AD-7'4.

829

FM/CW RADAR SIGNALS AND DIGITAL

PROCESSING
Donald E. Barriok
National Ousanlo and Atmospheric
Administration
Boulder, Colorado
July 1973

DISTRIBUTED BY:

No"lu T10whlu m"tl SOMN

U.S.DEPARTMENT OF CMMERCE
5285 Port Royal Road, Spdnlefiltd Va. 22151
 ENVIRONMENTAL RESEARCH LABORATORIES

The mis~ion of the Environmental Research Laboratories isto study the ocean , inland
water%, the lower and upper atmosphere, the ipece environment, and the earth, in search of the
undqr~ tandinio nee.ded to provide more useful services in imorovino man's prospects for strvival
as I n luenced by the physical environment, Lahoratories contributing to these studies are:
N~orth i~~oea La~nr~ pda (00.SL) eismology, Qeomaqnetism, geodesy, and related earth
sciences% earthquake procesges, Internal structure and shape of the earth and aistribution of
the earth's mass.
and Noto *vloU~il Wbortore,'v (AtWL),: Cvtoloov and qeoohysics
At lant u' LV\ opmjr'o~h,
of ocean basins, oceanic orocetsss and sea-air interactions (Miami, Florida),
mop~ QceanoohiDo tbtNr'.itorioma k'pol.': Oceanoora~h with emlnhO-t on the or-4
processes and dynamlesi tiunani generation, propaqation, modification, detection, and monitoring
(Seattle, Wahinqon).
Atmephorv'o Phpoiec' md Chomitt Wl-vrtory A'~) Processes of cloud and procipitation
the lower atmosphere-, and laboratory
Physics,, chemical composition and nuclea ting substance& isi
and field exptriments toward developing feasible ,nehods of weather hiodification.
Ai~r Ro'o~o t~hmee (ARL). Diffusion, tran~vnort, and dissipation of atmospheric
contavrinants, dcvvin ui motliuds for prediction and control of atiiospheric pollution, geo-
physical monitoring for climatic change (Silver Spring, Maryland).

('00phyeia Fluid In~vrtioa IAIve~apr (SrrIP1J Dynamic, and plhysics of geophysical fluid
systems; development of a toretical basis, through mathematica, Modeli ng and computer Jerley),
simula-
tion, for the behavior and properties of the dtm~oiphere and the o ceans (Princeton, Now
Natio'nal 9*e'r* .tomwi".b.or'titry (A'S0f), Tornadoes, squall line%, thunderstorms, and
other severe local convective phenomena directed toward Improved method% of prediction an(;
detection (Norman, Oklahoma).
spaoq' Atnuircni'nt Lahoritorij (Sirt). Solar-terrestriial physics, service and technique
developient in the areas of environmental monitoring and forecasting,
Aeroptmy1 Labo-tory (A0.- Theoretical, laboratoryv, rocket, and satellite studies of
the physical and chemical Processes controlling the ionosphere and exosphere of the earth and
other planets. and of the dynamics of their interections with hinh altitude meteorology.
Nave Pi'paqationi La, ritwry (mr'): Developmt_, of new methods for remote sensinog of
the gohysical environment with special emphasis on optical, microwave and acoustic sensing

Wcathie* Nodtfoat.,v *rm' lffiov (WWfF : Planib and directs ERL weather modification
activities, operats ERJ41crf fleet, and research on cumulu3 cloud modification, on hurricanes
t her
_tcci~vtrppt ems, and on hurricane modification.

I' .11NATIONAL OCEANIC AND ATMOSPHERIC ADMINISTRATION

*1

By

010, AI., . BOULDER, COLORADO 80302

L 7.0-
 A - 77'1 8 ,?

0J, %S, DEPARTMENT OF COMMERCE

A
P
Dent, Secretory
Fredsrick B.

~I NAtIONAL OCEANIC AND ATMOSPHERIC ADMINISTRATION

Robert M. White, Administrator
1
d oENVIRONMENTAL RESEARCH LABORATORIES
Wilmot N. Hess, Director

NOAA TECHNICAL REPORT ERL 283-WPL 26

FM/CW Radar Signals and Digital Processing

DONALD E. BARRICK
Thio work was sponsored In part by the
Defense Advanced Research Projects Agency

BOULDER, COLO,
July 1973
For %ale by the Superintendent of Documents, U. S. Government Printing Office, Woshlngton, 0. C. 20402
SN S J
 NOTICE

The NOAA Environmental Research Laboratories do not approve,

recommend, or endorse any proprietary product or proprietary
material mentioned in this publication. No reference shall
be made to the NOAA Environmental Research Laboratories, or
to this publication furnished by the NOAA Environmental
Research Laboratories, in any advertising or sales promotion
which would indicate or imply that the NOAA Environmental
Research Laboratories approve, recommend, or endorse any
proprietary product or proprietary material mentioned herein,
or which has as its purpose an intent to cause directly or
indirectly the advertised product to be used or purchased
because of this NOAA Environmental Research Laboratories
publication.

I!I
" ,,,~. .. .. . . . .

TABLE OF CONTENTS
Page
ABSTRACT v
1. OBJECTIVE 1

2. APPLICATION 1
3. TRANSMITTED WAVEFORM 1
4. RECEIVED WAVEFORM 2
5. DECHIRPED SIGNAL 3
6. DOUBLE-FFT DIGITAL PROCESSING 7
7. SINGLE FFT DIGITAL PROCESSING 12
8. NUMBER OF COMPUTER OPERATIONS REQUIRED 16
9. WINDOWING AND WEIGHTING 18
10. RULES FOR SIGNAL DESIGN 19
11. SATISFACTION OF REQUIRED ASSUMPTIONS 20
12. SUMMARY 21

13. REFERENCES 22

iti
 ABSTRACT

The use and processing of the FM/CW signal for radar and acoustic
sounder systems are examined in this note. This signal--along with real-
time digital processing via minicomputers--is currently being used by
several groups for HF over-the-horizon radars. A comparative analysis
of the different prccessing techniques for general radar applications
has yet to be undertaken. This note therefore attempts to promulgatc
details of these techniques so that they may find use in other systems.
An example involving an HF backscatter radar is used to permit the reader
to see how the techniques are applied to dn actual problem.

A linearly swept-frequency signal format is used in a 100% duty-

factor mode. In the receiver, a replica of the linear FM signal is mixed
with the received waveform at an offset such that the desired range window
is observed with the lowest possible IF frequency variation. This pulse
train is then analog-to-digital (A/D) converted and ready for computer
processing. Two techniques are described and analyzed for digitally
processing the signal via the Fast-Fourier-Transform (FFT) algorithm.
The first is a double-FFT process; the first FFT set is done within a
pulse-repetition-interval (PRI) to give range information. The next FFT
set is done over N PRIs to give Doppler information. In the second
technique, a single long FFT is used over N PRIs, giving simultaneously
both range and Doppler information. It is shown that both techniques are
identical, in that they produce the same information and require the same
number of computer steps in executing the required FFTs. Both techniques
yield Unanbiguous range and Doppler, for both discrete and distributed
targets; the note shows how and where this information is contained in
the processor output. The note also describes how two weighting functions
are normally applied to the pulse train time samples to reduce objection-
able range and Doppler sidelobes. Finally, simple "cookbook" rules are
given for obtaining the signal and processing parameters based on the radar
and target range/velocity specifications.

PI

Preceding page blan0kv

 FM/CW RADAR SIGNALS AND DIGITAL PROCESSING

Donald E. Barriuk

1. OBJECTIVE
The objective of this note is to present a simple and concise
analysis--backed by an example--of the application of an FM/CW signal
format in radar systems. It is shown how both time-delay (range) and
Doppler (radial velocity) information can be extracted unambiguously.
2. APPLICATION
For the sake of illustration throughout these notes, we pick the
following application and example. The HF radar car;-ier frequency is to
be 10 MHz. Sea scatter is to be observed from the radar out to a range
of 150 km fcorresponding to time delays up to 1 millisecond in a back-
scatter radar). It is known that HF sea scatter is confined spectrally
to frequencies within about 1/3 Hz of the carrier. Therefore a pulse-
repetition-frequency, fr' of 1 per second is selected so that all echo
Dopplers within + 0.5 Hz of the carrier will be displayed unambiguously.
To show sufficient detail, a Doppler processing resolution better than
0.02 Hz is desired, and a range resolution of the order of 1.5 km is
desired; the latter two requirements in an ordinary pulse-Doppler system
translate to a coherent integration time exceeding 50 seconds and a
signal bandwidth of 100 kHz, respectively,
3. TRANSMITTED WAVEFORM
We select a 100% duty factor signal whose frequency sweeps upward,
linearly, over one pulse-repetition-interval Tr (Tr = 1/fr = 1 sec for
our example). Since a 100 kHz signal bandwidth is desired, the signal
can be written

vT(t) - coslwct + ' Gfrt2 ] cos[OT(t)] (1)

for -Tri 2 < t < Tr/2. It is assumed that the signal is periodic, and
hence phase-coherent from one repetition interval to the next.

/
 Since the instantaneous frequency, fT(t), is the derivative of the
phase, we have

fT(t) -- c
d0ft c + Bfrt (2)

where here f* 10 MHz, fr - 1 Hz, and B =100 kHz. Thus it can be

seen that the frequency excursion of fTt over one pulse-repetition
interval is

AfT(t) - B = 100 kHz. (3)

The amplitude of the transmitted signal is taken to be unity. The

plot of signal frequency vs time is shown in figure 1.
4. RECEIVED WAVEFORM

I j fT(t) fRt
IC

f(W f2(t)

Figur'e 1. Frequency ye time of transmitted and delayed/Doppler ahifted

received signal.

2
 I
The received signal is both delayed in time and shifted in Doppler.
To illustrate the situation, we assume that we have a discrete target at
range 15 km and travelling radially away from the radar at v-5 m/s (e.g.,
an ocean wave). At time t-O, the target is exactly at R0-15 km from tie
radar. After that, its range is a function of time as

R(t) - Ro+vt (4)

The received signal from this discrete target is thus just a replica
of the transmitted signal, but multiplied in amplitude by a factor A
and delayed in position by a factor td, where td - 2R(t)/c. It is thus

vR(t) - AvT(t-td) * Acos[w (t-td)+ Bfr(ttd) ] ; (5)

its frequency is shown in figure 1 as the dashed curve.

5. DECHIRPED SIGNAL

Now after RF amplification, we mix the received signal with a replics

of the transmitted signal; this is represented mathematically by sub-
tracting a phase OT(t) from T(t-td) to give a signal

vI(t) - APT(t)cos[wc(t-td) - Wct+rTBfr(t-td)z - nBfrt2 . (6)

There is also a sum signal with phase OT(t)+OT(t-td), but it is

near 2wc (twice the carrier), and hence removed by filtering. The function
PT(t) denotes a pulse of unity amplitude and width T, where here,
T - Tr-t d .
Thus the mixture of the two sawtooth frequency waveforms and their
subtraction, as shown in figure 1, produces a signal whose frequency
format, fY(t), is as shown in figure 2. The two frequtncies 4re

fl I d ~ ; ' -C~) (7a)

and

3
 Id
f2 77iR T(t-td) - T(t+Tr)l (7b)

The intermediate, dechirped signal can be represented as the sum

of two pulse trains aF ehown in figure 2. One, vI(t), is at frequency
fI, and the width of these pulses is T - Tr-td* The other, v2(t), is
at frequency f2, and the width of the pulses is T-td. It will be
possible to eliminate v (t) by filtering if f >>f ;such will be the
22 1
case here.

2R f f

I Tr t

#"*A 1 /f2 4

Figur'e 2. Frequenoy and wVrpitude plots ve time o.freceived signal after

dechirping.
 Therefore, we are left with a single pulse train to analyze, as
represented by equation (6). It is possible to re-certer the time origin
so that it falls in the middle of the first pulse; this is done in
figure 2. The frequency and phase from pulse to pulse are changing very
slightly, however; we will analyze this now.
5.1 -Tr/ 2 < t < T,/2
Let us simplify the phase in the first pulse; denote internal time,t,
within this pulse as t1 . Using t=t i and td=2R/c=2Ro/c+2vt/c=to+2vtl/c,
(where to -2R/c is tie initial delay of the signal), we have
*l(ti)i4T(ti-td)-T(ti), or
t2 f!+frt Bfrtolt I
ltl) [-2-nfcto0+Bfrt ] + 2r[-2 c fc+Bfrto T r
2V M~t2
-2nTfr . (1 -c I " (8)

Thus we have three co-itributions to the phase: a constant, a linear

term in time t i , and a quadratic term in time, t(. For the parameters
of the example, however, the quadratic phase term is always small within
the interval -Tr/2 < ti < Tr!2; e.g., at ti=Tr/ 2 , it is of the order of
0.005 radian. Also, it can b, showri that the second term in the linear
factor is small compared to the irst term and is also much less than one
radian. Of course, in all cases under consideration here, v/c << 1, i.e.,
target velocity is small compared to propagation velocity. Therefore we have

2Y
i(ti ) _ 0 -2r
r[ T- fc+Bfrtolti (9)

hence within the first pulse, the frequency f, is

f, a2.v fc+Bfrto (10)

As can be seen, this frequency offset (also shown in the preceding

figure) consists of two terms: the first due to the target velocity and
the second due to the time delay (or range) to the target (t0 2R0 /cln 1

5
millisecond for Ro=15 kin). For the examiple selected here, the second
term (range term) is larger; i.e., !-fc= Hz, Bfrto=10 Hz. Thus
it is not possible to separato range from target velocity by measuring
frequency f, within a aingZe puZee.
5.2 (2 n-l)Tr/ 2 < t < (2n+l)Tr/2

Here we want to examine the phase in the n-th pulse, assuming that
the n=O pulse is the one centered at t=O. Again, we describe the time
within the n-th pulse (from its own center) as ti . The time delay to
the target, td, however is now given by

td = 2R/c = 2Ro/c+2vt/c = to+2v(nTr+ti)/c , (11)

where we describe time to the center of the n-th pulse as nTr. We can
now substitute this into the phase:

) 4ini
=T(ti-td) - OT(ti)
-. 2v 2 (12)
2vtt - nT + Bfr[to+ - (nTr+ti)]

-
2 rBfr[to+ L (nTr+ti)] •

After expansion and elimination of terms which are small cimpared

to others and also small compared to one radian, we have (assume n- 100)
I(ti ) 2fc-' c fnTr - 2t[ L fc+Bfrto + L Bn~ti ; (13)

hence the frequency in the n-th pulse is the quantity in square orackets,
i.e.,

f uvfc+Bfrt L Bn. (14)

ln 'c c ro0 + c

Comparison of (14) with (10) shows that the frequency in the n-th
pulse is identical to that in the first pulse, with the exception of the

6
third term. The explanation for the third term Is simpl. It merely
means that the target is moving from pulse to pulse, and Its range at the
center of the n-th pulse is RO+c(2v/c)nTr/2 m Re+vnTr , as we would
expect. Since we want to integrate over as many as 100 pulses, the third
term is not negligible as n increases; e.g., at n-lO0, F Bn I Hz
Two other effects occur within the pulse; its width, being TTr-td
changes very slightly from pulse to pulse. Since Tul second,
td M to+(2v/c)nTr, we have for n-l, T - 1-10 " s; for n-lO0 we have
T - I01-4 1 10" S. Thus the change in pulse width is negligible.
A very important second effect, however, is the change in phase from pulse
to pulse, as represented by the second term in (13). This phase change
shall in fact prove to be the basis for the Doppler processing. As stated
earlier, all of this assumes that the transmitted signal Is phase-coherent,
i.e., *T(t+Tr)-T(t) - non-varying consLant.
6. DOUBLE-FFT DIGITAL PROCESSING
Here we want to demonstrate how a double Fourier-transformation
process can be used--often in real time because of the discovery of the
digital fast-Fourier-transform (F-r) algorithm--to produce a time-delay
(range) and Doppler (velocity) display of the radar target data*. The
first Fourier transform process is performed over a pulse repetition
period, Tr (i.e., within a pulse) to obt-in target range. The second
Fourier transform is then performed over several pulses of these data to
obtain target Doppler or velocity.
First, let us perform a Fourier transform on a single pulse. This
is shown in figure 3. We have a pulse of width T = Tr-td , amplitude A,
and frequency f, given by (14). To perform this Fourier transform
digitally, one must sample the pulse M times within the time period
Tr. The number M depends upon the maximum value f, can assume, and
M/Tr must be at least twice this value, i.e., 2fl.x , according to the
Nyquist theorem. For the problem considered earlier where we want to
*This technique is currently being used by the Stanford Research Institute
for recl--time processing of HF ionospheric radar signals at their Wide
Aperture Research Facility (WARF); (Sweeney, et aZ., 1971).

7
dtspl possible termes at all rages frm aero to 160 km, this ~lofld5
to a frVqVncy variation in f from 0 to 100 Hi hence I must be
gweater then MOO since Trl sec. Since VFT !cassors r5Vequir that
"Pik, whe, k is an Intver Not"E Would suffices

vIOtM %.afp, *1 I/T

I In

14- T

PigW. 3. Si~fZ. puts*. OW its Pok'ie~r *N.'X~

The Fourier transform of the pulse is then

T/2 -1i2ff?
vI1(f uf Alcos*1 n(tint it
4/2
or -T1

S1n[2t(f+fln)T/ 2)]+1$012iyfc" P" nTr * (15)

8
 This Fourior transform is show in figure 3, Sinn we started with
*IfibuTr samples (HolOO minimum), we obtain eamples in the freuency
domain from -tie" to +fI%"# i..0at 14/let IM Y, positive values ot fre-
quency, These samples are complex in as evidenced by the exponential
1,nea
phase factor containing *6and 10~ nT1. I Thus we conceptually have
M/1 ran"e bins (M/10lOO here), permitting ut to realize the 151 km range
resolution over a 150 km window, as initiolly stipulated. Note that each
N/a resolution element after tho first FF1' can be consideMe a range bin
so long as the Doppler terms (Iv/c)fco Is small compared to the range
term, Bfrtgj this is true for the example considered here. Since each
pulse is approimately 1/1 wide at the half-power point (T oTr a I5c
here), we should be able to resolve 100 targets in range because the width
of each FF1' pulse in this 100 Ha window is 1 Hs. Hance after one FFT
process within a pulse we have range information, but no Doppler infov'mtioni
we turn now to extraction of Doppler.
Note that If we start with the first pulse at nul and do this FF'
process on each pulse, we obtain a Fourier transform n times, where we
ar~sume n~s N (somt upper value). Since the frequency, fin and phase,
2Wf J nTr. shifts slightly from pulse to pulse due to thrget velocity
(as given In(14)), this slrla/40 pulse in the frequency domain will change
very slightly after each Fourier transformation. Since our digital FF1'
iUcapable of producing numbers at M/2 discrete points, (18) should really
be written with f replaced by i - #1f,. where -M/2.sMN/2.
Thus the first FF1' process on M samples within a pulie gives M/2
range bins for each pulse. For each successive pulse, this FF1' gives M4/2
additional positive frequency samples. Digitally, we store each M4/2
samples Inrows of a matrix, as shown infigure 4,until we have N rows.
Thus, we have an 1/2-by-N matrix whose columns so far represent range
bins.
Now, we perform another FF1' over each column, or range bin. This
will require N points altogether. Each matrix element is a complex
number whose value changes In a column because the frequency, f , and
the phase, 2wfce 4 nT,,, are changing from sweep to sweep. Snce each
of the N vertical elements comes from a different pulse T r Sac apart,

9
NT, sec are required to fill this matrix, Also, n can be related to tim from
the first pulse by use Cf te"T r, or n.t/T r (aglin, 14 nt N). Hence each
col1o is really a function of time, end the N column elements can be
considered (digital) samples of this tin function,
To Fourier transform over a typical column (say the m-th), let us
again refer to our example for the target at R1,06 kmi this target
will appear in the mlO bin for K/2100, As we saw before, this produces
fin . 10 . 0" aM. Thus for n running from I to 100--
corresponding to tim running between I and 100 seconds--two things happen
to the positive pulse in the
m-th range bin t its li - . R g
tude changes slightly due to 4.- Range Pins -.

the shift of the sinW/v S i S In S10

pulse because of fil and
isSm , , s . SmX,
its phase changes. The St* Sam A/
amplitude variation from nl
to n-lO0 is slow. For the
example given, the shift n
the pulse due to In Hi S Sna n/a
in 3
over N-l0 pulsesi the 3 dB
width of the sinx/x pulse i,
I/Tl Hz while the total $ s . Sf.. SM/
width between the first nulls
Is 2/12 Hz. Hence the Figgure . Matrix oontaning nmg@-DoppZer
amplitude variation within a nwb.r. obtainad ith double-
column is slight, and can be
represented in most cases by a constant or, for more accuracy, by a constant
plus a small linearly varying term; the 'esults will not differ signi-
ficantly for either case. Hence we represent the amplitude by a constant
(i.e. sin[2(f-flN/2 )T/2]/[2w(f-flN/2 )T/2], its value midway down the
column where n=N/2) and leave the second representation as an exercise
to the interested reader.
Thus the only variation now within the column (at the positive
frequency corresponding to m) is the phase factor, i.e.,

10
 s * K(f)o i
a # fc
KMO 'it tn & where (16)

in the rightmost expressions nTr has been replaced by t n to represent the

disrete flow of time from pulse to pulse. The Fourier transform of this
quantity over tn from 0 to NTr is

m KNTr, where
,hr2 (7
sRnr(f, fc)NTr/1]
here again we should note that our digitAl FFT does not really give a
continuous variation over f (frequency), but will compute values at N
discrete frequency points. The question arises as to how we should choose
these N frequency points, i.e., how wide a frequency window do we want
to display. Since our PRFs fr(fr-1 Hz here) results in an unambiguous
Doppler of I. Hz, we would logically select "1 Hz here) so
as to display all of the unambiguous Doppler window, Then the frequency
window In Doppler will be from "fD., to +fDma at a spacing fr.,,/N,
which turns out to every 2fwex/N Hz. or 1/100 Hz here. Note also in
(16) that If 4t
fc, i.e., the Doppler shift, exceeds 2 fr-l/ZTr then
from pulse to pulse we will be sampling at iess than the required Nyquist
sampling rate. Hence our pulse-repetition frequency (PRF), fr, must
always be at Least twice as great as the maximum expected Doppler frequency.
Observe now an important fact in (17): the displacement of the
sinz/x pulse resulting from the second Fourier transformation over the
columns occurs at m T f . This is precisely the Doppler shift that results
from a target at (radia) velocity v with a backscatter radar having
carrier frequency fc' Furthermore, the 3 dB width of the pulse represented
by (17) is l/NTr Hz, as shown in figure 5. Thus we produce N (or 100)
Doppler frequency points every fr/N Hz (or .01 Hz here) having a Doppler
resolution of lI/NTr Hz (- .' Hz hero). Since NTr is the coherent
integration time (in this scheme, it is the time required to fill the
matrix), l/NTr is exactly the Doppler resolution one would expect from
any coherent pulse-Doppler radar.

11
 '.'.. -N.r .. ,

Theefore, in sum % we have done two sets of MFTs. One set

within each pulse it N nts to give N/2 range binst these bins
an the elements of a raw of a matrix, The second set is over N, pulses,
or over the N column elements of the matrixt to give N Doppler bins
for each range bin. Note that the original target rings also containod
a small offset due to Doppler. If this offset is objectionable, It can
now be rtmoved--in the case of a discrete target--by using the Doppa
information to correct the target range
A little thought
will show that this
process also work~s j I /NT.
for distributed tar-
gets such as rain or
sea waves, If one
has many targets In
a range bin (say L f
targets), he has L D=mof
terms in (15), and
each element in the
matrix isreally Figure 5. Doppler epeotrwun after sooond trans-
the sum of L such om on wthin a given range bin.
terms. The second
FFT over the columns, therefore by superposition, gives L terms in (17);
if each of the L scatterers in the bin (representing the distributed
target complex) has a different velocity, then the resulting Doppler
spectrum for the L ..rgets will consist of L-sinx/x pulses at different
positions, as given by (17) and shown in figure 5. Thus a continuous
Doppler spectrum represented by the sum of L scatterers with many differ-
ent velocities and scattering amplitudes will result, as would be expected
in any coherent pulse-Doppler radar system.
7. SINGLE-FFT DIGITAL PROCESSING
Now we examine another technique for extracting range and Doppler
information from the same signal. This involves a single, long FFT over

12
 the some N pulses. This vachnique is used by the Rome Air Development
Center for same of Its HF over-the-horison radars (Eddy, 1973). It
Involves the same number of computer operations as that described In the
preceding section.
Here we will drew heavily on much of the material in the preceding
section. Since we have a maximum frequency f, In our pulse Tr(sT)
seconds long, wc rcqurs Mt42f U.,
1 Tr samples per pulse, as before.
Performing the FFT over N pulses gives a total of M x N samples per
transform. Let us analytically find an expression for the Fourier trans-
form of this pulse train first. To do this, we can use superposition to
express the Fourier transform of the pulse train as the Fourier transform
of each pulse as though it were all alone:

V ,T
Nf) "i2Wft
MVn(tle dt . (18)
nwO
(m-I)Tr/2

Here we reexpress the phase *In(tt) appearing in VIn--as given

in (13)--ln terms of continuous time, t, rather than time within a pulse,
ti . This is done by substituting tit-nTt into (13) to give

(t ) 01 + 2w(Bto+ 2 BfnTr]n'2w[ 2v f +Bf to+ 2v enlt. (19)

Using this in (18) and performing the indicated integration, we

obtain (with the approximation Tr oT)

VIMf) X1 VIn(f)e'2ifnTr .(20)

where Vzn(f) Is given in (15) an discussed in that section

Now, to perform the summation, we make the same assumptions as
before; i.e., that of both the amplitude and phase variations over n
which occur in V1,(f), only the phase variation is Important. Also
we use only the first sinx/x function in (15) since it represents positive

13

....
..............
"n e tv L.. ?.- -- -
 frequencies; an Identical result obtains for the second term representing
the negative frequencies, Therefore, (20) becomes
K-1 2

V1 (f). K(f)F e-12i(f- fc)nTr (21)

where we have used (16). The above summation can be performed by using
the identity:
Neina e N/2 sn[(N+l)2]
no sln[ot/2 12 (22)

Thus we obtain

Vi(f) - K(f) VI n -2 vf-_r)NTr/'2]1 e-i2r(f- - fc)(N-)Tr/2

sine2(23) (23)
sin[2 (f- r-fc)Tr/2]

In the above final result the complex exponential factor--as well

as the residual phase factor eI 0 contained in K(f)--is not important
because it has unity amplitude. However we note that the sinNx/sinx is
much like the sinNx/Nx function. It contains a peak at x-O and side-
lobes away from the peak; it is, however, periodic whereas sinNx/Nx is
* not.
Now, we note that the FFT does not actually compute a continuous
function, F(f), but a transform at MN positive and negative frequency
points. Since the maximum frequency, fuix, is determined primarily by
the maximum target range desired, we have MN/2 positive frequency points,
and hence a value of VI(f) computed every f = 2f m../(MN) Hz along the
positive frequency axis.
To see how a discrete target will appear, we plot first in figure 6
the broad function representing the integration over a single pulse, i.e.,

si[27r(ff
n[2rr(f-f N 2 T/2 J
iN/2 )T/2]
Kf
K(f) [2(ffIN/2 )T/2] (24)

This gives the range bin, or location of the target in rang-. Its center
is slightly displaced, however, due to the Doppler term L fc in fiN/2

14
 K(f) " /'

Next we show t rfore

the sinN /stnr? factor (a])/""i X vn
00000,Envelope
of (23) infigure 6
for a moving target. fN

Notice that it sin Nx

repeats Itself at sin x 1/N.
the pulse-repetition IF2v_
frequency, fr, as
one would expect in (b)%
L- VrV1V
any pulse-Doppler
radar. Note that this
factor contains Product
only Doppler inform-
ation and would be CA

identical for a
target at any range
having the same
radial velocity.
Finally we -_Elf)
nvelope for
show the product of (d) 0stationary target
the two functions 44

for a moving target

sin x for non-moving target
sponding to fIN/2'
Notice that the
slow;y varying range (e)
function essentially - -
W1 mfr j* WRo1ge W
I<1" r -
envelope modulates
theDoplerfuntio,
4 ~r*~mf, -NRange bin center
the Doppler function, I for non-moving target
now isolating the v
target in both range Spectrum of -10 - f, Doppler shift
Doppler. The (for I( target in m-th
and Doh f range bin
range resolution is l M . A.
essentially the width mfr
Pigure 6. Piotc'riat deeoription of tong aingle
FFT proceaaing over N puleea.
15
 of the broader function, corresponding to fr l/Tr Hz (i.e., a 1.5 km
bin here), while the Doppler resolution is essentially the width of the
narrower line, corresponding to l/NTr Hz, the coherent integration time.
The frequency axis after the long FFT can thus be broken up into
M/2 coarse range bins of width lI/Tr Hz; within each range bin, finer
frequency divisions then correspond to the Doppler spectrum of the target.
In particular, there are N Doppler bins per range bin, corresponding to
a Doppler resolution of l/NTr Hz. It seems proper therefore to center
each range bin on a zero-Doppler line. The centers of each range bin--
as shown in figure 6--are thus located at mfr(multiples of the PRF) along
the frequency axis, and extend + fr/ 2 away from this central, zero-
Doppler position. Thus we can take the plot along the positive frequency
axis and divide it into M/2 pieces, each centered at mfr where
0 < m < M/2. Each piece then represents the Doppler spectrum of an indi-
vidual range bin. Or, we can have the computer do the "dividing" for us,
displaying each range bin however we choose. For example, range bi.ns
could be lined up behind each other, closely spaced, to give a 3-dimension-
al range-Doppler-intensity display. Note also that each range bin--and
the resulting Doppler spectrum thus obtained--is similar to the Fourier
)..transform over a given column in the preceding section; both are range
bins containing a Doppler spectrum -0th the same resolution and width.
THEREFORE THE TWO PROCESSING TECHNIQUES YIELD IDENTICAL RESULTS.
A little thought will also show that this technique will work for
distributed targets. For example, if we have many targets over several
range bins but at the same velocity, we will effectively have several
K(f) functions in (23), but centered on slightly different positions.
The sinNx/sinx functions for the Doppler will be identical. Thus in effect
the target at a given Doppler will appear in several range bins, as it
should, but at the same discrete velocity in each.
8. NUMBER OF COMPUTER OPERATIONS REQUIRED
The possibility exists with present day computers--especially
"minic.rauters" of the NOVA and HP 2110/2115 variety--that the range-
Doppler processing described above can be done in real time. Such
processing for HF radars has in fact been done digitally in real time by

16

. ...-.. .......: ........ ,.. ..

several groups for both discrete targets and sea scatter, using no more
than a single HP 2115 minicomputer. To ascertain whether such is possible
for a given application, we must know the number and size of the digital
words to be stored and processed per second.
The FFT process is known to require Llog 2 L operations for a linear
array of L numbers. Let us first analyze the total number of operations
required by the double FFT. We first do an FFT on a pulse, using M
sawples; this requires Mlog 2M operations. Next we begin transforming
over each of the M/2 columns; each now contains a real and imaginary
word for a total of M words. With N elements in a column, Nlog 2 N
operations are required for the FFT on each column. For M column words,
this gives MNlogN operations. Thus the sum of operations required in
the first and second sets of FFT processing is

MN1og2 M + MNlog12 N = MN(log 2 M + log 2 N) - MNlog12 MN (25)

operations.
The number of operations required in the single long FFT is simple
to calculate. With N pulses and M samples per pulse, we have MN
total samples per transform. This therefore requires MNlog2 N operations.
THIS IS IDENTICALLY THE SAME NUMBER AS FOR THE DOUBLE FFT!
Normally the FFT requires that the number of samples to be transformed
be an integer power of 2. For the double FFT process therefore, both
M and N must be powers of 2 (e.g.. 256 and 128, 32 and 64, etc.; just
so M and N Individually are greater than the number required by the
sampling rate and Doppler resolution). For the single, long FFT, the
product MN must be a power of two, and hence again M and N must
individually be powers of two.
In both cases, MN elements must be accumulated and stored for
processing; this dictates the size of the required core and/or disc
storage. The entire number of MNlog MN operations must be performed
every NTr seconds if the process Is to be done Its real time. This
requires that (Mlog2 MN)/T r computer operations per second be done (not
Including time for buffering and display functions). Thus the obvious
way to reduce the requiried data rate--if such is necessary--is to lower
M, the number of range bins. Since M is equal to 2fiTmTr, we must

17
 V ~ "'yi.........jx . ...

reduce f1,.o the maximum IF frequency per pulse. This does not
necessarily require one to reduce the range resolution. For example,
suppose for our example that instead of observing all ranges from 0 to
150 km with a 1.5 km resolution (giving M-200), we decided that we only
wanted to observe the window between 126 km and 150 km, but still with
1.5 km resolution. This gives conceptually M132 or M/2-16 range bins.
To achieve this, one merely slides the linear sweep delay in the receiver
so that instead of varying between 84 and f max= 100 Hz, f now runs
between 0 and f :16 Hz. Then the M=32 samples are adequate for
imax
the Tr=1 cecond pulse repetition interval.
Finally, the number of bits required per word also affects the data
rate to some extent. The processor dynamic range depends upon the bits
per word because of quantization error. Thus the dynamic range is optimally
6b decibels, where b is the number of (biiary) bits per word. Currently
about 80 dB dynamic range can be realized by digital processors without
too much difficulty, requiring 14 bit words and a 14-bit A/D convertor.
9. WINDOWING AND WEIGHTING
*In all of the preceding sections, we assumed a square pulse at
frequency f,, and N such pulses all with the same amplitude. As a result
we arrived at sinx/x and sinNx/sinx functions in the frequency domain
for the target echoes. Joth functions have rather high, objectionable
sidelobes: the first sidelobe of the sinx/x function is only 13 dB down
from the main lobe, while the average sidelobe level of the sinNx/sinx
function between main lobes is only down 20 dB. Thus some of the side-
lobes from a single target--as illustrated in figure 6--are quite high
and could be mistaken for other targets.
The remedy for this is the same as that taken by antenna designers
to reduce sidelobes: use an amplitude taper across the original function
before Fourier transforming. This technique is currently being used in
nearly all radar digital processing schemes. The common amplitude
taper--or weighting--used across the time window is the Taylor weight
(although Hamming and cosine-squared weights (Blackman,1958; Nathanson,
1969) are sometimes used). This results in average sidelobes down 40-50 dB

18
below the main lobe. The only bad effects of such weight.ng are the slight
broadening of the main lobe (by as much as 40% in some cases at the 3-dB
point) and a drop of 1-2 dB in signal-to-noise ratio due to attenuation of
the original received signal at the edges of the window.
For both types of processing described above, two weighting functions
are normally performed digitally. The first is to weight the M samples
within the pulse according to the selected function (e.g., Taylor weight-
ing). The next is to weight the N pulses to be used in the coherent
integration by the selected technique. Both weighting processes across
the two respective windows of Tr and NTr seconds are normally required
to keep both the range and Doppler sidelobes unobjectionable.

10. RULES FOR SIGNAL DESIGN

Here we give a simple, stepwise procedure for calculating the signal

parameters required for a given set of backscatter radar or sounder specifi-
cations. We assume that the following parameters describing the system are
given: (i) fc, the carrier frequency, in Hertz; () R, the range
window width to be calculated and displayed, in meters; (iii) vM, the
maximum target velocity in m/s; (iv) AR, the range resolution desired,
in meters; (v) Av, the velocity resolution desired, in m/s.
With these parameters given, the following four steps are to be used
to calculate the following four FM/CW signal and processing parameters:
() B, the signal bandwidth, or frequency excursion, in Hertz; (ii) Tr,
the pulse repetition interval, in seconds; (iii) N, the number of pulses
of period Tr needed for a single coherent processing operation; and
(iv) M, the number of samples needed per pulse interval, T r.
(1) B - c/(2AR), where c is the wave propagation velocity in the
medium.
(2) Tr - 1/fr , where fr a 2fl , fiM being the maximum target
Doppler shift, given by flM - (2vM/c)fc.
(3) N a Tc/Tr , where Tc, the total coherent integration time is
the reciprocal of the desired Doppler resolution, AfD, where
AfOD (2Av/c)fc.

19
 (4) M a 2Rw/AR samples per pulse interval, Tr.
In the above, we have assumed that fc' Rw, VM9 AR, and Av were
all given and that B, Tr , N, and M were to be found. In practice, the
size of the computer and data handling rate will often limit M and N.
Thus one usually iterates until an acceptable compromise is achieved, i.e.,
he varies his requirements for R, AR, and Av untlel values of M and
N are obtained within the tapacity of his machine.
11. SATISFACTION OF REQUIRED ASSUMPTIONS
In the course of the analysis herein, certain assumptions were made,
upon which the desired output is dependent. If these are not satisfied,
quadratic and other types of distortions will result which reduce or limit
the achievable signal-to-noise ratio. Having derived B, Tr M, and N
from the rules of the preceding sections, one can quickly check thE follow-
ing critiria to see whether the optimum processing gain will be realized.

(1) BTr(vr/c)2N2 < 1,

B(2vM/c)(2RW/c)N << 1,

B(2vM/c)Tr/ 4 << 1

Satisfaction of the above conditions was assumed in going from (12)

to (13) for the phase; if one or more of these conditions are not
satisfied, distortion will reduce the achievable processing gain.

(2) vMNTr < AR.

This merely means that the target is not traveling so fast that it
moves through several range bins within one coherent integration period,
NTr* If the inequality fails, it simply means that the echo will appear
in several range bins, but with a proportionately reduced amplitude in
each.

(3) dv/dt NTr < Av.

20

- .- . .
 This assumption--heretofore unmentionad--concerns the rate of
change of radial target velocity (or radial acceleration), It his been
assumed throughout the analysis that the targets under consideration have
a constant, nonaccelerating velocity. Small radial accelerations can be
tolerated, but if dv/dt is sufficiently large that the above Inequality
fails, then the echo will appear spread into several Doppler bins with
proportionately reduced amplitude in each.
12. SUMMARY

Despite statements often seen concerning "chirp" (ie., linearly

swept frequency) signals used with microwave radars, ther ie no conbigt
between target range and velocity for processing done n the straight-
forward digtael manner described thss
pr note* Furthermore, two seemingly
different digital processing schemes are described and analyzed herein,
which will produce exactly the same pulse-Doppler (range-velocity) output.
Both employ the FFT; the first uses a shorter FFT many times, while the
second uses only one long FFT to produce the same coherent pulse-Doppler
map. Both techniques work equally well for discrete targets (such as an
aircraft), as well as for continuous or distributed target complexes
(such as ocean waves, rain, atmospheric tu'.bulence, etc.), and display the
targets in their appropriate range-velocity perspective.
Identically the same total number of FFT operations is required for
both techniques; the same data rate (A/D convertor rate) is required in
each case also, i.e., (2Rw/AR)x2x(2vM/c)fc words per second. Here, Rw
is the range window length to be examined, AR is the range resolution
desired, vM is the maximum target velocity to be encountered, c is the
free space wave propagation velocity, and fc is the carrier frequency.
The choice of whether to use the multiple vs the single FFT processing
technique then rests with the availability of appropriate equipment. For

*Perhaps the ambiguity occurring in the microwave systema is attributable

to the analog pulse compression techniques commonly employed there, such
as the dispersive delay line. Here -the technique used is more pruperly
described as a coherent correlator followed by pulse-Doppler processing,
rather than time-domain pulse compression. The difference betwepn the
two techniques results in the elimination of the ambiguity for type of
proce'sing described here.

21

,. , . , f la
f, .. - . .... t. . . ... . . . . . -..
 examples small computers may L4 limited in the size of a tingle FFT they
can handisi In this cases the multiple FFT technique having smaller
unit site may be requiredl On the other hand, special hard-wired FFT
computers are currently available (called "PFT boxes"), The&# can perform
* a fairly large, fixed-length transform very rapidly because of their
* specialited constructions and are used al one component In the overall
digital processing system, Here, the single long FF1 isusually more
efficient be~iuse the need for continual, Interactive storage/rttrieval
of elements in matrix/fashion demanded by the multiple FFT scheme is
eliminated.
13. REFERENCES
Blackman, R,B. and J. W.Tukoy (1958), The HOOSU"WOee of Pow speot.
(Dover Publications: Now York, pp, 14-16).
Eddy, F. N. (1973)6 Analysis of repeated chirp system (in preparation),
The MITRE Corporation, Bedf'ord, Mass. (prepared for Rome Air
Development Center).
Nathanson, F.E. (1969), Hadar'Doettp P no'ploo. (McGraw-Hill: New York,
pp. 520-530).
Sweeney, L. E., Jr., W, 0. Zavoli, and D.E.Westover (1971), Processing
of swept-frequency-CW HF backscatter signals using the two-dimensional
FFT, Technical Report No. 6, Ionospheric Dynamics Laboratory, Stanford
Research Institute, Menlo Park, Calif., Contract No. N00014-70-C-0413.

22 *0, 1973 -744-572/134 RIGION NO, I