Spectral Moments Explain
Spectral Moments Explain
Spectral Moments Explain
i
I
-
May
1989
Signal Processing
for Atmospheric Radars
R. Jeffrey Keeler
Richard E. Passarelli
TBSIE OF COTENTS
iii
..
v
.
..
.vii
..
i.......................
1.
2.
2.1
2.1.1
2.1.2
2.1.3
2.1.4
2.2
2.2.1
2.2.2
2.2.3
2.2.4
2.2.5
2.2.6
3
3
3
4
5
6
6
7
8
8
9
10
11
3.
3.1
3.2
3.2.1
3.2.2
3.2.3
3.2.4
3.3
3.3.1
3.3.2
3.3.3
13
14
18
18
20
23
25
27
27
28
3.3.4
3.4
3.4.1
3.4.2
3.4.3
4.
4.1
4.1.1
4.1.2
4.1.3
4.2
4.2.1
sampled data. .
.............
Poly pulse pair techniques .....
Uncertainties in spectrum moment estimators .
Reflectivity. ...
............
Velocity. . .
.....
. ......
Velocity spectrum width ...........
31
33
.
35
35
36
37
43
43
44
45
46
50
51
4.2.2
4.3
55
56
5.
5.1
5.1.1
5.1.2
5.1.3
5.1.4
5.2
5.2.1
5.2.2
5.3
5.4
57
57
58
59
61
63
63
64
68
69
70
6.
6.1
6.2
6.3
6.4
71
71
74
76
78
7.
7.1
7.1.1
7.1.2
7.1.3
7.1.4
7.2
7.3
7.3.1
7.3.2
7.3.3
7.3.4
7.3.5
7.3.6
7.4
7.4.1
7.4.2
7.4.3
7.4.4
7.4.5
81
81
81
82
83
83
84
85
85
86
86
87
87
87
87
88
88
88
88
89
8.
8.1
8.2
8.3
8.4
Conclusions. . ...................
Assessment of our past. .............
Recommendations for our future .
........
Acceptance of new techniques ...........
Acknowledgements. ..............
91
91
92
93
93
95
...........
....
iv
97
TIST OF FJIGRES
Fig 3.1
of
Doppler power spectrum (128 point periodogram)
Estimated
typical weather echo in white noise.
parameters are velocity ~ 0.4 Vax velocity spectrum
10 dB.
width ~ .04 Vmax, and SNR
15
Fig 3.2
of the complex
representation
dimensional
Three
Radius of helix
autocorrelation function as a helix.
Ps;
proportional to total signal power,
Rs(0) is
rotation rate of helix is proportional to velocity, V;
width of envelope is inversely proportional to velocity
Delta function Rn(0) represents
spectrum width, W.
noise power.
29
Fig 3.3
32
Fig 3.4
34
Fig 3.5
39
Fig 3.6
42
vn=W/2Vmax=2Wrs/X.
is
and
47
Fig 4.lb
48
Fig 5.1
62
Fig 5.2
65
Fig 5.3
Adaptive filter
structure.
The desired response (dk) is
determined by the application.
The adaptive filter.
coefficients (Wk) and/or the output signal (Yk) are the
parameters used for spectrum moment estimation (Widrow
and Stearns, 1985).
66
Fig 6.1
26
vi
LSTr OF TAHBI
Table 1
Table 2
20
Table 3
38
Table 4
41
vii
PiRFACE
of
signal
processing
and
the
technology
that
are
and the references have been slightly expanded over the version published in
Radar in Meteorology.
We hope that this Technical Note will assist the many individuals who want a
R. Jeffrey Keeler
Richard E. Passarelli
March 1989
ix
1.
frequency
(typically in-phase (I) and quadrature phase (Q) components from a linear
channel and often a log receiver). These "time series", in their raw form,
convey little information that is of direct use in determining the state of
the atmosphere.
storage for later analysis is impractical except for limited regions of time
and space.
The data must be processed in real time to reduce its volume and
to convert it to more useful form.
In this paper the current state of signal processing for atmospheric radars
(weather radars, ST/MST radars or wind profilers, and lidars) shall be
discussed along with how signal processing is currently optimized for
various applications and remote sensors.
The focus shall be on signal
processing for weather radar systems but the techniques and conclusions
apply equally well to ST/MST radars and lidars. Zrnic (1979a) has given an
excellent review of spectral moment estimation for weather radars and
Woodman (1985) has done the same for MST radars.
Problem areas and
promising avenues for future research shall be identified.
Finally, we
shall discuss the scientific and technological forces that are likely to
shape the future of atmospheric radar signal processing.
We will differentiate between "signal processing" (the topic of this review)
and "data processing" in the following way. "Signal processing" is that set
1
dBZ
Radial velocity
Velocity spectrum width1
V
W
ms- 1
ms-l
confidence that can be placed on the estimates (e.g., the SNR). Note that
signal processing is primarily used in atmospheric remote sensing as an
estimation procedure as well as a detection process as in some aviation
applications. The emphasis is on making estimates of atmospheric parameters
or meteorological events.
"Data processing", on the other hand, takes up where signal processing
leaves off -- although the line of demarcation is not razor sharp. Data
processing algorithms take the base parameter estimates and further process
them so that they convey information that is of direct use to the radar
For example, data processing techniques imply display generation,
user.
data navigation to a desired coordinate system, wind profile analyses, data
syntheses from several Doppler radars or other sensors, applying physical
constraints to the measured data, and forecasts or "nowcasts" of severe
weather hazards.
Many aspects of data processing are covered in other
chapters.
2.
classes of
Electromagnetic radars
systems.
GENRA
"radar"
--
electromagnetic
and acoustic.
Acoustic radars
The signal
processing techniques employed for all these systems are similar (Serafin
and Strauch, 1978).
2.1
ClARACTERISTICS OF IRDCESSING
are nearly identical
radars,
the
way
in
which
this
backscattered
or
partially
reflected radiation is sampled, the principle noise sources, and the nature
of the scattering mechanisms are different.
2.1.1 Sampling
Because electromagnetic
less than 1 im,
constraints
from several
meters
to
The first is that the backscattered signal should be coherent from sampleto-sample, i.e., the motion among the scatterers should be small compared to
the wavelength so that their relative positions produce highly correlated
echoes from sample-to-sample.
T s < tcoh =
(2.1)
/4rW
where the true velocity spectrum width W in ms-1 is a direct measure of the
The coherence time is a measure of the
for
must be
microwave system.
such as
rapidly than
longer wavelength
The second constraint on sampling is that for regularly spaced pulses, the
sampling frequency must be at least twice the maximum desired Doppler shift
In this case
Ts < tNyq =
(2.2)
4V'
where tNyq is the minimum time between samples such that the desired
velocity V' is at least the so-called Nyquist velocity.
Since V' is
typically much larger than W, the latter constraint usually dominates the
sampling requirement.
+ 25 ms-1 , then T s
2.1.2 Noise
One of the goals of signal processing is to suppress the effects of noise.
The main source of noise in microwave radar is thermal in nature. This noise
power is simply
Pn = k Tsys Bsys
(2.3)
system temperature, and Bsys is the total system bandwidth including effects
of preselector filters, IF filters, and all other amplifiers in the signal
path
With recent
from environmental,
sources,
and
is not easily
quantified (Rottger and Larsen, Chap 21A). Therefore, antenna design and
the specific radar location and frequency band of operation define the
system noise.
4
Atmospheric radars respond to a variety of scattering targets-precipitation, cloud particles, aerosols, refractive index variations,
chaff, insects, birds, and ground targets.
1k12
Ze L
(2.4)
Pr=
1024 ln2 X 2 R 2
This equation includes L, the product of several small but significant loss
terms which are necessary to accurately estimate radar reflectivity factor,
e.g. receiver filter loss, propagation loss, blockage loss, and processing
bias. Zric (1978) defines the receiver filter loss as that portion of the
input signal frequencies not passed by the finite receiver bandwidth,
typically 1-3 dB. The other losses depend on atmospheric conditions and
antenna pointing and are enumerated in Skolnik (1980). This equation is
correct for Rayleigh scattering of a distributed target that completely
fills the resolution volume.
Non-Rayleigh targets or partially filled
resolution volumes will give received power estimates that cannot accurately
Huffaker, et al. (1976, 1984) give similar received power expressions for
returns from refractive index variations and from lidar aerosol returns,
respectively.
required
The
dynamic
range
for
measuring the
1.
The R
2.
Microwave
systems
sum of these
two effects
and
typically can achieve a dynamic range of order 100 dB for power measurements
using either a log receiver,
or some combination
of these.
2.1.4
The ratio
of the received
signal
power
to the measured
noise power
is
(2.5)
SNR = Pr/Pn
is
analytic
2.2
Table 1
Table 1
Remote Sensor Sampling Comparison
Sensor
Pulse
Beamwidth Duration
(deg)
(~sec)
Wavelength Scatterers
Radar
S-band
Ka-band
mm-band
10cm
1 cm
1 mm
Precipitation
Precipitation
Cloud
ST/MST (profilers)
UHF
75 cm
Refractive
VHF
6 m
index
0.5-3
0.5-2
0.2-1
3-10
3-10
Sample
Rate
(Hz)
0.25-4
0.25-1
0.25-1
103
104
105
0.2-5
0.2-5
104->102
103->10
Lidar
IR
Optical
2.2.1
10 gmi
<1 im
Aerosols
Molecules
0.01
0.1-3
(near field) <1
107
Microwave radars
1979).
precipitation,
cloud particles.
fluctuations, and
0.5 to 3 degrees and the maximum usable ranges for storm observation is 200-
500 km. After a few kilometers range, the pulse volume is "pancake" shaped,
i.e., the pulse depth in
beam.
Attenuation effects
range
from
severe
Doviak and
Zrnic (1984) and Strauch (1988) have shown that since only the second pulse
of a double pulsing radar may be contaminated by overlaid echo from the
first pulse of the pair, only random errors occur in the pulse to pulse
correlations.
These random errors may change very slowly with time so they
2.2.2
and/or the
troposphere are called ST/MST radars and sometimes known as wind profilers,
describe the
details
of
ST/MST
radar processing
2.2.3
FM-CW radars
FM-CW
(CW) to achieve
planetary radar.
2.2.4 Mobile radars
Airborne and spaceborne radars are an important class of atmospheric remote
sensors covered by Hildebrand and Moore (Chap. 22A). Special problems are
evident when a moving platform supports the remote sensor. Many of the
signal processing problems have well known solutions but have not been field
tested.
Lidar
2.2.5
or
Optical
infrared
radars,
cammonly
known
as
lidars,
scatter
et al.,
dominates
scattering
1984 and
Molecular
(Huffaker,
from
at
the
shorter
Lidar
is
most useful
in
wavelengths.
is
"clear air" applications (Lawrence, et al., 1972; McWhirter and Pike, 1978).
achieve
atmospheric
antenna
the
necessary
lidars,
gain
both ground
(or telescope)
and
Consequently,
sensitivity.
many
near field operation is the collimation of the optical energy into the "near
field tube" with minimal "sidelobe" radiation.
beamwidths are measured in milliradians.
The expected Doppler shifts and coherence times require sampling at rates of
10 -
100 MHz.
Current laser
duty cycle constraints limit PRF's to about 100 Hz, which produces data
rates that can easily be processed and recorded (Hardesty, et al.,
1988;
a single pulse is
the
degraded range resolution that results when the pulse propagates outward
during the data collection interval.
particles are appearing at the leading edge of the illuminated volume, while
10
This creates an
2.2.6
Acoustic sounders
for
the
boundary
layer
(Little,
1969).
Acoustic
-1
waves
are
Scattering is from
are
avoided.
possible.
Thus,
complex
from a real
(single channel)
(dual channel)
data
data
processing
is
than the carrier frequency of the sodar so that zero Doppler shift is offset
from
zero
frequency.
In
this
manner unambiguous
11
and
signed
velocity
DOPLER PE
3.
It is well established that the first three moments of the Doppler power
spectral density or the "power spectrum" (incorrectly termed the "Doppler
spectrum" in the community) are directly related to the desired atmospheric
base parameters:
power spectrum and moment estimation, we shall find it useful to define the
input waveform.
Since
the return
scattering
from
refractive index
from
a
individual
range cells
typically
is generated by
central
limit
process
(Parzen, 1957;
theorem)
a very
good
Swerling,
approximation to a
1960;
Mitchell,
(by the
Gaussian
random
Thus,
signal
1976).
estimation theory framework wherein one seeks to make the best estimate of
the ensemble parameters given a particular sample function
Davenport and Root, 1958).
(Wiener, 1949;
particular importance when one wishes to scan a phenomenon quickly since the
random process nature of
a certain
single
reproduces
the transmitted
z(t,R) = A exp[j2rf(t-2R/c)
where
function
is
that
depends
on
the
W(t-2t-2R/c)
pulse
length
(3.1)
a range weighting
and
the
receiver
Actual targets in
The received
We desire to
sample functions. The vector sum of the return complex voltage from the
individual scatterers is
z(t,R) = Z Ai exp[j2fi(t-2Ri/c)] W(t-t2Ri/c)
i
(3.2)
where the subscript i represents the individual particle. Each particle has
a complex voltage return (Ai), a Doppler shifted frequency (fi), and a range
(Ri). At any given sampling instant for the kth pulse the received waveform
can be represented in the complex signal plane by a vector (or "phasor")
Ik(R) + j Qk(R)
(3.3)
where Ik(R)=IVk(R) Icos ek(R) is the in-phase and Qk(R)=|Vk(R) Isin Ek(R) is
the quadrature phase component (Rader, 1984).
that
only
0
-10
dB -20
-30
-40
-0.5 -0.4 -0.3
-0.2
-0.1
0.1
0.2
0.3
0.4
0.5
VELO)CITY/2 Vmax
Fig 3.1
15
Vmax = \/4Ts
the so called
(3.4)
commonly the "Nyquist velocity interval" and all possible velocities are
measured within this interval. The reality of sampling theory dictates that
sampled Doppler spectra exist on a circular frequency domain rather than a
frequency line extending both directions from zero (Gold and Rader, 1969).
Thus, as a target velocity increases beyond Vmax, it aliases or "folds" onto
Gaussian shape.
For ST/MST radars the spectrum is often assumed to be Gaussian, but spectra
measured at near vertical antenna beam directions (zenith angles less than
about 10) very regularly show one or more strong spectral spikes superposed
on a Gaussian shaped base. The spikes result from a corresponding number of
16
frequency
domain processing using the power spectrum, and time domain processing using
the
autocorrelation
function.
Each
approach
has
its
advantages
and
since the power spectrum of the sampled signal and its autocorrelation
function comprise a Discrete Fourier Transform (DFT) pair, (Oppenheim &
Schafer, 1975; Tretter, 1976):
N-1
(3.5a)
N-l
Z S(nfo) exp [+j27rmn/N]
(3.5b)
n=0
where S(nfo) is the Doppler spectrum in multiples of the fundamental
frequency shift fo=l/NTs and R(rTs) is the autocorrelation function in
multiples of the sample time Ts . This is the discrete version of the
celebrated Wiener-Khinchine theorem (Wiener, 1930; Khinchine, 1934).
The
the
17
The primary
is that
3.2
Estimating the Doppler power spectrum and its moments directly are
straightforward techniques (Haykin and Cadzow, 1982). However, some basic
questions must be answered first. We implicitly assume a data model for
weather and clutter spectra when we choose a spectrum estimation technique.
A specific data model such as a sum of sinusoids or white noise passed
through a narrowband filter is best analyzed by a spectrum analysis
technique compatible with that data model. Robinson (1982) emphasizes this
Marple (1987)
point in his historical review of spectrum estimation.
stresses the importance of using an appropriate model fitting analysis and
gives a very well organized discussion of classical and modern spectral
estimates using digital techniques.
3.2.1 Fast Fburier transform techniques
The Doppler power spectrum may be estimated from the Discrete Fourier
Transform (DFT) of the complex signal. The DFT decomposes the observed data
into a sum of sinusoids having amplitude and phase that will exactly
It is easy to show that these N discrete
reproduce the observed data.
components are adequate to reconstruct the entire continuous spectrum so
long as the complex data samples {zk} are taken at a rate equal to or
greater than the bandwidth of the signal. The advantage of measuring the
full Doppler spectrum is that spectral impurities such as ground clutter,
18
bi-modal
spectra
or artifacts
can be
suppressed by
intuitive
(if non-
optimal) algorithms.
Usually
The periodogram
where
an estimate.
(3.6)
In general,
This tapering
spectrum
energy
smearing,
"leakage"
of
spectral
reduces the
introduced
by
Windowing
window
is
hk =
(or no windowing).
the
The
periodogram of a single point target has the first side lobe only 13 dB down
from the peak.
of the designed signal, but if strong clutter is present, then the sidelobe
power from the clutter that leaks throughout the Nyquist interval can mask
weaker weather echoes.
windows.
description
Harris
(1978)
of window
and Marple
functions.
(1987) both
In general,
of several
give
the
19
extraordinary
lower the
an
common
sidelobes
This broadening
Table 2
Characteristics of time series data windows (Marple, 1987).
Window
Name
Rectangle
Triangle
Hann
Hamming
Gaussian
Equiripple
Highest
Sidelobe
-13.3
-26.5
-31.5
-43
-42
-50
Sidelobe
Decay Rate
dB
dB
dB
dB
dB
dB
-6
-12
-18
-6
-6
0
interval.
dB/octave
dB/octave
dB/octave
dB/octave
dB/octave
dB/octave
Equivalent
Bandwidth
(Bins)
1.00
1.33
1.50
1.36
1.39
1.39
1/2 Power
Bandwidth
(Bins)
0.89
1.28
1.44
1.30
1.33
1.33
(Welch, 1967).
Cooley and Tukey (1965), a recent historical investigation into the history
of the Fast Fourier Transform by Heideman, et al.
(1984) attributes an
Apparently the
3.2.2
sequences produce
spectral resolutions better than the inverse duration of the data sequence,
which is characteristic of classical spectrum estimators.
digital
Furthermore, fast
This interest in
alternative
spectrum estimators
Maximum
entropy
(ME) spectrum
analysis
estimates
the
spectrum
using
ME
technique
was
developed
by J.P.
Burg
(1967, 1968,
1975)
as
model
can
approximate
Makhoul
any
arbitrarily
closely
by
ratio of the estimated spectrum to the true spectrum integrated over the
Nyquist
interval.
(Klostermeyer,
1986)
and the
lidar community (Keeler and Lee, 1978) have used the maximum entropy method
for characterizing atmospheric targets.
Sweezy (1978) and Mahapatra and
Zrnic
weather radar data and compared them with Fourier transform and pulse pair
estimators.
including weather
essential
spectral
features
well
although
little
quantitative work
is
(ARMA) model
21
is more
appropriate
(Cadzow,
R(L) for the Lth order estimator. Additional lags are realized by requiring
that the entropy (in an information theoretic sense) of the probability
density function having the extended autocorrelation function be maximized.
This extended autocorrelation function allows computation of coefficients
for a whitening or linear prediction filter.
from these filter coefficients which are defined by the matrix equation
A = R-1 P
(3.7)
22
^2 (L)
SME (f)
(3.8)
L
I1 -e
am exp[-j2rfm] 12
m=l
where am are the elements of A and ao2(L)
(1967)
gives
the
"forward-backward"
is
technique
of
estimating
the
Burg
linear
prediction coefficients directly from the data which frequently permits more
detail to be shown in
Kesler
(1976)
Friedlander
give
(1982)
the spectrum.
the
complex
Smylie,
and Makhoul
form
(1977)
of
et al.
the
describe
ME
(1973)
spectrum
estimator.
Papoulis
(1987)
presents a
more readable
exposition.
Cadzow
(1980,
1982)
pulse pair frequency estimator is identically the mean (or the peak, in this
special case) of the first order maximum entropy spectrum.
The atmospheric
remote sensing community has been using the simplest form of ME for almost
two decades!
radars,
ST/MST profilers
and
lidar
signals
is an
active research
area
(Haykin, 1982).
3.2.3
estimation is
of
known probabilities.
"efficient",
variance.
i.e.
there
ML estimates
of
is no other unbiased
spectral
parameters
estimator having
are
lower
shape is known or assumed when neither a priori knowledge nor a valid cost
function associated with moment estimator error is known (Van Trees, 1968).
Zrnic (1979a) uses ML techniques to derive the minimum variance (Cramer-Rao)
23
shape
and SNR.
Shirakawa
and Zrnic
(1983) evaluate
the ML
estimator for sinusoids in noise and find a slight improvement over the
pulse pair estimator at low SNR's.
exact ML mean velocity estimator for Gaussian shaped spectra using more than
one autocorrelation lag. Their technique is similar to Lee and Lee's (1980)
poly pulse pair velocity estimator.
for independent pairs, the pulse pair estimator of mean frequency is ML for
prior knowledge of the spectral shapes. Woodman (1985) shows that this
technique is a ML estimator of the spectral characteristics.
It is
gratifying that the simple pulse pair estimators approach the minimum
variance bound over a wide range of SNR's.
If the spectrum shape is completely unknown, the ML spectrum gives the most
probable estimate which concentrates the spectral energy at the input signal
[SML(f) -
L
= L1 Z [SME,m(f)m=l
24
(3.9)
It has
many of the same properties as ME spectra but the details are obscured by
combining all order ME spectra.
3.2.4
Computing these
on techniques
signal
(Denenberg, 1971,
"sampling theorem"
1976).
imposes certain
For
requirements on moment
the
and transform
Then the
Mn =
fn S(f)df .
(3.10)
signal power, so the clutter and noise powers must be estimated and removed.
Noise
power
is
generally
easy
to
remove,
but
clutter
removal
causes
The classical normalized first moment represents mean velocity and is given
by the linear weighting of S(f) over the Nyquist interval
fc =
f S(f)df / Mo
V = (X/2) fc
(3.11a)
(3.1Ib)
25
Note that white noise biases the velocity towards zero and for a pure noise
spectrum the mean velocity is identically zero.
been described for mitigating this bias, most of them requiring manipulation
of the power spectra.
the
noise
spectral
is
but
common,
some
sensitivity
is
lost
The "spectral balancing technique" rotates S(f) until the signal spectrum is
near
zero
velocity.
so that
the signal
and the
noise share
the same
zero mean
(3.12)
(f-fc) S(f)df = 0
The
normalized
(Nathanson and
Reilly,
1968).
contribution
from antenna
(Nathanson,
af2 = I
(f-fc)2 S(f)df / M0
W = (>/2) of .
(3.13a)
(3.13b)
Spectrum estimation algorithms are fairly time consuming to invoke, and once
the frequency domain is entered, there is still substantial computation to
accurately extract the meteorological moments.
the frequency domain lies in the ability to more easily filter spectral
26
artifacts
unimodal
or
identify multi-modal
from
spectra.
artifacts,
are
3.3
The basis for time domain moment estimation is the transform relationship of
the autocorrelation function of the complex signal to the power spectrum.
R(m) = (N-m)
N-m-1
2 Zk* Zk+m
k=0
(3.14)
interest.
Mn =
where
R[n] (0)
is
the
evaluated at lag = 0
nth
R[n](0)/(j2w)n
derivative
(Papoulis, 1962;
of
(3.15)
the
Bracewell,
autocorrelation
1965).
function
spectral moments are used to estimate the reflectivity, radial velocity, and
velocity dispersion or width respectively (Miller, 1970; Miller, 1972).
3.3.1
Gecmetric interpretatins
an
alternative
dimensional
representation
allows
better
tapered toward zero radius at the ends having a Gaussian shaped envelope.
27
Figure
3.2
The radius at lag 0 represents the signal power and the real
component.
delta function
at lag 0
or dispersion.
Thus,
interval which is the "pulse pair estimator" used almost universally for
mean velocity in weather radar and lidar processors.
useful
geometric
classical
Passarelli
3.3.2
The advent of the so-called pulse pair, double pulse, or complex covariance
technique (Rummler,
1972;
Berger
Woodman
and Guillen,
1974)
for mean
velocity estimation was revolutionary since the algorithm arose at about the
signal
processors
covariance processing
and weather
for velocity
apparently was
(Woodman
Rummler's work.
1972; Sirmans,
et al.,
1972;
Real axis
x
R(O)
Pn
PS
R(I)
Imaginary
axis
lag
m
v1
J
Fig 3.2
29
autocorrelation
proportional to
proportional to
proportional to
represents noise
Other time domain algorithms such as the "vector phase change" (Hyde and
Perry, 1958) and the "scalar phase change" (Sirmans and Doviak, 1973) are
closely related
inferior.
estimator, but
(1975b)
their performance
capare these
and
is
other mean
frequency estimators.
It is well known that the first few lags of the autocorrelation function are
sufficient to deduce spectrum parameters of interest.
1984),
Bracewell
(1965),
Papoulis
(1965 and
expansion
spectrum with the low order moments being the leading terms.
In other
words, the first few lags of the autocorrelation function contain the moment
information of interest.
the form
(3.16)
The even function A(mrTs) is determined primarily by the even central moments
(e.g.,
power, variance
and kurtosis),
function 0(mTs)
is
determined primarily by the mean velocity and the odd central moments (e.g.,
skewness).
Estimators
can be generated
Therefore,
sufficient
Pr = R(0) - Pn
(3.17)
The pulse pair mean velocity estimator is not biased by white noise and is
obtained by taking the argument of the first
V = (
autocorrelation lag,
1
/2) (2Ts)tan[Im R(T)/Re R(Ts)] .
30
(3.18)
[1 - p(Ts) (1 + SNR-1)]
(3.19)
where p(Ts) =|R(Ts) |/R(O) is the normalized first lag and the noise power
must be determined independently.
3.3.3
S(f) =
N-1
z R(mTs) z-m
m=O
Iz=exp [-j 27rf]
(3.20)
in understanding
digital signal processing. Figure 3.3 shows the z plane and the frequencies
associated with various points on the unit circle. Zero frequency, where
ground clutter usually appears, corresponds to z=l and the Nyquist frequency
(where velocity spectra alias into the next Nyquist velocity interval)
corresponds to z=-l. Thus, the z plane representation of spectral space
allows an immediate and simple geometric interpretation of velocity aliasing
and the velocity ambiguity arising from sampling too slowly. Analysis and
synthesis of digital filters requires heavy application of z transform
theory, thus easily allowing visualizing the effect of various types of
ground clutter filters, for example.
It is natural to compute spectral moments on the unit circle rather than
along the frequency line in the Nyquist velocity interval. The zeroth
moment or total receiver power, is still that area under the spectrum
31
Imag
f = (2Ts)-'
Real
I0
f = -(4Ts )-I
Fig 3.3
Z plane
32
A simple geometric derivation shows that the first circular moment estimate,
fc, of the estimated spectrum, S(f),
tan-1
(3.21)
Z S(27m/N) cos(2nm/N)
this equation to
N-1
Z S(27m/N) sin[27(n/N - fc)] = 0
n=0
Thus, fc is the sinusoidal weighted mean of S(f)
(3.22)
(Zrnic, 1979a).
Further,
we see that the numerator and denominator of (3.21) are the imaginary and
real parts of R(mrTs) and that the circular first moment is identically the
pulse pair frequency or velocity estimator.
Two points are clear from this discussion: 1) white noise does not bias the
pulse
pair
particular
frequency
estimate
frequencies
on
the
because
the
circle,
and
noise
2)
does
not
symmetric
weight
spectra
any
have
identical first moments using either the classical (linear weighting) or the
circular
prefer linear weighting over the more common sinusoidal weighting (the pulse
pair estimator).
Indeed,
3.3.4
received
signal,
velocity estimates by using, not only R(1), but R(2), R(3), etc.
33
The
Classical:
2Ts
freq -
2Ts
Circular:
Zsin[27r (f
fci)]S(f)
0
-t~~~~~i,,]scr, o~~~~~~~~~~~~
fcir
2T5
Fig 3.4
Comparison of classical
and circular
(pulse pair)
first moment
estimators.
Classical estimate is determined by linear weighting
of spectrum estimate and circular estimate,
by sinusoidal
weighting.
34
are given higher weighting since the correlations are higher. Poly pulse
pair velocity estimates (using a few lags) produce lower variance estimates
than the pulse pair estimates when the spectrum width of the signal is only
a few percent of the sampling frequency (Lee and Lee, 1980).
Strauch, et al. (1977) evaluated poly pulse pair for 3 cm radar processing.
They concluded that for typical velocity spectrum widths and PRF's (sample
rates) used with X-band Doppler radar, the coherence time was frequently too
Furthermore,
these samples.
UNCERTAINTIES IN SPECThUM
UMENT ESTIfM4AI
35
distribution. Thus, the mean received signal power is Ps with variance Ps2
and the coherence time is determined by the spectrum width of the signal.
The number of independent signal samples in a given integration time Td
seconds is approximately M I = 2/WT
d
(3.23)
samples is smaller than about 20 and a log receiver is used, the bias in the
Ulaby, et al.
(1982),
Velocity
(1969)
Zrnic (1977b)
Two
M >>
/47
(SNR- 1
W Ts
(3.24a)
+ 1)2 /
p2 (Ts)
(3.24b)
pairs
typical
of
and for
var(V) = f W/
8/7 M Ts .
(3.25)
Table 3 summarizes the velocity uncertainties at high SNR for three cases:
1) contiguous samples,
variance bound.
width in ms
(Wn).
-1
2)
Figure
3.5
shows
the
standard
deviation
of velocity
estimates
normalized to the Nyquist velocity interval and to the square root of the
number of samples M as a function of the normalized spectrum width.
is
parameter
for
the
two
sets
of
curves
--
those
for
The SNR
the typical
contiguous pairs and for less typical spaced pairs of pulses (Campbell and
Strauch,
1976;
1984).
Woodman
(1985) discusses
errors
in
estimates in an optimum fashion, he concludes that for high SNR only a few
(2 or 3) lags are necessary.
3.4.3
Benham, et al. (1972) and Berger and Groginsky (1973) applied a perturbation
analysis to the spectrum width estimator and Zrnic
their results to arbitrarily spaced pulse pairs.
high SNR, contiguous pairs, and narrow, Gaussian shaped spectra is that the
variance of the velocity width is
var(W) = 3 XW / 64/T M Ts
37
(3.26)
THBLE 3
Var(V) using W
Contiguous
samples
(typical case)
167r MTs 2
8]/7w MT
wn
X2
Independent
pairs
2M
Minirum
variance
bound
Var(V) using Wn
8Mr5 2
48 TS 2
2
-MX
3\
Wn4
W4
X2
MTs
38
Wn2
C\M
<>
V(
' 0
1.
u
Q
0
0.5
Q
0
0.1
0.2
0.3
0.4
Spectrum
Velocity error as function of spectrum width and SNR.
M is
=2WTs/X.
vn=W/2Vn
interval,
Nyquist
width is normalized to
1
velocity
Nyquist
to
is
normalized
error
and
number of sample pairs
interval, 2va = 2Vmax . Small circles represent simulation values
(Doviak and Zrnic, 1984).
39
Table 4 summarizes the width uncertainties at high SNR for three cases:
1)
contiguous samples, 2) independent sample pairs, and 3) the minimum variance
bound.
Expressions are given both in terms of the actual spectrum width in ms-1
(W) and normalized to the Nyquist velocity interval (Wn). Figure 3.6 shows the
normalized standard deviations of the width estimates as a function of normalized
spectrum width for a range of SNR's. The width estimator is relatively good if
the normalized width is between 0.02 and 0.20 of the Nyquist interval and the SNR
> 5 dB.
40
TABLE 4
Expressions for variance of width estimators at high SNR.
Assumes Gaussian
spectra in white noise, low normalized velocity width (Wn=W/2Vmax) and large M.
Var(W) using W
Contiguous
samples
(typical case)
Independent
pairs
3 X2
3
w
64W,/
642z MT
128J7r MTS 2
w2
x 2
8MT2
8MTs2
81]?
2M
Minimum
variance
bound
2880 T 4
4
Var(W) using wn
45 X2
W6
M X4
MTS2
41
2
Wn2
Wn
C
I
z 0.5
a
cn
Fig 3.6
.1
0.2
0.3
NORMALIZED SPECTRUM WIDTH, Ovn
0.4
Spectrum
Width error as a function of spectrum width and SNR.
width is normalized to Nyquist interval, vn=W/2Vmax=2WTs/X- M is
number of sample
interval, 2Vma
x .
42
4.
unbiased
estimates
of
the
characteristics
of
meteorological
range
Indeed, this
Ground clutter poses a significant problem for both coherent and incoherent
radar applications.
area of coverage at close range where the azimuth resolution is best. Even
weak clutter can frequently mask clear air echoes.
Fortunately, signal
processing can greatly reduce the effects of clutter.
(1981), Zrnic, et al.
is possible
and
clutter
filter
specifications
43
and
design
concepts
4.1.1
The first line of defense against clutter is an antenna with low sidelobes
and a good radar site. Main lobe clutter is very difficult to suppress
because clutter targets are usually much stronger than weather targets.
However, since sidelobes are usually down at least 20 dB (one way) from the
peak power, signal processing is effective in suppressing resulting clutter
P-2
for P<<1
(4.1)
better than 0.1 degree rms phase error which corresponds to 55 dB CNR.
44
A well
In many systems it is the dynamic range of the linear receiver that poses
the fundamental limit on the ability to separate weather signals from strong
clutter signals. If the linear receiver has a dynamic range of 50 dB, then
this will be the order of the maximum clutter-to-signal ratio that can be
handled.
near zero.
assigned the background color. More sophisticated processors use either time
Which approach is
ms-1)
Clutter is typically a
step in frequency domain filtering is to enter the frequency domain via some
spectrum estimation technique. This is usually done via an FFT. The choice
of the time-domain window is critical since the window sidelobes should be
matched to the dynamic range characteristics of the transmitter/receiver
system. For example, a 57 dB Blackman window (Harris, 1978) might be used
in a klystron system but it would not be justified for a magnetron system
that has a phase noise limited CNR of 25 dB.
Removal of clutter in the frequency domain is easily performed by the human
eye, and it
is
suppression.
Passarelli, et al. (1981) discuss several algorithms for
frequency domain clutter cancellation and point out the adaptive nature of
the general technique, i.e., both the notch width and depth of the filter
can be adjusted to remove only the clutter that is present, with minimal
distortion of overlapped weather or noise. On the other hand, time domain
filters usually, but not necessarily, have a fixed notch width and stop band
attenuation.
4.1.3 Time domain filtering
Time domain digital filtering has been an active research area for over 20
years (Kaiser, 1966; Gold and Rader,
Rabiner and Gold, 1975; Tretter, 1976; Roberts and Mullis, 1987).
1975;
Precise
the parameters that are adjusted are the stopband attenuation, the stopband
46
I
I I cVs-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~___
~~~~~~~~~~~~I
-10
Z
z
TV__-
-20
-30
I
-50
I\
0
I
I
I, I
VELOCITY (m s')
Fig 4.la Clutter filter frequency response for a 3 pole infinite impulse
For ground clutter
response (IIR) high pass elliptic filter.
width of 0.6 ms-1 and scan rate of 5 rpm this filter gives about
40 dB suppression. V s = stop band. Vp = pass:band cutoff, Vmax= +
47
Xk
Fig 4.lb
1981).
48
width, the transition band width and the passband ripple which if too large,
can bias the mean velocity.
The IIR filter is computationally more efficient to implement than a
FIR filter but, because of its transient response
comparable
is best run in a continuous mode with minimal
characteristics, it
perturbation such as those caused by slow AGC changes or PRF changes.
Initialization of the filter can improve the transient response
characteristics. Hamidi and Zrnic (1981) and Groginsky and Glover (1980)
evaluate IIR filters for weather radar systems.
FIR filters offer linear phase performance and are well suited for batch
processing of pulses since they operate on a finite number of pulses. This
makes them well-suited to slow AGC or multiple PRF techniques (i.e., where
the PRF is held constant while a batch of pulses is collected and then
changed for the next batch).
There are other types of clutter suppression algorithms that should be
Anderson's (1981) test of the mean block level subtraction
mentioned.
The parametric
technique offers 20 to 30 dB of clutter cancellation.
clutter cancellation techniques described by Passarelli
physical
models
of
clutter
and weather
along
with
of
the
Clutter suppression of 30
processing scheme to fit the observed clutter spectrum and extract the
spectral moments when clutter is about 50 dB stronger than the signal.
49
+ 10 log r.
(4.2)
When multiple PRT measurements are made for the purpose of extending the
4.2
(4.3)
(4.4)
to
the
Nyquist
measurements
coherency).
50
to minimize second trip echoes, will make the velocity aliasing problem even
(Doviak, et al.,
1978).
Fortunately,
there
are several
Bergen
NEXRAD algorithms.
There are several signal processing techniques for extending the unambiguous
range/velocity.
1.
techniques.
2.
velocity/range processing.
Batch PRT.
interlaced
PRT
sampling
whereby
short
PRT
is
used
for
velocity
followed by a clearing period (no transmission) and then one or two pulses
separated by a long PRr for the reflectivity estimate. The basic assumption
is that the PRT for the reflectivity estimate is sufficiently long so that
there are no second trip reflectivity echoes. The short PRr velocity
estimates will have two classes of range aliased echoes,- those that are
overlaid with the first trip echoes and those that are not overlaid with the
first trip echoes. When there is no overlap, the velocity estimates can
actually be assigned to the correct range. When first and higher trip
echoes are overlaid and one dominates the others in power by 10 dB or more,
The
1.
Loss of velocity data where first and second trip echoes are overlaid
2.
52
3.
The data acquisition time is increased because the long PRT pulses are
The
Dazhang, et al.
In the dual PRT (or staggered PRT) method the two PRT's
first lag complex autocorrelation for each PRT, averaging over a number of
pulses. Then, the expanded velocity is calculated from
=
(981 -2)/47r(T2-T
1)
(4.5)
4/4(T2-T1
) .
(4.6)
4/3, the expanded velocity range is 3 times. Why not expand further? Since
the variance of the expanded range velocity estimate is based on the
difference between the two fundamental estimates, its variance is roughly
proportional to twice that of each fundamental estimate.
53
Fortunately, the
expanded velocity estimate need be used only to roughly dealias the two
fundamental estimates. The velocity estimate can be improved by averaging
the two velocity estimates to get the final estimator provided they have
been correctly dealiased. This averaging technique provides an estimator
that uses all available pairs of consecutive pulses, rather than half the
available pairs.
Since the dual PRT technique dealiases velocities by a large factor, one can
operate the radar at a lower PRF and thus have a larger unambiguous range.
Doviak and Zrnic (1984) point out that another advantage of the multiple PRT
This can be
overcome for some filtering schemes by using a dual PRF technique wherein a
sequence of pulses is collected at each of two PRF's and then each sequence
is processed separately. The data processing is identical to the standard
pulse pair processing except that the velocity from the previous sequence is
used along with the velocity from the current sequence to dealias the
current velocity. The sampling statistics are similar to the pulse pair,
except that for this technique to be viable the mean velocity change between
54
4.2.2
Low PRF radars minimize overlaid echo but require sophisticated velocity
dealiasing techniques. If we promote the occurrence of overlaid echoes by
using a higher PRF to provide a large unambiguous velocity, then the range
aliased echoes must be resolved.
Most range dealiasing techniques use phase codes to distinguish between
first and second trip echoes. The simplest is the "magnetron" technique for
which each transmitted pulse has a random phase. A typical magnetron is
This means that
coherent-on-receive only for the current pulse.
contributions from multiple trip echoes are not coherent so that they appear
to the first trip or the second trip, etc., by using the appropriate phase
shifts so that Doppler spectra can be evaluated for each trip (Laird, 1981).
This technique offers information for both the first and second trip
returns, but does not solve the problem of reduced sensitivity for overlaid
echoes.
Siggia (1983) addresses this issue by filtering the first trip echo from the
The
second trip echo and vice versa, to reduce noise contamination.
technique works well as long as the two Doppler spectra (1t and 2 nd trip)
are not so broad that they occupy a large fraction of the Nyquist interval.
Zrnic and Mahapatra (1985) have evaluated this technique.
55
are inserted to cause the second trip Doppler spectrum to be a split bimodal
spectrum whose autocorrelation for lag 1 is zero.
second trip echo does not bias the first trip velocity estimates.
All of these
clutter
"phase diversity"
filtering
techniques.
techniques
However,
are well
there
are
substantial
signal
4.3
Bringi and Henry (Chap. 19A) describe various polarization techniques which
provide
valuable
velocity
scattering
target
dealiasing
and
more
information
but
difficult.
instrumental
make
clutter
Differential
effects
preclude
use
suppression
phase
of
and
propagation,
simple
Doppler
However, it is possible in
Processing
velocity
techniques
and
range
to
simultaneously
dealiasing,
and
provide
polarization
56
clutter
suppression,
processing
are
just
5.
TECHXNIQUES
PUISE OCfMPRESSICN
the linear FM (chirp) pulse which has been widely used in military radars.
Reid (1969) described a CW meteorological radar using pseudo-random coding.
Barton (1975) has edited a collection of pulse compression papers which
et al.
The
STORMY
weather
group
at
McGill
University
implemented
England which was modified to generate 4 Js, 5 MHz chirp pulses and measure
Doppler shifts from ice crystals at 8 km range.
Woodman
(1980b) shows how a continuous wave phase coded waveform was used in the
he
has
obtained
full
spectrum
military and aviation radar communities and has been used in the ST/MST
radar community
Woodman,
(Crane,
1980;
Gonzales
and Woodman,
1984;
Sulzer and
The reasons
Range resolution and transmit power using standard high peak power
5.1.1
The driving force for exploring pulse compression in weather radars is the
desire for ground based and airborne Doppler radars to rapidly sample the
volume at a spatial resolution adequate for mesoscale or cloud physics
analyses.
Dwell
Rayleigh
fluctuations
to reduce
the variance
of parameter estimates.
(1953) pointed out that frequency separations greater than the inverse pulse
and power.
Keeler and Frush (1983a) describe how this tradeoff relates to the "timebandwidth product" (TB) of the compressed pulse.
For the same average
transmitted power the increase in independence is TB and the decrease in SNR
is TB. For example, a chirp waveform 1 microsecond long sweeping 10 MHz of
bandwidth has a TB = 10.
independent and have a SNR ten times lower than the uncompressed 1
microsecond pulsed waveform. Frequently, the independent range samples can
be averaged to provide estimates having a reduced variance while allowing
much faster scan rates.
The primary disadvantage is a contribution to the backscatter from range
time sidelobes. Because the receiver filter output is the cross-correlation
59
of the received waveform and the time reversed transmit waveform (a matched
filter), range time sidelobes will cause data "blurring" in range space
similar to that caused by antenna sidelobes in the transverse spatial
dimension.
Range time sidelobes (and antenna sidelobes) are especially
troublesome in high reflectivity gradients.
using
Schmidt, et al.,
is the case of
(Golay, 1961;
However, more
show good results in practice when non-linearities in the system distort the
compatible with the shorter wavelength weather radar and lidar system.
are unacceptable
solution is
to extend the scan time using a short pulse mode for short
ranges and use a long pulse mode for long ranges.
Other techniques also
exist.
A third disadvantage relates to the availability of bandwidth.
Research
systems are not seriously constrained, but operational systems may require
bandwidths which do not fit into the channelized frequency assignments.
60
5.1.3
Ambiguity function
so much flexibility in
"ambiguity function"
fact
(1953) to study the effects on range and velocity ambiguities for a specific
waveform.
Weather
targets are distributed in range and velocity by their very nature and are
especially sensitive to these undesirable responses.
The
ambiguity
function
defines
the
ability
of
waveform
to
resolve
Skolnik,
1980;
Brookner,
1977).
Figure
5.1
shows the
ambiguity
diagram for a single FM chirp waveform in range (r) and velocity () space.
Note that targets having non-zero velocities at ranges different from the
desired range
function
(r=0)
evaluated
contribute significantly
along
the
axis
to the
(i.e.,
0=0)
filter
is
output.
identically
The
the
Atmospheric
radars
involve
estimation
of the
of
the
ambiguity
function
autocorrelation function.
order only
103
Hz
along
the
axis
is
simply
the
of
107
Hz
allow us to
small
velocity.
to
avoid contamination
Known waveform
61
Ijx
(r, ) I
Io1O6
/
I
Fig 5.1
62
ambiguity region acceptably small (Deley, 1970; Kretschmer and Lewis, 1983;
Costas, 1984; and Lewis, et al., 1986).
5.1.4
CCpariso
time-bandwidth
product TB has
characteristics
similar to a multiple frequency radar using the same time and bandwidth
factors.
This
scheme yields the same number of independent samples for the same total
pulse duration and total bandwidth.
5.2
At
in
the early
colleagues
(Widrow and Hoff, 1960) developed a class of filters that could "learn"
their received signal environment and, in time, adapt their characteristics
to optimally
filter
an
incoming
signal.
Initial
applications
were
in
1967)
quickly
followed.
Griffiths
instantaneous
Atmospheric
radar applications (i.e., non-military) have been sparse mainly because the
computational load associated with constantly changing filter coefficients
could not be accommodated until recently.
(1977) have
With the advent of fast programmable signal processors, we can expect to see
a rash of new applications
in radar
for adaptive
filtering techniques.
(Alexander, 1986b).
common
index for optimization is the minimum mean squared error (mmse) between the
processed output signal and a known desired output (or at least one which is
correlated with the desired signal). Figure 5.2 depicts a 2 dimensional (2
weight) error surface. Widrow's (1970) popular Ieast Mean Square (IMS)
algorithm estimates the gradient of the quadratic error surface and steps
yet to be defined desired response dk, and the error signal, ek = dk - Yk.
This error signal drives an adaptive algorithm which controls the transfer
function of the adaptive processor, and its output yk. Various closed loop
structures are possible as are a variety of adaptive algorithms. Widrow and
Stearns (1985), Honig and Messerschmitt (1984), Alexander (1986b), and
Haykin (1986) give excellent overviews of these structures and algorithms.
Widrow, et al. (1976) describe the learning characteristics of IMS adaptive
filters in both stationary environments when the filters converge to an
optimal setting and non-stationary environments where the filter continues
to adapt to the time variable input signal statistics. Adaptive filters
have found application in data prediction schemes, system identification or
modeling, parameter tracking, deconvolution and equalization, and
interference (clutter) cancelling (Alexander, 1986a). Usually the
64
LU
()
.0
W2
WI
WI
Fig 5.2
W* =
(0.65,-2.10).
change so that the error surface varies with time, the adaptive
weights will track this change (Widrow and Stearns, 1985).
65
Inni it
Xk
Error
k
Fig 5.3
is
(dk)
The desired response
structure.
Adaptive filter
coefficients
filter
The
adaptive
the
application.
by
determined
(Wk) and/or the output signal (Yk) are the parameters used for
spectrum moment estimation (Widrow and Stearns, 1985).
66
application determines the origin of the reference signal and the specific
adaptive algorithm to be used.
Sibul
filters.
networks and fault tolerant computing are being explored (Lippmann, 1987;
Shriver, 1988).
As
an example
of an atmospheric
radar application, an
adaptive
linear
prediction filter will improve the SNR of the received signal so that the
moment estimation will yield improved estimates.
the prediction filter acts as a narrow band pass filter having time variable
center
spectral
noise
components.
Tufts
(1977)
and Anderson,
et
al.
(1983)
is the difference between Xk=dk and its estimate Yk, i.e. ekx-yk.
The
filter is adjusted using the IMS algorithm so that the mean squared error
signal <ek2> is minimized.
the predicted signal will have an improved SNR over the input itself.
In
this sense, we have an adaptive matched filter which can track the input
signal as
its characteristics
change
Probing
deeper
into
the
mathematics,
we
find
that
the
algorithm
is
in the L
weight vector Wk to step towards the minimum mean squared prediction error
with
every
defines
the
iteration.
highly
This operation
efficient
steepest
plus some
gradient
supporting mathematics
descent
IMS
adaptive
67
(5.1)
It is easy to show that the one step prediction structure leads to the Lth
order maximum entropy
Griffiths, 1975).
Keeler and Lee (1978) have shown how the complex, first
order, one step prediction filter yields the pulse pair frequency estimator,
which has been made adaptive.
Keeler
What
makes
difference
adaptive
in
prediction
correlation
time
and
SNR
of the
enhancement
desired
possible
narrow band
is
the
signal
(or
Doppler shift
and sea
clutter may be sufficiently offset from zero Doppler that an adaptive scheme
may provide adequate suppression in both cases.
5.2.2
direction
while
simultaneously
nulling
interfering
sources
and
maximizing the signal to interference plus noise ratio at the output of the
adaptive beamformer (Haykin, 1985a; Compton, 1988).
troubled by interfering ground clutter returns and could benefit from using
an adaptive antenna.
place a line of nulls along the dominant ground clutter return angles near
0 elevation.
interfering sources
UHF
for
Forming nulls in the array antenna patterns in real time as the interferers
become active or as the antenna elevation increases may be feasible in many
cases.
Constraints on the adaptation speed and antenna scan rates may limit
68
thousand array elements and a digital control system for a truly adaptive 3
dimensional
beam.
Cost is a limiting
(Mailloux,
1982).
Recently, Sachidananda,
contributions
to velocity
5.3
UITI-CHANNEL PRDCESSING
achieve
algorithms
will
channels
greater
become
computational
more
can be thought
common.
of as
power, multi-channel
The
vector
signals
from
processing
separate
input
(Marple, 1987).
are found by solving a set of linear equations similar to the single channel
equations used in linear prediction filtering and associated applications.
Wiggins and Robinson
"normal" equations.
parameters. Each input data point can be thought of as a 2x2 matrix, the
polarization matrix, rather than a complex I and Q estimate. The set, or
vector of these matrix inputs is then processed using complex matrix
algorithms which are designed to optimally and jointly estimate target
parameters.
Processing both channels simultaneously yields additional
information that could not be obtained if they were processed independently.
Integrated sensor systems can benefit by multi-channel processing schemes.
A multi-channel algorithm might make use of 10 minute wind profiler data and
1 minute radar or lidar data.
produced with lower error than either system operating alone. Application
of coherence functions to these multi-channel sensors provides an analytic
tool for correlated data which improves the analysis.
5.4
A IPRICI INFR4ATICN
70
6.
serve
as the interface between the radar system and the radar data
processing system (usually a host computer). These control tasks include:
1.
2.
Polarization switching
3.
Phase sequencing
4.
5.
6.
7.
generates the PRF, although good practice dictates that the basic clock be
derived from a reference oscillator that is shared between the processor and
PRF control by the processor minimizes the possibility of range
bin jitter caused by timing uncertainties in the A/D sampling and is
the radar.
71
and phase characteristics in the last few range bins for each transmitted
pulse or during antenna repositioning intervals with the transmitter off.
These bins are then processed identically to all other bins. The output
values can be monitored in real time to verify that the system is
functioning properly, and for system power calibration.
In addition, the
injected signal can be made coherent so that the Doppler processing can be
checked.
range capabilities (90-100 dB). The linear channel measurements are used
for extracting information related to the phase of the signal, i.e., mean,
velocity, spectral width and clutter measurements, and can provide power
estimates as well. The linear channel measurements operate over a more
restricted dynamic range, typically z40-60 dB, that is usually shifted by
means of an automatic gain control (AGC) loop over a range of ~100 dB. It
is the linear channel gain control problem where digital signal processing
methods:
1.
IF limiting
72
2.
3.
Slow AGC
4.
Fast AGC
5.
Multiple receivers
(Nathanson, 1969;
Zeoli, 1971;
Frush, 1981 ).
retrieval.
clutter mixed with a weather spectrum, this technique tends to "capture" the
detectability.
For the slow AGC, the log receiver measurements from the previous ray are
used to optimize the linear receiver gain for the targets that are actually
present at each range. The samples for an integration period are collected
If the log receiver is used for
while the gain is held constant.
quantitative power measurements, the actual gain does not need to be known
The fast AGC, or instantaneous AGC (IAGC) approach, for which the gain of
the linear receiver is adjusted for each range and each pulse, is used where
there is a high degree of phase purity in the transmitted pulse
(e.g.,
klystron systems). The power measurement for either the previous pulse, or
the current pulse (in which case a delay line is required) is used to set
the receiver gain. This is the most complicated form of AGC to implement
since it requires a very accurate calibration of both the amplitude and
phase response of the receiver as a function of gain and the input power.
Figure 6.1 shows a block diagram of a typical digital, time domain Doppler
signal processor.
The digital signal processor provides the interface to
the radar I, Q and log signals, and connects to a host computer that
provides the
communications.
user
interface,
data
processing,
display
and
After analog phase detection, the I, Q and log values are digitized.
data
In the
case of a fast AGC, a digital AGC value may also serve as an "exponent" for
g
ication
Fig 6.1
75
signals
and one bit denotes the sign of bipolar data, the usable
instantaneous dynamic range is limited to ~54 dB for 12 bit samples. This
range provides a margin for an AGC that may not optimize its use of the
receiver dynamic range and offers reasonable clutter rejection. For the log
channel, the quantization of the digitized signal determines, to some
extent, the accuracy of the final power estimates. However, it is usually
the inherent large fluctuation of 30 dB for Rayleigh signals (Nathanson,
1969) that imposes the more fundamental limit.
The A/D converter values should not saturate.
increases the SNR by N, but also increases the dynamic range by 10 log N.
The discussions above illustrate the need for time series and power spectrum
displays
Just as
important,
the host
6.3
REEIETlVlTY PROCESSING
(6.1)
where Pk is the current estimate of average power based on the new sample pk
and the previous estimate Pk-1 . C is a weighting constant between 0 and 1.
When C is close to 1, the current pulse is more strongly weighted. This
technique is extremely simple to implement in real time and provides a new
estimate for each pulse. Since real time digital processing capabilities
have improved, and analog CRT displays are rapidly being replaced by color
displays, this technique has been largely replaced by a simple uniformly
weighted average over a fixed number of pulses.
Averaging of independent samples is required to obtain accurate reflectivity
estimates.
imposes a fundamental constraint on the scan rate for data collection. For
example, at 3 rpm and 500 Hz PRF, one can average only 27 pulses per degree
of antenna rotation. Depending on the wavelength and the spectrum width of
the scatterers, not all of these pulses will be independent.
A technique
a range bin spacing that is greater than the pulse width. This requires
somewhat more processing power, but results in more accurate reflectivity
estimates. Also, averaging can be adjusted as a function of range so that
the resulting
dimension.
average range
interval
77
The conversion from dBm to dBZ is done via the radar equation which involves
the radar constant and range normalization. The term "STC" is sometimes
inappropriately used to refer to the digital range normalization that is
performed in the processor.
technique that was used in the past to represent the radar reflectivity on
CRT display. Digital range normalization merely adjusts the output values
appropriately without causing the loss of sensitivity at short range.
6.4
'THRESHOEIDfING
R DATA QUALITY
The goal of thresholding is to have the signal processor flag data that
transmitting radar data and products and reduce the archive resources that
are required to store them.
There are numerous thresholding criteria and variables that are employed in
modern radars:
1.
2.
3.
4.
Clutter-to-signal power
5.
Zero velocity
6.
Geometric criteria
7.
Statistical criteria
This criterion is
most commonly used to threshold the wide dynamic range power measurements
(e.g., from a log receiver). The coherent signal-to-noise power is the area
78
under the signal portion of the power spectrum divided by the total noise
power (S/N). It can be calculated directly from the spectrum, or using the
measured autocorrelations. Similarly, the spectrum width itself can be used
as an indicator of the accuracy of the Doppler mean velocity and spectrum
width.
Both a low coherent signal-to-noise ratio and a large spectrum width
contribute to a large variance in the velocity and width estimators.
Ideally, thresholding should be made at a constant variance level, e.g.,
velocity is accepted if it's expected error is less than 1 m/s.
Unfortunately the relationship that governs the effect of SNR and width on
the variance of the velocity estimator is not a simple one (Zrnic, 1977b),
hence it is usually not implemented as a real time thresholding criterion.
Instead, the typical approach is to use either the coherent SNR and/or the
width separately and adjust the threshold until the displays are reasonably
free of speckles.
A popular measure of the quality of velocity and width estimates, which
accounts for the effects of both the coherent SNR and the spectral width is
conveniently
bounded
between
0 and
It is
and thresholds
The measured
clutter-to-signal ratio
purpose of correcting the log receiver power for the effects of clutter.
When the actual CSR exceeds the dynamic range capabilities of the receiver
or the ability of the clutter filter to accurately remove clutter, then the
data should be discarded.
thresholding criterion.
Another method of thresholding range bins that are affected by clutter is to
simply not display bins that have a mean velocity within a narrow band about
zero velocity. This technique is effective for Doppler radars that have no
clutter filter, or Doppler radars of limited linear dynamic range available
for
cancelling
clutter.
Both
the velocity
79
and reflectivity
can be
a "speckle remover" that eliminates all isolated range bins that have no
nearest neighbors in range or azimuth. Use of a speckle remover eliminates
aircraft and point clutter targets. It also allows other thresholds to be
set to lower values for greater sensitivity since only double speckles will
be passed.
Finally, statistical criteria involve considerations of local continuity and
rejection of data that are a few standard deviations away from local mean
values.
Strauch,
et
al.
(1984) utilize
a very
effective
"consensus
80
7.
7.1
IN SIGNAL PROCESSING
ITREND
REALIZATIC[N FACT
Several key components comprise a realizable signal processing system-chips, memory, and a large bandwidth output device. This digital technology
has found wide applications in modern radars (Rabinowitz, et al., 1985).
7.1.1
In the last 5 years integrated circuit chips specially optimized for digital
signal
(DSP)
processor
and
memory,
the
operations
logic
supporting
multiply-accumulate, on-chip
such as
have
developed
computational
power
on
commercially available high speed busses, such as VME and Multibus II.
As
available
from
variety
of
and
manufacturers
can
be
to
allow
caching
and
make
DSP
algorithm's
more
efficient.
and
can achieve clock rates of tens of MHz and execution rates of a few Million
Instructions Per Second (MIPS). The next generation of microprocessor and
DSP chips will be fabricated from gallium arsenide (GaAs) and will allow
several processors to be attached to a single chip component. Clock rates
for these advanced devices will be a few hundred MHz with instruction rates
81
chips that are compatible with previous versions. Thus, a relatively simple
redesign of the processor board using the same basic architecture, combined
with reprogrammed algorithms, offers greatly enhanced processing power at
low cost.
The ready availability of the processing power obviates a move towards more
The half inch tape is the standard, but various other tape-based media and
technologies are being explored. These include special high density tapes
such as NCAR's obsolete TBM (terra-bit memory), magnetic tape cartridges,
video cassettes, and the digital audio tape (DAT) devices using helical scan
technology.
All of these tape storage media suffer from serial access
delays and are undesirable for on line, fast access storage. However, they
are extremely well suited for "write-once" archiving applications such as
radar data acquisition. Storage capacities of two or more gigabytes can be
achieved today. Higher capacity and faster transfer rates will continue to
evolve. Winchester disks using "vertical recording" techniques allow high
density and fast access and fit many applications which require fast, random
access storage.
The thrust in storage media development now seems to be in optical recording
techniques.
82
7.1.3
Display technology
Real time color radar displays have become an important component of remote
Intensity modulated PPI and RHI scopes show high resolution reflectivity
displays, but digital color displays show all the directly measured
variables (e.g., velocity) as well as derived variables such as differential
reflectivity, phase, and depolarization quantities.
Plotting data from
multiple sensors in real time, zooming into specific areas of interest,
generating time lapsed images, and defining special overlays provides a
measure of flexibility not available only a few years ago. Special purpose
programmable graphics processors allow these new, yet fairly simple, image
processing capabilities. The next generation of graphics processors will
accommodate 3 dimensional real time image generation, color images with
engineering
improvements
in
of
signal
hardware
processors
technology
and
is
changing
architectures
because
(Allen,
of
the
1985).
engineering methodology.
(e.g., POSIX), for computer language (e.g., ANSII standards, Ada, etc.),
run-time environments
bus standards,
VME)
(e.g.,
and
Data
7.2
IREN1&S IN I
The
new generation
sensors
EGRA4TBILr[TY OF DSP
of digital
is programmable.
processors
in
which
signal processors
for atmospheric
remote
the
algorithms
could
be modified
only
with
great
difficulty and most often resulting in the loss of the original capability.
Programmable
processors
allow
algorithm
modifications,
processing
Modern
Schmidt,
programmable
signal
et
al.
(1979)
processors
and Woodman,
for VHF
et
al.
Doppler wind
(1980) describe
profilers.
These
present day DSP systems are directly programmable in modern languages, such
as "C".
processing
algorithms
using
adaptive
and
data
compression
techniques (Haykin, 1985a; Kay, 1987; Marple, 1987) can be coded and tested
on
line
in
real
implementation.
be modified.
As
destroying
the
original
algorithm
easier
algorithms,
without
becomes
time,
for
support
to
higher
achieve.
example
in
level languages,
Reproducability
a
radar
network,
of
is
algorithm portability
clone
processors
feasible.
and
However,
problems.
for
trouble-shooting
radar
malfunctioning
required
level
Board
processor.
Programmability
is
the
application
of
multiprocessor
operating
systems.
management.
purpose
general
processor chips
68030),
but
future
Software
system of choice for many applications programs and for development of real
time software, which then typically run under a UNIX compatible real time
operating system (e.g., VxWorks, PDOS).
7.3
SHIOR
TERM EXPECCTATICNS
We have lived by
explored that in same instances provide better parameter estimates but have
not been feasible to implement in the past.
disappearing.
7.3.1
Rarge/velocity ambiguities
new
dealiasing.
pulsing
processing
schemes
for
and
range
velocity
clutter
overlaid
echo
suppression
suppression
simultaneously
algorithms.
85
with
There
velocity dealiasing
will
be
exploration
and
of
The
mechanical solution of simply scanning faster and using short dwell times is
of the return signal. Some research groups are testing short dwell time
waveforms (Keeler and Frush, 1983b; Strauch, 1988) on both airborne and
ground-based weather radars.
86
7.3.4
Data cmpression
encoding"
or
for strings
Typically
parameter
"compaction"
estimates
not
and
"run length
passing
some
threshold
test
are
7.3.5
Artificial intelligence
from
airborne
and ground
based meteorological
stations.
Graphics computers with large video memories allow time lapsing of high
resolution 3D images and arbitrary cross sections to be displayed using a
variety of techniques currently being developed.
Transparency of data
elements near the viewer allows observation of the storm interior.
7.4
Several years from now we can expect revolutionary changes in the way signal
processing will increase our ability to understand atmospheric dynamics in
87
real time.
Combining new hardware forms and more efficient software
development techniques with evolving communications technology and the
tumbling cost of computing power will allow remote sensor systems to present
interactive user interface taking forms that are only dreamed about today.
Advance hardware
The present development of GaAs (gallium arsenide) computing elements may
replace silicon dominated chips if the promised five fold speed increases
and higher reliability in thermal and radiation extremes are realized.
7.4.2
Fiber
Ctumunicatians
Military budgets have financed the development of highly efficient, very low
sidelobe,
multiple
beam, two
dimensional
electronically
scanned
array
The
very
for
high
cost
of
this
performance
increase
must
be
justified
7.4.5
Self
Adaptive systems
learning,
optimization that is
systems will
allow a
degree
of
load
so
that
the
overall
computational
efficiency
is
89
8.
8.1
CoNCrI3SIONS
least, have taken a somewhat narrow view of signal processing in the past.
A large effort has been dedicated to using the pulse pair algorithm for
estimating the first two or three spectral moments, largely because the
existing processing power has been rather limited to these simple algorithms
and because for an important class of signals the pulse pair algorithm is
optimum. Advances have been made in the ST/MST radar community in pulse
compression, coherent averaging, and non-linear least squares parameter
fitting techniques, and in the lidar community in multiple lag processing.
Other techniques have been ignored or rejected simply because the scientific
need for these advances did not exist, or if it did, the risk of undertaking
such a development was not warranted.
The operational radar community and many researchers have been unable to
explore weak echoes because of inadequate sensitivity. There are better ways
of improving radar sensitivity than brute force techniques of more power and
larger antennas. Advanced signal processing techniques must be explored
more thoroughly to achieve these sensitivity gains.
Modern spectrum
analysis methods for modeling distributed target echoes in strong clutter
and multi-channel processing techniques to extract better information from
collections of remote sensors is an area ripe for extensive research.
The digital boundaries of the signal processor are being extended in both
directions. Digital IF quadrature mixers are presently available which will
accept IF and local oscillator analog signals and put out digitized I and Q
samples.
91
8.2
RECEMMENDATIONS
R CU
FOR RE
these sessions (and vice versa) since our target models, our propagation
medium, our processing problems, and our techniques are nearly identical.
As noted before members of these communities have successfully explored
modern algorithms and predated weather radars use of the pulse pair and poly
pulse pair velocity estimators as well as use of pulse compression and
complementary coding schemes.
Finally, as R.W. Lee of the Signal Processing panel stated once, we can now
build processors with "megaflops to burn". We can use them very easily by
implementing new processing algorithms, for example, using a priori
knowledge to improve estimates. Computing special diagnostic outputs which
have no bearing on the data collected, but simply allow the operator to
adjust processing parameters, is an effective use of processing power.
92
8.3
New techniques are not usually accepted easily by any scientific community.
Twenty years ago, Doppler processing using the now standard pulse pair
estimator was not readily accepted.
1.
2.
4.
8.4
ACKNoICMrEDGMENT
The authors wish to thank the panel members for their verbal and written
contributions to this report.
helpful with comments
on various drafts.
V. Chandrasekar, J. Evans, G.
and R. Woodman
and have provided helpful written comments that have been incorporated into
this signal processing review.
93
ACIHNYM isr
A/D
analog to digital
AGC
CNR
CSR
DFT
DSP
FFT
FIR
FM-CW -
IIR
IF
intermediate frequency
I/Q
in-phase / quadrature
IMS
ME
maximum entropy
ML
maximum likelihood
PRF
PRT
SNR
95
BI BLT OQ"A BI Y
Alexander,
S.T.,
Applications.
1986a:
New York:
1986b:
IEEE Trans. Acoust.,
_______,
Adaptive Signal
Springer Verlag.
Processing:
Theory
and
signal
Alldritt, M., R. Jones, C.J. Oliver, and J.M. Vaughan, 1978: The processing
of digital signals by a surface acoustic wave spectrum analyzer.
J.
Phys. E: Sci. Instrum., 11, 116-119.
Allen, J.,
1985:
Computer architecture
Proc. IEEE, 73, No. 5, 852-873.
for digital
signal processing.
Anderson, C.M., E.H. Satorius, and J.R. Zeidler, 1983: Adaptive enhancement
of finite bandwidth signals in white Gaussian noise.
IEEE Trans.
Acoust., Speech, and Sig. Proc., ASSP-31, No. 1, 17-28.
Anderson, J.R., 1981: Evaluating ground clutter filters for weather radars.
Preprints 20th Conf. Radar Meteor., AMS, 314-318.
, 1987:
The measurement of Doppler wind fields with fast
scanning radars:
signal processing techniques.
J. Atmos. Ocean.
Tech., 4, no. 4, 627-633.
Anderson, N., 1978:
Comments on the performance
algorithms. Proc. IEE, 66, 1581-1582.
of
maximum
entropy
MA:
Dedham, MA:
Artech
Battan, L.J.,
1973:
Radar Observation
University of Chicago Press.
of
the
Atmosphere,
Chicago:
Bello, P.A.,
1965:
Some techniques for the instantaneous real-time
measurement of multipath and Doppler spread.
IE:E Trans 9 on Comm.
Tech., COM-13, No. 3, 185-292.
Benham, F.C., H.L. Groginsky, A.S. Soltes, and G. Works, 1972: Pulse pair
estimation of Doppler spectrum parameters. Final Report, Contract F19628-71-C-0126, Raytheon Company, Wayland, Mass.
Bergen, W.R., and S.C. Albers, 1988: Two and three dimensional de-aliasing
of Doppler radar velocities.
J. Atmos. Ocean. Tech., 5, No. 2, 305319.
Berger, T., and H.L. Groginsky, 1973: Estimation of the spectral moments of
pulse trains. Paper presented at the Int. Conf. on Information Theory,
Tel Aviv, Israel.
Bilbro, J.G., D. Fichtl, D. Fitzjarrald, M. Krause and R. Lee, 1984:
Airborne Doppler lidar wind field measurements. Bull. Amer. Meteor.
Soc., 65, 348-359.
, C. DiMarzio, D. Fitzjarrald, S. Johnson, and W. Jones, 1986:
Airborne Doppler lidar measurements. Appl. Opt., 25, 3952.
Blackman, R.B., and J.W. Tukey, 1958:
New York: Dover.
Bode, H.W., and C.E. Shannon, 1950: A simplified derivation of linear least
squares smoothing and prediction theory. Proc. IRE, 38, 417-425.
Boren, T.A., J.R. Cruz, and D. Zrnic, 1986:
An artificial intelligence
approach to Doppler weather radar velocity de-aliasing. Preprints 23rd
Conf. Radar Meteor., AMS, JP107-110.
Bracewell, R., 1965:
McGraw-Hill.
98
New York:
Radar Technology.
Dedham, MA:
Artech House.
Browning, K.A., P.K. James, D.M. Parkes, C. Rowley, and A.J. Whyman, 1978:
Observations of strong wind shear using pulse compression radar.
Nature, 271, 529 - 531.
Burg, J.P., 1967: Maximum entropy spectral analysis. 37th Annual Intl. Soc.
Exploration Geophysicists Meeting, Oklahoma City, OK, Oct. 31.
, 1968: A new analysis technique for time series data.
NATO Advanced Study Institute on Signal Processing with Emphasis on
Underwater Acoustics Vol 1, paper #15, Enschede, Nederlands.
, 1972:
The relationship between maximum entropy spectra and
maximum likelihood spectra. Geophysics, 37, No. 2, 375-376.
, 1975:
Maximum entropy spectral analysis. Ph.D. Dissertation,
Dept. of Geophysics, Stanford University.
Cadzow, J.A., 1980:
High performance spectral estimation -- a new ARMA
method.
IEEE Trans. Acoust., Speech, Signal Processing, ASSP-28, No.
5, 524-529.
, 1982:
Spectral estimation: An overdetermined rational model
equation approach.
Proc. IEEE, 70, No. 9, 907-939 (Special issue on
Spectral Estimation).
Campbell, W.C., and R.C. Strauch, 1976:
Meteorological Doppler radar with
double pulse transmission. Preprints, 17th Conf. Radar Meteor., AMS,
42-44.
Capon, J., 1969:
High-resolution frequency-wavenmber spectrum analysis.
Proc. IEEE 5,7, 1408-1418.
Caton, P.A.F., 1963:
213-222.
P.1
.4 a
.4.,A........
Chandrasekar, V., G.R. Gray, V.N. Bringi, and R.J. Keeler, 1989:
Efficient
differential
receivers.
J.
reflectivity
processing
using
logarithmic
IEEE
Press
Selected
Adaptive Antennas.
and M.
Bernfeld,
1967:
J.W.,
and J.W.
Tukey,
Key
to
Radar Signals:
more
Prentice
efficient
radar
An
Introduction-- to
I
McGraw-Hill.
1965:
Radio Science,
100
detection.
Uncertainties in
, R.J. Serafin, and L.C. Peach, 1972:
coherent measurement of the mean frequency and variance of the Doppler
Preprints 15th Conf. Radar
spectrum from meteorological echoes.
Meteor., AMS, 216-221.
Doppler radar with polarization
Doviak, R.J., and D.S. Sirmans, 1973:
diversity. J. Atmos. Sci., 30, 737-738.
Considerations
, D. Sirmans, D. Zrnic, and G.B. Walker, 1978:
for pulse-Doppler radar observations of severe thunderstorms. J. Appl.
Met., 17, No. 2, 189-205.
, D.S. Zrnic, and D.S. Sirmans, 1979:
Proc. IFFE, 67, No. 11, 1522-1553.
Doppler
radar.
weather
1987:
Optimised Radar
Processors.
London,
Peter
pulse
(RANDOP).
Fischler, M.A., and R.C. Bolles, 1981: Random sample consensus: A paradigm
for model fitting with applications to image analysis and automated
cartography. Commun. of ACM, 24, 381-395.
Foord, R.,
R.
101
1983:
Precise
SNR's
in
C02
laser
Proc.
Adaptive arrays -
an introduction.
Proc. IEE,
64,
1988:
1978:
Doppler radar probing of the clear
59, No. 9, 1074-1093.
A dual
Glover, K.M., G.M. Armstrong, A.W. Bishop and K.J. Banis, 1981:
Preprints 20th Conf. Radar
frequency 10 cm Doppler weather radar.
Meteor., AMS, 738-743.
Golay, M.J.E.,
82-87.
1961:
Complementary series.
Griffiths,
L.J.,
1975:
Rapid measurement of digital instantaneous
frequency.
IEEE Trans. on Acoustics, Speech and Signal Processing,
ASSP-23, 207-222.
Groginsky, H.L., 1965:
3, 7-13.
9, No.
, 1966:
Digital processing of the spectra of pulse Doppler
radar precipitation echoes. Preprints 12th Conf. Radar Meteor., AMS,
34-43.
,
parameters.
1972:
Pulse pair estimation of Doppler spectrum
Preprints 15th Conf. Radar Meteor., AMS, 233-236.
on Doppler
_____
spectral
Communication Systems.
, editor, 1985a:
Prentice Hall.
New York:
103
Cliffs, NJ:
1986:
Prentice
Hall.
The crmplex form of the maximum
, and S. Kesler, 1976:
method for spectra estimation. Proc. ITEE, 64, 822-823.
____
entropy
Hjelmfelt, M.R., A.J. Heymsfield and R.J. Serafin, 1981: Combined radar and
aircraft analysis of a Doppler radar "black hole" region in an Oklahoma
thunderstorm. Preprints 20th Conf. Radar Meteor., AMS, Boston, 66-70.
Adaptive Filters: Structures,
Honig, M.L., and D.G. Messerschmitt, 1984:
Algorithms, and Applications. Boston: Kluwer Academic Publishers.
Explorations in fixed and adaptive resolution at GE
Howells, P.W., 1976:
and SURC. IEEE Trans. on Antenna Prep., AP-24, 575-584 (Special issue
on Adaptive Antennas).
C02 laser Doppler systems for the measurement of
Huffaker, R.M., 1974-75:
and
turbulence. Atmospheric Technology, NCAR, 71.
atmospheric winds
Pulsed coherent lidar
, D.W. Beran, and C.G. Little, 1976:
Proc.
based wind field measurement.
systems for airborne and satellite
7th Conf. Aerosp and Aeronaut. Meteor., AMS, 318-324.
,_T.R. Lawrence, M.J. Post, J.T. Priestley, F.F. Hall, Jr.,
R.A. Richter, and R.J. Keeler, 1984: Feasibility studies for a global
Analysis of simulated
wind measuring satellite system (Windsat):
performance. Appl. Opt., 23, 2523-2536.
104
:1969:
IEEE
MIT
An approach to the
Trans.
Audio
and
Proc.
Chapter 24.
E.
, 1981b:
Estimation.
Englewood Cliffs,
N.J.:
and
C.L.
Frush,
1983a:
105
--
Coherent
wideband
processing
of
considerations.
1983b:
AMS, 284-290.
Khinchine, A.J.,
1934:
Correlation theory
processes. Math. Ann., 109, 604-15.
of
stationary
stochastic
Klauder, J.R., A.C. Price, S. Darlington, and W.J. Albersheim, 1960: The
theory and design of chirp radars.
Bell System Technical Journal, 39,
745-808.
Klostermeyer, J., 1986:
Experiments with maximum entropy and maximum
likelihood spectra of VHF radar signals.
Radio Science, 21, No. 4,
731-736.
Krehbiel, P.R., and M. Brook, 1968: The fluctuating radar echo.
Weather Radar Conf., AMS, Boston, 2-11.
Proc. 13th
, 1979: A broad-band noise technique for fastscanning radar observations of clouds and clutter targets. IEEE Trans.
on Geoscience Electronics, GE-17, 196-204.
Kretschmer, F.F. Jr., and B.L. Lewis, 1983: Doppler properties of polyphase
coded pulse compression waveforms. IEEE Trans.on Aerosp. Electron.
Syst., AES-19, .No. 4, 521-531.
Lacoss, R.T., 1971:
36, 661-675.
Geophysics,
Lee,
R.W., 1978:
Performance of the poly-pulse-pair Doppler estimator.
Lassen Res. Memo 78-03.
106
Aspects of Radar
Ihermitte, R.M.,
1960:
The use of special "pulse Doppler radar" in
measurements of particle fall
velocities.
Proc. 8th Wea. Radar Conf.,
269-275.
, 1972: Real time processing of meteorological Doppler radar
Preprints 15th Conf. Radar Meteor., AMS, 364-367.
signals.
radar.
Measurements,
Oct.
1987.
A.
Deepak
Ligthart, L.P., L.R. Nieuwkerk, and J. van Sinttruyen, 1984: FM-CW Doppler
radar signal processing for precipitation measurements. Preprints 22nd
Conf. Radar Meteor., AMS, 538-543.
Lippmann, R.P., 1987: An introduction to computing with neural nets.
Acoustics, Speech and Signal Processing Magazine, 4, No. 2, 4-22.
IEEE
Environmental Sensing).
Mahapatra, P.R., and D.S. Zrnic, 1983:
Practical algorithms for mean
velocity estimation in pulse Doppler weather radars using a small
number of samples. IEEE Trans. Geosci. Remote Sens., GE-21, 491-501.
Mailloux, R.J., 1982:
246-291.
107
, 1977:
Stable and efficient lattice methods for linear
prediction. IEEE Trans on Acoustics, Speech and Signal Proc., ASSP-25,
423-428.
, 1986:
Maximum confusion spectral analysis.
IEEE Third ASSP
Workshop on Spectrum Estimation and Modeling, Boston, 6-9.
Marple, S.L., 1987:
Digital Spectral Analysis with Applications, PrenticeHall, Englewood Cliffs, N.J.
Marshall, J.S., 1971:
Peak reading and thresholding in processing radar
weather data. J. Appl. Meteor., 10, 1213-1223.
,
and W. Hitschfeld, 1953:
Interpretation of the
fluctuating echo from randomly distributed scatters (Part
I).
Can.
Jour. Phys., 31, Pt. 1, 962-995.
McCaul, E.W. Jr., H.B. Bluestein, and R.J. Doviak, 1986: Airborne Doppler
lidar techniques for observing severe thunderstorms. Appl. Optics, 25,
No. 5, 698-708.
McClellan, J.H., 1982:
Multidimensional spectral estimation.
70, 1029-1039 (Special issue on Spectral Estimation).
Proc. IEEE
of
spectral
moments
of
time
series.
moment estimation.
, 1972:
A covariance approach to spectral
IEEE Trans. Inform. Th., IT-18, No. 5, 588-596.
108
Dedham, MA:
Proc.
Artech House.
Recursive multichannel
Trans. on Geoscience
IEEE
AIEE
IEEE Trans.
Signal
Processing,
Part I and
Digital
New York:
, 1981:
Maximum entropy and spectral estimation: a review.
IEEE Trans. on Acoustics, Speech and Signal Processing, ASSP-29, No. 6,
1176-1981.
clutter
Quart J.
Rabiner, L.R., and B. Gold, 1975: Theory and Application of Digital Signal
Processing. Englewood Cliffs, N.J.: Prentice-Hall.
Radoski, H.R., P.F. Fougere, and E.J. Zawalick, 1975: A comparison of power
spectral estimates and applications of the maximum entropy method. J.
of Geophysical Research, 80: 619-625.
Rastogi, P.K., and R.F. Woodman, 1974:
Jicamarca incoherent-scatter radar.
1217-1231.
Ray,
P.S.,
and C. Ziegler, 1977:
De-aliasing
estimates. J. Appl. Met., 16, 563-564
Reid, M.S.,
radar.
first-moment
Doppler
1969:
A millimeter wave pseudorandom coded meteorological
IEEE Trans. Geosci. Electron., GE-7, No. 3, 146-156.
Radio
Rife, D.C., and R.R. Boorstyn, 1974: Single tone parameter estimation from
discrete-time observation. IEEE Trans. on Info. Theory, IT-20, No. 5,
591-598.
Rihaczek, A.W., 1969:
McGraw-Hill.
Principles
of high resolution
110
Reading,
_____
Rottger, J.,
, J.
Klostermeyer, P. Czechowsky, R. Ruister, and G. Schmidt,
1978:
Remote sensing of the atmosphere by VHF radar experiments.
Naturwissenschaften, 65, 285-296.
Rummler, W.D., 1968a:
Introduction of a new estimator for velocity spectral
parameters.
Tech. memo. MM-68-4121-5, Bell Labs, Whippany, N.J.
, 1968b:
Accuracy of spectral width estimators using pulse
pair waveforms.
Tech. memo. MM-68-4121-14, Bell Labs, Whippany, N.J.
, 1968c:
Two pulse spectral measurements.
4121-15, Bell Labs, Whippany, N.J.
Tech.
memo MM-68-
, 1989:
Efficient processing of alternately
J. Atmos. Ocean. Tech., 6, in press.
Crosscorrelation properties of
Proc. IEEE, 68, No. 5, 593-619.
Proc. IRE,
1987:
IEE
Preprints 20th
five mean
Extension of maximum
, D. Zrnic, and B. Bumgarner, 1976:
unambiguous Doppler velocity by use of two sampling rates. Preprints
17th Conf. Radar Meteor., AMS, 23-28.
112
1980:
Radar Handbook.
New York:
McGraw-Hill.
New York:
McGraw-
Hill.
Smith, P.L., 1987:
The relationship between coherent integration
Doppler spectral processing of weather radar echoes.
J.
Atmos.
Ocean. Tech., 4, 541-544.
, 1986:
On the sensitivity of weather radars.
Tech., 3, No. 4, 704-713.
J. Atmos.
and
and
Ocean.
R.C.
Carbone,
1969:
Statistics
of instantaneous
Radio
R.G.,
1988:
meteorological radars
1985:
1146-1154.
airport
New
Proc.
Ulrych, T.J., and T.N. Bishop, 1975: Maximum entropy spectral analysis and
autoregressive decomposition. Rev. Geophys. and Space Phys., 13, 183200.
Van
entropy
a review.
Proc.
IRE WESCON
Adaptive
Englewood
moment estimation
in
MST radars.
the
Radio
__,
R.P.
Kugel,
and J.
Rottger,
115
1980:
and
31
A coherent integrator-
, 1980a:
High altitude resolution stratospheric measurements
with the Arecibo 430-MHz radar. Radio Science, 15, no. 2, 417-422.
, 1980b:
High altitude resolution stratospheric measurements
with the Arecibo 2380 MHz radar. Radio Science, 15, no. 2, 423-430.
Woodward, P.M., 1953:
Probability and Information Theory with Applications
to Radar. New York: Pergamon.
Zeidler, J.R., E.H. Sartorius, R.M. Chabries, and H.T. Wexler, 1978:
Adaptive enhancement of multiple sinusoids in uncorrelated noise. IEEE
Trans. Acoust., Speech, and Sig. Proc., ASSP-26, 240-254.
Zeoli, G.W., 1971: IF versus video limiting for two-channel coherent signal
processors. IEEE Trans. Inform. Theory, IT- 17, 579-587.
Zrnic, D.S., 1975a:
Moments of estimated input power for finite sample
averages of radar receiver outputs.
IEEE Trans. Aerosp. Electron
Syst., AES-11, No. 1, 109-113.
, 1975b: Simulation of weather-like Doppler spectra and signals.
J. Appl. Meteor., 14, No. 4, 619-620.
, 1977a:
Mean power estimation with a recursive filter.
Trans. Aerosp. Electron. Syst., AES-13, 281- 289.
IEEE
, 1977b:
Spectral moment estimates from correlated pulse pairs.
IEEE Trans. Aerosp. Electron. Syst., AES-13, 344-354.
radar.
, 1979a:
Estimation of spectral moments for weather echoes.
IEEE Trans. on Geoscience Electronics, GE-17, 113-128.
, 1979b:
Spectrum width estimates for weather echoes.
Trans. Aerosp. Electron. Svst., AES-15, Sept., 613-619.
IEEE
, 1980:
Spectral statistics for complex colored discrete-time
sequence. IEEE Trans. Acoust. Speech, and Sic. Proc., ASSP-28, No. 5,
596-599.
, and S. Hamidi, 1981:
Considerations for the design of ground
clutter cancelers for weather radar. Interim report, Systems Research
& Development Service, Report No. DOT/FAA/RD-81/72.
, S. Hamidi, and A. Zahrai, 1982: Considerations for the design
of ground clutter cancellers for weather radar. Final report, Systems
Research & Development Service, Report No. DOT/FAA/RD-82/68.
116
-----
117