The paper also shows how the process
can be integrated into a SAR processing algorithm named Fast Backprojection
(FBP). In the proposal, the filtering step takes advantage of the fact that FBP
as the first stage performs a Doppler filter. The STAP filtering is applied on the
Doppler filtered beams. Results are presented from a measurement campaign
with the UHF UWB SAR system LORA. The field experiment shows good
results with a gain of about 20 dB in SCNR, while even up to 23dB in gain
when compared to the strongest object in the image before filtering. It can
thus be concluded that the method combined with the LORA system has good
chances for success with SAR GMTI in combination with forest clutter.
68 Part II





Target track

Airplane track

© OpenStreetMaps contributors
Data available under Open Database License.

Figure 7: The experiment area with the flight track indicated on the bottom
and target path indicated along a road on the west side of Visingsö island.

Acknowledgment
The authors would like to thank the KK-Foundation for making this research
project possible, and the support from Saab Bofors Dynamics, Saab Electronic
Defence Systems and RUAG Space. Also, very special thanks to Magnus Her-
berthson for very valuable feedback in the process of finishing this manuscript.
Suppression of Clutter in Multichannel SAR GMTI 69

Figure 8: The target truck with corner reflector on the roof.

70 Part II

Figure 9: In this figure, the original SAR image over the test scene is shown.
The image covers 4096m in azimuth and 2896m in slant range.
Suppression of Clutter in Multichannel SAR GMTI 71

20

18

16
SCNRgain [dB]

14

12

10

4
0.95 1 1.05
γt

(a)
18.6

18.5

18.4

18.3
SCNRgain [dB]

18.2

18.1

18

17.9

17.8

17.7

17.6
1.024 1.026 1.028 1.03 1.032 1.034 1.036 1.038
γt

(b)

Figure 10: In this figure, the SCNR of the target in the SAR image after
focusing the target for different normalized relative speed is shown. In Fig. 10a,
the step size is 0.01, and the measurements span from 0.96 to 1.04. At 1.03
there is a clear peak. A smaller step size of 0.001 is used around 1.03 and
shown in Fig. 10b. At 1.031 a clear peak can be distinguished. For these plots,
the target bearing was chosen to 120 degrees.
72 Part II

21

20

19
SCNRgain [dB]

18

17

16

15
90 95 100 105 110 115 120 125 130 135
αt [degrees]

Figure 11: In this figure, the SCNR of the target in the SAR image has been
measured by searching in bearing, αt from 90 to 135 degrees. The gain increases
monotonically until 104 degrees, after which the gain is reduced gradually.
Suppression of Clutter in Multichannel SAR GMTI 73

Figure 12: In this figure, the SAR image after focusing the target for relative
speed is shown. The target signature is highlighted with a red square and a
zoom-in can be viewed. It must be noted that the truck is now the strongest
of all reflections in the SAR image. Even the very strong specular reflections
found in the original image are successfully suppressed to the noise level and
can no longer be detected in the image.
References

[1] Baumgartner, S.V., and Krieger, G. (2008) SAR traffic monitoring using
time-frequency analysis for detection and parameter estimation. In
Proceedings of IGARSS, Boston, MA, 7-11 July, (2008).
[2] Marques P.A.C., and Dias, J.M.B. (2005) Velocity estimation of fast
moving targets using a single SAR Sensor. IEEE Transactions on
Aerospace and Electronic Systems, 41, 1 (2005), 75-89.
[3] Vu, V.T., Sjögren, T.K., Pettersson, M.I., and Gustavsson, A., and
Ulander, L.M.H (2010) Detection of Moving Targets by Focusing in
UWB SAR - Theory and Experimental Results. IEEE Transactions on
Geoscience and Remote Sensing, 48, 10 (2010), 3799-3815.
[4] Barbarossa, S. and Farina, A. (1992) Detection and imaging of moving
targets with synthetic aperture radar. 2. Joint time-frequency analysis by
Wigner-Ville distribution. IEE Proceedings in Radar and Signal
Processing, 139, 1 (1992), 89-97.
[5] Zhou, F., Wu, R., Xing, M., and Bao, Z. (2007) Approach for single
channel SAR ground moving target imaging and parameter estimation.
IET Radar, Sonar and Navigation, 1, 1 (2007), 59-66.
[6] Zhou, H., Huang, X., Chang, Y., Zhou, Z. (2007) Single-channel UWB
SAR ground moving targets detection method using change detection
based on single-pass sub-aperture images. In Proceedings of Asian Pacific
Conference on SAR, Huangshan, China, 5-9 November, (2007).
[7] Sjögren, T.K., Vu, V.T., Pettersson, M.I., Zepernick, H-J., Gustavsson,
A. (2007) Speed estimation experiments for ground moving targets in
low frequency UWB SAR. In Proceedings of IET International Radar
Conference, Edinburgh, UK, 15-18 October, (2007).
[8] Jao, J.K. (2001) Theory of synthetic aperture radar imaging of a moving
target. IEEE Transactions on Geoscience and Remote Sensing, 39, 9
(2001), 1984-1992.

75
76 REFERENCES

[9] Vu, V.T., Sjögren, T.K., and Pettersson, M.I. (2008) Moving Target
Detection by Focusing for Frequency Domain Algorithms in UWB Low
Frequency SAR. In Proceedings of IGARSS, Boston, MA, 7-11 July,
(2008).

[10] Gierull, C.H., Ground moving target parameter estimation for

two-channel SAR. IEE Proceedings on Radar Sonar and Navigation,
153, 3 (2006), 224-233.

[11] Axelsson, S. (2004) Position correction of moving targets in

SAR-Imagery. In Proceedings of SPIE - The International Society for
Optical Engineering, Barcelona, Spain, 5236, 80 (2004).

[12] Suchandt, S., Runge, H., Breit, H., Steinbrecher, U., Kotenkov, A., and
Balss, U. (2010) Automatic Extraction of Traffic Flows Using
TerraSAR-X Along-Track Interferometry. IEEE Transactions on
Geoscience and Remote Sensing, 48, 2 (2010), 807-819.

[13] Ender, J.H.G. (1999) Space-Time Processing for Multichannel Synthetic

Aperture Radar. Electronics and Communication Engineering Journal,
11, 1 (1999), 29-38.

[14] Ward, J. (1998) Space-time adaptive processing for airborne radar. In

IEE Colloquium on Space-Time Adaptive Processing, London, UK, 6
April, (1998).

[15] Pettersson, M.I. (2004) Detection of moving targets in wideband SAR.

IEEE Transactions on Aerospace and Electronic Systems, 40, 3 (2004),
780-796.

[16] Neumann, C., Meyer-Hillberg, J., Senkowski, H. (2008) STAP/GMTI

analysis using SAR data from simulated scenarios. In Proceedings of
International Radar Symposium 2008, Wroclaw, Poland, 21-23 May,
(2008).

[17] Ulander, L. M., Blom, M., Flood, B., Follo, P., Frölind, P.-O.,
Gustavsson, A., Jonsson, T., Larsson, B., Murdin, D., Pettersson, M.,
Rääf, U., Stenström, G. (2003) The VHF/UHF-band LORA SAR and
GMTI System. In Proceedings of SPIE Algorithms for Synthetic Aperture
Radar Imagery X, Orlando, FL, 21-23 April, (2003).
REFERENCES 77

[18] Yegulalp, A.F. (1999) Fast backprojection algorithm for synthetic

aperture radar. In IEEE Radar conference, Waltham, MA, 20-22 April,
(1999).
[19] Ulander, L.M.H., Hellsten, H., Stenström, G . (2003) Synthetic-aperture
radar processing using fast factorized back-projection. In IEEE
Transactions of Aerospace and Electronic Systems, 39, 3 (2003), 760-776.
[20] Frolind, P.-O., Ulander, L.M.H. (2006) Evaluation of angular
interpolation kernels in fast back-projection SAR processing. In IEE
Radar, Sonar and Navigation, 153, 3 (2006), 243-249.
[21] Sjögren, T.K., Vu, V.T., Pettersson, M.I., Gustavsson, A, Ulander,
L.M.H. (2012) Moving Target Relative Speed Estimation and Refocusing
in Synthetic Aperture Radar Images. IEEE Transactions on Aerospace
and Electronic Systems, 48, 3 (2012), 2426-2436.
[22] Brennan, L.E., Reed, L.S. (1973) Theory of Adaptive Radar. IEEE
Transactions on Aerospace and Electronic Systems, 9, 2 (1973), 237-252.
[23] Melvin, W.L. (2004) A STAP overview. IEEE Transactions on Aerospace
and Electronic Systems, 19, 1 (2004), 19-735.
[24] Hellsten, H., and Ulander, L.M.H. (2000) Airborne array aperture UWB
UHF radar - Motivation and system considerations IEEE Aerospace and
Electronic Systems Magazine, 15, 5 (2000), 35-45.
[25] Ulander, L.M.H. and Hellsten, H. (1999) Low-frequency ultra-wideband
array-antenna SAR for stationary and moving target imaging In
Proceedings of Radar Sensor Technology IV, held in Orlando, FL, 5-9
April 1999, SPIE 3704 (1999), 149-158.
[26] Ulander, L.M.H. (1991) Accuracy of using point targets for SAR
calibration. IEEE Transactions on Aerospace and Electronic Systems,
28, 1 (1991), 139-148.
 Part III

Linear Apodization for UWB SAR

Images
 Linear Apodization
for UWB SAR Images

Thomas K. Sjögren, Viet T. Vu, Mats I. Pettersson.

Abstract

Ultrawideband Ultrawidebeam (UWB) Synthetic Aperture Radar (S-

AR) refers to SAR systems utilizing large fractional bandwidths and an-
tenna beamwidths associated with large integration angles. Suppressing
efficiently both orthogonal and non-orthogonal sidelobes, minimizing the
loss in SAR spatial resolutions, and preserving phase information of SAR
images are basic criteria for a UWB SAR apodization method. This pa-
per proposes a linear apodization method for UWB SAR imaging. The
proposal is based on applying windows in frequency domain of SAR im-
ages. In order to find suitable windows the paper shows how optimization
can be employed for systems with arbitrary fractional bandwidth and in-
tegration angle. In the paper the proposal is illustrated by applying
2D-Cosine-On-Pedestal windows, thus having a two dimensional search
space. Also two different shapes of windows are used. The results in-
dicate that for systems with bandwidth and beamwidth matched, the
same window can be used in both azimuth and range frequency, or in
both angle and radial frequency. Furthermore, the use of windows with
annular-sector shape show better results with regard to suppressing side
lobes and maintaining minimal main lobe broadening. This improvement
is largest for systems with large integration angle. In the paper, side lobe
suppression is applied to a UWB SAR image from the CARABAS-II sys-
tem and the result indicates that the simulation results are correct.

81
82 Part III

List of Symbols

A13dB Main Lobe area broadening, measured

as percent in relation to non-windowed.
Br Relative bandwidth.
F Cost function measuring sidelobes.
H Set of pixels.
HM L Set of pixels in main lobe, i.e. within
13dB area of peak value.
|HM L | Cardinality of HM L , or number of pixels
in main lobe.
|HM L,lim | Upper limit for cardinality of HM L , i.e.
upper limit for number of pixels in
main lobe.
k Wavenumber.
kc Centre wavenumber.
kmin Wavenumber corresponding to smallest
radar signal frequency.
kmax Wavenumber corresponding to largest
radar signal frequency.
kx Azimuth Wavenumber.
ky Slantrange Wavenumber.
S (X, Y ) Value of complex SAR image in
the position (X, Y ).
W Window.
Wsep (kx , ky ) Window separable in kx and ky .
Wnonsep (kx , ky ) Window not separable in kx and ky .
Wcart Window described in Image Fourier
Domain by cartesian coordinates.
Wpolar Window described in Image Fourier
Domain by polar coordinates.
Wcart,CoP (kx , ky )Window given as a product of two
Cosine-on-Pedestal windows in kx and ky
respectively.
Wpolar,CoP (ω, φ) Window given as a product of two
Cosine-on-Pedestal windows in ω and φ
respectively.
φ Angle in clockwise direction from positive
ky axis to arbitrary point in
Linear Apodization for UWB SAR Images 83

Image Fourier Transform.

φ0 Integration angle.
ξ Trade off parameter for
Cosine-On-Pedestal windows.
ξx Trade off parameter for Cosine-On-Pedestal
windows in azimuth direction.
ξy Trade off parameter for Cosine-On-Pedestal
windows in slant range direction.
ξω Trade off parameter for Cosine-On-Pedestal
windows in radial direction.
ξφ Trade off parameter for Cosine-On-Pedestal
windows in angular direction.
a∗b Convolution between a and b.

1 Introduction
Over the last decades, Synthetic Aperture Radar (SAR) has attracted consid-
erable interest as the number of applications in geosciences, remote sensing,
surveillance and reconnaissance increases. The ability to effectively collect
data in severe conditions such as rain, clouds and/or darkness is considered
to be the main advantage of SAR systems compared to other imaging sensors.
Ultrawideband Ultrawidebeam (UWB) SAR systems utilizing large fractional
bandwidths and antenna beamwidths gives images with very high resolution in
range as well as in azimuth. Examples of the experimental UWB SAR systems
are CARABAS-II operating in the lower VHF-band from 20 to 90 MHz [1]
and P-3 with a bandwidth of 515 MHz in the VHF/UHF-bands [2]. Having
systems with large beamwidth related with large integration angle, sidelobes
will appear not only in slantrange and azimuth directions in the image, but
also in other directions [3],[4].
Let us have a look at the currently used sidelobe control or apodization
methods reserved for Narrowband and Narrowbeam (NB) SAR imaging. Ba-
sically, they can be classified into two categories: linear and non-linear. Linear
apodization methods are based on two-dimensional (2D) weighting functions
applied to the spectrum of the SAR image. These methods allow orthogonal
sidelobe suppression, preserving phase information of SAR images, however,
always accompanied with loss in SAR spatial resolutions. With a rectangular
approximation for the support of the NB SAR spectra, the loss in resolutions
84 Part III

depends strongly on the 2D weighting function in use. Mostly used weight-

ing functions are Rect, Hamming, Hanning, Kaiser and Blackman. Nonlin-
ear apodization methods have been built on the linear apodization methods
and a class of nonlinear operators. Nonlinear apodization methods has shown
their ability to control sidelobe level and retain resolutions, simultaneously.
The resolution is preserved thanks to the nonlinear operators which are ap-
plied to linear-apodized images on a pixel-by-pixel basis. Well-known nonlinear
apodization methods are Dual Apodization (DA), Complex Dual Apodization
(CDA) and Spatially Variant Apodization (SVA) [5]. The main disadvantage
of the nonlinear apodization methods is that the phase information of SAR
images is destroyed.
Applying the NB SAR apodization methods to UWB SAR is introduced
recently. A possibility to use a linear apodization method for UWB SAR
is mentioned sparsely in [6]. The 2D weighting function used in [6] is not
clearly defined. Effects caused by the linear apodization method is also not
investigated. Another discussion on apodization for UWB SAR imaging is in
[4]. Investigations in [4] show that both the linear and non-linear apodization
methods can be applied to UWB SAR. However, the effects caused by the
shortcomings of the NB SAR apodization methods are much more serious when
applying to UWB SAR. For example, using linear apodization method can
lead to an extremely high loss in SAR spatial resolutions or using nonlinear
apodization methods such as Tri Apodization (TA) [5] or Multi Apodization
(MA) [5] is followed by damage to the phase information. An apodization
method applicable for UWB SAR imaging, which can suppress efficiently both
orthogonal and non-orthogonal sidelobes, minimize the loss in SAR spatial
resolutions, and preserve the phase information of SAR images, is still of high
interest.
In [7] sidelobe suppression in UWB SAR images of point targets were in-
vestigated. Using windows with rectangular shaped support and windows with
annular-sector shaped support, the ability so suppress side lobes was compared.
A set of systems are tested with windows having the two kinds of support. Us-
ing windows with annular sector-shape, it was possible to suppress sidelobes
even for UWB systems. However, relevant is not only the ability to suppress
sidelobes, but also the ability to maintain a small mainlobe. Furthermore it is
important to study many different windows to see if it is possible to find some
optimal window and find ways for defining an optimality criterion with regard
to choice of window. With this background, the objective of this paper was
formed, which is to propose an optimum apodization for UWB SAR imaging.
The proposal is based on the linear apodization technique to retain the phase
Linear Apodization for UWB SAR Images 85

ky

k max
ky'
kc kr'
k min
'

0

kx
k x'
Figure 1: Plot showing the general shape for the spectrum of a SAR image. The
azimuth and slantrange wavenumbers are denoted by kx and ky respectively,
the wavenumber corresponding to highest, centre, and lower frequency of the
radar signal is denoted by kmax , kc and kmin respectively. Two main system
parameters are relative bandwidth, Br = kmaxk−k c
min
and integration angle, φ0 .

information of SAR images and an optimization of the sidelobe suppression

and loss in SAR spatial resolutions. The optimization is developed on a 2D
Cosine-On-Pedestal weighting function.

The paper is organized as follows. A review of the NB SAR linear apodiza-

tion technique and its inherent limitations when applying to UWB SAR is
given in Section II. Section III presents an optimization of the sidelobe sup-
pression and loss in SAR spatial resolutions. Simulation results and evaluations
is provided in Section IV. Section V summarizes the contribution of this paper.
86 Part III

Slantrange frequency [Hz].

fc

−prf/2 0 prf/2
Azimuth frequency [Hz].

Figure 2: Plot showing the spectrum of a SAR image for a system with Br =
0.05 and φ0 = 5◦ . This corresponds to a conventional narrowband SAR system
having a characteristic shape of its spectrum similar to a rectangle.

fs
Slantrange frequency [Hz].

fc

0
−prf/2 0 prf/2
Azimuth frequency [Hz].

Figure 3: Plot showing the spectrum of a SAR image for a system with Br = 0.5
and φ0 = 50◦ . This corresponds to an Ultra Wideband SAR system for which
the shape of the spectrum is not similar to a rectangle.

2 Linear Apodization in SAR

Conventionally in SAR, in order to suppress sidelobes a window is multiplied
with the spectrum of a SAR image, or the raw data. The support of the
window is traditionally rectangular, while the actual windows applied in range
and azimuth frequency varies. The reason why the windows are of rectangular
shape is partly because of simplicity, and partly because the support of the
spectrum of the SAR image traditionally is very close to a rectangle. As shown
Linear Apodization for UWB SAR Images 87

ky

k max W rect
ky'
kc kr'
k min
'
kc 0

kx
k x'
Figure 4: Plot showing the general shape for the spectrum of a SAR image,
and overlaid a window as defined by (1).

in Fig. 1, the spectrum of a UWB SAR image has in general not a rectangular
shape. If the integration angle and the bandwidth is small, it can be understood
that the support is approximately rectangular. Often in SAR, the integration
angles are less than 5 degrees and the relative bandwidths less than 0.05, thus
making the assumption reasonable. Looking at a spectrum for such small
integration angle and relative bandwidth, the similarity to a rectangle becomes
obvious as can be seen in Fig. 2. However, if we increase the integration angle
to 50 degrees and the relative bandwidth to 0.5 as in Fig. 3, a clear difference
can be seen. And the shape can no longer be considered to have the shape of
a rectangle.
Since the shape of the spectrum does not fit well with a rectangular window,
it has been attempted to combine many rectangular windows that cover from
small up to the full part of the spectrum of the SAR image with the use of
non-linear apodization [4]. It showed the ability to suppress the non-orthogonal
sidelobes which are characteristic of Ultra Wideband SAR. However, if we are
88 Part III

ky

k max W polar
ky'
kc kr'
k min
'

0

kx
k x'
Figure 5: Plot showing the general shape for the spectrum of a SAR image,
and overlaid a window as defined by (2).

to use linear apodization, we can only choose one window for the spectrum. In
[7] a preliminary study was made in which different windows were applied to the
spectrum for systems with increasing integration angle and relative bandwidth.
Two types of windows were chosen in [7]. The first windows were of rectangular
shape and the spectrum of the SAR image is shown in Fig. 4. It is in general
given by

 
 
kx ky − kc
Wrect (kx , ky ) = W (kx , ky ) boxcar     boxcar (1)
2kc sin φ20 kmax − kmin

Where the parameters can be found in Fig. 1. However, if instead we choose

windows defined on a polar grid in the frequency domain of the SAR image
and limited in radial and angular direction, as shown in Fig. 5, the window
will fit with the shape of the spectrum of the SAR image. These were also
investigated in [7] and can be formulated mathematically as
Linear Apodization for UWB SAR Images 89

q 

φ
 kx2 + ky2 − kc
Wpolar (kx , ky ) = W (kx , ky ) · boxcar · boxcar   (2)
φ0 kmax − kmin

The results in [7] indicated that suppression of sidelobes for UWB-systems

was more successful using windows of the type (2). One example to illustrate
this can be seen in Fig. 7. It is clear that for Fig. 7(a) much more sidelobes
can be seen compared to Fig. 7(b). In section 4 these two types of windows
are further compared. Also in section 4, the optimization approach proposed
in section 3 is evaluated, thus comparing if different optimal coefficients are
found using the different window types.

3 Optimum Linear Apodization Technique Adapted For

UWB SAR
In this section, an optimum linear apodization technique for SAR imaging is
proposed. The proposal is based on the linear apodization technique mentioned
in Section 2.
When we talk about optimization it is important to define a cost function
which is to be minimized or maximized. Such cost function can be written
generally as

arg min F (W (kx , ky ) , X, Y, Br , φ0 ) (3)

W (kx ,ky )

Here, we suggest that the two main problems to be mitigated is total side lobe
energy as well as the level of the highest side lobe. We can mathematically
write these two optimization alternatives as
where

X
F (W (kx , ky ) , X, Y, Br , φ0 ) = kS (X, Y, Br , φ0 ) ∗ F −1 (W (kx , ky )) k2
(X,Y )∈HSL
(4)
or
90 Part III

Range

Azimuth

(a) NB SAR system

(b) UWB SAR system

Figure 6: Plots of the main lobe and side lobes a for point target in a SAR
image. An NB system generated (a) and a UWB system generated (b). In (b)
one can see side lobes appearing in all directions instead of traditionally only
in two main directions. This phenomena is typical for UWB SAR.

F (W (kx , ky ) , X, Y, Br , φ0 ) = maxkS (X, Y, Br , φ0 ) ∗ F −1 (W (kx , ky ))

for (X, Y ) ∈ HSL (5)

In the above equations, SL refers to sidelobe, HSL refers to the set of pixels
in the sidelobe area, and (X, Y ) refers to pixel coordinates.
Having formulated a cost function, it must be noted that it is just as impor-
tant to consider the size of the mainlobe. We thus have two objectives given by
(4) and (5) and third objective may be the main lobe size. Each objective has
MxN input variables, where M and N are number of points, in the frequency
Linear Apodization for UWB SAR Images 91

Range

Azimuth

(a) Rectangular Window

Range

Azimuth

(b) Polar Window

Figure 7: Plots showing point targets after applying windows in frequency

domain of a SAR image for a UWB SAR system. A window of rectangular
support such as shown in Fig. 4 was applied and (a) was obtained. After
applying a window of the shape shown in Fig. 5, (b) was obtained.

domain of the SAR image, in kx and ky direction respectively. In order to

keep the problem easier to analyze, we make some limitations in the problem
formulation.

First, we study a cartesian window. The window is symmetric about kc in

ky direction, and symmetric about 0 in kx direction. The following expression
is found
92 Part III

 
 
k x ky − kc
Wcart (kx , ky ) =Wkx     Wky
2kc sin φ20 kmax − kmin
 
 
k x ky − kc
boxcar    boxcar
 (6)
2k sin φ0 kmax − kmin
c 2

Secondly, we study a window on a polar grid, symmetric about the radius kc

in radial direction, and symmetric about 0 in angular direction, the following
simplified expression is found.

    q 
arctan kkxy   kx2 + ky2 − kc 
 
Wpolar (ω, φ) =Wφ  φ0
 Wω  
2
kmax − kmin
    q 
arctan kkxy   kx2 + ky2 − kc 
 
boxcar   boxcar   (7)
φ0 kmax − kmin

leads to roughly a reduction of the search-space with a factor of 4, due to a

factor approximately of 2 for the symmetry for each of the first two factors in
(6) and (7). By making a further restriction on the windows we can chose a
window commonly used in SAR images. Here we chose to use a product of two
Cosine-On-Pedestal windows, thus the following two expressions are obtained

Wcart,CoP (kx , ky , ξx , ξy ) =
" !#   
πkx 2π (ky − kc )
0.5 + ξx cos 0.5 + ξy cos (8)
kc tan φ20 ∆ky

and

Wpolar,CoP (ω, φ, ξω , ξφ ) =
   " !#
(ω − ωc ) πφ
0.5 + ξω cos 0.5 + ξφ cos φ0
(9)
B 2
Linear Apodization for UWB SAR Images 93

ξx and ξy are weighting factors in azimuth and range. Possible values for these
weighting factors are in the interval from 0 to 0.5, this corresponds to the
window changing from a Rect to Hanning window.
Thus, we can summarize the limitations in our investigation to:

1) Make a number of optimizations of W, for a specific set of Br and φ0 .

2) Windows that are separable in kx and ky or in φ and ω.
3) Symmetric windows.
4) Windows that are a product of two Cosine-On-Pedestal windows.

Using this information, we can simplify (3) to (5) and obtain

arg min F (ξω , ξφ , X, Y ) (10)

(ξω ,ξφ )

where

S (X, Y ) ∗ F −1 W (ξω , ξφ )2
X  
F (ξω , ξφ , X, Y ) = (11)
(X,Y,)∈HSL

or

2
F (ξω , ξφ , X, Y ) = max S (X, Y ) ∗ F −1 W (ξω , ξφ ) for (X, Y ) ∈ HSL (12)


As mentioned, it is also important to keep as small size of the mainlobe as

possible in order to get sharp images. For this reason, we can add the mainlobe
size constraint to (11) and (12). If we choose (12), we obtain

2
F (ξω , ξφ , X, Y ) =max S (X, Y, ) ∗ F −1 W (ξω , ξφ )


for (X, Y ) ∈ HSL and |HM L | ≤ |HM L,lim | (13)

and for the case when using windows of the type (1)
2
F (ξx , ξy , X, Y ) =max |S (X, Y, ) ∗ W (ξx , ξy )|
for (X, Y ) ∈ HSL and |HM L | ≤ |HM L,lim | (14)
94 Part III

in which |HM L | refers to the cardinality of HM L or in other words, the

number of pixels within the mainlobe. Also, |HM L,lim | refers to an upper limit
on the broadening of the mainlobe.
Finding the optimal values of ξx and ξy or ξω and ξφ with the need of a
search algorithm is possible in the case of choosing a constant Br and φ0 . Such
search may be a parametric search. We will in this study limit the search to
the set of windows which can be obtained from (8) and (9). This means we will
evaluate the shape of a windowed point target for different values of ξx and ξy
and see what is the resolution loss, and total side lobe energy.

4 Sidelobe suppression results

Table 1: Simulated SAR System Parameters

System Br φ0
A 0.1 10◦
B 0.3 30◦
C 0.8 80◦
D 0.2 130◦
E 1.2 20◦

In this section, results for applying the sidelobe suppression technique is

presented. The sidelobe suppression is conducted using Cosine-On-Pedestal
with the two parameters ξx and ξy , or ξω and ξφ varied from 0 to 0.5. The
set of systems chosen for evaluation of the suppression techniques are shown in
Table 1. For the Cosine-On-Pedestal window given in (8), the results in Fig.
8 were obtained, and using the window given in (9), the results in Fig. 9 were
obtained. Three different kind of investigations were made. First of all, the
side lobe suppression was measured, using Integrated Sidelobe Ratio (ISLR)
in dB, as a function of the ξ-parameters. Then for each window, the main
lobe broadening was measured by how many percent the area had grown due
to windowing compared to no window. Based on the two first measurements,
a search was conducted based on the earlier two measurements to find for
Linear Apodization for UWB SAR Images 95

different values of main lobe broadening, what values of (ξω , ξφ ) gives best side
lobe suppression.
Normally, one would be interested in resolution measurements, but as can
be seen when inspecting Fig. 6a a point target for a NB-SAR system and in
Fig. 6b a point target for a UWB-SAR-system, the mainlobe shape of the
two images are not the same. For this reason it may not be best suited to
use resolution measurements only based on the 3-dB resolutions in x and y
direction. In this paper we choose instead to use the number of pixels within
a 13-dB area surrounding the peak of the impulse response.
The results are presented in Fig. 8(a)-(e) correspond to side lobe suppres-
sion and main lobe broadening for system A to system E. In Fig. 8(a) there
appear to be no difference between suppressing in angle or radial direction, it
appears that a minimum loss in mainlobe area size for a certain ISLR sup-
pression is obtained when ξω = ξφ . It can also be observed that below an
ISLR of 26 dB, the mainlobe area increases considerably. Similar results can
be seen for Fig. 8(b)-(c), but as the bandwidth and integration angle increases,
the relationship between sidelobe supression and mainlobe area becomes more
complex, and in Fig. 8(d), the main lobe broadening measurements are very
large while still not being able to suppress side lobe energy to as large degree
as for Fig. 8(a)-(c). In Fig. 8 (e), the results look very similar to those of Fig.
8(a)-(c), with the only seen difference being higher main lobe broadening is
needed to obtain same side lobe suppression. In Fig. 8(f)-(h), optimal values
of the tradeoff coefficients for the side lobe suppression of system A, D and
E can be seen. For system A and E, it appears both coefficients should be
approximately equal to each other. However (g) shows that for system D it is
optimal to keep ξy higher than ξx .
As suggested in section II, non-separable windows may improve the results
of sidelobe suppression.To investigate this, windows given by in (9), a Cosine-
on-Pedestal in radial and angular direction is applied with the two parameters
ξω and ξφ in the range between 0 and 0.5. For these windows, the contour plots
in Fig. 9 are obtained. In comparison to Fig. 8, there is an improvement with
regard to A13dB for all systems.
For the same level of sidelobe suppression, the mainlobe area is clearly
smaller. There appear to be no difference between suppressing in angle or radial
direction, however it appears that a minimum loss in mainlobe area size for a
certain ISLR suppression is obtained when ξω = ξφ . It can also be observed
that below an ISLR of 26 dB, the mainlobe area increases considerably. As the
bandwidth and integration angle increases, the relationship between sidelobe
supression and mainlobe area becomes more complex. In Fig. 9(c) it now
96 Part III

appears the result that as ξω is above approximately 0.25, the suppression

of sidelobe becomes worse. With regard to as low ISLR as possible given a
certain increase in mainlobe area, it still appears as best result is obtained
when ξρ and ξφ are the same. A large improvement can be seen for Fig. 9(d)
compared to Fig. 8(d). While still the side lobe suppression is not great,
however the main lobe broadening is greatly reduced, thus making possible
for useful images. In the same manner as for Fig. 8(g), the results for Fig.
9(g), show that one coefficient should optimally have larger value, in this case
ξω > ξφ . Furthermore, it can e noted a difference for the area with higher
acceptable main lobe broadening in Fig. 9(g), compared to Fig. 8(g).

4.1 Evaluation on UWB SAR data

Table 2: CARABAS-II SAR system parameters

System Br φ0
CARABAS 1.3 130◦

In this subsection some results are shown in which the simulation results
are verified on measurements of a real UWB SAR scene in which the side lobe
suppression is performed. The suppression is performed on SAR data from
the CARABAS-II systems developed by the Swedish Defense Research Agency
(FOI). This system was chosen as it is one of the system with highest relative
bandwidth and integration angle of all SAR systems, thus a suitable test for
applying the side lobe suppression on. In the experiment, both parameters
ξφ and ξω are set to same values, going from 0.1 to 0.5. Measurement was
made on a part in the SAR image in which the object is a wind power plant.
In the large image in Fig 10, the object can be seen at the position Local
Azimuth 1000m and Local Range 3500m. The wind power plant is clearly
visible in the zoomed in image shown in Fig. 11. Measuring the ISLR and
A13dB , the results are shown in Fig. 13 and Fig. 12 respectively. One can note
that the ISLR is not as good as in simulation. This can be understood by
the noise and/or clutter level in the SAR image surrounding the wind power
plant. With regard to the A13dB , one can not that for low ξφ and ξω values
are worse compared to simulation, while for larger ξφ and ξω , the A13dB is not
worse than in simulation. To further illustrate the side lobe suppression ability,
Linear Apodization for UWB SAR Images 97

the windowing was applied inside the small harbor area in the image Fig 10.
Original image is shown in Fig 14a, while result after windowing in Fig 14b.
The window coefficients were set to 0.2 in this example. In the image can be
clearly seen that the ringing effect is mitigated.

5 Conclusions
In this paper, we present results on the linear apodization technique and its
inherent limitations when applying to UWB SAR imaging. Means of finding
an optimum apodization for UWB SAR imaging is proposed, which is based
on the linear apodization technique and an optimization of the sidelobe sup-
pression and loss in SAR spatial resolutions. The optimization is illustrated by
simplifying the general case to a 2D Cosine-On-Pedestal weighting function.
The proposal is tested and evaluated with simulated UWB SAR data given
a set of different system parameters. Optimal parameters for the Cosine-On-
Pedestal weighting functions are also obtained. The results show that the use
of rectangular shaped windows on the spectrum of a SAR image leads to the
main lobe to be broaden significantly. The use of windows with annular-sector
shape however show better results with regard to suppressing side lobes and
maintaining minimal main lobe broadening. The improvement is largest for
systems with large integration angle. Furthermore, to evaluate on real data, a
high resolution UWB SAR image from CARABAS-II was used. On real data,
side lobe suppression was not as successful, this can however be understood by
the underlying noise and clutter in a real SAR image. Measurements of loss
in resolution area follows quite well the results from the simulations. One can
conclude that the real measurements fit rather well with conducted simulations.

Acknowledgment
The authors would like to thank the KK-Foundation for making this research
project possible, and the support from Swedish Defence Research Agency
(FOI), Saab Bofors Dynamics, Saab Electronic Defence Systems and RUAG
Space. Also, very special thanks to FOI for providing the CARABAS-II SAR
image used in the research.
98 Part III

    
       
   
      
    










   


    
     










    





  
    
  
 


   
 
  





  









 
     

 
 


   
  




   




    


 

 
 








 


  


  





 


     
         
 

 

     

  



  
  
  






   
 
 

       
 
  


   


 


  


     


      

     
 
   

  




 







 
    
  



    

 

   


 



     


   



 
     







    





  
    
    
 

 
   
  


  
    





 
     





 
    
   


   
  










 




   








  
   





   



  

         
 

 
 
 
   

 


 

 

 
          
 

 

Figure 8: In (a) to (e), the plots show the result given by contour plots for
systems A to E given in Table I. The contour plots show ISLR and A13dB , for
point targets, after applying windows given by (8) in frequency domain of a
SAR image. The contour plot is a function of two variables, ξx and ξy , the
Cosine-On-Pedestal parameters for azimuth and slant range wavenumber. In
(f) to (h), the optimum coefficients, ξx and ξy are found for different values of
A13dB for the systems A, D and E.
Linear Apodization for UWB SAR Images 99

    
     
 
      
     





 


 
   
     













  


  




     
 
     
     
 
  













 

     


 









  



  

 
   







   


   
 






     

         
 
 
 
  
   

 
  


   











   
 



  
   
   


   


   








  

  






   
   

 
    

 











 







 



 


 
 




 



  

  






 




   





 







  

 





 
 





 
         
 
 

  
  


    


    
 


  

 
 


  

    


    



 








 

    






 



       
 

  


 




  
         
 

 
 
 
   

 


 

 

 
          
 

 

Figure 9: This figure shows the corresponding plots as shown in Fig. 8, but
when applying windows given by (8) in frequency domain of a SAR image.
The contour plot is a function of two variables, ξφ and ξω , the Cosine-On-
Pedestal parameters related to angular and radial wavenumber. In (f) to (h),
the optimum coefficients, ξφ and ξω can be found for different values of mainlobe
area broadening A13dB for the systems A, D and E.
100 Part III

500

1000

1500
Local Range [m]

2000

2500

3000

3500

4000

4500

5000
1000 2000 3000 4000 5000
Local Azimuth [m]

Figure 10: A UWB SAR image of the city of Simrishamn on the coast of the
Baltic Sea, located in south Sweden.

50

100
Local Range [m]

150

200

250

300

350

400
100 200 300
Local Azimuth [m]

Figure 11: in the centre of this figure, the wind power plant which was used
for the measurements is shown.
Linear Apodization for UWB SAR Images 101

80

70

60

50
A13dB

40

30

20

10
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
ξφ,ω

Figure 12: This figure shows the A13dB measurements for the UWB SAR image
of the CARABAS-II system.

−11.5

−12

−12.5
ISLR

−13

−13.5

−14

−14.5
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
ξφ,ω

Figure 13: This figure shows the ISLR measurements for the UWB SAR image
of the CARABAS-II system.
102 Part III

(a)

(b)

Figure 14: In Fig.14a, a non-apodized SAR image from the harbor area of the
image Fig. 10 is displayed. On the image spectrum, a Cosine-On-Pedestal
window with the parameter values ξφ = 0.2 and ξω = 0.2 was applied and then
Fig.14b was obtained. One can note the positive effects of the windowing. Not
only is the strong ringing is suppressed, but also weaker side lobes which was
hiding some scatterers.
References

[1] A. Gustavsson, L. M. H. Ulander, B. H. Flood, P.-O. Frölind, H. Hellsten,

T. Jonsson, B. Larsson, and G. Stenström, “Development and operation of
an airborne VHF SAR system-lessons learned,” in Proc. IEEE IGARSS’98,
vol. 1, Seattle, WA, Jul. 1998, pp. 458–462.
[2] D. R. Sheen, C. M. Strawitch, and T. B. Lewis, “UHF wideband SAR design
and preliminary results,” in Proc. IEEE IGARSS’94, vol. 1, Pasadena, CA,
Aug. 1994, pp. 289–291.
[3] V. T. Vu, T. K. Sjogren, M. I. Pettersson, and H. Hellsten, ”An Impulse
Response Function for Evaluation of Ultra-wideband SAR Imaging,” IEEE
Trans. Sig. Process., vol. 58, no. 7, Jul. 2010, pp. 3927–3932.
[4] V. T. Vu, T. K. Sjögren, and M. I. Pettersson, “On apodization techniques
for ultra-wideband SAR imaging,” in Proc. EURAD’2009, Roma, Italy,
Sep. 2009, pp. 529–532.
[5] H. C. Stankwitz, R. J. Dallaire, and J. R. Fienup, “Nonlinear apodization
for sidelobe control in sar imagery,” IEEE Trans. Aerosp. Electron. Syst.,
vol. 31, no. 1, pp. 267–279, 1995.
[6] X. Xu and R. M. Narayanan, “Enhanced resolution in SAR/ISAR imag-
ing using iterative sidelobe apodization,” IEEE Trans. Image Processing,
vol. 14, no. 4, pp. 537–547, 2005.
[7] T. K. Sjögren, V. T. Vu, and M. I. Pettersson, “2D apodization in UWB
SAR using linear filtering,” in Proc. IEEE IGARSS’2011, Vancouver,
Canada, July. 2011.

103
 Part IV

Performance of Target Parameter

Estimation Using SAR
Performance of Target Parameter
Estimation Using SAR 107

Performance of Target Parameter Estimation Using

SAR
Mats I. Pettersson, Thomas K. Sjögren, Viet T. Vu
Blekinge Institute of Technology
SE-371 79 Karlskrona, Sweden
[email protected]
Abstract- Synthetic Aperture Radar (SAR) gives not only the systems, there are also methods for UWB such as the one
opportunity to image targets with high resolution but also to published in [7] and demonstrated in [5],[8]. For NB SAR
measure their position and velocity. Without acceleration, the systems, the methods published in [9]-[11] use a single
position and velocity both in range and azimuth can be channel NB SAR system to detect the target and estimate
estimated using a multi-channel SAR system. This paper
introduces a method to find the lower bound on variance of
the target speed and acceleration. The basic idea is to use
estimate of position and velocity for targets. The method is the phase of target range migration in combination with the
based on the assumptions needed for the Cramér-Rao Lower antenna footprint. The illustration of the method is based on
Bound (CRLB). The method works for both narrowband (NB) simulated and experimental data. However, detection based
and ultrawideband (UWB) SAR systems. For demonstration, a on dual- or multi-channel systems is even more interesting
monostatic single-channel UWB SAR system operating at low due to the possibility to suppress stationary targets, i.e.
frequency is used. Thus, only lower bounds for estimate clutter. Estimating radial speed in dual- or multi-channel
variance of azimuth, range and relative speed are derived. SAR has been introduced in [12],[13]. In [14], the
acceleration of the target is shown to be estimated using
I. INTRODUCTION dual-channel SAR-GMTI and in [3], radial speed, relative
speed and acceleration are determined. That means that all
Ground Moving Target Indication (GMTI) using a multi- speed components can be determined if flat earth and no
channel Synthetic Aperture Radar (SAR) has been of acceleration are assumed.
interest to the scientific society for the last decades. It has Let us have a look at estimation accuracy. In [15], the
been shown that a SAR system can detect moving targets Cramér Rao Lower Bound (CRLB) of a moving target
and measure their properties [1], [2]. The measurements do position at minimum range is derived and the main interest
not only give the information of a moving target but also the is the radial speed component. In [16], [17], the radial speed
target surrounding such as e.g. houses and trees with very estimation accuracy is introduced for an along-track
high accuracy. The performance of moving target estimation interferometric SAR system. The derivations in [16] also
based on a dual channel SAR system has been reported in consider the bound for two different models of the moving
[3] and the traditional solution of Space Time Adaptive target response, i.e. deterministic and a zero-mean Gaussian.
Processing (STAP) with low resolution can be found in [4]. The lower bound takes target strength, clutter and thermal
However, it has been shown that long integration time does noise into account. In [17], the CRLB for radial speed is
not only increase detectability but also adds the capability to calculated.
measure radial speed and along-track components. This is In this paper, we derive the CRLB for single- or multi-
of interest to all types of radar systems, especially to the channel UWB SAR systems. The derivation is based on a
lower-frequency band SAR systems operating at, e.g. point-like scattering moving target. The CRLB is found
VHF/UHF bands. Parameter estimation for low frequency from scattering amplitude, scattering phase, range position,
systems is associated with long integration time. In this azimuth position and velocity. Depending on the number of
case, small target accelerations do not destroy the coherence channels of the considered SAR system, i.e. single or multi-
due to the long radar wavelength [5]. The clutter density is channel, the relative speed or the range and azimuth speed
also low in comparison to the system resolution. The clutter CRLB is determined. In the published papers mentioned
backscattering is therefore very stable between two above, backscattering amplitude, phase and radial speed
measurements occasions, even if they are separated by days CRLB are mainly aimed at, whereas in this paper we derive
[6]. This means that statistical properties measured on one up to six parameters of CRLB, depending on the number of
day can be used for the day after. channels of the SAR system. The method presented, in
For the latest years, there has been a large interest in SAR comparison to the previous publications, is not restricted by
processing of moving targets, demonstrated by the NB systems. The idea of the method has been first
considerable number of publications of simulation and mentioned in [18]. For simplification, we consider the
experimental results. In most publications, dual- and multi- correlation between intermediate band (between NB and
channel systems are taken into account. Besides this, single- UWB) SAR image samples instead of UWB ones to
channel systems are also used to detect and estimate moving illustrate the proposed method. A motivation of this
target properties. Although the moving target detection and approach is given in the paper. Also, the illustration is based
estimation methods are developed for narrow-band (NB)
 108 Part IV

on a single-channel SAR system. This allows us to compare Since the SAR image formation is seen as a linear process,
our results to the results published in [5], [8]. the moving target parameter vector can be found from the
The paper is organized as follows. Moving target´s image reconstructed by the l-channel SAR data [7],
parameter vector and motion model are presented in section
II. The signal model, which will be used to develop the θ = [ X 0 Y0 ΔX 0l ΔY0l St φt ] (4)
method, is given in section III. The measurement signal is
given in IV and the clutter and noise model is discussed and where X 0 and Y0 are the azimuth and range positions at t0
motivated in section V. In VI, the measured signal model is
motivated. Section VII aims at SAR image sampling and of the moving target on the SAR image formed by the first
correlation between the image samples. The CRLB of channel data and ΔX 0l and ΔY0l indicate the displacements
moving target parameters are derived in Section VIII. between the images reconstructed from the first channel and
Section IX provides numerical calculation of lower bounds the l-th channel data. The relationship between displacement
in position and relative speed. and location can be found, e.g. in [7]. Depending on the
focusing parameter, i.e. the processed relative speed, a
moving target is either focused, when the processed relative
II. MOVING TARGET´S PARAMETER VECTOR AND MOTION speed is the same as the true relative speed of the target, or
MODEL smeared as a curvature, when an incorrect relative speed is
When a target is moving on the Earth’s surface, its used for processing. This may result in a large reduction in
motion can be described in a local Cartesian coordinate the target resolution and the target peak energy level [20].
system by (ξt ( t ) ,ηt ( t ) , ζ t ( t ) ) . Since the Earth is The mathematical description of the moving targets smeared
as a curvature is presented, e.g. in [21] and this smeared
approximately flat, the target location is simplified to effect is clearly shown by the experimental data in [5].
(ξt ( t ) ,ηt ( t ) ,0 ) . Detection of low signature targets or
In this paper, we propose the lower bound for target
detection in low frequency SAR is associated with a large
parameter estimation variance. The proposal presented in
integration angle and in most cases with a long integration
the next sections is based on a measurement vector, given
time. Detection of very slow moving targets are also
by stacking all channels and all SAR image samples. In the
associated with a long integration time. At low frequencies,
illustration of the proposal, we will use only one channel
the signal wavelengths are long and therefore not very
data, i. e. only one SAR image. This makes the illustration
sensitive to target vibration and acceleration. Thus, a fairly
simpler with only five parameters to estimate instead of six
good approximation, at least at low frequencies, for a
in case of a multi-channel SAR system. The lower bound is
moving target is to assume that there is neither vibration nor
therefore independent of channel separation, a parameter
acceleration. The target position in Cartesian coordinates is
that is very system dependent. The selection of one channel
expressed by
data is also for comparison purpose. Thus, the lower bound
can be related to the relative speed estimates in [5], [8].
(ξt ( t ) ,ηt ( t ) , ζ t ( t ) ) = (ξ0 + vξ ( t − t0 ) ,η0 + vη ( t − t0 ) , 0 ) (1) Hence, the parameters that can be estimated in one channel
data are given by
where ξ 0 and η0 are the ground positions at time of the
minimum range t0 , and vξ and vη are the constant speed θ = [ X 0 Y0 γ t St φt ] . (5)
components. The platform motion is often described in a
similar way as where γ t is the normalised version of the relative speed or
the Normalised Relative Speed (NRS) and given by
(ξ ( t ) ,η ( t ) ,ζ ( t ) ) = ( v
pl pl pl pl t ,0, h ) (2)
(v )
2
γt = pl − vξ + vη2 v pl (6)
where v pl denotes the speed and h is the altitude over
ground of the platform. If more antennas are used, the true The fundamental principal in SAR processing is to handle
position of each antenna channel needs to be considered, range migration. The distance between the moving target
especially in case of UWB [19]. Let us consider a point-like and the platform is given by
scatterer with the backscattering strength S t and phase φt .
The target parameters are represented by a vector as
(ξ ( t ) − ξ ( t ) ) + (η ( t ) − η ( t ) ) + (ζ ( t ) − ζ ( t ) )
2 2 2
rt ( t ) =
θ = ⎣⎡ξ 0 η0 vξ vη St φt ⎦⎤ (3) pl t pl t pl t

(7)
Performance of Target Parameter
Estimation Using SAR 109

In the image coordinates and with assumption of no

acceleration, the distance refers to (
μ s ( ms ( n, m ) , γ p , θ ) = hs n, m, γ p , θ = )
⎛ (
rp k , n, m, γ p ⎞ ) (12)
∑
K

rt ( t ) = γ t2 (ξ pl ( t ) − X 0 ) + Y0 2
2
(8) = g ⎜ ts ( k ) , ,θ⎟
⎜ c0 ⎟
k =1 ⎝ ⎠
Similarly, the distance to any image sample (X, Y) in the where ms ( n, m ) is the stacked index given by n and m, and
SAR image processed at a relative speed γ p is shown to be ( ) (
rp k , n, m, γ p = rp ts ( k ) , X ( n ) , Y ( m ) , γ p is the discrete )
distance for pulse k to image pixel ( X ( n ) , Y ( m ) ) .
rp ( t , X , Y , γ p ) = γ p2 (ξ pl ( t ) − X ) + Y 2
2
(9)

IV. MEASUREMENT SIGNAL

where γ p is the processed NRS.

The measured samples will not only contain a signal from

III. MOVING TARGET SIGNAL MODEL the moving target but also backscattering from clutter and
thermal noise. An image sample is given by a superposition
To determine the lower bound of the variance for a of the pulses along the synthetic aperture and can therefore
parameter estimate, it is necessary to build an appropriate be rewritten to a sum of measurements as
model of the signal, clutter and noise. Under a point-like
scattering approximation and a non-acceleration
⎛ (
rp k , n, m, γ p ) , θ ⎞⎟
∑ p ⎜⎜⎝ k ,
K
assumption, one channel SAR image (reconstructed with no
clutter and noise) is represented by the integral
(
μ x ( ms ( n, m ) , γ p , θ ) = hx n, m, γ p , θ =) x
c0 ⎟
k =1 ⎠
(13)
rp ( ts , X , Y , γ p )
T2
⎛ ⎞
(
hs X , Y , γ p , θ =) ∫ g ⎜⎜⎝ t ,
T1
s
c0
, θ ⎟dts
⎟
⎠
(10)
( )
where hx n, m, γ p , θ is the sampled SAR image processed
at relative speed γ p , px ( k , t f , θ ) is the pulse-compressed
where T1 and T2 are the start and stop of the measurement received signal at aperture position k. The pulse repetition
time in slow time ts . Equation (10) is therefore used as the interval (PRI) or frequency (PRF) is a design parameter that
signal of the target in parameter estimation. Measurement will affect the correlation.
time is often considered to be infinite in SAR processing.
Conversely, a finite measurement time is set by the impulse V. CLUTTER AND NOISE MODEL
response of g ( ts , t f , θ ) in ts and in fast time t f . The
The accuracy in the estimation of the moving target’s
impulse response depends on system parameters such as parameter vector depends on clutter and noise. For radar
selected wave form, antenna gain function, range distance systems, the signal interference is commonly divided into
and target scattering characteristics. In this paper, we use a two independent groups: Clutter generated by scatterers in
simple model with a transmitted chirp pulse, flat receiver the illuminated area and noise originated from thermal
characteristics, an isotropic antenna element, no acceleration processes or other radio sources. In this study, we
on target and platform, point-like scattering characteristics concentrate on the lower bound of variance for parameter
for stationary scatterers (clutter) and the moving target. For estimation and therefore the covariance between SAR
L-channels, a vector is formed by stacking L × N × M image processed image samples. In the stacked measurement
samples together where N is the number of range positions
and M is the number of azimuth positions. As discussed
( )
vector μ x γ p , θ , the correlation between image samples is

above we illustrate the method for one channel, and in this defined by
case the stacking results in a steering vector given by
C x ( γ p , θ ) = E ⎡⎣ μ x ( γ p , θ ) μ xH ( γ p , θ ) ⎤⎦ (14)
μ s ( γ p , θ ) =

⎡ hs (1,1, γ p , θ ) " hs (1, M , γ p , θ ) hs ( 2,1, γ p , θ ) " hs ( N , M , γ p , θ ) ⎤

T Let us first in the derivation of the clutter and noise model
⎣ ⎦ assume that the target influence on the clutter and noise is so
(11) weak that the target can be neglected. Because the clutter
and noise are independent [7], the correlation can be found
With K transmitted pulses along the synthetic aperture, the from
steering vector is given by
 110 Part IV

Cx ( γ p , θ ) = Cc ( γ p , θ ) + Cn ( γ p , θ ) (15) backscattering straight so low that the weak scatterers could
be hidden in the thermal noise. In these cases there is no
model available for UWB SAR clutter modelling. However
Starting with the thermal noise statistics, neglecting target if the backscattered area is homogenous, each clutter
and clutter, the measured noise sample would be scatterer could be assumed to be normally distributed in its
real and imaginary parts.
⎛ (
rp k , n, m, γ p ) , θ ⎞⎟
∑ p ⎜⎜⎝ k ,
K
In radar backscattering models, the radar backscattering
μ n ( ms ( n, m ) , γ p , θ ) = n
c0 ⎟ appears from a homogenous area with a large amount of
k =1 ⎠ (16) scatterers. For one pulse along the synthetic aperture the
resolution area is given by the area of a circular annulus on
(
A common model for μ n γ p , θ consists of white Gaussian ) the ground limited by the impulse bandwidth of px ( k , t f , θ )
noise with zero mean at the antenna. The measured pulse- and the antenna footprint. The resolution in a UWB SAR
compressed thermal noise pn k , t f , θ will therefore be ( ) system is for one pulse very low due to the wide antenna
footprint. Adding two pulses together selecting the proper
coloured and set by the bandwidth of the receiver and the
range will make a large overlap between the pulses causing
antenna. Therefore, the correlation in range will be limited
many clutter scatterers to correlate. As the number of pulses
by the bandwidth of the thermal noise and in azimuth by the
increases, resolution increases in the processed SAR image
PRF of the system. The spectrum of thermal noise is
if a proper range compensation is made, depending,
assumed to be flat in the frequency domain. The correlation
however on the possibilities to perform range compensation.
is therefore set by the system impulse response which is a
If the focusing parameter is set to a moving target, i.e.
sinc function in azimuth and in range of the SAR image.
γ p = γ t ≠ 1 , the resolution for clutter decreases in
Hence, the correlation of μ n ( γ p , θ ) will be given by the
comparison to the case γ p = 1 and therefore the amount of
impulse function set by the receiver bandwidth and the PRF
of the system. clutter scatterers increases in the SAR image resolution cell.
For the clutter excluding the noise and the target, the The lower resolution of the clutter will cause a larger
correlation between pixels in the SAR image, because the
correlation Cc ( γ p , θ ) is given by the statistical properties of sampling criteria of the SAR image are set by moving
targets that are focused. We therefore conclude that change
⎛ (
rp k , n, m, γ p ) , θ ⎞⎟ of the processing speed will effect the correlation between
∑ p ⎜⎜⎝ k ,
K

μ c ( ms ( n, m ) , γ p , θ ) = pixels in the SAR image.

⎟
c
c0
k =1 ⎠ (17)
In range direction the SAR image spectra of the clutter is
The measured pulse-compressed clutter signal set by the bandwidth of the transmitted and received pulse.
pc ( k , t f , θ ) will be correlated between pulses at distance
This is the same limitation as for the thermal noise if the
received signal has the same band limitation as the
( )
rp k , n, m, γ p with γ p = 1 . For one clutter scatterer on the transmitted signal. In the azimuth direction of the SAR
image, the clutter however behaves differently compared to
ground, the backscattering has a very strong correlation the thermal noise. There are two reasons for this: First, the
between k-th samples along the synthetic aperture. Doppler frequency depends on the speed of the platform,
Traditionally, the SAR resolution cells are assumed to PRF and the transmitted wavelengths. As the wavelength
include a large number of scatterers which results in speckle decreases, the Doppler increases within the same pulse.
processes between the scatterers inside the resolution cell. Secondly, the range frequency in the SAR image decreases
The most commonly used models assume large number of as the Doppler increases due to the coupling between range
scatterers that by support by the central limit theorem cause and azimuth in the SAR process. This means that the
the real and imaginary backscattering part to be normally spectrum of the thermal noise is rectangular, while the
distributed [22]. However the number of scatterers in a spectrum of clutter is a circle segment limited by the
resolution cell in a UWB system will decrease dramatically integration angle and the range components of the
and the traditional clutter model will fail. Especially at low transmitted and received pulse. The conclusion is that the
frequency the amount of scatterers in the resolution cell will spectrum of the clutter is approximately a square when the
be close to one when the resolution gets smaller than the SAR system is a NB system, but not when it is a UWB
operating wavelength. It is only in heavy forested areas that system. As for the noise correlation, the clutter correlation
the amount of scatterers could be more than one (if volume can be found from the SAR image impulse response. The
scattering occur), while in open areas the amount of impulse response for a NB system is a sinc function while it
scatterers per resolution cell will often be much less than is much more complicated for a UWB system [26].
one. For UWB low-frequency systems used over open areas
the amount of scatterers may be extremely low and the
Performance of Target Parameter
Estimation Using SAR 111

smeared to either an elliptic or a hyperbolic function [24].

The smearing effect of a moving target is clearly illustrated
VI. MEASURED SIGNAL MODEL both in simulation and on real data in [5]. In [22] the
curvature i.e. the image coordinates ( X tp , Ypt ) along which
In this paper we do not exclude long integration time in the
the target energy smears is given by
estimation phase, and therefore the algorithm has to handle
the non-linear range migration over the aperture. In the SAR
(Y ) γ p2γ t2 ( X 0 − X p )
2 2
image plane the stationary clutter will be unfocused as γ p
t t
p
+ =1 (19)
separates from one, and the dependency between the Y02 γ t2 − γ p2 Y02
elements in the covariance matrix increases. The noise is
independent of the clutter and the target. The question is the This function is an elliptic function if γ p < γ t and a
dependence between clutter and target. Electromagnetic
scattering is superimposed from different scatterers. If hyperbolic function if γ p > γ t . On one hand, the trajectory
scattering of each scatterer do not interact with another in (19) illustrates a decrease in resolution as the incorrect
scatterer they can be considered to be independent. This will speed is used in processing. On the other hand, it shows an
be justified if the Born approximation [23] holds for both increase in correlation between image samples. There are
the clutter and the target scatterers. However, in many cases two natural choices to generate a focused image of a moving
there are multi-path backscattering and shadowing effects in target or to generate an image of stationary targets such as
the backscattering process. Despite of that, the Born clutter. In this paper we select the clutter speed, i.e. γ p = 1 .
approximation is a good model in an resolution cell with
The reason is that the energy of the clutter is more focused
many scatterers and therefore
and therefore the SAR image samples are less correlated.
The independency between samples is often formed by
μ x ( γ p , θ ) = μ s ( γ p , θ ) + μ c ( γ p ) + μ n ( γ p ) sampling data on the Nyquist limit [25]. This can easily be
(18)
made on a SAR image that is processed from white noise
is often used. Note that if clutter and noise are where the spectrum is quadratic and the impulse response is
independent of the target they are also independent of target a sinc function. For a SAR image consisting of only
parameter vector θ , and therefore the clutter and noise only stationary clutter it is, however, different. As mentioned
before, the clutter spectrum is more of a circle sector
depends on the focusing parameter. For UWB SAR, this
segment and therefore the sinc function approximation will
approximation may be more questionable as the resolution
fail. At NB, however, the spectrum is approximately
of a SAR system is in the size of the scatterer. Backscattered
quadratic. This is adequately illustrated in [26] where three
clutter in the neighbouring resolution cell to the target will
spectrums are shown with different integration angles and
therefore be influenced by that target through multi-path
different bandwidths. For the intermediate SAR system, the
scattering or by shadowing effects. For a moving target the
scattering process will be even more complicated. If the relative bandwidth Br of 0.35 and an integration angle φ0 of
integration time is long, the target will pass many clutter 35 degrees gives a close to sinc-like impulse response.
scatterers during illumination. Many clutter scatterers will However, this rather wideband system has a tendency of
therefore multi-scatter with the moving target or be non-orthogonal side lobes that will for higher bandwidths
shadowed by the target for a short period of time as the break up to non-orthogonal sidelobes that are not sinc-like
target passes. To our knowledge, these effects have not been [26,27]. Due to the non-orthogonal structure and the non-
investigated in any publication. We therefore use sinc-like shape, the correlation will increase between the
independency as the best available model. However, in the samples even if the sampling is chosen in a way considered
later simulations, the approximation is correct. For real optimum in NB SAR. The non-orthogonality and the non-
data, we believe that the error caused by the approximation sinc-like shape are caused by the coupling of range and
is rather small and therefore conclude that the measured azimuth. The increased effect of coupling as the bandwidth
signal model is given by (18). We have then made use of the in azimuth and range increases is shown in [27]. In that
circumstance that the SAR process is a linear system that paper, broadening factors in range and azimuth are derived
will not change the independency. that relate the real impulse response to a sinc function. It
therefore gives a good indication of in which ranges of
VII. SAR IMAGE SAMPLING AND CORRELATION bandwidths and integration angles the NB approximation
BETWEEN THE SAMPLES will be appropriate for the correlation. The article also
shows that if the resolution is the same in azimuth and range
To form the SAR image of a moving target, the data is
the NB approximation holds for larger bandwidths. This
processed with γ p = γ t . The target is focused in the SAR
means that as the resolution increases the range and azimuth
image whereas the clutter gets unfocused and smeared. If should increase in both directions, for a low correlation. The
the clutter speed is used, γ p = 1 , the moving target will be highest broadening factors, and therefore assumably the
 112 Part IV

highest correlation, are given at low range resolution and IX. NUMERICAL CALCULATION OF LOWER BOUNDS IN
high azimuth resolution. A low correlation between image POSITION AND RELATIVE SPEED
samples is found (at Nyqvist) if corresponding broadening To illustrate the method, we estimate the lower bound
factors are close to one. As we have not determined the for a simple case with a simulated moving point-like
UWB correlation in this study, we limit the evaluation to the scatterer (target), true clutter and true noise, based on the
area where Br ≤ 1 and φ0 ≤ 32 D . By evaluating the impulse white noise approximation at antenna input. As mentioned,
response of the simulated systems, we found that pixels the measurement vector is formed by stacking the image
have low correlation in this interval using Nyqvist limits. samples together in a long vector given in (13). The
However, the highest correlation occurs when Br illustration of the proposed method uses single-channel data.
mismatches φ0 (in resolution) and in particular when the However, this should not be seen as a limitation and the
range resolution is low and the azimuth resolution is high. method can as discussed preferably be used on multi-
channel data (4). The extension is only to stack the antenna
VIII. CRAMÉR RAO LOWER BOUND (CRLB) OF MOVING channels in the measurement vector. The estimation results
TARGET PARAMETERS are very dependent of radar system parameters. In this
Under the assumption that the clutter backscatter and the illustration, we select radar system parameters related to
thermal noise are normally distributed at the antenna input, published results in [5],[8]. These typical cases are selected
due to the long integration time associated with these
and the SAR image formation is a linear process, the image
samples are also normally distributed. With this statement, experiments. This is also not a restriction and the method
the estimation performance can be evaluated from the well works well with NB data. In fact, the low correlation
between samples, as discussed in section V, is even a better
known CRLB. In its complex form, the lower bound of the
variance for the parameter estimate is found to be [25] approximation for NB data.
In CLRB, the covariance and derivation of the steering
vector are needed. The clutter covariance is found using the
C θˆ = I − 1 ( θ ) (20)
assumption of independent samples described in section
VII, which relates to the clutter correlation and is set by the
where each element in the Fischer information matrix is impulse response in the SAR image. As mentioned earlier,
the correlation length is affected by the separation between
⎡ ∂C x ( γ p , θ ) −1 ∂C x ( γ p , θ ) ⎤ the processed NRS and the NRS of the imaged target. The
I ij ( θ ) = tr ⎢C−x 1 ( γ p , θ ) C x ( γ p , θ ) ⎥+
⎢⎣ ∂θi ∂θ j ⎥⎦ (21) extension of the correlation is minimum at γ p = 1 for
⎡ ∂μ s ( γ p , θ )
H
∂μ s ( γ p , θ ) ⎤ stationary scatterers, i.e. clutter. To use independency
+2 Re ⎢ C x ( γ p , θ )
−1
⎥ between pixels in the SAR image, γ p = 1 is selected for the
⎢⎣ ∂θi ∂θ j ⎥⎦
processing. The PRF is set to be equal to Nyquist of the
In this study, we assume that the clutter backscattering is maximum clutter Doppler (the target speed is assumed to be
independent on the target motion as discussed in the much less than the platform speed and the antenna is
previous sections. With the knowledge of thermal noise approximately isotropic). Doppler is limited by operating
which is independent of target and clutter, the Fisher wavelength, platform speed and antenna footprint. The
information is reduced to selected PRF also sets the limit for the thermal noise
described in section V. Solving (20) and (22) numerically
⎡ ∂μ sH ( γ p , θ )
gives us the uncertainty in azimuth, range, and relative
∂μ s ( γ p , θ ) ⎤ (22)
I ij ( θ ) = 2 Re ⎢ C−x 1 ( γ p ) ⎥ speed given by the diagonal elements of the inverted Fisher
⎢⎣ ∂θi ∂θ j ⎥⎦ information. The parameters lower bound of estimate
variance is
where the target, clutter and thermal noise independency
cause the first part of (21) to be zero, i.e. σ X2ˆ = ( Cθˆ )
0 11

⎡ ∂C x ( γ p , θ ) −1 ∂C x ( γ p , θ ) ⎤ σ Y2ˆ = ( Cθˆ ) (24)

tr ⎢C −x 1 ( γ p , θ ) C x ( γ p , θ ) ⎥ = 0. (23) 0 22

σ = ( Cθˆ )
⎢⎣ ∂θ i ∂θ j ⎥⎦ 2
γˆt 33

As mentioned in the previous sections, this is an outcome of

the Born approximation and also consistent with lower where σ 2
X̂ 0
is the lower bound for the target position in
bound estimates based on STAP [4].
azimuth, σ Y2ˆ is the lower bound for target position in range,
0

and σ γ2ˆt is the lower bound for target relative speed.

Performance of Target Parameter
Estimation Using SAR 113

smaller at narrower integration angles. In this illustration,

The main steps, for this example, in obtaining the bounds we have selected SCR to 10 dB and SNR to 20 dB at a
for variance of the parameter estimates, is described in maximum integration angle of 32°. For each clutter
Algorithm 1. scatterer, the SAR process will add the scatterers
backscattered signal coherently for each pulse and at the
same time the resolution will increase. If the target is a
Procedure GenerateEstimationBound single point-like scatterer and the clutter consists of many
Input : θ, SCR, SNR, γ p , φ scatterers, an increase in resolution will raise the SCR. Since
Output : σ θ2ˆ noise is added incoherently between pulses, the SNR
increases as the number of pulses increases. Since the
Image = GenerateSARImage ( γ p , θ, φ )
processing speed is γ p = 1 to estimate the bound, the
μ s ( γ p , θ, φ ) = GenerateVector ( Image )
moving target is unfocused. However, we use the peak
Cx = GenerateCovarianceMatrix ( γ p , φ ) signal power of the focused target, i.e. γ p = γ t , to estimate
SCR and SNR. This means that the used peak signal power
For par1 = 1: Length ( θ ) is much higher than that in the bound estimate images when
For par2 = 1: Length ( θ ) γ p = 1 . The peak signal power corresponding to γ p = 1 is
⎡ δ μ Hs δ μ s ⎤
quite low due to the smearing effect. However, the total
I ( par1, par2 ) = 2 Re ⎢ Cx ⎥ signal energy of the target should be the same in both cases.
⎢⎣ δ par1 δ par2 ⎥⎦
end
end TABLE I.
THE RADAR SYSTEM PARAMETERS USED IN THE EVALUATION

Cθˆ = I −1 Minimum distance rmin 7570 m

( )
σ θ2ˆ = diag Cθˆ Platform speed vp 100 m/s
Centre frequency fc 55 MHz
Algorithm 1. The main steps in the procedure to obtain
estimation bounds. Relative bandwidth Br 1.0
Target normalized relative speed γt 1.03
Integration interval φ 2°-32°
Although the example for illustration of the method is
selected according to the published results given in [5],[8],
the simulation uses a smaller relative bandwidth Signal to Noise Ratio (at φ = 32 )
D
SNR 20dB

f max − f min
Br = (25) Signal to Clutter Ratio (at φ = 32D ) SCR 10dB
fc

where f max , f min and f c are the system maximum,

minimum and centre frequency, respectively. In Fig. 1, the The results from the simulations are shown in Fig. 1. In
relative bandwidth Br = 1 has been used in comparison to Fig. 1, the lower bound in range is shown. Even at low
Br = 1.3 in [5],[8]. The reason for the lower bandwidth is integration angles, the target can be measured with high
non-orthogonal side lobes of the targets [26] causing the range accuracy, i.e. σ Yˆ = 2.5 meters. As the integration
0
dependency between diagonal samples. This effect always angle increases, the accuracy increases . At the integration
appears in the following cases: extremely high relative angle of 32°, the lower bound of the standard deviation is
bandwidth, i.e. UWB, small signal bandwidth in
down to σ Yˆ = 0.7 meter. By making a log-log plot of the
combination with wide integration angle, and vice versa, in 0

large signal bandwidth in combination with a narrow result, the slope is found to be almost a straight line at low
integration angle. Hence, with the selected radar parameters, integration angles with an increasing deviation at higher
this effect occurs due to the narrow integration angle and angles. The slope of the curve shows that the lower bound is
large signal bandwidth i.e. the resolution is mismatched. In almost inversely proportional to the integration angle
the following calculations, we set the Signal to Noise Ratio ( σ Yˆ ∝ φ −0.5 ). This relates to the known relation of all pulse
(SNR) and Signal to Clutter Ratio (SCR) at the maximum 0

integration angle. This means that the SCR and SNR are
 114 Part IV

radars; variance in range is inversely proportional to the σ Yˆ = 1.00 and σ γˆt = 0.0007 which refers to a standard
square root of pulses or SNR [28],[29]. 0

In Fig. 1, the lower bound of variance is shown for the deviation in speed of 0.09m/s. In [8], the error between
azimuth direction. In this case, the resolutions at narrow estimated speed and the true speed was found to be 0.07m/s.
integration angles are low in comparison to the range Although one sample of the estimator performance does not
resolution. At the narrowest integration angle, the lower give any significance to the estimator performance, we can
conclude that the estimate error has the same order as the
bound of the variance is found to be σ Xˆ = 57 meters.
0 lower bound.
However, the slope is very steep in this case. Thus, the In the second case [5], the moving target is a real ship
lower bound of the variance is down to σ Xˆ = 1.2 meters at moving on the sea surface. The radar parameters are
φ = 300 ,
0
rmin = 7200m , f c = 52MHz , Br = 1.15 ,
the integration angle of 32°. Described in a log-log plot, the
slope of this curve is also found to be almost a straight line. SNR = 35.7dB . At these frequencies, the scattering from the
Also in this case there is an increasing deviation at the water is very low actually much below the thermal noise
higher integration angles with respect to the straight line. floor, and therefore SNR>>SCR. The bandwidth is higher
The found slope indicates that the lower bound is also than that in Table I but the integration angle is the same. In
almost inversely proportional to the integration angle, this case, we do not know the true speed. However, from the
( σ X̂ ∝ φ −1.5 ). It is a stronger slope than that in range broadening of the target given by Fig. 15 in [5] and from the
0 SNR, we can at least say that the lower bound is realistic.
direction. However, this is reasonable because the resolution Performing lower bound estimates are found to be
increases proportionally to φ −1.0 and the SNR to φ −0.5 at low σ Xˆ = 0.064m , σ Yˆ = 0.023 and σ γˆt = 8.2 ⋅10−5 which
0 0
integration angles.
refers to a standard deviation in speed of 1cm/s. The error is
In Fig. 1, the lower bound of the NRS is presented. At
very small because the SNR is high. With a SNR of 35.7dB,
low integration angles, the estimation performance is almost
the noise fluctuation at the peak given by Fig. 15 in [5]
useless. Hence, the value σ γˆt = 0.61 is equivalent to 61m/s
would refer to a value of −8.2 ⋅10 −5 dB. The broadening at
in uncertainty of the target’s relative speed. However, the this value is hard to investigate based on Fig. 15 in [5].
slope in this case is much steeper than the others. At 32°, the However, the standard deviation seems to be reasonable
lower bound of the variance is down to σ γˆt = 0.0015 which because the broadening is almost flat surrounding the peak
in the Figure. We therefore believe that the derived
corresponds to a speed of 0.15 m/s. Using a log-log plot, calculated lower bounds for the real data are a good estimate
the slope of this curve is also found to be almost a straight if compared to the results from the experiments.
line and with an increasing deviation at the higher
integration angles. The slope of the curve shows that the
lower bound is inversely proportional to the integration
angle and the slope is a little bit above a factor of 2.5, i.e.
σ γˆt ∝ φ −2.5 .
Finally, the lower bound was calculated according to the
measurements in [5],[8]. The radar parameters used in these
measurements are similar to the ones listed in Table II with
some discrepancies, however, that will be mentioned later.
The data set are of the real measurements made over open
sea, land and forest. The measurements are performed by
the CARABAS II system and they illustrate two cases of
moving targets, one boat at open sea and one simulated
target in the forest. In the forest, a target has been
implemented in the data according to [8]. The radar
parameters are similar to Table I and given by rmin = 9922 m
, f c = 52MHz , Br = 1.15 , γ t = 1.0155 , φ = 600 ,
SNR = 15.7dB and SCR = 5.7dB . This bandwidth and
integration angle result in an increase in correlation between
pixels in comparison to simulations presented in Fig. 1.
However we still use the independency between pixels as
discussed in previous sections. In the calculations, the lower
bound of the estimates are found to be σ Xˆ = 1.67 m ,
0
Performance of Target Parameter
Estimation Using SAR 115

Range position XI. REFERENCES

2.2
[1] M. Soumekh, “Moving Target Detection in Foliage Using
2 Along Track Radar Monopulse Radar”, IEEE Transactions
on Image Processing”, vol. 6, no. 8, pp.1148-1163, 1997
1.8 [2] J.H.G. Ender, “ Detection and estimation of moving
target signals by multi-channel SAR”, International Journal
1.6 of Electronics an d Communications, vol. 50, no. 2, p 150-
156, 1996
σ [m]

1.4
[3] C.G. Gierull, “Ground moving target parameter estimation
y

for two channel SAR”, IEE Proceedings of Radar Sonar and

1.2
Navigation, vol. 153, No. 3, pp. 224-233, 2006
1 [4] J. Ward, “Cramer-Rao bounds for target angle and Doppler
estimation with space-time adaptive processing radar”,
0.8 Conference Record of The Twenty-Ninth Asilomar
Conference on Signals, Systems and Computers, pp. 1198-
0 5 10 15 20 25 30 35
1202, 1996
φ [Degrees] [5] V. Viet, T. Sjögren, M.I. Pettersson, A. Gustavsson, L.M.H.
Ulander, ”Detection of Moving Targets by Focusing in
Azimuth position
60 UWB SAR − Theory and Experimental Results”, IEEE
Transactions on Geoscience and Remote Sensing, Vol. 48,
No. 10, 2010, pp.3799-3815
50
[6] L.M.H. Ulander, M. Lundberg, W. Pierson, A Gustavsson,
“Change detection for low-frequency SAR ground
40 surveillance” IEE Proceedings-Radar, Sonar and
Navigation, vol. 152, no. 6, p 413-420, 2005
[7] M.I. Pettersson, ”Detection of moving target in wideband
σ [m]

30
SAR”, IEEE Transactions on Aerospace and Electronics
x

Systems, Vol. 40, No. 3, July 2004, pp. 780-796

20 [8] T.K. Sjögren, V.T. Vu, M.I. Pettersson, A. Gustavsson,
L.M.H. Ulander, ”Moving Target Speed Estimation and
10
Refocusing in SAR images”, IEEE Transactions on
Aerospace and Electronics Systems, submitted 2010
[9] Dias, J.M.B., Marques, P.A.C., “Multiple moving target
0
0 5 10 15 20 25 30 35 detection and trajectory estimation using a single SAR
φ [Degrees] sensor”, IEEE Transactions on Aerospace and Electronics
Normalized relative speed
Systems, Vol 39, Issue 2, 2003, pp. 604 - 624
0.7
Normalized relative speed
[10] Zhou F., Wu R., Xing M., Bao Z., “Approach for single
0.05 channel SAR ground moving target imaging and motion
0.6
0.045
parameter estimation”, Radar, Sonar & Navigation, IET,
0.04
Vol. 1, Issue 1, 2007, pp. 59-66
0.5
0.035
[11] P.A.C. Marques, J.M.B. Dias, “Velocity estimation of fast
0.03
moving targets using a single SAR sensor”, IEEE
Transactions on Aerospace and Electronic Systems, vol.
σγt

0.025
0.4
41, no. 1, pp. 75-89, 2005
0.02
γt
σ

[12] Zhiyong Suo; Zhenfang Li, Zheng Bao, “Multi-Channel

0.015

0.3 0.01

0.005
SAR-GMTI Method Robust to Coregistration Error of SAR
Images”, IEEE Transactions on Aerospace and Electronics
0.2 0
5 10 15 20
φ [Degrees]
25 30 35
Systems , Vol. 46, Issue 4, 2010, pp. 2035 - 2043
[13] Cerutti-Maori D., Gierull C.H., Ender J.H.G.,
0.1
”Experimental Verification of SAR-GMTI Improvement
Through Antenna Switching”, IEEE Transactions on
0
0 5 10 15 20 25 30 35 Geoscience and Remote Sensing, Vol. 48, Issue 4, 2010, pp.
φ [Degrees] 2066 – 2075
[14] Sharma J.J., Gierull C.H., Collins, M.J., “Compensating the
effects of target acceleration in dual-channel SAR-GMTI”,
Figure 1. The lower boands of range, azimuth and NRS calculated using IEE Proceedings Radar, Sonar and Navigation, Vol. 153,
parameters of Table 1 as a function of the integration angle. Issue 1, 2006, pp. 53-62
[15] Ender, J.H.G., Gierull C.H., Cerutti-Maori D., ”Space-
X. ACKNOWLEDGEMENT based moving target positioning using radar with a switched
aperture antenna” IEEE International Geoscience and
The authors would like to thank the KK-Foundation for Remote Sensing Symposium, IGARSS 2007, 2007
the funding that makes this research project possible, and [16] Budillon, A., Pascazio, V., Schirinzi, G., “Estimation of
for the support from Saab Bofors Dynamics, Saab Radial Velocity of Moving Targets by Along-Track
Interferometric SAR Systems”. IEEE Geoscience and
Electronic Defence System, the Swedish Defence Research Remote Sensing Letters, Vol. 5, Issue 3, 2008, pp. 349-353
Institute (FOI) and RUAG Space Sweden. [17] Budillon A., Pascazio V., Schirinzi G., “Amplitude/phase
approach for target velocity estimation in AT-InSAR
systems”, IEEE Radar Conference, 2008. RADAR '08.
IEEE
116 Part IV

[18] M.I. Pettersson, “Lower bounds of moving target estimation [24] Jao, J.K., “Theory of synthetic aperture radar imaging of a
in low frequency UWB SAR”, International Radar moving target”, IEEE Transactions on Geoscience and
Symposium (IRS), Wroclaw, Poland, 2008, pp.151-155 Remote Sensing, Vol. 39, Issue 9, 2001, pp. 1984-1992
[19] M.I. Pettersson, “Extraction of moving ground targets by a [25] S.M. Kay, “Fundamentals of Statistical Signal Processing-
bistatic ultra-wideband SAR”, IEE Proceedings - Radar, Estimation Theory”, Prentice-Hall, New Jersey, 1993, USA
Sonar and Navigation, vol 148, no. 1, pp. 35-40, 2001 [26] V.T. Vu, T.K. Sjögren, M.I. Pettersson, H. Hellsten, ”An
[20] M.I. Pettersson, “Optimum relative speed discretisation for Impulse Response Function for Evaluation of UWB SAR
detection of moving objects in wide band SAR,” IET Radar Imaging”, IEEE Transactions on Signal Processing, Vol.
Sonar and Navigation, vol. 1, no. 3, pp. 213–220, 2007. 58, Issue 7, 2010, pp. 3927 – 3932
[21] M.I. Pettersson, V. Zetterberg, I. Claesson, “Detection and [27] V.T. Vu, T.K. Sjögren, M.I. Pettersson, “On synthetic
imaging of moving targets in wide band SAS using fast aperture radar azimuth and range resolution equations”,
time backprojection combined with space time processing”, Accepted for publication in IEEE transactions on Aerospace
Proceedings of MTS/IEEE OCEANS, 2005, Washington, and Electronic systems
USA, 2005, pp. 2388-2393 [28] Skolnik, M. I.: ‘Radar Handbook’, (McGraw-Hill, New-
[22] M.I. Pettersson, J. Askne, D.J. Cavalieri, “SAR York, 1970)
observations of Arctic freeze up compared to SSM/I during [29] Whalen A. D., Detection of Signals in Noise, Academic
ARCTIC’91”, International Journal of Remote Sensing, Press Incoperation. 1971.
Vol. 17, Issue 13, 1996, pp. 2603-2624
[23] N.S. Chauhan, R.H. Lang, K.J Ranson, ” Radar modeling of
a boreal forest”, IEEE Transactions on Geoscience and
Remote Sensing, Vol. 29, No. 4, 1991, pp. 627-638
 and image processing for moving target indication and side lobe suppression
Synthetic Aperture Radar signal
ABSTRACT
This thesis summarizes a selection of my re- algorithm on an image. The results from LORA
search contributions within Synthetic Aperture data are satisfactory especially with regard to
Radar (SAR). The research is mainly aimed at forest clutter suppression. These methods can
Synthetic Aperture Radar signal and
developing signal processing methods and app- also be applied in other fields. The ability to image processing for moving target
lying them on single channel and multi channel detect and focus images of ships allows for sur-
wideband SAR systems. veilling coastal areas to help rescuing persons
indication and side lobe suppression
SAR systems can generate images similar to at sea. Detecting and locating of cars and trucks
optical pictures, i.e. photos, and for satellite bor- allow for traffic monitoring to obtain statistics
ne systems, sometimes with much finer resolu- of the number of cars traveling on the roads and
tion compared to optical systems. SAR has also their speed.
been used for instance to obtain fine resolution
images of the moon, Venus and the satellites of In the thesis, two more important aspects for
Saturn. Other applications of SAR are detecting SAR processing are presented. One paper pre-
e.g. changes in ice sheets and deforestation. sents windowing of UWB SAR images. A stron- Thomas Sjögren
In this thesis, SAR systems capable of ima- gly reflecting object such as a power line in a
ging with high relative resolution are conside- SAR image causes strong side lobes, i.e. ringing
red. Data from such systems, namely the VHF on its both sides. This ringing can cause a small
system CARABAS-II and the UHF system house to be covered by the side lobes. Applying
LORA, are used. The term high relative resolu- a window can suppress these side lobes in the
tion imaging is used with regard to high relative image. However, the power line will broaden in
bandwidth. the image and may cover the small house if an
Two of the topics in this thesis are related unsuitable window is chosen. The last topic in
to detecting and estimating parameters of mo- the thesis is on the theoretical limits on accura-
ving targets in SAR, with CARABAS-II data cy in measuring parameters of a moving target
in the first topic and LORA data in the second. in a SAR image. The parameters are position,
A speed estimating and refocusing method is velocity, radar cross section and phase shift in
introduced for the CARABAS-II data. The re- the reflected waves. The theoretical expressions
sults show good estimation accuracy as well as of the limits on target position and relative
good ability to focus the moving targets and to speed are verified using simulation of a single
suppress forest clutter by applying a refocusing channel system.

Thomas Sjögren

Blekinge Institute of Technology

Doctoral Dissertation Series No. 2012:14
2012:14

ISSN 1653-2090 School of Engineering

2012:14 ISBN: 978-91-7295-243-0