ECE 6272

Fundamentals of Radar Signal Processing

Lecture #15:
Ambiguity Resolution,
Blind Zones, & Pulse Pair
Processing
School of Electrical and Computer Engineering
Georgia Institute of Technology
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture# 15, Slide 1

ECE6272

Fundamentals of
Radar Signal Processing

Module #50
Staggered PRF in Pulse Doppler
and Blind Zones

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 2

Reminder: Blind Speeds

MTI filters are digital filters,
so frequency response is
periodic

2-Pulse Canceller Frequency Response

2

1.5

H (F )

0.5

Nulls at multiples of PRF Hz

0
-1.5

Because sampling rate in slow time is

PRF

Blind to targets at
corresponding radial velocity:
Doppler coverage = PRF Hz
Could fix by raising PRF, but

We fixed this in MTI by using

pulse-to-pulse staggered PRFs
We cant do this in pulse
Doppler
Why Not?
Fall 2010

vblind
=

-1

0.5
1
-0.5
0
Doppler Frequency (multiples ofPRF)

1.5

PRF

c PRF
=
2
2 F0

T1 = 1/PRF1

T2 = 1/PRF2

T1 = 1/PRF1

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 3

Staggered PRFs for Pulse Doppler

0
M-1
slow time (pulse #)

L-1

fast time (range bin #)

Independent FFT on each row

fast time (range bin #)

The first step in pulse Doppler

is to convert to range-Doppler
with a slow-time DFT
DFT requires data sampled
at a constant sampling
interval (PRI in slow time)

L-1

0
K-1
frequency (Doppler) bin

So MTI-style staggered PRF is out

Instead, we stagger the PRF on a CPI-to-CPI basis

Constant within a CPI DFT still meaningful
Different from one CPI to the next
Alters the spacing of clutter replicas in the spectrum
Data on different CPIs is on different Doppler frequency scales

Detection from multiple CPIs combined

Common in airborne pulse-Doppler systems
Expensive in time, computation, and energy
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 4

Blind Speeds in CPI-to-CPI Stagger

Blind @
PRF #1

replication of
clutter

Two or more CPIs

transmitted and aliases of
moving target
processed
noise
In this example,
PRF2 < PRF1
Typically repeat
3PRF1
PRF1

for 3 to 8 PRFs
2
Values of F that are
obscured by clutter
are called Doppler
blind zones
Intervals of PRF Hz
Therefore different
for different PRFs

Fall 2010

moving
target

clutter

PRF1
2

PRF1
2

+ PRF1 + 3PRF1
2

Not blind
@ PRF #2
F

3PRF2
2

PRF2 PRF2
2

Copyright Mark A. Richards, All Rights Reserved.

PRF2
2

+ PRF2 + 3PRF2
2
Lecture#15, Slide 5

Range Blind Zones: Eclipsing

Monostatic radar cannot receive while a
pulse is being transmitted, an effect called
eclipsing
Periodic blind zones in range at all Doppler
shifts result
Transmitted
pulse

0
receiver on

Fall 2010

T
cT/2

2T
cT

time,
range

receiver on

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 6

Near-In Clutter Eclipsing

Strong near-in clutter may lengthen the
effective blind zone in range
e.g., altitude line
Transmitted
pulse

T
cT/2
receiver on

Fall 2010

2T
cT

time,
range

receiver on

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 7

Blind Zone Map

Transmitted
pulse

1
T

T
cT/2

T=

1
PRF

2T
cT
time,
range

Fall 2010

Changing PRF
expands the map in
one dimension,
contracts it in the
other

PRF =

receiver on

Blind zone map

receiver on

Eclipsing, near-end
clutter, and blind
speeds combine to
produce a 2-D
pattern of ranges and
velocities where
targets are
undetectable

F
PRF

Copyright Mark A. Richards, All Rights Reserved.

PRF
2

PRF
2

+ PRF

Lecture#15, Slide 8

Parameters for Example

Blind Zone Map
Assume pulse compression waveform with = 1
MHz, = 11
range resolution = range bin spacing = c/2 = 150 m
uncompressed pulse is 11 s long = 11 range bins

Assume PRF = 10.4 kHz

so PRI = 96 s = 96 range bins

Assume 104 Doppler filters

either explicit bandpass filters, or 104-point DFT
so Doppler bins are 100 Hz wide

Assume two-sided clutter spectrum width is 3.4 kHz

34 Doppler bins
Relatively fat clutter spectrum (~1/3 of spectrum width)
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 9

Blind Zone Map for 1 PRF

No near-in
clutter
eclipsing
considered
96
cells

Positive
Doppler
frequencies
only shown
17
cells

Slide courtesy of N. Levanon

Fall 2010

11
cells

34
cells

PRF=10.4 kHz
Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 10

M of N Blind Zone Map

Goal is to design a set of staggered PRFs
that collectively provide a large space in
range and Doppler where detection is not
prevented by eclipsing or clutter
Usually based on multiple PRFs with an M
out of N detection logic
if the target is eclipsed or covered by clutter in 1 or
2 PRFs, it will hopefully be in the clear on other
PRFs
8 PRFs with detection on at least 3 is common in
multimode airborne radar
Mainly in medium PRF modes, where ambiguities
in both range and Doppler are expected
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 11

Example: 1 of 2 Detection - 1

Fall 2010

10 s pulse length
100 Hz Doppler resolution

20 m/s clutter spread

Eclipsing due to pulse length only

Blind zone map for PRI=120 s

Blind zone map for PRI=100 s

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 12

Example: 1 of 2 Detection - 2
Combined blind zone map

Regions that are clear on at

least 1 of 2 PRIs

Note first pulse duration and DC clutter are always blind

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 13

BlindZone GUI

blindzone.zip available on T-Square in Extra Goodies section

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 14

Parameters for Blind Zone Example

Fc

=
10
GHz,
3 cm

10.4 KHz PRF 20 KHz

PRI / t p

{50,51,52,....,95,96}

1 s range cell =
150 m
tp =
=
=
Number
of range cells
1000,
max range 150 km
Doppler resolution =
100 Hz velocity cell =
1.5 m/s
Number=
of velocity cells
200,
=
max velocity 300 m/s
Clutter notch = 17 velocity cells
Pulse eclipsing = 1 range cell
Near clutter masking = 10 range cells

8 different PRIs are

used. A cell in the
blind-zone map is
considered clear if it is
not blind on 3 or more
PRIs
3 out of 8 detection

Design question:
Which set of 8 PRIs
yields the clearest
blind zone map?

Slide courtesy of N. Levanon

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 15

Typical 3 of 8 Blind Zone Map

PRI = { 50 53 55 58 61 64 68 72 } s

Slide courtesy of N. Levanon

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 16

Genetic Algorithm Blind Zone Map

PRI = {51 53 60 63 67 84 89 93 } s

Davis, P.G. and Hughes, E.J.

Medium PRF Set Selection Using
Evolutionary Algorithms, IEEE
Trans on AES, vol. 38(3), pp. 933939, July 2002.
Slide courtesy of N. Levanon

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 17

ECE6272

Fundamentals of
Radar Signal Processing

Module #51
Ambiguity Resolution

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 18

Ambiguous Range of Targets

Target range Rt > unambiguous range Rua
Once steady state reached, target appears at a shorter,
ambiguous apparent range Ra after every pulse
2 Rt c
2 Rua c

pulse #1

t
target from pulse #1

pulse #2

t
target from pulse #2

2 Ra c

t
pulse #4

target from pulse #3

t
pulse #5
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 19

Ambiguous Range of Targets

Target range Rt > unambiguous range Rua
Once steady state reached, target appears at a shorter,
ambiguous apparent range Ra after every pulse
2 Rt c
2 Rua c
pulse #1

Fall 2010

pulse #2

2 Ra c
pulse #3

target from pulse #1

target from pulse #2
target from pulse #3
pulse #4

pulse #5

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 20

Ambiguous Range Formula

If true range Rt > Rua, then apparent range is

Ra = ( ( Rt ) )

Rua

(())x notation means modulo x

Equivalently, given the measurement Ra and

known ambiguous range Rua,

R=
Ra + kRua for some k
t
Normalizing to the range bin spacing R,
Rua
Rt
, nt =
, etc.)
nt =
na + kN for some k (N =
R
R
Equivalently, na = ( ( nt ) )
N

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 21

Range with Multiple PRFs

Number of range bins in PRF #i is Ni (i=0,1,):
so Rua= N i R
i
If we use multiple PRFs (thus multiple values
of Rua):
assume range bins are the same size for each PRF

nt =
na0 + k0 N 0 =
na1 + k1N1 =
Equivalently,

( )N

nai = ( nt )
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 22

Aside: Chinese Remainder Theorem

Given pairwise relatively prime integers
N0, N1, ,Nr-1 and the system of equations (congruences)

, , na
( nt )) N , na (( nt ))=
(=
(( nt )) N
N

na0

r 1

r 1

There is a unique solution for nt (modulo N = N0N1Nr-1)

given by
nt
=
=

(( k n

( ( n

0 0 a0

0 a0

+ k11na1 + + kr 1 r 1nar 1

+ 1na1 + + r 1nar 1
r 1

(( ))

1
,

=
ki =
N Ni =
N
k
j i
i
=j 0, j i

Fall 2010

) ) N ,

) ) N

i ki i

Ni

( ( i ki ) )

Copyright Mark A. Richards, All Rights Reserved.

Ni

1
=

Lecture#15, Slide 23

Solution Using the Chinese

Remainder Theorem
For 3 PRFs, to make it more specific:
and because 3 PRFs is a common choice in
airborne radars for ambiguity resolution

nt=

( ( n

0 a0

+ 1na1 + 2na2

))N N N
0

=
i =
i N j
i ki
j =0
j i

1
( e.g.,=

1N 0 N 2 )

i are smallest integers such that

=
( ( 0 N1N 2 ) ) 1,=
( ( 1N0 N 2 ) ) 1,=
( ( 2 N0 N1 ) ) 1
N0

Fall 2010

N1

Copyright Mark A. Richards, All Rights Reserved.

N2

Lecture#15, Slide 24

CRT Example

True range cell # = nt = 19

# range cells per PRF N0=11, N1=12, N2=13,
=
na0 8,=
na1 7,=
na2 6
0 = 6, 1 = 11, 2 = 7
1
e.g., 0 satisfies ( ( 0 12 13) )11 =
=
=
check: ( ( 6 12 13
) )11 ( ( 936
) )11 1
0 = 936, 1 = 1573, 2 = 924
e.g., 0 = 61213 = 936

nt=
Fall 2010

( ( n

0 a0

+ 1na1 + 2na2

))N N N =
1

19

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 25

CRT Problems
Measurement errors in data can cause
very large range errors
Same CRT example, but suppose we
erroneous!
na0 8,=
na1 7,=
na2 7
measure =
Then it turns out nt = 943!
instead of 19

A number of algorithms exist in the

literature to solve the problem in the
presence of errors
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 26

Example Graphical Equivalent of CRT - 1

True ranges Rt corresponds to range bin #s 6 and 11

Measure with a PRF corresponding to 7 range cells per

PRI
Then the data will appear like this:

Individual pulse data

PRI #1

PRI #2

PRI #3

PRI #4

PRI #5

Sum of all responses

(actual detection data)

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 27

Example Graphical Equivalent of CRT - 2

True ranges Rt corresponds to range bin #s 6 and 11

Assume three PRFs with, respectively, 7, 8, and 9 range

cells per PRF
Then, using process similar to previous page, the targets
will appear (in steady state), respectively, as follows:

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 28

Measurements
True cells= #6 (unambiguous), #11 (ambiguous) result in
these measurements when N0 = 7, N1 = 8, N2 = 9

Extend by concatenation and find coincidences:

#6
Fall 2010

#11
Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 29

Other Algorithms
The graphical algorithm will also fail
given a measurement error
But it serves as the basis for more
sophisticated approaches based on
finding the tightest clusters of
measurements
There are also techniques based on
coding theory that seem clever

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 30

Why Dont We Just Figure It Out From

the Pulse-by-Pulse Detection Pattern?

Individual pulse data

In other words, pay

attention during the
build-up to steady
state:

PRI #1

PRI #2

PRI #3

PRI #4

PRI #5

Sum of all responses

(actual detection data)

We rarely make the basic detection test on single, individual pulses

Instead, we integrate multiple pulses to improve SIR and then do
detection test

pulse Doppler DFTs

Noncoherent integration, or
Coherent integration (this include Doppler processing), or
M-of-N detection processing

So we mix several PRIs together before we make the detection test

Measurements in example would really be detections on a full CPI at the stated PRIs,
not just one pulse at each PRI
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 31

Ghosts
In general, need N+1 PRFs to resolve N targets
Insufficient number of PRFs can lead to
ghosting
Same example using only 1st and 3rd PRF:
target #2
target #1

#6

#11

#20
ghost!

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 32

Resolving Velocity Ambiguity

Its the same problem
Doppler shift is measured modulo the PRF
DFT size and PRF establish the size of the
Doppler (velocity) bins

Some techniques published based on

Nyquist sampling/reconstruction theory
usually put constraints on the PRFs that
can be used

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 33

ECE6272

Fundamentals of
Radar Signal Processing

Module #52
Pulse Pair Processing

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 34

Pulse Pair Processing

Technique used widely in
meteorological radar for estimating
echo
power
mean velocity
spectral width

These measures used in turn in

algorithms for detecting severe
weather, accumulated rainfall, etc.
Often referred to as PPP
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 35

PPP Spectral Model

PPP assumes the Doppler spectrum of the
data for a given range consists of a single
moving target component, plus a noise
floor
no clutter
no multiple targets
Usually (always?) assume Gaussian shape
Sy(F)
area =
power

Sn(F)

spectrum
width
F

Sw(F)

white noise floor

F0
Fall 2010

mean velocity

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 36

Autocorrelation Function and

Power Spectrum
Given a finite sequence of complex data samples
from a given range, y[m], m (0,,M-1), the
(deterministic) autocorrelation function is defined
as
M k 1

sy [k ]

m =0

y [ m] y [ m + k ]

Its Fourier transform is

=
S y ( ) F=
{s y [ k ]} Y ( )
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 37

Estimating Power
In the time domain:
Px s=
=
y [ 0]

M 1

M 1

[ m] y [ m]
y [ m] y =

m 0=
m 0
=

this is really the energy in the finitelength signal

dividing energy by time duration
would convert it to power
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 38

Mean Frequency, Time Domain

Suppose the signal is an ideal complex
sinusoid, and consider sy[1]:
y [ m ] = Ae j 2 F0Tm
s y [1]

M 2

m =0
M 2

y [ m ] y [ m + 1]
j 2 F0T ( m +1)

Ae
Ae
=

+ j 2 F0Tm

M 2

j 2 F0T

m 0=
m 0
M 2
2 j 2 F0T
2
j 2 F0T

A e

=
(1)

( M 1) e

m =0
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 39

Mean Frequency, continued

s y [1]
=

M 2

j 2 F0T

e =

j 2 F0T
M
1
e

(
)

m =0

amount of phase
rotation per sample

By inspection, we can obtain

arg s y [1]
F0 =
2 T

The sum over m doesnt appear to contribute

anything
but in practice it provides some averaging of noise
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 40

PPP Mean Estimator

1

F0 =
arg s y [1]
2 T

This is the pulse pair processor for estimating

mean frequency
works for more general signals than the pure
sinusoid, provided there is a single dominant
frequency component with adequate SNR
assumes no aliasing
of gives you aliased frequency

Multiply by /2 to convert to velocity units

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 41

Gaussian Spectrum
Work with continuous-time case
Assume that Doppler mean is estimated and
removed from data to give new zero-Doppler
data sequence y'(t) (sampled version y'[m])
Assume spectral energy is Gaussian shaped:

S y ( F ) =

2 F

F 2 2 F2

Then so is the continuous-time autocorrelation

function:
2 2 2 F2 z 2

s y ( z ) = A e
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 42

Sampled Version
Sampled autocorrelation function:

=
s y [ k ] s=
y ( z ) z = kT

2 2 2 F2 k 2T 2

A e

Corresponding DTFT, still in units of Hz:

S y ( F ) =
Fall 2010

2 F T

F 2 2 F2

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 43

PPP Spectral Width Estimator

Assuming a Gaussian spectrum, we can
estimate its variance from the
autocorrelation lags:

s y [1]
2
F

Fall 2010

A
e
s y [ 0 ] e
=
2 2 2 F2 T 2

2 2 F2 T 2

s y [1]
= 2 2 ln

2 T
s y [ 0]

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 44

Simplified Width Estimator

Use a series expansion of ln{ } to get
rid of the logarithm calculation:
x 1 1 x 1 1 x 1
ln x = +
+
+
x
2 x 3 x
x 1
1
1

=
x
x

Apply to width
0
s
[
]
1

y
2
2 2 1
F =

estimator:
2

2 T

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

s y [1]

Lecture#15, Slide 45

Frequency-Domain PPP
There are equivalent estimators for
power, mean frequency, and spectrum
width that operate on the power
spectrum instead of the autocorrelation
function
See text for details

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 46

NEXRAD Processing Algorithms

The WSR-88D
(NEXRAD) radar of the
National Weather Service
uses pulse pair processing
for power and spectral
moments:
power: average between 24
and 512 samples, 1 dB
accuracy
velocity: 40 to 200 samples,
v ~= 4 m/s
spectral width: 40 to 200
samples, v ~= 4 m/s
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 47

Nice NEXRAD Display

Reflectivity (power)
Fall 2010

Velocity

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 48

Tornado Signatures

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 49

Hurricane Katrina reflectivity 8/29/05 08:49:26 CDT

NEXRAD data, KLIX radar (New Orleans)

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 50

Hurricane Katrina velocity 8/29/05 08:49:26 CDT

NEXRAD data, KLIX radar (New Orleans)

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 51

Hurricane Katrina spectral width 8/29/05 08:49:26 CDT

NEXRAD data, KLIX radar (New Orleans)

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 52

Doppler Stuff I
Havent Covered
Read the book for
Clutter mapping and the moving target
detector for air surveillance, air traffic
control
MTI and pulse Doppler transients
Non-DFT Doppler spectrum estimators

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 53

Assignments
Reading:
6.1 6.2

Homework
HW #5
DL: due Tuesday 10/19

HW #6
On-campus: due Thursday 10/21
DL: due Thursday 10/28

Computer Project #3: Pulse Doppler processing

On-campus students: due Tuesday 10/26
DL students: due Tuesday 11/2

Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture#15, Slide 54

End of Lecture 13

Next:
Detection
Fall 2010

Copyright Mark A. Richards, All Rights Reserved.

Lecture# 15, Slide 55