10/14/2017

Introduction to Synthetic Aperture Radar

Dr. Armin Doerry

Detailed contact information at

www.doerry.us

Major Sections
• Introduction
• Electromagnetic Roots
• Signal Processing
• Image Formation
• Radar Equation (Performance)

This presentation is an informal communication intended for a limited audience

comprised of attendees to the Institute for Computational and Experimental Research
in Mathematics (ICERM) Semester Program on "Mathematical and Computational
Challenges in Radar and Seismic Reconstruction“ (September 6 ‐ December 8, 2017).

This presentation is not intended for further distribution, dissemination, or

publication, either whole or in part.
2

1
 10/14/2017

Synthetic Aperture Radar ‐ Introduction

SAR is first and foremost a radar SAR allows resolving target scenes to much finer
mode that allows creation of angles/locations than other real‐beam techniques;
images or maps. effectively synthesizing an antenna much larger
than what the platform might otherwise carry.
Each pixel in a SAR image is a
measure of radar energy reflected
from that location in the target
scene.

All the usual advantages of radar

apply; penetration of weather,
dust, smoke, etc.

Images are formed taking

advantage of coherent processing
of radar echoes from multiple
pulses, or over extended
observation intervals.

Synthetic Aperture Radar ‐ Introduction

Radar frequencies from VHF SAR images can be formed from aircraft,
through THz have been used for spacecraft, and ground‐based systems.
SAR.

Lower frequencies offer better

penetration of weather, foliage,
and even the ground.

Higher frequencies offer easier

processing to finer resolutions.

Pulse radars as well as CW radars

can be used.

We will hereafter assume generally

airborne microwave/mm‐wave
pulse radars.

2
 10/14/2017

Brief History
• Late 19th century
– Heinrich Hertz shows radio waves can be reflected by metal objects
• November 1903
– Christian Hülsmeyer invents “Telemobiloscope” to detect passing ships
• Reichspatent Nr. 165546, initially filed 21 November 1903.
• June 1951
– SAR idea Invented by Carl A. Wiley, Goodyear Aircraft Co.
• April 1960
– Revelation of first operational airborne SAR system
• Airborne Subsystem – Texas Instruments AN/UPD‐1
• Ground processor – Willow Run Research Center
• February 1961
– First publication describing SAR
• L. J. Cutrona, W. E. Vivian, E. N. Leith, and G. O Hall, "A High‐Resolution Radar
Combat‐Intelligence System," IRE Transactions on Military Electronics, pp 127–131,
April 1961.
• June 1978
– First orbital SAR system
• SEASAT
5

Select References
• Synthetic Aperture Radar
– Optics & Photonics News (OPN), November, 2004

3
 10/14/2017

Electromagnetic Roots for Radar

Outline (more‐or‐less)
• Maxwell’s Equations
• Wave Propagation Equation
• Plane‐Wave Propagation
• Plane‐Wave Reflection
• Radar Range/Delay
• Dielectrics
• Point Sources and Reflections
• Complicated Scattering
• Born Approximation
• Antenna Basics
8

4
 10/14/2017

Maxwell’s Equations
Maxwell’s equations relate electric fields and magnetic fields. Let there be light.
They underpin all electrical, optical, and radio technologies.

 D   (1) Gauss’ Law E = Electric Field

 B  0 (2) Gauss’ Law for magnetism H = Magnetic Field
 = charge density
B
 E   (3) Faraday’s Law J = current density
t  = permittivity
D  = permeability
 H  J (4) Ampere‐Maxwell Law
t

D   E  Electric Displacement field

B   H  Magnetic Induction field

Everything starts here.

Vector Calculus Identities/Formulae

A B  C  B C  A  C  A  B
A   B  C   B  AC   C  A B 
 A  B   B   A  A  B
   A  B   A B  B A   B   A   A  B
   2
  A  0
    0
    A     A    2 A
  A B    A  B   B   A  A     B   B     A 

S   AdS  l Adl Stokes theorem

V  A dV   S AdS Divergence theorem

10

5
 10/14/2017

Free‐Space Propagation
In free‐space there are no currents or charges, We further identify
and no losses.
1
Maxwell’s equations can be manipulated to c  Propagation velocity

2 E 
    E      Characteristic
t 2  wave impedance

and in turn, using some identities, to In free‐space

2
 E    0  8.854  1012 F m
 2 E  
t 2   0  4 107 H m
Similarly, for the magnetic field c  c0  299, 792, 458 m s
  0  377 ohms
2 H
 2 H  
t 2
Note that these are second‐order
In Cartesian coordinates, each component of the differential equations, with
vectors E and H satisfy a scalar wave equation. solutions that are sinusoids.
11

Free‐Space Propagation
Taking the Inverse Fourier Transform of both Poynting’s theorem shows that the
sides yields the Helmholtz equations direction and magnitude of energy
flow is
 2E  k 2E  0 P  EH
2H  k 2H  0 As seen in the next few slides,
Maxwell’s equations reveal that E and
where we also define H are perpendicular to each other,
f  Temporal frequency in Hz (cycles/sec.) and both are also perpendicular to the
direction of travel.
  2 f  Angular frequency in radians/sec.
 The orientation of E defines the
k  Wavenumber in radians/meter
c “polarization” of the plane‐wave.
We further define
These ‘waves’ travel, with a free‐
c 2 space velocity of propagation
   Wavelength in meters
f k Solutions have phase that is a
function of both time and space.
12

6
 10/14/2017

Sinusoidal Plane‐Wave Propagation

A propagating wave with a planar
E and H fields are related as
wave‐front is a plane‐wave.

The electric field of a linearly polarized kˆ  E =  H

plane wave is given by
The Poynting vector is in the
E  t , r   E0 cos t  k r 
direction of k̂

where

E0  E Eˆ 0  Polarization vector
k  k kˆ  Direction of propagation
r  r rˆ  Field observation point

13

Sinusoidal Plane‐Wave Propagation

If traveling in the direction of the z‐axis,
with an electric field oriented parallel to Wave front (right travelling)
the x‐axis, our field reduces to simply x E
E  Ex xˆ P
with z
y H
kˆ  rˆ  zˆ
Eˆ  xˆ
0

and the field equation reduces to E and H fields are related as

2 1 2 zˆ  E   H
2
Ex  Ex
z c02 t 2 1
zˆ  H   E

with a solution

Ex  t , z   e1 cos  t  kz 
Forward/right travelling
and another solution
Backward/left travelling
14 Ex  t , z   e2 cos  t  kx 

7
 10/14/2017

Propagation in a Dielectric
In a lossless dielectric, it remains true, but
In a dielectric
with different numerical values, that
   r 0
1
   r 0 c  Propagation velocity

where  Characteristic
 
r  relative permittivity  wave impedance
r  relative permeability with E and H fields still related as
1
These relative quantities are typically zˆ  E   H and zˆ  H   E

greater than one.
with comparable electric field solutions
Complex values denote propagation is lossy.
Ex  t , z   e1 cos  t  kz 
Frequency‐dependence implies a
Ex  t , z   e2 cos  t  kx 
“dispersive” media, where the echo may
‘not’ be a faithful reproduction of the
incident signal. Note additionally that k and  are affected.
15

Fields in a Perfect Conductor (Time‐varying fields)

In a conductor, we observe electric fields At the conductor boundary:

causing charge motion, i.e. a current,
with density calculated by Ohm’s law Applications of Stokes’ theorem to a perfect
conductor boundary shows that tangential electric
J =E field must be zero at the boundary.

Application of Gauss’ law for magnetic fields yields

where  = conductivity. that the normal electric field may exist, but must do
so with a corresponding surface charge density.
Actually a part of Maxwell’s
original set of equations Applications of Stokes’ theorem to a perfect
conductor boundary also shows that a tangential
magnetic field may exist, but must do so with a
In a perfect conductor corresponding surface current density.

  Application of Gauss’ law for magnetic fields yields

that the normal component is zero.
Free charges placed within a conductor
will disperse towards the conductor
surface, instantaneously, leaving none in Bottom line: the surface of a perfect
the interior, until the total electric field conductor cannot support tangential
inside the conductor is zero. electric fields, or normal magnetic fields.

16

8
 10/14/2017

Plane‐Wave Reflection from Perfect Conductor

For these boundary conditions to be
E met, we must also have generated at
the surface a backward travelling wave
0 z0
z such that
Let the forward‐travelling H
field encounter a perfectly Ex  t0 , z0   Ex  t0 , z0   0
conducting planar surface
at normal incidence after This backward travelling field will take
distance z0. an additional time t0 to reach the
forward wave starting point z  0
At distance z0 the forward‐travelling
Consequently, the backward travelling
wavefront will have travelled a time
field is related to the forward‐travelling
z
t0  0 field by
c  z 
Ex  2 0 , 0    Ex  0, 0 
 0 
c
Recall that at the conductor surface, the tangential  z 
electric field must be zero, and the normal magnetic Ex  t , 0    Ex  t  2 0 , 0 
field must also be zero.  c0 
Boundary
17 Conditions Radar Echo

Plane‐Wave Reflection from Perfect Conductor

We observe that the incident and These observations combine to yield
reflected fields are the following:

Ex  t , z   e1 cos  t  kz  A fundamental tenet of monostatic

Ex  t , z   e1 cos  t  kz  2kz0  radar is that any generated/transmitted
field in free‐space that encounters a
z  z0 reflecting boundary will echo a faithful
Note that the ratio of the magnitudes of reproduction (in shape) of the incident
these fields is constant (unity), and signal, to arrive at its origin with a
independent of frequency round‐trip time delay of
z0
Furthermore, the field equation is linear, tdelay  2 In free‐space
c
meaning that any signal that can be
written as the sum of sinusoids will
exhibit the same reflection characteristics, True for all frequencies.

which means pretty much any signal we

can realistically create.
18

9
 10/14/2017

Plane‐Wave Reflection from Dielectric Boundary

incident
Ei Note that for a perfect “match”
Pi
Et transmitted 1  0
Hi
 0 ,  0 Pr Hr Pt 1, 1 and for a perfect conductor
Ht
Er 1  0
reflected

The boundary conditions are that the

tangential components of E and H must With respect to power, we observe
be continuous at the dielectric interface.
2
  Relative reflected power
We define
0 2
Exr  z0  1  0 Reflection   Relative transmitted power
   1
 z0 
Exi 1  0 coefficient
Et  z  21 Transmission
  xi 0  
Ex  z0  1  0 coefficient
19

Plane‐Wave Reflection from Dielectric Boundary

For oblique angles, and lossy dielectrics,

reflections and transmission properties
are readily calculated.

Furthermore, familiar optical properties

of reflection, refraction, and Snell’s law
apply.

Similarly, for interfaces other than a The “index of refraction” is still

plane, diffraction applies. defined as

c0
n  r  r
c

20

10
 10/14/2017

Spherical Wavefronts
Consider a radiating field in a lossless E
dielectric driven by a forcing function;
a “ping” in both time and space, H
namely
2
 2 E  t , r    E  t , r     r    t 
t 2
Observations:
This has a solution
• Recall that fields are perpendicular to
 r
 t   the direction of travel
E  t , r   E0  rˆ   c for r  0 • A small finite‐dimension area becomes
4 r more planar as r increases
• Power/Energy density diminishes as 1/r2
This field is travelling in a radial direction,
• Total power/energy crossing the
with diminishing field strength.
sphere’s surface remains constant;
independent of sphere size.
Furthermore, recall that fields are
perpendicular to direction of travel.
21

Mythical Point Target

Consider a reflecting object that Consider the point reflector intercepting a
• Occupies a point in space propagating field with power density  W/m2.
• Intercepts a portion of a radiated
field, and Let the point reflector reradiate a field with a
• Emanates a reflected field from power density as seen by a receiver of
that point towards a receiver with
finite total power   4 r 2 
Furthermore, that point reflects all The point then has a “Radar Cross Section” of
frequencies equally, and instantaneously,

without generating a delay more than its RCS    m2
range from the wave emitter and receiver. 

More complicated targets are often

presumed to be merely collections Using real targets that
(clouds) of point reflectors. approximate point
target reflectors is an
These assumptions allow tractable indispensable tool for
processing algorithms to be developed. radar performance
evaluation. Courtesy NASA
22

11
 10/14/2017

Mythical Point Target

Trihedral
For real radar‐hardware testing and “corner”
evaluation, we like to use targets that reflector
mimic a point reflector to some extent.

These targets are typically large with

respect to wavelength, so geometrical
optics principles apply.

The RCS of these “canonical” targets can

be calculated with relatively high
accuracy and precision… with some
caveats.
Tophat reflector

Sphere

23

Complicated Scattering
We now presume that some incident The wave equations can be manipulated to
electric field results in a scattered, or the Lippmann‐Schwinger integral equation
reflected, electric field, with
 r ρ 
  t   
E i  t , r   Incident field E r
t, r      c  2
  ρ  2 E tot  , ρ  d dρ
4 r  ρ 
E r  t , r   scattered/reflected field

The total field is the sum of both, namely

problem
E tot  t , r   E i  t , r   E r  t , r 
Since the total field also contains the
Scattering occurs from dielectric changes,
scattered field, this becomes an equation that
which causes changes in propagation
needs to be solved, which is not tractable
velocity. For convenience we acknowledge
except for the simplest of geometries.
this with the model

1 1
 2  r 
c2 r  c0

24

12
 10/14/2017

Born Approximation See development by Cheney &

Borden, and Cheney & Borden.

To make the problem tractable, we ignore the

scattered field on the right side of the equation
and approximate the problem as
Incident field only
 r ρ 
  t   
 c  2
E r t, r       ρ  2 E i  , ρ  d dρ
4 r  ρ 

Born Approximation

This is the equivalent to assuming that the

scattered/reflected field is generally
small/weak compared to the incident field.

While this makes the problem tractable, it leads to some errors

in rendering radar data, often called multipath ‘artifacts.’
25

Born Approximation ‐ Artifacts

Direct return,
Near range Single bounce

Ray trace
Double bounce

Side of monument
Triple bounce

This image of a tank seems to

Jet engine inlets often Ground
suggest 3 cannon barrels. However
exhibit characteristic careful analysis shows that along
multipath effects. with the direct return, we have
multipath effects of double and
Far range triple bounces involving the ground.
26

13
 10/14/2017

Antenna Basics – Hertzian Dipole See development in Ramo,

Whinnery, & Van Duzer

The task of creating propagating E and M Spherical field components,

fields from signal voltages/currents is the using phasor notation, are
function of an antenna.
I 0 h  jkr  jk 1 
H 
4
e  r  2  sin 
Typical antenna design/analysis begins  r 
with establishing a current density J as a
I 0 h  jkr  2 2 
forcing function to generate the fields. Er  e  2   cos 
4  r j r 3 
Consider a short linear current element I 0 h  jkr  j  1 
of length h and current strength I0 E  e   2  sin 
4  r r j r 3 

In the far‐field, where r is large,

h    r
j I 0 h sin   jkr
E  e
4 r
h I0
E
H 
 Hertzian  Field components are
perpendicular to each other,
dipole and to the direction of travel
27

Antenna Basics – Hertzian Dipole

For our Hertzian dipole More complicated antennas can be
analyzed by treating them as collections
2
 k 2  I0h  of infinitesimal Hertzian dipoles, and
P r   2 2
sin 2   Power density superposing the results.
32 r
2
  I 0 h  As a practical matter, at large distances,
W  Total radiated power
3 2 we may assume the following

The power density related to that of an 1. Differences in the radius vectors to the
isotropic antenna is calculated as elemental dipoles are unimportant in
their effect on magnitudes.
3 2. All field components decreasing faster
p  r   sin 2 
2 than 1/r are negligible.
3. Differences in the radius vectors to the
The power radiated in some directions has elemental dipoles ‘are’ important for
been enhanced at the expense of other their phase, but may be approximated.
directions
 Antenna Gain

28

14
 10/14/2017

Antenna Basics – Linear Aperture

Consider a line‐antenna that is long The main lobe of this response has a
compared to a wavelength, but with nominal angular beamwidth of
constant‐strength current

r  h   bw 
h
 r
I0
h
The power density within this main beam
 has been enhanced with respect to an
isotropic antenna by
2 2h
The field pattern can be calculated as an p r    
 bw 
integral of a line of infinitesimal Hertzian 2
dipoles, resulting in the form
Longer antenna, shorter wavelength,
h 
E  r   e jkr E peak sinc  cos   mean more/higher gain.
 
sin   Wave‐fronts are still spherical, but
where sinc    strength varies with direction.
29 

Antenna Basics – Far‐Field Pattern

Note that the current density with shape

1 l h 1 2
l 
x  l   rect    1 2 l h 1 2
h 
0 l h 1 2
has far‐field pattern shape

 cos    cos  
X   h sinc  h
     

These constitute a Fourier Transform pair

 cos   It is generally true that the far‐field

x l   X  
   antenna pattern shape is the
Fourier Transform of the current
distribution on the radiator.
and can be shaped accordingly
30

15
 10/14/2017

Antenna Basics – Area Aperture

The main lobe of this response has
In two dimensions nominal angular beamwidths of
r  w  

r  h   bw 
h
J0 r 
 bw 
w

h The power density within this main beam

 has been enhanced with respect to an
isotropic antenna by
w
4 4 hw 4
p  r   2    2A
  2 bwbw 2 
h  w 
E  r   e jkr E peak sinc  cos   sinc  cos   Actual aperture area
   
Larger‐area antenna, shorter wavelength,
Shape is 2D Fourier Transform of mean more/higher gain, narrower beam
31 current density

Antenna Basics – Gain and Effective Area

Real antennas radiate only a fraction of the Just as current densities can cause
power with which they are supplied. radiated fields, so too can radiated fields
cause current densities. This is the
The ratio of total radiated‐power to supplied‐ duality nature of antennas.
power is the antenna efficiency.
The sensitivity of a receiving antenna
The antenna power gain in the center of its versus direction is the same as the field
main beam is approximated as shape for field generation.

The power generated by an antenna,

4 4
GA    A useable to subsequent processing, is
bwbw 2 based on the power density incident.
Specifically, received power is the
efficiency incident power density multiplied by
2
Ae  G A  Antenna effective area
4
A

32
Bigger antennas are more sensitive

16
 10/14/2017

Antenna Basics – System Parameters

From a systems standpoint, the important
parameters of an antenna are

• Frequency of operation
• Bandwidth
• Gain versus angles
• Mainlobe beamwidths
• Pattern shape in all dimensions
• Sidelobes
• Efficiency
• Phase center

Where on the physical

structure is the center of
the spherical wavefront? Answers the question
“range from where?”
33

Various internet sources

Antenna – Examples

34

17
 10/14/2017

Section Summary
• Maxwell’s equations are the root of all radar behavior
and operation
• Radar is about how radiated fields interact with, and
reflect from, dielectric boundaries
• A “point target” is a useful fiction, and can be physically
approximated for radar analysis
• The Born approximation makes radar analysis tractable,
but comes at a price of artifacts in the data rendering
• Antennas are the transducer between signal
voltages/currents and EM fields
• The far‐field pattern is related to aperture current
distribution by Fourier transform
35

Select References
• Margaret Cheney, Brett Borden, Fundamentals of radar
imaging, Society for Industrial and Applied Mathematics,
2009.
• Margaret Cheney, Brett Borden, “Theory of Waveform‐
Diverse Moving‐Target Spotlight Synthetic‐Aperture
Radar,” SIAM J. IMAGING SCIENCES, Vol. 4, No. 4, pp.
1180–1199, 2011.
• Constantine A. Balanis, Advanced Engineering
Electromagnetics, ISBN‐13 978‐0471621942, John Wiley
& Sons, Inc., 1989.
• Martin A. Plonus, Applied Electromagnetics, ISBN 0‐07‐
050345‐1, McGraw‐Hill, Inc., 1978.
• Simon Ramo, John R. Whinnery, Theodore Van Duzer,
Fields and Waves in Communication Electronics, ISBN 0‐
471‐87130‐3, John Wiley & Sons, Inc., 1984.
36

18
 10/14/2017

Mathematics of Signal Manipulation and

Processing for Radar

37

From EM Fields to Signals

The antenna is a transducer that allows
converting propagating fields to/from
voltages and currents on transmission lines.

Our time‐domain “signal” is the voltage

function with respect to time.

We manipulate these signals to extract

useful information from them, subject to the
constraints of the components we employ.

This manipulation is nothing more than

applying desired mathematical functions to
the signals of interest.

Some of these manipulations will be to

analog signals by electronic circuit elements,
and others may be to digital data after
conversion.
38

19
 10/14/2017

Time Domain and Frequency Domain

Using Maxwell’s equations, the propagating fields are

described by a second‐order differential equation, which
has at its natural solution sinusoids.

Consequently, it becomes useful to decompose our time‐

domain signals into sinusoidal constituents, and assess
signal behavior, and perhaps any modifications to the
signals, in terms of those sinusoidal basis functions.
Customary definitions
Hence the prominence of the Fourier Transform in for electrical engineering
electrical engineering.

 
 j 2 ft  j 2 ft
Xf   x t  e dt x t    X f e df
 
Forward transform Inverse transform

X  f  , x  t  are Transform pairs, for which we use the shorthand X  f   x  t 

39

Time Domain and Frequency Domain

Signal processing is often described in terms

of manipulating sinusoidal signals, and/or
sinusoidal signal components.

Both purposeful Linear and Nonlinear

manipulations are used.

40

20
 10/14/2017

Common Radar Signals

A typical radar signal as transmitted is time‐limited, and effectively

band‐limited.

It has a finite time‐duration.

It has the bulk of its energy in some finite frequency band.

Xf
B

* *
 f0 f0
f

Typically B  f0
41

Manipulating Signals
Amplification/Attenuation This is a Linear processing step.

y t    x t 

if  1 then “attenuation”

if  1 then “unity gain”

We often use a logarithmic scale
when talking gain or attenuation,
if  1 then “gain”
with units “decibel,” or “dB.”

dB  10 log10  power gain 

dB  20 log10  voltage gain 

(Caveat: same reference impedance)

42

21
 10/14/2017

Manipulating Signals
Linear Filtering This is a Linear processing step.

Linear filtering is effected to a time‐domain signal by convolution.

 If x(t) is a Dirac delta function, then

y t    x  t  h   t  d y(t) yields the filter function itself.

h(t) is the “impulse response” of
In the frequency domain the filter

H(f) is the “transfer function,” also

Y  f   X  f H  f 
known as the “frequency response”
of the filter.

It is important to us that a linear filter exhibits “linearity,” where

no mixing products are generated.

43

Manipulating Signals
Linear Filtering

The linear filter can often be decomposed into sums of

derivatives and integrations of the input signal. This may
be manipulated to a transfer function of the form

zeros

  f  m 
H f C m
  f  p 
p
poles

In fact, it is often much more common (convenient) to

describe linear filters in terms of their transfer function
(frequency response) characteristics.
44

22
 10/14/2017

Manipulating Signals
Linear Filtering
Xf
Low‐pass
f
Xf
Band‐pass

f
High‐pass Xf

Band‐stop Xf

45

Manipulating Signals

Mixing = signal multiplication This is a Nonlinear processing step.

x1  t   a1 cos  2 f1t  y  t   x1  t  x2  t 
x2  t   a2 cos  2 f 2t   a1 cos  2 f1t  a2 cos  2 f 2t 


a1a2
2
aa
2

cos  2  f1  f 2  t   1 2 cos 2  f1  f 2 1 t 

Mixing allows us to translate

Sum frequency Difference
a band‐limited signal to a
frequency
different center frequency to
allow easier processing. The desired component is selected by
filtering.

Selecting the sum = up‐conversion

Selecting the difference = down‐conversion

46

23
 10/14/2017

Manipulating Signals

Frequency Multiplication = signal to some power

This is a Nonlinear processing step.

y  t    x  t  
p
x  t   a cos  2 f t 
  a  cos  2  p f  t   ...
p

Frequency multiplication
allows us to multiply the Frequency multiple
bandwidth of a signal;
generate wideband signals The desired component is selected by
from narrow‐band signals. filtering.

47

Signal Processing Diagrams/Schematics

We use visual representations of the sequence
of mathematical operations to communicate
signal processing chains.

X  x2

mixer summer frequency doubler Signal source

Hf

h t 
amplifier Low‐pass filter Band‐pass filter general linear filters

Much variability in symbols

exists in practice. xxx

Symbols often labelled with Special purpose blocks can

48 relevant parameters be labelled as appropriate

24
 10/14/2017

Signal Processing Diagrams/Schematics

Example: Typical AM Broadcast Radio Receiver

455 kHz center

10 kHz bandwidth
Antenna Square‐law
input X detector

535 – 1700 kHz

Local
Oscillator 5 kHz

990 – 2155 kHz

(this architecture is called a
superheterodyne receiver.)

The arrows mean “tunable”

The dotted lines mean ganged, or synchronized.

Amplifiers would also typically be labeled with their gain. Audio output
to speaker

49

Modulation/Demodulation
We generally do not deal exclusively with m  t  = modulation signal, m t   1
single‐frequency sinusoidal tones.
cos  2 f0t  = carrier
Rather we generate/create signals with
desired properties and then modify them Double‐sideband
with signal processing suitable for m  t  cos  2 f0t 
modulation
transmission.
1  m  t   cos  2 f0t  Amplitude
modulation
This tone modification with interesting
signals is called “modulation.”
cos  2 f0t   m  t   Phase
modulation
Recovering the original signal (modulation)
of interest is called “demodulation.”

cos 2 f0t    m  t  dt  Frequency
modulation

others

techniques can be combined

50

25
 10/14/2017

Modulation/Demodulation

For high‐power radar transmitters, we A common model for the amplitude

prefer phase/frequency modulations, envelope is

with amplitude modulations typically  t 

rect  
limited to on/off.  TTX 

TTX  Pulse width of TX pulse

This is because high‐power

transmitter components will 1  1 2

typically distort the amplitude. where rect    1 2  1 2
0  1 2


rect  

1 2 12 

51

Linear FM Chirp Modulation

Envelope
For a transmitted (TX) pulse

 t  tn 
 cos  T  t  tn  
t
xT  t , n   AT rect 
 TTX 

Amplitude Frequency T    2 f0  T   T   d

modulated with modulated with
T
a pulse envelope a linear ramp  2 f0  T  2
2
= quadratic phase function
AT  Amplitude of TX pulse
tn  Reference time of nth TX pulse
This is a very popular radar
TTX  Pulse width of TX pulse
signal modulation, offering
T  t   Phase function of TX pulse some very nice characteristics.

More later.
52

26
 10/14/2017

Linear FM Chirp Modulation

T
T    2 f 0  T  2 Customary Properties
2
Recall that the instantaneous frequency Relatively narrow‐band
is the time‐derivative of phase Bc  f 0
d   
T    T    2  f0  T  
d  2  Relatively large time‐bandwidth product

Recall also that the finite duration of TTX Bc  1

the chirp stipulates
 TTX 2    TTX 2
The chirp bandwidth is merely the span of
This implies the frequency ramp between instantaneous frequencies over which the
waveform chirps.
 Bc  T    Bc 
 f 0  2   2   f 0  2  It is ‘not’ necessarily the width of the frequency
    span of significant energy of the Fourier Transform
of the waveform, although large time‐bandwidth

where Bc  T TTX = “chirp bandwidth” products approach this.
2
More later.
53

Quadrature Demodulation
This is the only part of the spectrum
In a real bandpass signal, negative in which we are interested
frequencies contain the same
information as positive frequencies. Xf
B

* *
 f0 f0
f

For easier processing, we will shift the portion of the

spectrum in which we are interested to lower, more
easier to process, frequencies.

We do this with “mixing,” specifically “down

conversion,’ and filtering.
Easier processing means lower
sampling frequency for data
sampling, lower data rates, and
lesser processing burden.
54

27
 10/14/2017

Quadrature Demodulation

First we mix with a signal 2e j 2 f0t X  f  f0 

This shifts the positive spectra to baseband

* *
2 f 0 0
f

X  f  f 0  H LP  f 
Then we use a linear low‐pass filter
to trim away undesired energy

*
0
f

The essential information

is preserved.
55

Quadrature Demodulation
We accomplish this with the following signal processing

j 2 m t 
x  t   cos  2 f0t   m  t   X y t   e

2e j 2 f0t Note that we began with a real signal,

mixed with a complex signal,
and output a complex signal, with
phase modulation preserved.

Although y(t) contains negative frequencies, we say

that we are working with the analytic signal of x(t).

The problem seemingly is “How do

we generate and represent complex
signals with real hardware?”

Once the signal becomes data, there

56 is no issue.

28
 10/14/2017

Quadrature Demodulation
We base our architecture on the Euler formula e j  cos   j sin 

In‐phase channel
Analog to digital
converter

x  t   cos  2 f 0t   m  t   X ADC
j 2 m t 
y t   e
2 cos  2 f 0t 

2sin  2 f 0t 

X ADC X
j
Quadrature‐phase channel Analog signal Digital data

Output data is often called “I/Q” data.

57

Quadrature Demodulation
Alternate architecture – Digital Quadrature Demodulation

In‐phase channel
x  t   cos  2 f0t   m  t  

X ADC X
j 2 m t 
y t   e
2 cos  2 f1t 
2 cos  2  f 0  f1  t  
2sin  2 f1t 

f1 is a convenient frequency
to make the subsequent
X X
digital implementation easier.
j
Analog signal Digital data
This implementation
Quadrature‐phase channel has some advantages
with respect to
It is particularly convenient when keeping real hardware
f1 is ¼ the ADC sampling frequency channels balanced.
58

29
 10/14/2017

Noise
The bane of processing real signals is “noise.”

These are the perturbations to measured signals

resulting from natural random currents and fields,
superimposed onto the signals.

These obfuscate and yield error to measurements.

With some notable exceptions, the noise is

modelled as a Gaussian random process,
sourced with little or no time correlation.

Additive White Gaussian Noise (AWGN)

The amount of noise relative to the signal

levels we seek to measure is frequently the
performance‐limiting factor for radar systems.
59

Signal Similarity
The issue is measuring the similarity of
two signals. Parseval’s Theorem

Overwhelmingly, the similarity measures If X  f   x t 

used in radar signal processing are based
on Euclidean distance, or the L2 norm. Then
 
2 2
For signals, this manifests as measures of
 x  t  dt   X  f  df
Mean‐Squared Error (MSE).  

This is tractable because the square of

signal magnitude is a power/energy
measure. Note that energy is a
conserved quantity per the first law of 2
thermodynamics. Xf = Energy Spectral Density (ESD)

So MSE is a measure of average, or

expected, power/energy in the error.

60

30
 10/14/2017

Spectral Density and Autocorrelation

2
We identify the Energy Spectral Density as SX  f   X  f 

This measures the distribution of energy with frequency


We identify the autocorrelation function as RX  t    x   x   t  dt
*



This measures the similarity of a signal with a time‐shifted version of itself

The Wiener–Khinchin theorem states that these quantities

constitute a Fourier Transform pair

S X  f   RX  t 
This is incredibly useful for
signal processing, especially
of random signals.
61

Matched Filter
The question is “Given a signal in AWGN, Derivations can be readily found in the
what filter gives us a minimum MSE in literature. A typical derivation uses the
determining the presence of that signal?” Cauchy–Schwarz inequality. In any case,
the result is that the optimum filter is
With no noise, we have
h  t   x*  t 
x t  h t  y t  This is just the time‐reversed conjugate
of the signal we wish to measure, i.e.
With AWGN input noise, we have “matched” to the desired input.

x  t   ni  t  h t  y  t   no  t  y t    x   h  t    d


 x   x   t d
*
Input Output 
AWGN Noise 
Additive and
Gaussian, but This is an autocorrelation
no longer white function, maximum at = 0.
62

31
 10/14/2017

Matched Filter

If we use the same matched filter, but

now delay our input, then we calculate The output of a matched filter,
when input with a signal to
 which it its matched, is the
y t    x   td  h  t    d signal’s autocorrelation function.


 x   td  x   t d
*
 This fact is important in radar
 waveform design.

This function is maximum at time t  td Corollary – The Energy Spectral

Density of a radar waveform
This allows us to identify the best defines the autocorrelation
estimate of delay time for a signal, function, which also defines the
perhaps a radar return. output of a matched filter.

Shape the ESD  shape the MF output.

63

Matched Filter ‐ Correlator

In practice, the matched filter is often implemented as a correlator.

We create a set of correlation kernels with different time offsets, that is

g  , tk   x   tk 

Then we correlate against the set of kernels


y  tk    x   td  g  , tk d
*


Or, with explicit input noise

y  tk     x   td   n   td   g  , tk d
*



The time offset tk that yields the maximum

output is our best match to td.

64

32
 10/14/2017

Spectral Density and Autocorrelation

Scaling property of the Fourier Transform

S X  f   RX  t 

1  f 
S X    RX   t 
 

This means that if we make the energy spectrum wider,

the consequence is a narrower autocorrelation function;
‐ sharper autocorrelation peaks
‐ more localized output of the matched filter

Wider bandwidths yield finer time‐resolution.

65

Linear FM Chirp Processing

Consider a unit‐amplitude Linear FM For typical large time‐bandwidth products
chirp waveform 
TTX Bc  1 where Bc  T TTX
 t   T 2  2
x  t   rect   cos  T  2 f0t  2 t 
 TX 
T  
For small time‐offsets, where the bulk of
the output energy will be found
Quadrature‐demodulating to baseband
(essentially the stationary‐
and ignoring the constant phase term t  TTX phase condition)
T
 t  j t2
y  t   rect  e
2 Then the matched filter output is
 TTX  adequately approximated as
This signal gets applied to a filter that is
sin  Bct 
matched to it. The output becomes z  t   TTX  TTX sinc  Bct 
 Bct 

z t    y   y   t d
*


sin  
where sinc   

66

33
 10/14/2017

Linear FM Chirp Processing

Recall that the output of a
matched filter, when input with SX  f 
a signal to which it is matched, TTX
is the autocorrelation function, Bc

f
Consequently, Bc
RX  t   TTX sinc  Bc t 

As the time‐bandwidth product shrinks

This means that the ESD is
so that
calculated to be approximately
TTX Bc  1
TTX  f 
SX  f   rect   then the spectrum model degrades
Bc  Bc 
SX  f 

f
67

Spectral Density and Matched Filter Output

SX  f  Recall the Matched Filter output,
TTX which we now write as
Bc
z  t   TTX sinc  B t 
f
Bc
TTX

In this case, where TTX Bc  1

the chirp bandwidth also describes
the width of the spectrum itself,
i.e. the signal bandwidth. 1B

B  Bc As previously discussed, as bandwidth

increases, the match filter output
Note that there are many different becomes more localized.
definitions for various kinds of
signal bandwidth.  Finer time‐resolution

68

34
 10/14/2017

Spectral Density and Matched Filter Output

Furthermore, the Energy Spectral Density Regardless of the waveform,

does ‘not’ uniquely define the waveform, note that we began with a pulse
i.e. a Linear FM Chirp. of width TTX and ended with a
Matched Filter output with
width 1 B  TTX
In fact, ‘ANY’ waveform with the same
ESD will generate the same Matched
We refer to this as
Filter output.
“pulse‐compression.”

Additionally, we may even modify the

ESD shape in order to effect desirable
properties in the Matched Filter output. The compression ratio TTX B
will often reach several orders
of magnitude, depending on the
radar and/or its application.

69

Sidelobes
An artifact of limited bandwidth
signals, is that the linear processing of sidelobes
those signals generates “sidelobes” in 13 dB
the Matched Filter output.

These sidelobes can mask other

legitimate and desired Matched Filter
output responses, like a secondary
radar target. First sidelobes for a sinc function are only ‐13 dB
with respect to the mainlobe peak.

We don’t like them.

Hamming
We can reduce the sidelobes with linear window
filtering. response

This effectively ‘un‐matches’ the Matched Filter

somewhat, with some degradations, but does
desirably impact the problematic sidelobes. Matched Sidelobe
Filter Filter
70

35
 10/14/2017

Section Summary
• Radar echo signals may be “processed” to more
convenient frequencies by mixing and
demodulation
• The optimum (minimum MSE) filter for measuring
radar echoes is the “Matched Filter”
• The Energy Spectral Density (ESD) of a transmitted
waveform ultimately defines the nature of the
Matched Filter output
• Wideband signals allow finer time resolution at the
Matched Filter output

71

Select References
• Noise and Noise Figure for Radar Receivers
– Sandia National Laboratories Report SAND2016‐9649
• Catalog of Window Taper Functions for Sidelobe Control
– Sandia National Laboratories Report SAND2017‐4042
• John R. Klauder, A. C. Price, Sidney Darlington, Walter J.
Albersheim, “The theory and design of chirp radars,” Bell
Labs Technical Journal, Vol. 39, no. 4, pp. 745‐808, 1960.
• Generating Nonlinear FM Chirp Waveforms for Radar
– Sandia National Laboratories Report SAND2006‐5856
• Shaping the Spectrum of Random‐Phase Radar
Waveforms
– Sandia National Laboratories Report SAND2012‐6915

72

36
 10/14/2017

SAR Image Formation

73

Imaging Geometry Model

rc,n = vector defining radar location for nth pulse SAR is about “space”
s = vector defining target scatterer location more than “time.”
Consequently, we
need a good definition
SRP = Scene Reference Point
of the geometric
(nominal center of geometric coordinate frame)
relationships between
radar and targets.
z
The radar will emit
and collect pulses
y along its flight path.

s rc,n
SRP
x
n  g ,n

74

37
 10/14/2017

Envelope
Signal Model – TX Signal
Everything begins with a signal model. Phase
function
We shall assume a transmitted (TX) signal
of a single pulse of the form
In addition, we have assumed an
 t  tn  arbitrary phase function. Many phase
xT  t , n   AT rect   cos  T  t  tn   functions can be used. It is desirable
 TTX  that the phase function be a modulation
AT  Amplitude of TX pulse that provides a desired Energy Spectral
tn  Reference time of nth TX pulse Density for the waveform. One
common waveform is the Linear
TTX  Pulse width of TX pulse Frequency Modulated (LFM) chirp
T  t   Phase function of TX pulse signal.

T  t   T ,n  T ,nt  T ,n t 2
We have tacitly assumed a rectangular 2
pulse envelope. This doesn’t have to be
so. Other envelopes can be used. T ,n  Reference phase
However, rectangular is a popular pulse T ,n  Reference center frequency
shape model for systems where TX power
 T ,n  Reference chirp rate
amplifiers are driven into compression.
75

Signal Model – RX Echo Signal

The received (RX) echo signal from the target scatterer is simply an attenuated and
delayed version of the transmitted signal, namely

 t  t n  t s ,n 
xR  t , n, s   AR rect  
 cos T  t  tn  ts ,n   This is one of the most
 TTX  fundamental presumptions
in radar, and the real
AR  Amplitude of RX pulse starting point for algorithms.
ts ,n  Time delay of RX echo pulse for nth pulse

The echo time delay is related to the c  Velocity of propagation

geometry by the velocity of propagation
of the waveform rs ,n  rc,n  s  Vector from target
scatterer to radar
2
t s ,n  rs ,n
c

The factor of 2 accounts for a round trip

76

38
 10/14/2017

Signal Model – RX Echo Signal

Target scene
The collection of pulse Each location within the
locations along the flight target scene generates a
path define the synthetic unique range profile rs,n
aperture. along the synthetic
aperture.
Pulse spacing is chosen to
meet anti‐aliasing It is the uniqueness of the
requirements for the profile that allows us to
illuminated scene – separate these locations
another way of saying from each other.
that radar PRF is chosen
to adequately sample  ap
Doppler bandwidth of
scene illuminated by Flight path
antenna.
x x x x x x x x x
Pulse locations (synthetic aperture)
77

The Doppler effect (or Doppler shift) is

What is Doppler in SAR? the change in frequency of a wave (or
other periodic event) for an observer
moving relative to its source. It is
named after the Austrian physicist
Consider a CW pulse echo Christian Doppler, who proposed it in
(pulse of fixed‐frequency signal) 1842 in Prague. It is commonly heard
when a vehicle sounding a siren or
horn approaches, passes, and recedes
Received signal from a static range rs,n  rs 0 from an observer.
– Wikipedia, 21 September 2017

  2 
xR  t , n, s   AR rect  ... cos  2 f0  t  tn  rs 0  
  c 

Received signal from a linearly changing range rs,n  rs 0  vs  t  t0 

  2 
xR  t , n, s   AR rect ... cos  2 f 0  t  tn   rs 0  vs  t  t0    
  c 
  2  4 f 0 4 f0 
 AR rect ... cos  2 f 0 1  vs   t  tn   rs 0  vs  tn  t0  
  c  c c 

Time/Frequency scaling inside of a pulse Pulse‐to‐pulse phase change

78 Often ignored Principal exploited effect

39
 10/14/2017

What is Doppler in SAR?

A tacit assumption in SAR is that tn increases linearly
with pulse index n, which increases linearly with
distance flown along a synthetic aperture.

So, a pulse‐to‐pulse phase change becomes a

spatially‐dependent phase change,
i.e. a wavenumber measure.

For SAR, ultimately the velocity, and hence the

times at which data are collected, is immaterial.
The important factor is “where” the synthetic
aperture has been sampled.

The concept of “Doppler” is just a means to an end.

79

Signal Model – Baseband Video Signal

The RX signal is typically demodulated to baseband We want to translate
for easier processing. Accordingly, we define a Local the echoes to lower
Oscillator (LO) signal of the form frequencies where they
are more easily sampled
 t  tn  tm 
xL  t , n   2 rect   exp  j   L  t  tn  tm  
and processed.
 TLO 

TLO  Pulse width of LO pulse

tm  Reference time delay of LO pulse
 L  t   Phase function of LO pulse

Quadrature demodulation yields a video signal of the form

 t  tn  t s ,n   t  tn  tm  T  t  tn  ts,n  
xV  t , n, s   AR rect   rect  T  exp j 
 T TX   LO   L  t  tn  tm  

This is just a product of the envelopes and a difference of the phases.

80

40
 10/14/2017

Signal Model – Baseband Video Signal

If the radar position changes as a function of pulse index n, mainly by changing the
aspect angle , then we may write our model as

 2    2 
 t  tn  c rs,n   t  tn  tm  T  t  tn  rs,n 
xV  t , n, s   AR rect   rect  T  exp j  c 
 TTX   LO    t  t  t  
   L n m 

Or more explicitly in terms of radar position and target position as

 2    2 
 t  tn  c rc,n  s   t  tn  tm  T  t  tn  rc,n  s  
xV  t , n, s   AR rect   rect  T  exp j  c 
 TTX   LO    t  t  t  
   L n m 

Note that everything is known in our video signal except the target amplitude AR
and target scatterer location s. Recall that this is for a single point target scatterer.

81

Signal Model – Distributed Target Scene

More generally, a target scene is composed of many scatterers, and the echo video
signal is a superposition (integration) of all of them. We can write this as

 2    2 
 t  tn  c rc,n  s   t  tn  tm  T  t  tn  rc,n  s  
xV  t , n, s      s  rect   rect   exp j  c  ds
s  TTX   TLO     t  t  t  
   L n m 
Where the single target amplitude is now replaced by

  s   Target scene reflectivity function

If the reflectivity function were a single impulse, then the previous model would apply.

The question now becomes “If we pick some arbitrary scene location,
how much echo signal can we measure as coming from that location?”

82

41
 10/14/2017

SAR Processing – Matched Filter

Let us now pick a test location and create the following reference filter function for
a particular scatterer location

 2    2 
 t  tn  c rc,n  sˆ   t  tn  tm  T  t  tn  rc,n  sˆ  
h  t , n, sˆ   rect   rect  T  exp j  c 
 TTX   LO    t  t  t  
   L n m 
where
sˆ  Test location

All quantities in the filter function are known or specified.

Note that this is just the expected response from a unit amplitude target at the test
location. We want to see how well the actual data matches this.

The task at hand is to “filter” or correlate the

input video signal against this function.

83

SAR Processing – Matched Filter

The output of this filter (for a particular scatterer location) is calculated as

y  sˆ     xV  t , n, s  h*  t , n, sˆ dt
n t

where “*” denotes complex conjugate. Note that we do this for all pulses and over the
entire pulse for each pulse.

Doing so for an array of interesting/desired locations ŝ will yield a SAR image.

A direct implementation of this algorithm is very computationally

intensive, and generally prohibitively so.

As a result, we attempt to select waveforms and algorithms that get us

very close to this result, but with a more tolerable computational load.

Nevertheless, all SAR image formation algorithms attempt to at least

approximate this “matched‐filter” output.
84

42
 10/14/2017

SAR Processing – Resolution

The question is now “If the target is indeed an impulse, how well can we localize the
target’s response?”

The overall function is called the “Impulse Response” (IPR) of the SAR image.

The IPR width is its resolution. This is most often characterized in cardinal directions
of ‘range’ and ‘azimuth’. Neglecting any sidelobe filtering, we can calculate

c
r   Slant‐range resolution BT  Bandwidth of TX signal
2 BT


a   Azimuth resolution   Nominal wavelength of TX signal
2 ap cos g ,0

 ap  Nominal angle subtended by

synthetic aperture in ground plane
Azimuth resolution depends how
much data you collect, not the  g ,0  Nominal grazing angle
size of the antenna you carry
85

SAR Processing – A Zoo of Algorithms

There are many different SAR image formation algorithms, and usually many variations
of each. They all have strengths and weaknesses. A partial list might include

Doppler Beam Sharpening

Simple 2D‐DFT (2D‐FFT) Processing
Polar Format Algorithm
Overlapped Subaperture Algorithm Forgive me if I left out
Range Migration Algorithm your favorite one
Chirp‐Scaling Algorithm
Wavenumber‐Domain Processing
Backprojection Processing

Picking one over the other is done by first examining the following

• parameter space of the data, including frequency, bandwidth,

waveform, imaging geometry, etc.
• Processing constraints, including need for real‐time, image size,
processing hardware, etc.
86

43
 10/14/2017

SAR Processing – A Zoo of Algorithms

A fairly typical trade space for algorithm selection might be described by

Less More
computationally computationally
intensive intensive
Image Formation Algorithm

Lower Higher fidelity

fidelity image image
Application
More dependent Less
approximations design point approximations
to data model to data model

87

Spotlight vs. Stripmap Processing

Spotlight Spotlight SAR creates a
single full 2D image from a
SAR single synthetic aperture

Synthetic aperture

Stripmap SAR

…
Classic Stripmap SAR processes a column of pixels from a single More typically today, Stripmap SAR images are formed by
synthetic aperture, and then adds/drops data to form the next mosaicking individual Spotlight SAR images formed from non‐
column of pixels, for an arbitrarily long composite image. overlapping distinct synthetic apertures.

88 Images Courtesy of Sandia National Laboratories, Airborne ISR

44
 10/14/2017

Pulse Compression – LFM Chirp Correlation

near range far range
f echo echo

LO pulse Pulse compression via

Microwave correlation filters is
TX pulse signals indicated when the
t range swath is a large
f fraction (or larger) of
Baseband the TX pulse width, or
t video signals for coarser range
resolutions.
f
Note that the
Correlation baseband video signal
t filters bandwidth is the same
as the microwave
correlator signal bandwidth.
Range‐
output
compressed This model applies to
signals other waveforms, too.

89
range

Pulse Compression – LFM Chirp Stretch Processing

near range
f echo
far range
echo
Pulse compression via
LO pulse Microwave stretch processing is
TX pulse signals indicated when the
t range swath is a small
f fraction of the TX pulse
Baseband width, or for finer
t video signals range resolutions.

After de‐ramping the echo energy with a LO Note that the

pulse that is also a chirp, range is encoded in baseband video signal
baseband video frequency. Consequently a DFT bandwidth is less than
will separate frequencies, and hence ranges. the microwave signal
bandwidth.
DFT
Range‐ Stretch processing
output
compressed generally assumes LFM
signals chirp waveforms.

90
range

45
 10/14/2017

Doppler Processing
Geometry
with targets Target locations displaced
displaced in in azimuth exhibit
azimuth different range profiles,
that manifest as different
echo phases. This causes
f Target different phase
Doppler fluctuations as a function
t frequency of pulse position, i.e.
responses Doppler.
Doppler
focus
Consequently, an analysis
f function of Doppler characteristics
Focused will yield a measure of
Doppler resolution in azimuth.
t
frequencies

Focusing Doppler entails removing the Doppler “chirp”

characteristic, and causing targets to correspond to
unique constant Doppler frequencies.
91

Stretch Processing the Video Signal

We now choose to narrow the scope of our processing to a TX signal that is a LFM
chirp. Furthermore, we will assume that the LO signal is also a LFM chirp with the
same constant reference phase, constant center frequency, and constant chirp rate.
Furthermore, we will digitize (sample) the video signal such that

 t  tn  tm   Tsi
where
Ts  ADC sample spacing
i  ADC sample index within any one pulse  I 2  i  I 2

The video signal model then becomes a function of the new index i, which we denote
 T ,0   T ,0Tsi  ts ,n  tm  
 
xV  i, n, s   AR exp j   
t  t 
T ,0 2
 
 2 s ,n m 
Note that all quadratic phase terms have disappeared. The signal has been “de‐chirped.”

92 In fast‐time, index i

46
 10/14/2017

More Geometry
We expand the terms that involve time‐delay into ranges as follows

  
 ts,n  tm   2c rs,n  rm  2c rc,n  sˆ  rm 
where rm  Reference range associated with tm
We may further expand the relative delay term to

 ts,n  tm   2c  
2 2
rc,n  sˆ  2rc,n  sˆ  rm 
 
For small horizontal target offsets with respect to range, this may be approximated as

 ts,n  tm   2c   rc,n  rm   s y cos g ,n cos n  sx cos g ,n sin  n 


where the target location has coordinates s  s x , s y , s z 
This simplification is tantamount to presuming planar wavefronts (instead of spherical)
at the SRP. It is often called the “far‐field” approximation. Terms ignored in this
simplification generate errors, and limit the SAR scene size and resolutions achievable
with adequate fidelity. But it is often “good enough.”
93

Processing – Simple 2D‐DFT

If we further simplify the geometry to allow small‐angle approximations for  n

 
 ts,n  tm   2c rc,n  rm  s y cos g ,0  sx cos g ,0 n 
and assume the pulses are sampled such that

 n  d n with  N 2  n  N 2

and make further simplifying assumptions about the video signal, including no
motion during the pulse, the we can arrive at a model
 2T ,0 
 c 
rc,n  rm  
 
 2T ,0 cos g ,0d 2 T ,0Ts cos g ,0 
xV  i, n, s   AR exp j  sx n  s yi 
 c c 
  inconsequential error terms  
 
 

We will assume that rm is a constant.

94

47
 10/14/2017

Processing – Simple 2D‐DFT

We observe that the first term is a phase term that varies with the range to the SRP. If
the flight path were straight, this would be predominantly a quadratic term.
Compensating for this term will “focus” the SRP, and some neighborhood around it.

Ignoring the inconsequential error terms, and compensating the first term, yields the
expression
 2T ,0 cos g ,0d 
 sx n 
 This suggests
xV  i, n, s   AR exp j 
c
 that the data
2
 T ,0 sT cos
s yi 
g ,0
samples are on a
 c 
rectangular grid
Assuming, of
Note that the first phase term here is linear exclusively in index n, course, lots of
and the second phase term here is linear exclusively in index i. simplifications

This suggests then that a 2D‐DFT can determine the frequencies of the corresponding
indices, and therewith identify target spatial coordinates. The independence of the
indices allow us to perform the 2D‐DFT as orthogonal independent 1D‐DFTs.
95

Processing – Simple 2D‐DFT

We perform the 2D‐DFT conventionally with the summations

y  v, u    xV Tsi, n, s  exp j 

i n
 2
N
un
2
I
vi 
where u and v are image pixel indices in the azimuth and range dimensions.

Performing the 2D‐DFT yields the “image” with pixel values described by

  N  2T ,0 cos g ,0 d 2   
WN   sx  u  
 2
   c N  
y  v, u   AR  
This is the complete
2 T cos g ,0
 W  I   T ,0 s sy 
2   
v
2D Impulse Response
 2 
I
 c I    (IPR)

where the “shape” of the single point‐target response is given by

M 21 2 f f Has peak at f = 0,
j m j  sin  f  
WM  f    e M Me M
 
and mainlobe
m M 2  M sin  f M   nominal width of 1
96

48
 10/14/2017

Processing – Simple 2D‐DFT – Extensions

Recall that we made the assumption that that rm is a constant in the following signal model

 2T ,0 
 c rc,n  rm   
 
 2T ,0 cos g ,0d 2 T ,0Ts cos g ,0 
xV  i, n, s   AR exp j  sx n  s yi 
 c c 
  inconsequential error terms  
 
 
If we had allowed the reference range to equal the range to the SRP, namely rm  rc,n
on a pulse‐to‐pulse basis, then the signal model would be

 2T ,0 cos g ,0 d 2 T ,0Ts cos g ,0 

 sx n  s yi 
xV  i, n, s   AR exp j  c c 
  inconsequential error terms  
 

and no independent focusing operation would be needed. The data would already be
focused to the SRP.
Real‐time motion compensation.
97

Processing – Simple 2D‐DFT – Extensions

Recall that we assumed that the shape of the IPR in cardinal directions was given by
M 21 2 f f
j m j  sin  f  
WM  f    e M Me M
 
m M 2  M sin  f M  

This function exhibits fairly high processing sidelobes. These sidelobes can be
diminished by tapering, or “windowing,” the data prior to DFT processing. For example
M 21 2 f
j m
WM  f    wM  m  e M 0

IPR plots
no window
Hamming

m  M 2 -10 Hanning

-20
where
-30

wM  m   Window function
dBc

-40

Popular window functions include -50

Hamming, Hanning, Taylor, and -60

Blackman. There are many others. -70

The cost is a wider mainlobe. This is

-80
generally considered a good trade. -10 -8 -6 -4 -2 0
pixels
2 4 6 8 10

98

49
 10/14/2017

Processing – Simple 2D‐DFT – Extensions

If we now include a target with some height offset, that is, s z  0 then

 ts,n  tm   2c   rc,n  rm   s y cos g ,0  sx cos g ,0 n  sz sin g ,0 

This can be rewritten as

 ts,n  tm   2c   rc,n  rm    s y  sz tan g ,0  cos g ,0  sx cos g ,0 n 

The SAR image is then modified to

  N  2T ,0 cos g ,0 d 2   
WN   sx  u  
  2  c N  
y  v, u   AR  
  2 T ,0Ts cos g ,0 2  
I
 W    
s y  s z tan g ,0  v  
 2 
I
 c I   

This indicates that a target that is above the nominal “Layover”

ground‐plane will manifest energy at the equivalent of
a ‘nearer’ ground‐range. (foreshortening)
99

Processing – Simple 2D‐DFT – Limitations

The previously termed “inconsequential error terms” in the signal model represent
errors can in fact become “consequential” under the right conditions. These errors
tend get worse for target scene locations farther away from the SRP, and at finer
resolutions.

The errors cause excessive blurring in the SAR image. This is due to uncompensated
phase errors, as well as residual range migration errors. Degradation tends to be
gradual as distance from SRP increases.

For simple 2D‐DFT processing, the scene diameter limits are due to residual migration,
and are often expressed as

4
Dx  a r  Azimuth scene diameter limit

4 2
Dy  a cos g ,0  Ground‐range scene diameter limit

For 2 cm wavelength, 1 m resolution,
and shallow grazing angles, the scene
100 diameter limit is on the order of 200 m.

50
 10/14/2017

Wavenumbers
Recall that our focused signal model was expressed as
 2T ,0 cos g ,0 d 2 T ,0Ts cos g ,0 
xV  i, n, s   AR exp j  sx n  s yi 
 c c 
This can be written as


xV  i, n, s   AR exp j k x  n  sx  k y  i  s y 
where These are spatial‐frequency terms, known
as “wavenumbers.” The fact that they
2T ,0 cos g ,0d
kx  n  n each depend on an index indicates that
c they each represent a ‘band’ of
2 T ,0Ts cos g ,0 wavenumbers.
k y i    i
c
The width of the bands is what allows
resolution of the spatial variables.

Processing is considerably easier when

each is dependent solely on an
independent index. We like this.
101

Processing – Polar Format Algorithm

The Polar Format Algorithm (PFA) for SAR image formation “fixes” the principal error
sources that limited the 2D‐DFT processing algorithm.

We recall our signal model for a sampled constant‐waveform LFM chirp where we are
employing stretch processing as
  2
xV  i, n, s   AR exp j  T ,0   T ,0Tsi  ts,n  tm   T ,0  ts,n  tm  
 2 
We employ motion compensation such that the reference range tracks the range from
the radar to the SRP as rm  rc,n
We also then make the less severe (than for 2D‐DFT processing) approximation

 ts,n  tm   2c   rc,n  rm   s y cos g ,0  sx cos g ,0 n  Very simple

 ts,n  tm   2c  s y cos g ,n cos n  sx cos g ,n sin  n  sz sin g ,n  Less simple

and sample with pulse positions at equal angle increments  n  d n

102
This is still a far‐field planar‐wavefront approximation.

51
 10/14/2017

Processing – Polar Format Algorithm

 T ,0
 t s ,n  tm 
2
The phase term is known as the Residual Video Phase Error (RSVE)
2
and is responsible for the “skew” between near‐range and far‐range in the video
signal plot of frequency versus time.

It can often be ignored, but in any case can be precisely removed with some
preprocessing of the data. The preprocessing involves a DFT followed by a quadratic
phase‐error correction, followed by an IDFT. This is sometimes called “deskewing.”

In either case, the signal model for the target scene data then becomes
   s x cos g ,n sin  n  
 2   
xV  i, n, s   AR exp j   T ,0   T ,0Ts i    s y cos g ,n cos  n  
 c   
   s z sin g ,n  
where samples are taken at pulse positions  n  d n
This is the starting point for PFA processing.

103

Processing – Polar Format Algorithm

In terms of wavenumbers, we may write the signal model as


xV  i, n, s   AR exp j k x  i, n  sx  k y  i, n  s y  k z  i, n  sz 
Note that all
where the orthogonal wavenumber functions are identified as wavenumbers are
2
c

k x  i, n    T ,0   T ,0Tsi  cos g ,n sin  n  functions of both
indices.
2

k y  i, n    T ,0   T ,0Tsi  cos g ,n cos  n
c
 This cross‐coupling
of indices causes
2
k z  i, n    T ,0   T ,0Tsi  sin g ,n
c
  problematic range
migration.

A careful study of these wavenumbers shows that for each data index pair (i,n), the
wavenumber triplet (kx,ky,kz) describes a specific location in the 3D Fourier‐space
of the target scene.

The grazing angle and aperture angle also describe the polar angle in Fourier space, and
the radius from DC is described by  2 c  T ,0   T ,0Tsi 

104

52
 10/14/2017

Wavenumbers Again
A 3D world nets a 3D wavenumber domain.

3D‐DFT
z kz

y ky
Inverse
3D‐DFT
x kx
3D target space 3D wavenumber domain
(Fourier‐space of the scene)

105

Processing – Polar Format Algorithm

Individual data samples are associated with specific locations in the 3D wavenumber
domain. The collection of all data samples describes a sampled surface in the 3D
wavenumber‐domain of the image.

If we know (or presume) that sz  0 , then we may project these samples onto the kx-ky
plane, turning the problem from a 3D problem to a 2D problem.

Wavenumber domain

106

53
 10/14/2017

Processing – Polar Format Algorithm

Recall that efficient 2D‐DFT processing
requires, and assumes that, the samples
be on a rectangular grid.

In the 2D wavenumber plane, however,

the samples are typically not on a
rectangular grid. For straight‐line flight #1
with constant waveforms, the samples
are on more of a polar grid.

However, the data may be resampled

from this onto a rectangular grid, usually
in two stages, thereby allowing more #2
efficient processing by fast 2D‐DFT
techniques.

This resampling is called “polar

reformatting,” giving name to the
algorithm.
107

Processing – Polar Format Algorithm

This resampling creates new separated indices such that the new resampled signal
model can be written for a flat target scene as
Wavenumbers are

xV  i, n, s   AR exp j k x  n  sx  k y  i  s y  now functions of
new independent
single indices

If we assume that sample spacing is approximately the same as in the center of the
original wavenumber domain data, then performing the 2D‐DFT on this resampled data
again yields the “image” with pixel values described by

  N  2T ,0 cos g ,0 d 2   
WN   sx  u  
  2  c N  
y  v, u   AR  
2 T cos g ,0
 W  I   T ,0 s sy 
2   
v
 2 
I
 c I   

108

54
 10/14/2017

Processing – Polar Format Algorithm – Limitations

The complete far‐field planar‐wave approximation to the relative delay term yields a
higher‐fidelity signal model than for simple 2D‐DFT processing, but is nevertheless still
an approximation with some residual phase‐error terms. These phase‐error terms are
more pronounced for targets farther away from the SRP, at nearer ranges, and for
finer resolutions.

These errors also cause excessive blurring in the SAR image, in this case due to mainly
uncompensated phase errors, predominantly a quadratic phase‐error. Degradation
tends to be gradual as distance from SRP increases.

For simple PFA processing, the scene diameter limits are often expressed as

rc,0
Dx , D y  4  a  Azimuth/Range scene diameter limit


An increase from a For 2 cm wavelength, 1 m resolution,

simple 2D‐DFT and 10 km range, the scene diameter
109 limit is on the order of nearly 900 m.

Processing – Backprojection
The Backprojection (BP) image formation algorithm for SAR has its roots in
tomography. It mitigates errors that limit PFA and 2D‐DFT processing. In fact, it is
essentially a matched‐filter algorithm albeit for range‐compressed data.

We recall our stretch‐processing signal model as

  2
xV  i, n, s   AR exp j  T ,0   T ,0Tsi  ts,n  tm   T ,0  ts ,n  tm  
 2 
For convenience, we will again employ motion compensation such that the reference
range tracks the range from the radar to the SRP as rm  rc,n

However, now we will abandon any approximations for the relative delay  ts,n  tm 
as was done for PFA and 2D‐DFT processing, and specify exactly

 ts,n  tm   2c  rs,n  rc,n   2c  rc,n  s  rc,n   2c sr ,n

where

sr ,n  rc,n  s  rc,n  = relative slant‐range offset

110

55
 10/14/2017

Processing – Backprojection
The RVPE term can be compensated as with PFA processing. Doing so yields the signal
model we will use for processing, namely


xV  i, n, s   AR exp j 
2
c
T ,0   T ,0Tsi  sr ,n 
Performing range‐compression with a range DFT yields the resulting range‐
compressed signal model
 I  2 T 2    2 
z  m, n   AR WI    T ,0 s sr ,n  m   exp j  T ,0 sr ,n 
 2  c I   c 

Range profile Doppler term

We note that m and n are integer indices.

This is really the starting place for BP processing.

111

Processing – Backprojection
The data from any
. . . . . . . . .
The basic idea is to one pulse, at any
. . . x. . . . . .
1. select a pixel one range, is a
location, . . . . . . . . . superposition of all
2. interpolate the . . . . . . . . . targets along that
data to the . . . . . . . . . constant‐range arc.
exact range of . . . . . . . . .
the pixel, . . . . . . . . . It is the different
3. align the aspect angles that
phases, and allow separation of
4. accumulate the targets along that
result. arc.

* * *

112

56
 10/14/2017

Processing – Backprojection
First we select a pixel location ŝ that corresponds to image index pair (v,u).

Then we calculate for this pixel a corresponding range‐offset from the SRP for each pulse as


sˆr ,n  rc,n  sˆ  rc,n 
Next we calculate a compressed range‐vector fractional index location for each pulse as
 I  2 T ,0Ts
m     sˆr ,n
 2  c

We then back‐project each pulse by interpolating the compressed‐range vector to the

fractional index location and compensating for the Doppler term at that particular range‐
offset. Finally this is summed over all pulses for that particular pixel. We calculate this as
 2 
y  v, u    z  m, n  exp j  T ,0 sˆr ,n 
n  c  (Window functions for sidelobe
control can be applied during the
This is then repeated for every pixel index pair
initial range‐compression and
(v,u) in the entire image.
before the final summation.)
113

Processing – Backprojection
If we choose a pixel‐spacing similar to what the processing model for PFA and 2D‐DFT
processing, then we will end up with a similar image model, namely

  N  2T ,0 cos g ,0 d 2   
WN   sx  u  
  2  c N  
y  v, u   AR  
2 T cos g ,0
 W  I   T ,0 s sy 
2   
v
 2 
I
 c I   

except that this image model will exhibit higher fidelity over a larger image area and
for finer resolutions, etc. In fact, there is no inherent scene‐diameter limitation to
the processing itself. This is effectively a matched‐filter result.

This fidelity improvement comes at a price of increased computational load over PFA
and 2D‐DFT processing.

114

57
 10/14/2017

Some Remaining Issues

• We have tacitly ignored some additional processing issues. These
include

– Compensating for antenna beam pattern roll‐off in the image

• Both azimuth and elevation
– Compensating for range‐loss across the image
– Effects of imprecise/inaccurate radar position and motion information
• Employment of “autofocus”
– Effects of time quantization
• Waveforms
• ADC samples
– Effects of Doppler within the pulse itself
• We assumed that the radar was effectively ‘stationary’ during the pulse
– Effects of non‐ideal components/circuits
• Phase errors, amplitude errors
• System delays
– Effects of velocity‐of‐propagation errors
• Atmospheric variations in c
– Selecting more convenient pixel spacing
• Zero‐padding data, etc.
115

Section Summary
• SAR processing attempts to implement a matched‐
filter for each pixel location in the image.

• The different algorithms make various simplifying

assumptions that may or may not be valid,
depending on the imaging parameters and fidelity
required.

• The LFM chirp has some nice properties for SAR

processing, especially at fine resolutions, but other
waveforms can be used, too.
116

58
 10/14/2017

Select References
• Spotlight‐Mode Synthetic Aperture Radar: A Signal Processing Approach,
– Jakowatz, et al., ISBN 0‐7923‐9677‐4
• Spotlight Synthetic Aperture Radar, Signal Processing Algorithms
– Carrara, et al., ISBN 0‐89006‐728‐7
• Basics of Polar‐Format Algorithm for Processing Synthetic Aperture
Radar Images
– Sandia National Laboratories Report SAND2012‐3369
• Range‐Doppler Imaging of Rotating Objects
– J. L. Walker, IEEE Trans. on Aerospace and Electronic Systems, January 1980
• Basics of Backprojection Algorithm for Processing Synthetic Aperture
Radar Images
– Sandia National Laboratories Report SAND2016‐1682
• Computed Tomography – the details
– Sandia National Laboratories Report SAND2007‐4252
• A tomographic formulation of spotlight‐mode synthetic aperture radar
– D. C. Munson, et al., Proceedings of the IEEE, August 1983
• Catalog of Window Taper Functions for Sidelobe Control
– Sandia National Laboratories Report SAND2017‐4042

117

SAR Radar Equation

118

59
 10/14/2017

The Radar Equation

• This is the equation that relates basic radar

performance to basic radar parameters
– i.e. range, resolution, power, etc.

• We will focus on Synthetic Aperture Radar (SAR)

– Monostatic

• Used for performance trade studies

119

Power Radiation Towards Target

The emitted field expands in a spherical manner, diminishing its power density

 1   1  1  1 
PT   G A    2   L   TX power density (W/m2)
 TX   radome
L L   4 R   atmos  at target range

TX Power Loss from

amplifier TX power
output amplifier
Antenna Radome Range to Atmospheric
to
gain loss target loss
antenna

TX
Power dispersed over
an ever‐larger spherical
surface will diminish over a
fixed area.
120

60
 10/14/2017

Interaction With Target

The target captures some of the radiated energy, and then re‐radiates it back
towards the receiver, i.e. reflection
 1   1  1  1 
PT   G A    2   L     Power captured and radiated back
 TX   Lradome
L   4 R   atmos   towards receiver (W)

Target Radar Cross

Section (m2)

TX

121 Courtesy of Sandia National Laboratories, Airborne ISR

Re‐radiation Back Towards Receiver

The reflected field expands in a spherical manner, diminishing its power density
with range

 1   1  1  1    1  1  1 
PT   G A    4 R 2   L

   L   4 R 2   L  
 LTX   Lradome   atmos 
   atmos   radome  

= Power density at receiver (W/m2) with range

We are assuming a monostatic radar configuration, i.e. the path from TX to RX is just
the reverse of the path from RX to TX.
Power re‐radiated
back towards the
radar will also
expand in an ever‐
larger spherical
TX surface, and
therefor diminish
over a fixed area

122

61
 10/14/2017

Capture by the Receiving Antenna

The receiving antenna intercepts and collects some of the power in its
neighborhood, converting it to power at its terminals.
 1   1  1  1    1  1  1 
PT   G A    4 R 2   L       2   L   Ae
 TX   Lradome
L   atmos    Latmos   4 R   radome  

= Power captured by RX antenna (W)

Receiver antenna
effective area (m2)

TX

123 Courtesy of Sandia National Laboratories, Airborne ISR

Collecting Terms
Collecting some terms and simplifying yields

PT G A Ae  1 
Pr  2 4 L   Signal Power at RX antenna port
 4  R  radar Latmos 

where
Lradar  LTX Lradome  Combined radar system miscellaneous hardware losses

TX

124

62
 10/14/2017

Competing Noise
The received signal power must compete with noise that is also received and/or
generated by the radar… This is added to the signal that is received…

N r  kTFN BN  Noise Power at RX antenna terminals

where

k  1.38  1023 J/K = Boltzmann’s constant

T  290 K = system reference temperature (nominal scene noise temperature)
FN  System noise factor for receiver (referenced to antenna port)
BN  Noise bandwidth at antenna port

TX  To receiver

125

Noise Sources
• Thermal emissions from the scene to which the antenna is pointed
• Electronic noise in the radar component hardware
• Quantization noise due to the ADC
• Any additional purposeful noise sources used to perhaps ‘dither’ the
ADC data

Not included is “multiplicative” noise due to nonlinear nature

of radar, integrated sidelobe energy, spurious signals, etc.

126

63
 10/14/2017

Signal to Noise Ratio (SNR) at RX Antenna Port

An important measure of goodness is the ratio of power (energy) of
signal to noise.

Pr PT G A Ae
SNRantenna  
N r  4 2 R 4 Lradar Latmos  kTFN  BN

This is a generic form of the Radar Equation that

is true for all monostatic radar modes.

127

SNR in SAR Image

The SNR can often be improved by signal processing in a manner to “match” the
data to our transmitted signal(s). This involves pulse compression and coherently
combining multiple pulses, i.e. SAR processing.

PT G A Ae Gr Ga
SNRimage  SNRantennaGr Ga  2
 4  R Lradar Latmos  kTFN  BN
4
where
Gr  SNR gain due to range processing (pulse compression)

Ga  SNR gain due to azimuth (Doppler) processing (coherent pulse integration)

The product GrGa comprises the signal processing gain

TX  DSP SAR image

128

64
 10/14/2017

The Transmitter
The transmitter generally is constrained by 3 main criteria

1. frequency (wavelength) of operation, including bandwidth,

2. peak power output, and
3. maximum duty factor allowed.

We relate
PT TTX f p  PT d  Pavg  Average TX power during synthetic aperture
Courtesy of Sandia National Laboratories, Airborne ISR
where
TTX  TX pulse width
f p  Radar Pulse Repetition Frequency
(PRF)
d  TX duty factor

129

Survey of TX Power Amplifier Tubes

Power Amplifier Tubes

Higher peak power usually means lower duty factor

Higher power also usually means narrower bandwidth.

Solid‐state power amplifiers are generally lower‐power than their tube counterparts,
typically under 100 W, and more like 10 W to 20 W range (depending on frequency band).
However they do offer a possible efficiency advantage, and technology is advancing to the
point where these should be considered for relatively short‐range radar applications.

130

65
 10/14/2017

Antenna Details
For a monostatic radar, TX antenna “gain” and RX antenna “effective area” are
related to each other by Courtesy of Sandia National Laboratories, Airborne ISR

4 Ae
GA 
2
where

  Nominal wavelength of the radar

The “effective area” is related to the physical area of the

antenna aperture by an efficiency factor

Ae   ap AA
where
AA  Physical area of antenna aperture
ap  Aperture efficiency of the antenna For a dish antenna, the
(typically on order of 0.5) physical aperture is the
silhouette area of the dish
131

Active Electronic Steered Array (AESA) Antenna

While we have thus far explored traditional corporate fed antennas, many newer
systems are using AESA antennas.

While very promising and attractive on a number of fronts, some care needs to
nevertheless be exercised when considering AESA antennas for SAR application

• AESA have gain and beamwidths that are squint‐angle‐dependent

• Wideband waveforms require True‐Time Delay (TTD) steering, or equivalent. Phase

shifters are inadequate.

• AESA beam‐steering does not have the field of regard that a gimballed antenna has.

• AESA antennas still tend to be more expensive than more traditional corporate‐fed
gimballed antennas.

AESA antennas that use Digital Beamforming (DBF)

techniques on both transmit and receive offer the potential
of overcoming the wideband operation limitation. They
132 essentially use DSP techniques to implement TTD steering.

66
 10/14/2017

Signal Bandwidth
The signal bandwidth determines the achievable resolution of the
range‐compressed signal.

awr c
BT 
2r
where
c  Velocity of propagation
 r  Slant‐range resolution
awr  Range IPR broadening factor due to data
tapering (windowing) for sidelobe control

A fundamental rule from Signal Processing is that the Power

Spectral Density of a waveform is the Fourier Transform of
the Autocorrelation function.

The Autocorrelation function is the output of a matched

filter when the input is the signal to which it is matched.
133

Processing Gain – Range Compression

The range processing gain is due to noise bandwidth reduction during
the course of pulse compression.

T B
Gr  TX N
Lr
where

Lr  Reduction (loss) in SNR gain due to

non‐ideal range filtering; a result of
using a window taper function for
sidelobe control

This gain is based on matched filter performance.

Note that the gain is essentially a time‐bandwidth product.

However it is the input noise bandwidth that is important, not
the signal bandwidth. The typical presumption is that the input
noise bandwidth is wider than the signal bandwidth.
134

67
 10/14/2017

Processing Gain – Pulse Integration

The pulse‐integration processing gain is about coherent summation of
multiple pulses.

N f p  Rawa
Ga  
La 2  a vx La
where

N  The total number of pulses integrated

vx  Radar velocity component, horizontal and
normal to the direction to SRP
awa  Azimuth IPR broadening factor due to data
tapering (windowing) for sidelobe control
La  Reduction (loss) in SNR gain due to non‐ideal
azimuth filtering; a result of using a window
taper function for sidelobe control

This gain is based on matched filter performance

135

Taylor Window
Window functions are about
‘detuning’ the filters to get some
better sidelobe responses.

The downsides are

1. is slightly worse SNR
performance, because after all
the net filter isn’t precisely
matched, and
2. Slightly broadened IPR width,
i.e. slightly worse resolution.

One popular window for SAR

processing is the Taylor window,
for which some parameters must
also be specified.

awa, awr

Lr, La
136

68
 10/14/2017

Radar Cross Section (RCS)

Even for simple targets, a variety of frequency dependencies exist.

RCS frequency dependence

f2

Courtesy of Sandia National Laboratories, Airborne ISR

137

Radar Cross Section (RCS)

Size matters…
10
RCS of a sphere as a
function of the
RCS / silhouette area (dB)

circumference 0
normalized by the
Rayleigh Optical
signal wavelength.
region Resonance region
10 region
In fact, for very small
targets, shape
doesn’t matter.
20

30
0.1 1 10
Circumference / wavelength

138

69
 10/14/2017

Radar Cross Section (RCS)

SAR usually is interested in the RCS of “distributed” clutter, which is
resolution dependent.
 r 
   0 a  
 cos g
 
where

 0  Distributed clutter reflectivity, Distributed clutter is usually

measured in terms of RCS per m2 modelled as a Gaussian‐distributed
random process. The characteristic
Frequency dependence is typically reflectivity is the variance of the
characterized by associated random variable.
n
 f 
 0   0,ref  

 f ref 
where f is the nominal frequency of interest, and
0,ref is the reflectivity at reference frequency fref.

139

Typical Clutter Reflectivity Values

Reflectivity 0 values typically depend on frequency, grazing angle, polarization, etc.

Typical values at Ku‐band (16.7 GHz) might be

5 to 10 dBsm/m2 for urban areas or rocky areas

10 to 15 dBsm/m2 for cropland or forest areas

15 to 20 dBsm/m2 for grasslands

20 to 30 dBsm/m2 for desert areas or road surfaces

140

70
 10/14/2017

Geometry Effects
Typically, the radar is specified to operate at a particular height above the ground.
Consequently, grazing angle depends on this height, and the slant‐range of operation.
For a flat earth this is calculated as
2
h h
sin g  cos g  1   
R R
where

h = height of the radar above the target

Note that 0 also generally has a dependency on

grazing angle g. This is sometimes embodied in
a “constant‐gamma” model for clutter. In this
case the reflectivity is modelled as

 0   sin g
where
Typical effects of grazing angle on clutter
  Clutter “gamma” reflectivity
141

Miscellaneous Losses
• Signal Processing Losses
– Range & azimuth processing losses due to window functions
[typically 2‐3 dB or so]
– Straddling losses (target smeared across several pixels)
[typically ignored in SAR]
• Radar Losses
– Radar plumbing (between TX amplifier and TX antenna)
[typically 1‐2 dB or so]
– RX signal path losses often included in system noise factor
(figure) [typically 1‐2 dB or so]
– Radome [typically 0.5 dB or so]
• Atmospheric Losses
– Due to the less‐than‐clear atmosphere
• Worse losses in adverse weather
• Very frequency dependent
142

71
 10/14/2017

Atmospheric Losses
We identify the overall atmospheric loss as
R
Latmos  10 10

where
 = two‐way atmospheric loss rate in dB per unit distance

Atmospheric loss‐rates are very

nonlinear with frequency;
generally getting worse at higher
frequency, with notable peaks at
23 GHz (water absorption line) W
and about 60 GHz (oxygen Ka
absorption lines). Ku
X

There are some “windows” in the transmission

spectrum where we find favored radar bands.
143

Atmospheric Losses
There are some “windows” in the transmission spectrum where we find favored
radar bands.

100
H20 H20
02
Attenuation (dB/km)

10
02

1
H20
W
0.1
Ka
Ku
X
0.01
10 20 30 40 50 70 100 200 300 400
Frequency (GHz)
144

72
 10/14/2017

Atmospheric Absorption

145

Atmospheric Absorption

146

73
 10/14/2017

Atmospheric Absorption

147

Frequency Dependence
The optimal radar band
depends on operating
conditions, including
geometry and weather
conditions.

Optimum radar band as a function of range and altitude,

assuming 4 mm/Hr rain, n=1, and constant antenna aperture area.
148

74
 10/14/2017

Putting it All Together

There are a number of ways to write the Radar Equation.
Some of these include
n
 f 
 
2
Pavg ap AA  r 0,ref   f awa
SNRimage   f ref 
 2 R 
h
8 cvx  kTFN  Lradar Lr La   R3 1    10 10 
 R 
 
or perhaps
 2 3  r 

 Pavg G A    
cos g  0,ref
n
   1   awr awa   f 
SNRimage  3 3  R   
 2  4  R vx  kTFN  awr  L 10 10
  Lr La   f ref 
   radar 
 

Each has its own utility.

149

Comments
• SNR does not depend on azimuth resolution

• PRF can be traded for pulse‐width to keep Pavg constant

• For constant ground‐range resolution, there is no SNR overt

dependence on grazing angle, although 0,ref may itself exhibit
some dependence on grazing angle as previously discussed, and
atmospheric loss depends on height and range.

• Input noise bandwidth BN has no direct effect on ultimate image

SNR. Signal bandwidth does not explicitly impact SNR directly, but
rather through a somewhat looser dependence on range
resolution and perhaps window loss.

• The expressions in the square brackets are typically nearly unity,

or at least often presumed to be so, and so are often ignored. If
so, then processing losses should not be double‐counted
elsewhere.

150

75
 10/14/2017

Range Performance
At constant grazing angle and neglecting atmospheric losses, SAR image SNR depends
essentially on R3. This is because we allow the synthetic aperture to grow with range
to keep azimuth resolution constant.

However, within any one image formed from a constant‐length synthetic aperture, the
SNR remains dependent on R4.

image R3 dependence image

R4 dependence
R1 R2 R1 R2

Synthetic aperture

Synthetic aperture
151

Noise Equivalent Reflectivity (NER)

We define an entity that answers the question
“What equivalent clutter level does the noise look like?”

This goes by several names

• Sigma‐N
• Sigma‐Noise
• Noise Equivalent Reflectivity (NER)
• Noise Equivalent Sigma 0 (NES0)

3
0 2  4  R3vx cos g  kTFN  Lradar Latmos Lr La
N  
SNRimage Pavg G A2  3  r awa

In fact, the NER is the clutter reflectivity for which SNR goes to 0 dB.

 N   0 SNR
image 0 dB A typical minimum acceptable NER for
X‐band and Ku‐band is 25 dBsm/m2,
‐‐ lower for lower frequencies.
152

76
 10/14/2017

Courtesy of Sandia National Laboratories, Airborne ISR

‐30 ‐25
‐20 ‐15

NER Dependencies
We can write the NER equation as

 256 3kT  3  
 N  
 c

 R vx cos g

  BT FPN Lradar
2 3
G 
Latmos  Lr La 

  awr awa 
 avg A 

Radar operating Radar hardware Radar signal

Constants
geometry limitations processing

Some of these we can control, and some of

them we can’t control.

We can probably argue some of these.

154

77
 10/14/2017

Radar PRF
The Doppler bandwidth of the data depends on frequency (wavelength)
and antenna beamwidth.

2
BDoppler  vx az

where

 az  Antenna nominal azimuth beamwidth (presumed to be small)

We need a PRF that sufficiently oversamples the Doppler bandwidth.

f p  ka BDoppler

where

ka  Doppler oversample factor, typically about 1.5

(Accounts for antenna beam roll‐off)

155

Extending SAR Range

• Increase average TX power
• Increase Antenna area
– and/or efficiency
• Select better operating geometry
– Fly higher
• Operate at coarser resolution
• Select lower‐loss operating frequency
• Decrease velocity
– Squint mode vx  vaircraft cos  pitch sin  squint
• Decreasing Radar Losses,
Signal Processing Losses, and
System Noise Factor
• Easing Weather Requirements
• Changing Reference Reflectivity
156

78
 10/14/2017

Geometry Limits vs. Resolution

157

Blake Chart
num den

A tabular form of the 256*pi^3 7937.607 38.9969

Radar Equation is often k J/K 1.38E‐23 ‐228.601
T K 290 24.62398
referred to as a “Blake c m/s 3.00E+08 8.48E+01
Chart.”
Slant Range km 2.65E+01 1.33E+02
velocity m/s 36 15.56303
Some sources reserve altitude kft 2.50E+01
this term for the Radar grazing angle deg 16.68589 ‐0.18683

Equation rearranged to resolution m 0.1016

solve for maximum aw 1.1822
signal bandwidth Hz 1.74E+09 92.41584
range. noise figure dB 4 4
Radar losses dB 2 2
Atmospheric loss rate dB/km 0.2 5.31E+00
Ppeak W 320 25.0515
Antenna gain dBi 30.4 60.8
duty factor % 25% ‐6.0206
frequency GHz 16.7 ‐5.24E+01

Lr 0.92 0.92
La 0.92 0.92
awr*awa 1.45E+00

sigma_noise (NER) ‐25

158 dBsm/m2

79
 10/14/2017

A Note About Dynamic Range

Consider a SAR operating at shallow grazing angles at 0.1 m resolution,
with a NER of 35 dBsm/m2, very common for high‐fidelity images.

This equates to a noise floor with pixel values corresponding to about 55 dBsm.

Note that even in rural and suburban clutter, data indicates that we may expect
specular targets exhibiting RCS greater than +45 dBsm at least once per square
mile.

This suggests that the SAR image may easily render with more than 100 dB of
dynamic range, and indeed will be required to do so for some exploitation
techniques.

Consequently, SAR images are often

displayed with Dynamic Range compression
techniques, like Look‐Up Tables (LUTs). Note that the Human Visual
System can only perceive
A popular function is square‐root of magnitude. about 42 dB or so of dynamic
range in any single image.
159

Section Summary
• A common measure for image ‘goodness’ with respect
to SNR is the Noise‐Equivalent Reflectivity (NER)
– The clutter level that yields SNR of 0 dB

• The ‘range’ limit for SAR is often calculated when the

NER raises to 25 dB (for X and Ku bands)
– However images can still be useful with more noise than this

• Optimum radar frequency is range dependent

– Longer ranges favor lower frequencies for better penetration
– Shorter ranges favor higher frequencies for higher gains

160

80
 10/14/2017

Select References
• Performance Limits for Synthetic Aperture Radar –
second edition
– Sandia National Laboratories Report SAND2006‐0821
• Performance Limits for Exo‐Clutter Ground Moving
Target Indicator (GMTI) Radar
– Sandia National Laboratories Report SAND2010‐5844
• Performance Limits for Maritime Inverse Synthetic
Aperture Radar (ISAR)
– Sandia National Laboratories Report SAND2013‐9915
• Noise and Noise Figure for Radar Receivers
– Sandia National Laboratories Report SAND2016‐9649

161

162 http://static.panoramio.com/photos/large/39268735.jpg

81