04071995
04071995
04071995
Abstract-Estimators of echo signal power, mean Doppler velocity, Pulse-repetition time or separation between two
and spectrum width are investigated. Maximum likelihood (ML) solu- pulses of a pair.
tions can improve the estimate variances by an order of magnitude over w Doppler spectrum width (Hz).
those by the autocovariance or the Fourier method. However, the re-
quired computations are excessive for routine implementations on Wb Bias due to finite time window.
weather radars. Spectrum power estimation together with two con- z Signal-pulse-noise vector.
ventional methods of mean Doppler and width estimation are reviewed. Zk Sum of noise and the Doppler shifted signal.
Several results previously not available are presented. Output single sample estimate variance.
aQ Variance of a sample mean at the output.
2
LIST OF SYMBOLS Cramer Rao bound for the variance of signal
b A number between 0 and 1 related to the time power.
constant of a recursive filter. Uv Doppler spectrum width (m * s-1).
C Covariance matrix of the signal vector s. U9f2 Cramer Rao bound for the variance of the mean
D Diagonal matrix with elements e)i(kTs frequency.
DM (f) Fourier transform of the data window. Uw2 Cramer Rao bound for the variance of Doppler
f Mean Doppler frequency shift. spectrum width.
fij Elements of the Fisher information matrix. x Radar wavelength.
I Identity matrix. Xik Elements of the inverse covariance matrix K .
K Covariance matrix of signal-plus-noise vector z. X;k Elements of the inverse of the normalized co-
K' = K/S Normalized covariance matrix of signal-plus-noise variance matrix.
vector z. p Magnitude of the correlation coefficient.
km Index of strongest Fourier coefficient. Ps Correlation coefficient at the output of a square-
L Log likelihood function. law detector.
M Number of samples. p5 Mean Doppler radian frequency shift.
MI Equivalent number of independent samples. Kronecker delta symbol 6 T-TS,O = 1, if TS = T.
N Noise power per sample. A Normalized deviation of Doppler frequency esti-
N(f) Power spectrum of the noise. mate around the mean, also angular interval
n Noise vector. over whichM samples are averaged = Ts(f- tf).
nk Noise sample (complex).
P Signal-plus-noise power. I. INTRODUCTION
Pi Signal-plus-noise power of one sample. T HE pulsed Doppler weather radar should supply the three
p Probability density function. most important spectrum moment estimates. These are 1)
Q Output of a detector related to input signal-plus- the echo power or zero moment of the Doppler spectrum (this
noise power. is an indicator of liquid water content or precipitation rate in
Q(N ) Quadratic form that minimizes the log likelihood the resolution volume), 2) the mean Doppler velocity or the
function. first moment of the spectrum normalized to the zeroth mo-
R (Ts) = R1 Autocorrelation at lag Ts. ment (this is equal to the mean motion of scatterers weighted
S Signal power per sample. by their cross sections, which for near-horizontal antenna
S(f) Power spectrum of the echo signal. orientations is essentially the air motion towards or away
S (k) Power spectrum coefficient = Sk from the radar), and 3) the spectrum width ut, the square
s Signal vector. root of the second moment about the first of the normalized
Sk Echo signal sample (complex) with zero Doppler spectrum, a measure of velocity dispersion (i.e., shear or
shift. turbulence) within the resolution volume.
T Time separation between two consecutive pairs Moment estimates utilize samples of a randomly varying sig-
of pulses. nal. In the case of whether echoes, single sample estimates
have too large a statistical uncertainty to yield meaningful
Manuscript received April 12, 1979; revised July 13, 1979.
data interpretation. Thus a large number of echo samples
The author is with the National Severe Storms Laboratory, NOAA, (acquired during a few tens of milliseconds) must be processed
Norman, OK 73069. to provide the required accuracy which depends on both radar
system characteristics and meteorological conditions. These throughout the literature, are brought together and some new
include the signal-to-noise ratio (SNR), the distribution of results are presented. To that end the Appendix section con-
velocities within the resolution volume, the receiver-transfer tains a discussion of maximum likelihood (ML) estimators and
function, and the number of samples processed. Because the associated minimum variance bounds against which more
samples are collected while the antenna is rotating, the de- conventional estimators can be gauged. The Appendix is self-
sired precision that determines the collection time (dwell contained, and those readers not interested in the theory of
time) also restricts the rotation rate. optimum estimators need not read it.
Reflectivity estimation requires only echo sample averaging
to reduce statistical fluctuations. However, mean Doppler II. RADAR SIGNAL PROCESSING
velocity estimation involves the Fourier transform or com- For weather radars, where a large number (e.g., approxi-
plex covariance calculation, and requires a large amount of mately 1000) of range cells needs to be examined, the op-
data processing. Probably the long development and cost of timum estimation algorithms (see Appendix) become imprac-
Doppler processors (to estimate velocities simultaneously at tical. Rather, simpler methods, with less calculations and
all resolution volumes along the beam) were due to pursuit of smaller memory storage, are used. We concentrate on such
measurement of the whole Doppler spectrum, from which simpler techniques and compare them to optimum ones.
the most interesting moments (mean Doppler velocity and The block diagram (Fig. 1) is that of a typical meteorological
spectrum width) need to be extracted. Doppler radar receiver. Presented automatic gain control loop
One of the first Doppler spectrum analyzers that could in- (AGC) is a feed-forward type, such that previous time-series
deed, in real time, generate Doppler spectra for each con- power estimate controls the gain during acquisition of samples
tiguous resolution volume is described by Chimera [8] and for the next spectrum moment or periodogram calculations.
this machine, called Velocity Indicating Coherent Integrator, This gain, constant during dwell time, matches the input sig-
processed, with a single electronic circuit, the echo signals to nal's dynamic range to the range of the A/D converter. Alter-
generate Doppler spectrum estimates simultaneously at all nate schemes where the signal instantaneously controls the
resolution volumes. Another machine, called the Coherent gain of the linear amplifier have been used [25].
Memory Filter, employing the same principles, was devel- A matched filter is either in the video or IF part of the re-
oped [16] for weather radar observations and used by re- ceiver; dc offset control and gain balance circuits must be
searchers at the Air Force Cambridge Research Laboratories available in the two video amplifiers. Ground clutter suppres-
(AFCRL)1. This machine produced the first real-time maps sion and calculation of spectral moments are done in the last
of Doppler velocity shear fields on a plan-position indicator block, which contains a real-time hard-wired processor and/or
(PPI) [2]. In the early seventies, Sirmans and Doviak [34] an on-line computer. The signal power is sometimes obtained
described a phase change estimator which circumvents spectral from the output of the logarithmic amplifier.
calculations and digitally processes echoes to estimate directly The digitized echo is a sequence of complex video samples
the Doppler velocity in contiguous resolution cells. Zk separated by* the pulse-repetition time Ts and consists of
The need to obtain the principal Doppler moments economi- a signal part Skejwk
Te and a white noise part nk, i.e.,
cally, with minimum uncertainty, and in digital format (to jwkT, +l,k=0,l,1-,M-l
facilitate processing and analysis with electronic computers) Zk=Sk e (2.1)
has prompted use of covariance estimate techniques. Hyde Both nk and Sk are zero-mean Gaussian processes centered at
and Perry [17] reported an early version of this method, but zero frequency, but Sk is narrow band compared to the receiver
it was first used by ionospheric investigators at Jicamarca bandwidth. It is desirable to have a second spectral moment
[42]. Independently, and at about the same time, Rummler of Sk, narrow compared to the Nyquist interval squared Ts 2,
[30]- [32] introduced it to the engineering community. The in order to improve estimate accuracy. However, other con-
advantages of covariance processing coupled with the new siderations, such as unambiguous range, require compromises
technology (medium scale integrated circuits (MSI)) made pos- so that occasionally the second moments are not small com-
sible the implementation of this signal-processing technique pared to T2. The mean radian frequency Of Zk is co, such
on the pulsed Doppler radar [21], [25], [27], [36]. that IVlI<r/Ts.
It is the very nature of the weather echo that imposes limita- Let the true signal power be S while the magnitude of the
tions and tradeoffs on the Doppler radar. Weather targets are correlation coefficient p(wkTs) depends on sample spacing
distributed quasi-continuously over large spatial regions (tens and spectrum width w (square root of the second moment).
to hundreds of kilometers), the strength of significant weather Thus the autocorrelation function of the process Zk iS
echoes easily spans an 80-dB power range, and the signals
themselves are semicoherent, i.e., they are not purely sinusoi- R(kTs) =R =E(*ziz+k) =Sp(wkTs) eikTS +N5k (2.2)
dal [11]. where N is defined as the white noise power per sample and
In this article some proven methods of weather echo process-
8k iS 1, for k = 0, but zero otherwise. Implicit in (2.2) is that
ing are reviewed with emphasis on obtaining the first three the power spectrum is symmetric with respect to co and that
spectral moments. Several relationships, previously scattered the spectrum has finite width.
Good estimates of the three parameters: power S, frequency
IPresently the Air Force Geophysics Laboratory. f, and width w (via the correlation coefficient p (wkT,)) are re-
ZRNIC: SPECTRAL MOMENTS FOR WEATHER ECHOES 115
-LNLINEAR
1.0 P
LOGARITHMIC
.8
/4
(a.
.6
0
z e- 2 33
4 4
2 3 4 5 10 20 30 40 50 100
NUMBER OF SAMPLES Ml
(3.3) for various receivers (i.e., log, square law, etc.) are al- Square Law:
most identical to the ones obtained from a uniform average
of M samples if M = (2 - b)/b, [43]. Mi 1 = Y, L.d M-mM p,(mTsD. (3.8)
2) Number of Equivalent Independent Samples: The total -(M-1)
number of samples M obtained from the resolution volume is
determined by the PRT and dwell time. However, because The variance reduction factor for the other two receiver types
considerable correlation may exist from sample to sample, is shown by Walker et al. to be [40]
we must determine the equivalent number of independent Linear:
samples, in order to calculate the estimate variance reduction
achieved by averaging. The degree of correlation between
samples is a function of radar parameters (e.g., wavelength,
Mf' (:
Em
m=-(M-1) 4 T M' ) n=O
-'4 n 4(m Ts)
An
PRT, beamwidth, pulsewidth, etc) and the meteorological
status (e.g., degree of turbulence, shear, etc.) in the resolution (3.9)
volume. where
If the estimate Q of the true output mean Q is derived from
an average of MI uniformly weighted independent samples, the (2n - 1)!! (3.10)
output single-sample estimate variance a' is reduced by a n2n(n - 1)! (2n - 1)'
factor of 1/MI. That is Logarishmic:
2
or2U -UQ . (3.4)
-wI M7 =2m --(M-1) (nM ) L: (S).(3.11)
However, if we have M samples in which correlation exists
from sample to sample, estimate variance aQ does not vary as Correlation functions are also given for these two receiver
1/M. Instead, for a stationary process and equispaced samples, types by Walker et al. [40].
the estimate variance reduction factor for the M sample aver- The equivalent number MI of independent samples versus
age is given by [26] the product MwT, is shown in Fig. 3. This is a graph of (3.8),
(3.9), and (3.11). The logarithmic receiver decorrelates the
G~, M-1 M- Iml output samples most, and thus has the largest number of in-
aQ = MMI= p(mTO). (3.5) dependent samples
aQ -(M -1 ) MI (Fig. 3). But for equal MI the square-
law detection results in the least error among the three de-
The magnitude of the correlation coefficient can be expressed tector types (Fig. 2(b)). Having Figs. 2 and 3, we are in a
in terms of the Doppler spectrum, and the parameters of this position to compare the standard deviations of input power
spectrum (in particular spectral width) can be related to estimates from an M-sample average of correlated outputs.
atmospheric and radar system parameters. A Gaussian input Given a product MTsw, the equivalent number of independent
spectrum approximates reasonably signal spectra associated samples MI is read off Fig. 3(b) and (c) and then entered on
with precipitation echoes. To rigorously determine the auto- Fig. 2(b) to obtain the standard deviation. When done, the
correlation of the output Q, we must transform this spectrum comparison puts the square law ahead of the other two types.
by the receiver-transfer function (nonlinear for a logarithmic Specifically at large M the equivalent number MI for a
and square-law receiver). For example, a square-law receiver's square-law detector is 2MwTsVW while the log detector has
output correlation function p,(T,) for a noiseless receiver is Mir5/2wTs/3r(2.5), where t(2.5) denotes the Rieman zeta
derived as follows. function (sum of reciprocals Y1 n-2 5 t 1.34). These asymp-
Assume the input power spectrum to be Gaussian (A.24) totic formulas are valid for narrow spectral widths, for example,
with a corresponding autocorrelation function (A.25). The wTs = 0.1 corresponds to a width of 6.8 m * s-' on the National
post-detection width, i.e., the width of the amplitude fluctua- Severe Storms Doppler radars (mean spectrum width in torna-
tion spectra that determines the correlation, is v [4]. The dic storms is around 4 m * s-1). Because the standard devia-
predetection frequency spectrum width w is related to Dop- tion of the input power estimate (unbiased) from MI averaged
pler velocity spectrum width au, and to the radar wavelength output samples of the logarithmic receiver is 7T/IV 1.28
X, by the Doppler equation: times larger than for the square law, we conclude that correla-
tion decreases that value to /(2.5)t 1.16. Under similar
w2 = 4 U2V (3.6) conditions the Cramer Rao bound (A.38) is (4.5V wT5)_1
times lower than the variance corresponding to a square-law
Substituting this and accounting for the detection process, detector. Thus it appears that one can substantially reduce
we have the correlation the errors if wTs is extremely narrow. The aproximations in-
volved that MwT» >> 1, but it is not clear if the lower
p, (T,) = exp (-4r2w2T (3.7) bound isrequire
achievable. However, when samples are independent,
Combining (3.5) and (3.7), the equivalent number of in- the average from the square-law receiver is an efficient esti-
dependent samples MI is mator [47].
ZRNIC': SPECTRAL MOMENTS FOR WEATHER ECHOES 117
beamwidth 0 1.
B. Range-Time Averaging
Estimate variance can be reduced by averaging along range
time as well as sample time. Because the resolution volume's
range extent is usually small compared to its height and width,
/'MwT0,
e
averaging over a range interval Ar results in a more symmetric
(a) spatial resolution volume with little degradation of the spatial
resolution. Range-time averaging the output of a linear or
logarithmic receiver introduces a systematic bias of the esti-
m mate caused by reflectivity gradients, which will limit the
ZAJW
Z maximum Ar useful for averaging [29]. Nevertheless, we have
a a
a reasonable latitude available in choosing Ar.
The incremental spacing between samples multiplied by the
o
number of range samples being averaged equals Ar. The sam-
wo
pling increment is chosen by considering the autocorrelation
of the consecutive range samples of return signal plus re-
z -.. ceiver noise.
,ri MwT0, Quite often the bandwidth of the receiver may be about
we,
.t o
wz
2 to 3 times the reciprocal of the transmitter pulsewidth, and
(b)
since the noise statistics reflect this bandwidth, the noise
W
Z
samples will not be correlated as tightly as the return signal.
IL.
Noise samples become significantly correlated for a range
w
a sampling increment receiver bandwidth product less than
about 0.5 [35]. This suggests a range sample increment much
smaller than that based solely upon the correlation caused by
co
pulsewidth.
As for the case of sample-time averages, averages containing
range-correlated samples have an estimate variance that de-
pends not only on the number of samples, but also on sample
correlation [12]. The estimate variance reduction in this case
I/T MwTs , a
is given by (3.5), where T. is replaced by the range-time
(C) increment.
Fig. 3. (a) Equivalent number of independent samples for the three
receiver types. The number of samples is 10 and time or angle averag- IV. MEAN FREQUENCY ESTIMATORS
ing parameters of a Gaussian correlation are on the abscissa. (b) As with reflectivity estimators, we consider here only those
Equivalent number of independent samples for the square-law re-
ceiver. A very similar graph with slightly more independent samples methods of mean frequency extraction that have been im-
applies to a linear receiver. (c) Equivalent number of independent plemented on meteorological Doppler radars. They are
samples for the logarithmic receiver. based either on the complex autocovariance or on the dis-
crete Fourier transform of time samples.
3) Independent Samples Due to Resolution Volume Re-
plenishment: Acquisition of an ensemble sample series while A. Autocovariance Processing
the antenna is rotating results in continuous alteration in Autocovariance or pulse-pair estimator calculates the first
sample volume location. The echo samples for resolution two moments of the Doppler spectrum from estimates of the
volumes corresponding to two beam positions are correlated autocorrelation function at lag T. Assume that staggered
when the two radiation patterns overlap (we assume no sig- pairs are transmitted with interpulse spacing T. and distance
nificant decorrelation due to reshuffling of scatterers' posi- between pairs T, and let the complex envelope of the detected
tions). Walker et al. [40] have shown the correlation of echo radar echo be z (t) with Gaussian inphase and quadrature com-
118 IEEE TRANSACTIONS ON GEOSCIENCE ELECTRONICS, VOL. GE-17, NO. 4, OCTOBER 1979
ponents containing white noise and a colored signal (2.1). An imaginary parts of R (T5) are Gaussian and the density of A
estimate of the autocorrelation R (Ts) is obtained from samples can be found directly [28]:
Z2i = z(iT),Z2i+1 = z (iT + Ts):
1(A -Bc
(A )/2 exp ( A2pT I)
T) i EZ2i2+
M-i * (4.1) (A - Bcos 4iA) A +B
1 =o
{1 + afea [erf(a)+l+]} (4.6)
where M is the number of pairs.
This particular choice of pulse pairs has both practical and where
theoretical significance. It has been used on some meteoro- (A -B\1/2 Sp(T8) cos2irA
logical Doppler radars to reduce range ambiguities [7]. By (4.7)
A +B/ (A -Bcos4irA)1/2
letting T vary between T, and oo, all pairs from contiguous
(tightly correlated) to independent can be conveniently A =E(IR (T)12)- E(R (TS)12
treated [44].
The mean frequency is estimated as follows: S2M- 1 2SN +N 2
M2 p2(mT) (M - ImI)+ (4.8)
I
m2-(M- O)
Z
f= argR(T). (4.2)
2irT, S2
B Ivar (R (Ts)l = m2 p2 (Ts)
M-1
: p2(mT)(M Iml)
An alternate form for f (for contiguous pairs) is obtained -(M -l)
after expressing the autocovariance estimate as the inverse 2(M-1) ~ -(-1
Fourier transform of the periodogram [31 ]: + (M ) p(2Ts) SN8T Ts,0 (4.9)
M-1 A
j1 Sk sin (2irk/M) and
f= 1 tan-' 0 (4.3)
27rT, M- 1 A erf (a)= f e-t2 dt. (4.10)
E10 Sk cos (27rk/M) o1
I-
I.'
U)
z
m
4°
oCD .01
a.
\\\\
4
co .01
0
var (f ) a [82Mp2 (Ts)Ts1 {2r312wT/T+N2/s2
0-
\ \
\ TS/T-O
INDEPENDENT PAIRS
+ 2(N/S)[1 p(2Ts)65T-TS0]0} (4.14)
tiguous, spaced, and independent pairs is possible [44]: Note that this variance is about 1/48V'7T(wT,)3 times larger
var (f) = [8ir2Tp2(TS)f- {M2 [1 -
p2(T )] than the Cramer Rao bound (A.28). At low SNR's the op-
timum improvement (A.29) is ll16rr(wT,)2p2(T.). For a
M-1 typical width of wT, = 0.1 those two values are 12 and 2
, p2 (mT) (M -
Iml) + N2IMS2
times, respectively. Similar comparison, when the signal is
-(M- 1)
a sample of random sinusoids in noise with variance [18],
+ (2N/MS) [ 1 + p (2Ts)(1 M - 1)6 T- Ts, O} [44]:
(4.13) var (f) = (87r2MT )' [(21M)N/S + N2/S2 ] (4.16)
Plot (Fig. 5) of (4.13) reveals that variances for independent demonstrates that the Cramer Rao bound (A.18) is (for large
and contiguous pairs are not much different. At narrow M), M2N/12S times smaller than (4.16). This is very signifi-
widths and low SNR's contiguous pairs have a lower stan- cant at low SNR's because it suggests that there is consider-
dard deviation of estimates, but the opposite is true other- able room for improvement by using optimum estimators.
wise. Superimposed on Fig. 5 are results from simulations Incidentally, independent pairs asymptotically achieve the
[46]. Several 8192 point periodograms were generated with lower bound which is different from the one for contiguous
Gaussian shaped spectra. The inverse of such periodograms pairs [24].
120 IEEE TRANSACTIONS ON GEOSCIENCE ELECTRONICS, VOL. GE-17, NO. 4, OCTOBER 1979
C. Spectral Processing found wide use in radar meteorology. Although the spectrum
The advent of fast Fourier transform (FFT) and subsequent width of detected envelope is uniquely related to the Doppler
decrease of the cost in circuits for the required computation spectrum width, the envelope spectrum is broader, thus de-
make it attractive to use direct spectral methods for mean fre- grading more the width accuracy. Nevertheless, methods to
quency calculations. Still, the complexity and cost of FFT extract width used with coherent radars can also be employed
processors is an order of magnitude larger than their autoco- on the signal envelopes from incoherent radars [48].
variance counterparts. To calculate the mean frequency, one A. Au tocovariance Processing
first forms the periodogram (4.4) which is an estimate of the
aliased power spectrum. The next step is to obtain a rough Because weather echo spectra follow closely a Gaussian
mean frequency estimate, km IMTS, where km could be the shape, it is natural to estimate spectrum width from the
index of the strongest Fourier coefficient. Then the straight- estimate of the autocorrelation coefficient [30], [45]:
forward mean frequency estimate becomes
w =1 T Iln(S/IR11)I"2 sgn [ln (S/IR1 1)]. (5.1)
k 1 km+M12
Mf=-TTsP
+ Y3 (k - km)S modM(k) (4.17) The signal power estimate is obtained from the average of mag-
Ts km-M/2 nitudes squared after subtracting the known noise power:
where P is the total power in the periodogram. This form I M-i
eliminates biases due to aliasing. Variations of (4.17) are S =- , |Zkl4 -N. (5.2)
used to improve the accuracy; among the simplest ones are M 0
thresholding spectral powers above the noise level and/or The sgn term in (5.1) warrants some discussion. Since S is
around km. Standard error of estimates when spectra are an estimate, it is possible, especially at narrow spectrum
thresholded can be found in [37]. Berger and Groginsky [5] widths, that the logarithm of the argument in (5.1) becomes
have derived the variance of the mean analogous to (4.17). negative [45]. The sgn term tags these negative values and
In their estimate the noise spectral density is subtracted from allows recovery of data, i.e., assignment of a unique small
the periodogram, and the summation is replaced with the width to all such cases. It is worth noting that (5.1) is exactly
integral over the Nyquist interval. The variance reads ML when pairs are independent and spectrum shape is Gauss-
1/2 Ts ian. This result follows directly from the derivations of Miller
1
var(f) m2T2s 2T f2 [S(f+f)+N(f+f)] 2 df and Rochwarger [24].
M2SS /2 Ts A related estimator is obtained if the second derivative of
the autocovariance at zero is replaced with a finite difference
(4.18) at lag Ts. The estimate becomes
and f denotes the true mean frequency while S(f) and N(f)
are the signal and noise spectral densities. If Gaussian spectra 1 1IRI ~~1/2
Il -
sgnsgn I(1\
A }. (5.3)
(A.24) of narrow width (wT, << 1) are substituted in (4.18) VhrTs A
A^2 =pT1
w
km +M12 'k A\2 S
kmMI2 ( f Ts) (mod M(k)). (5.7)
Fig. 7. Bias of the width estimate due to the fmite time window associ-
ated with Fourier spectral processing.
ES
cr~~~~~~~~~~~~~~~
.CM. CO.-oOC CCOCO00 0000000 000000 0000000 CcCCCCC 0000000 0.-Cf' e'c SE
.G 0000C00>CQc00000000C0000C 00 00 0 00 0 0
Z 00000000000 -..OPacl .4'4
-n .0.COO0 0 00 0 0 0 0 0 0 00 0 0"A00 0 .ccco-oc-o~)"
4. ('4- @
o' .000000000000000000000000000000EO4E3-@
.4.N(J @
M .o c o c c c c 0 c 0 0 0 0 0 0 0 0 0 0 0 0 0 C 0 0 0 0-- 4 ' N ..0.
f c c c c c C c o c O cc o c O c c c c c c c cc c c c c nO - C e
MJ *00 00co00000
aCCc: 0
o,C
oc 0 0 0 0 00 0 0 00000000..000EC'CCQ
cOO C OC OD ,a .f%-.Q O-@4 OQQQ,cc et%
Or. Eo&O~ 1
M. -4 E'9 .
1')..0000000000000000000000000000000006
CC.-44(71-C)
COOCC CO Oo c: 40 E'-0 000006
CDO D040000000 CD ('9*
4M "WI'9%.@
M4. N- 4>
Om00 0 0 0 0 0 0 0 0 0 0 0 0 0.-0..0-0 0 0 0 0 0 0 0 0 0 0
0
It LL! ,-4flJ
0n4 000000000m0000O 0000.~.4h 000000000000@
_j 01 ~ ~ ~ ~ ~ ~ ~ ~ ~ N-.@,
o. OOmCooooo C 0 0 00 0 0 000'0 '-00 0 0 0 0 0 0 0 0 0 0 *
('C . .-N 1,
0 4M @00000 0000000000000. Om-p-00000 00000 0000
ES. I,-.E
LA4
Ir
hr
Z~1,9
hr -00 0 0 0 0 00 0 0 0 0 0 000 0 0 00 0 - 0 00 0N00 0 0 00 0 0
r N .0 0 0 0 0 0 0 00 0 0 0 0 0 0 0 0--4O..00 0 0 0 0 0 00 0 0 0 0 0 0@
0~~~~~~~~
7 hr4M.- 0. q P
4 4N. 0-N'-..
LA0 N. mm eN
000 0
0000
*a0 C0
a cC>:r cc) cDMCCCC00
00 1,
C)C:O 4.-- C O. O000o00000000 00o0000000
c-C=C C:C C 00000:)C c 9@40
trN.-
r c) Z
~CC: hr .0 0 0 0 o 000
0 0L) r0 rZ ) o C C 0o00000000
l, :-c)::C C oooooooo0o 0oooo0ooo0 0o-- @hr: o
LU hr 4 @0 0 0 0 00 0 0 0 0 00 0 0 0 O O O O OO o O C o oo o c o o .4
N Ip :C-O N.-or()f :()C~- c)co
O O C -O :C0COC~a() O4: c>O O -C c CC O0c)-C(-Co : c)C)o\
C! 1,9 @00C0000000000@,4 0-.04ooooooooooo0ooooo0o .,
a N.- -N 4N
C., *0000000000000004
x 0l hr
-,DC )c ) ~C )C)Cc )c ~c c )C Dc )c )c ,c CD0c000,000c00000CC0
0p-c.-4.-a00)
C0c00c00
CD CD000 .0,e
OL mmm ~ ~ ~ N
01c DC )O: .0 0 00001,I
00 0 l()L : )(0.-ES
)C 0
4P DC~C
~c 000
: : 000 000
C C00000000
:CD: 00000000
D CDD4> a C)4~c000000000c
0 C 0
CD L:
x -4u) M, 0 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~IC-
C:J CDp.V-0 .00000 oooooo ooooo
:) -j .-.ooCooo
C c~oo ooooooo: 000000000
C ooa
o oocD
oo o ao o o
C o4
oo ooooo .p.C 4' - 0-
-. .-.1,9
\.
c tJIC4 @O 0 0 0-0000Chra0(')4
= Dc~C ( 0.-C0o0CD0o()
Da C 00 0 0 0)C 0)aC)CC
) 0o0
0 o00)40 : C DC ooaoD oooo
C 0CDCDQ @4
I-
-. -4N-\mf\
0 >- -.~~~ It M.
1, @~~~~~~ol()OOr =OO Oc ohr~-- hr4hroOooC OOOOO-
O aa)a OO O O c Oc D=OO(4)OOOOOC CCO000000000000 000000cc1, :M .-
- N .00~~~~~~~fl 00 o o r O c9' 0 0 0 00 0 0 0 0 0 0 0 0 0 0 00 0 0 0 0 0 0 0 0 0 .
0 -. 1,~~~~~~~~~~~cc)c
,..C)C p.-c >c ~r-c-c
@- >cc~oC )c)C )C ,c cc )c ,Ccc Ccc( ,C ~C)C c ~c0 -
hr .-~~~~~~~~~1 t
.0000000-0.-.044lc~c c --C) ) D : c D D =rC 000000000000000000CD0CD00000C0oooooooooo
D c(~ CDO oo.ooo=ooc
- ...-&J
LU -.Z- ,. t
>: 0.00 0 0 0 0-.-.1 40 00 0 00 0 0a*0 00 0 000 0 000&0 000 00 00 00a00 00&0 000 0 00 0 0 0 00 0 00 000 0
al- N M- LI) %ol- 00 0,cl -41\ rw) It rlD F--CO
O' (O- (V M U),D r- CD Ok :~ -4M f ,* LI), P- 00 O cD -N -4 Lf) D 11- O Oh N fn4 V) NoP- co04(f -pN
a 0 a~~~~~~~~~~~~~
14J
Fig. 8. Scattergram of hardwired pulse pair velocity Vpp and power spectrum derived velocity Vfft for the same pulse volume. Only
velocities of echoes with SNR between 0 and 15 dB are included. The Nyquist velocity Va is 34 m - s-1.
124 IEEE TRANSACTIONS ON GEOSCIENCE ELECTRONICS, VOL. GE-17, NO. 4, OCTOBER 1979
where the optimum W' is found from the minimum of the fol-
Fig. 10. From top to bottom are displayed: (a) Mean spectrum width, lowing quadratic form:
vertical scale is 15 m s-1/cm. (b) Mean velocity, scale is 34 m s-1/
cm. (c) Mean reflectivity in log units; 40 dB/cm. (d) Raw reflectivity.
(e) One component of the Doppler channel. Horizontal scale is 11.5 Q(W) Z7iX, kZke i s (A.6)
km/cm. i,k
Because (A.5) and (A.6) are coupled, the solution for X is
APPENDIX complicated. Nevertheless, one is tempted to separate the two
by ignoring the exponential in (A.5) and taking a magnitude,
MAXIMUM LIKELIHOOD ESTIMATORS or to use iterative techniques between (A.5) and (A.6). In
It is characteristic of the ML method to provide estimates general, minimization of Q(w) alone is not trivial, and closed-
with minimum variance (for large number of samples AM). Al- form solutions can be obtained only in some special cases.
though the ML estimators will turn out to be quite complex, The difficulty of finding the minimum Q(w) is compounded
the motivation here is to determine the minimum variance by the fact that one needs the inverse covariance matrix. Note
bound and compare it to the presently used simple algorithms. that the Q(w) function can be thought to represent a "fre-
Investigators have analyzed the properties of the Fourier quency analysis" of data products at various lags with the
method [5], [9], [23] and of the covariance method for weighting function Xi,k- While (A.5) is an asymptotic ML
spectral-moment estimation [5], [24], [44]. Miller and solution for the autocovariance magnitude, (A.6) is valid for
Rochwarger have established the covariance estimators to be arbitrary M, i.e., even when the inverse covariance is not
ML when sample pairs are independent. Recently Brovko Toeplitz. For that reason the indexes on X in (A.6) are i, k,
derived the likelihood equation for the mean frequency esti- rather than i k.-
mate when the data base consists of contiguous correlated To summarize, the optimum mean frequency estimator calls
time samples [6]. We first treat a multiparameter estimation for 1) the knowledge or estimate of the covariance matrix at
and then proceed to show the minimum variance bound for all lags, 2) inversion of the covariance matrix, and 3) minimiza-
various limit cases of weather signals. Likelihood functions tion of a quadratic form (A.6).
ZRNIC': SPECTRAL MOMENTS FOR WEATHER ECHOES 125
In arriving at (A.5), it was assumed that the inverse elements and the ML estimate of S becomes
Xk are independent of each other; consequently, R k's are also 2
M-i
independent. If Rk's are related, as is the case with weather E Zi e jwiTS -
MN
echoes, maximization of the likelihood function with respect A 0
to signal power S and spectrum width w, becomes very tedious. S= ~M2 (A. 15)
However, when white noise power is negligible, the optimum
signal power estimate S is found from while c25 is found by maximizing
where xik's are elements of the normalized matrix K' = K/S. But S(X) is the magnitude squared of a discrete Fourier trans-
The correlation coefficients which determine K' are found as form (DFT). To find its maximum, one can computeM equally
spaced coefficients via the FFT, and then search for the maxi-
Pk= ARe [Z ZPi+kei-, kTs] (A.8) DFT's around
mum the candidate frequency, i.e., perform several
(M- Ikt)S ° in the immediate vicinity of the largest coefficients
obtained from the FFT. With XZ determined, the signal power
while the mean frequency is obtained from (A.6). Again itera- (A.15) can be directly calculated. Note that the estimate S is
tive procedure can be used to calculate 5 'Z Pk, but a lot of intuitively satisfying because it takes the fullest advantage of
labor is saved if Pk is estimated from the sample mean coherent processing.
I M-1 Ik
- With the help of perturbation analysis, it can be shown that
E Zi'Zi+k for large M, both estimators S and X, achieve the minimum
(M - IkI)S 0
variance bound. From the Fisher information matrix the
after which X^ and then S are obtained. If one had exact bounds are
knowledge of the covariance matrix, then the variance of the for signal power:
signal power estimate would be bound by
S2
M2 > M(MS +N)3
(MS+M2 N ::Z S 2 + 2SN/M (A.17)
M-1
SYS (A.10)
M- E (M- Ikl)pkXk for the frequency f = w/2a:
-(M-1)
But in practice only estimates Xk are available, and the mini- 6N(MS + N)
Gf2 > 47r2 T2S2M2(M2 6N (A.18)
mum variance bound is somewhat larger. - 1) 47T2 TsSM3
Exact solutions of the likelihood equations are possible The approximate equalities on the right are valid when the
when the signal process consists of sinusoids with random am- number of samples M is large. While the variance of signal
plitudes and phases whose power is S. Then the autocorrela- power estimate cannot be reduced below S2 (as expected since
tion function reduces to a simple form only one sinusoid of the lot is observed), the variance of f is
Ri,k =S+N6i,k inversely proportional to the number of samples cubed! There-
fore, when echoes are coherent, relatively small number of
where samples should suffice to provide accurate measurements of
1 if i=k the frequency and hence, radial velocity. Incidentally, our
tO,0 if i#k. 5i,k:- ~~~~~~~~~(A.lIl) result is very similar to Lank et al. who determined the mini-
mum variance bound (when the signal process is a constant
The inverse of the covariance matrix has elements [6], [41 1. amplitude but random phase sinusoid) to be [18]
6N
(A.19)
Xi,k N(N +MS) (A. 12) 47r2T2SM(M2 - 1)
and the determinant of K is expressible in closed form: Other signal processes that can be treated analytically occur
when the inverse covariance matrix is tridiagonal with equal
det K = (N + MS)NM-1. (A. 13) but negative off-diagonal terms. Suchis, for instance, a Markov
Since the essential ingredients for writing the log likelihood process in negligible white noise, for which the magnitude of
function have been determined, all that remains to be done is the correlation coefficient is a simple exponential. The spec-
find cZ and S. Thus trum width is unbounded for this process, but a ML estimate
of mean frequency is obtained from the argument of the esti-
I M-i12 S mated autocorrelation:
L(co,S) = -- EjZ + zZizk
* i e-)s-1N(MS+N) i,k M-1\
.eJiw(iJ(k)Ts_ In (N +MS)-_ (M-_ l)lInN cos = arg E: Z-I Zi.l . (A.20)
0
-rr/Ts [S/I 2';iw + NTs exp (X2/8ir2w2)] On the other hand, when noise is large, the estimate reduces to
(A.27)
The integral in (A.27) can be evaluated if the noise is zero or Z [S (k) -
NTS] exp [_ (# )2 ]
if it is large compared to the signal. Thus for N = 0 the mini- : ['(Ts ) /](A.33)
mum variance bound (the reciprocal of (A.27)) for the fre- Z: exp[ (t-- )/w2]
quency becomes
ZRNIC': SPECTRAL MOMENTS FOR WEATHER ECHOES 127
ations for pulse-Doppler radar observations of severe thunder- [32] -, "Two pulse spectral measurements," Bell Telephone Labs.,
storms," J. Appi. Meteorol., vol. 17, no. 2, pp. 189-205, 1978. Whippany, N.J. Tech. Memo. MM-68-4121-15, 1968.
[121 R. J. Doviak and D. S. Zrnic', "Receiver bandwidth effect on re- [33] W. Rutkowski and A. Fleisher, "R-meter: An instrument for
flectivity and Doppler velocity estimates," J. Appl. Meteorol., measuring gustiness," M.I.T. Weather Radar Res. Rep. no. 24,
vol. 18, no. 1, pp. 69-76, Jan., 1979. 1955.
[131 I. S. Gradshteyn and I. M. Ryzhik, Table ofIntegrals, Series, and [341 D. Sirmans and R. J. Doviak, "Pulsed-Doppler velocity isotach
Products. 4th ed. New York: Academic, 1965, pp. 19 and 939. displays of storm winds in real time," J. Appl. Meteorol., vol. 12,
[14] R. M. Gray, "Toeplitz and Circulent Matrices: A Review," Stan- no. 4, pp. 694-697, 1973.
ford Electronic Lab. Tech. Rep. no. 6502-1, June 1971. [35] -, "Meteorological radar signal intensity estimation," NOAA
[15] U. Grenander and G. Szego, Toeplitz Forms and their Applica- Tech. Memo ERL NSSL-64, 1973.
tions. Berkeley, CA: University of California Press, 1958. [36] D. Sirmans and W. Bumgarner, "Estimation of spectral density
[16] H. L. Groginsky, "The coherent memory filter," inElectron Prog. mean and variance by covariance argument techniques," in Pre-
(Raytheon Co.), vol. 9, no. 3, pp. 7-13, 1965. prints, 16th Conf. Radar Meteorol., (Apr. 22-24, Houston, TX),
[17] G. H. Hyde and K. E. Perry, "Doppler phase difference inte- Amer. Meteorol. Soc., Boston MA, pp. 6-13, 1975.
grator," M.I.T. Tech. Rep. p. 189, 1958. [37] -, "Numerical comparison of five mean frequency estimators,"
[18] G. W. Lank, I. S. Reed, and G. E. Pollon, "A semicoherent detec- J. Appl. Meteorol., vol. 14, pp. 991-1003, Sept. 1975.
tion and Doppler estimation statistic," IEEE Trans. Aerosp. [38] R. G. Strauch, R. A. Kropfli, W. B. Sweezy, and W. R. Moninger,
Electron. Syst., vol. AES-9, pp. 151-165, Mar. 1975. "Improved Doppler velocity estimates by the poly-pulse pair
[19] R. W. Lee, "Performance of the poly-pulse-pair Doppler estima- method," Preprints, 18th Conf. Radar Meteorol., (Mar. 28-31,
tor," Lassen Res. Memo 78-03, Aug. 1978. Atlanta, GA), Amer. Meteorol. Soc., Boston, MA, pp. 376-380,
[20] M. J. Levine, "Power spectrum parameter estimation," IEEE 1977.
Trans. Inform. Theory, vol. IT-ll, pp. 100-107, 1965. [391 P. Waldteufel, "An analysis of weather spectra variance in a
[21] R. M. Lhermitte, "Real time processing of meteorological Dop- tornadic storm," NOAA Tech. Memo ERL NSSL-76, 1976.
pler radar signals," in Preprints, 15th Conf. Radar Meteorol., [401 G. B. Walker, P. S. Ray, D. Zrnic', and R. J. Doviak, "Time angle
Amer. Meteorol. Soc., Boston, MA, pp. 364-367, 1972. and range sampling with weather radar," Preprints, 16th Conf.
[22] J. S. Marshall and W. Hitschfeld, "Interpretation of the fluctuat- Radar Meteorol., (Apr. 22-24, Houston, TX), Amer. Meteorol.
ing echo from randomly distributed scatterers," Pt. I, Can. J. Soc., Boston, MA, pp. 156-162, 1975.
Phys., vol. 31, pp. 962-995, 1953. [41] J. R. Westlake, A Handbook of Numerical Matrix Inversion and
[231 K. S. Miller and M. M. Rochwarger, "On estimating spectral mo- Solution of Linear Equations. New York: Wiley, 1968.
ments in the presence of colored noise," IEEE 7rans. Inform. [42] R. F. Woodman and T. Hagfors, "Methods for the measurements
Theory, vol. IT-16, pp. 303-309, May 1970. of vertical ionospheric motions near the magnetic equator by in-
[241 -, "A covariance approach to spectral moment estimation," coherent scattering," J. Geophys. Res., Space Phys., vol. 75, pp.
IEEE Trans. Inform. Theory, vol. IT-18, pp. 588-596,1972. 1205-1212, Mar. 1969.
[251 E. A. Mueller and E. J. Silha, "Unique features of the CHILL [43] D. S. Zrnic', "Mean power estimation with a recursive filter,"
radar system," in Preprints, 18th Conf. Radar Meteorol., Amer. IEEE Trans. Aerosp. Electron. Syst., vol. AES-13, pp. 281-289,
Meteorol. Soc., Boston, MA, pp. 381-382, Mar. 1978. May 1977.
[26] F. E. Nathanson, Radar Design Principles. New York: McGraw- [44] -, "Spectral moment estimates from correlated pulse pairs,"
Hill, 1969. IEEE Trans. Aerosp. Electron. Syst., vol. AES-13, pp. 344-354,
[27] L. R. Novick and K. M. Glover, "Spectral mean and variance esti- 1977.
mation via pulse pair processing," 16th Conf. Radar Meteorol., [45] -, "Spectrum width estimates for weather echoes," accepted
Preprints, Amer. Meteorol. Soc., Boston, MA, pp. 1-5, 1975. by the IEEE 7rans. Aerosp. Electron. Syst., (1979).
[28] A. Papoulis, Probability Random Variables and Stochastic Pro- [46] -, "Simulation of weatherlike Doppler spectra and signals,"
cesses. New York: McGraw-Hill, 1965. J. Appl. Meteor., vol. 14, no. 4, pp. 619-620, June 1975.
[29] R. R. Rogers, "The effect of variable target reflectivity on weather [47] -, "Moments of estimated input power for finite sample aver-
radar measurements," Quart. J. Roy. MeteoroL Soc., vol. 97, pp. ages of radar receiver outputs," IEEE Trans. Aerosp. and Elec-
154-167, 1971. tron. Syst., vol. AES-il, pp. 109-113, 1975.
[30] W. D. Rummler, "Accuracy of spectral width estimators using [48] -, "Radar measurement of turbulence in storms," Rep. RCA
pulse pair waveforms," Bell Telephone Labs., Whippany, N.J., Avionics Systems, Van Nuys, CA, 1977.
Tech. Memo, 1968. [49] D. S. Zrnic' and R. J. Doviak, "Matched filter criteria and range
[31] -, "Introduction of a new estimator for velocity spectral param- weighting for weather radar," IEEE Trans. Aerosp. Electron.
eters," Bell Telephone Labs., Whippany, N.J., Tech. Memo. MM- Syst., vol. AES-14, pp. 295-930, 1978.
68-4121-5, 1968.