Beruflich Dokumente
Kultur Dokumente
Abstract—A simple method for estimating parameters of the reflected by meteorological scatterers usually have a relatively
Doppler spectrum of atmospheric signals is described. The method narrow spectral band, which, in terms of the Doppler velocity,
is based on an adaptive filter processing that has been widely corresponds to the range of about 1–2 m/s. The width of the
used in telecommunications but rarely applied in measurements
using atmospheric radars. The method has been tested using signal spectrum is determined by turbulent fluctuations of the
both synthetic and real data obtained by the radar profiler— wind velocity, by the antenna beamwidth, and by other factors.
Transportable Atmospheric Radar (TARA)—and the weather The useful radar signals are mixed with the receiver broadband
radar—International Research Centre for Telecommunications noise so that the signal at the entrance of an estimator falls
and Radar (IRCTR) Drizzle Radar (IDRA)—both developed by within the frequency bandwidth of 10–50 m/s while the signal
IRCTR. The method is compared with the traditional pulse-pair
method and fast Fourier transform (FFT)-based methods, as well spectrum width is of 1–2 m/s. Methods based on FFT or on
as with the IRCTR FFT-based method implemented in the TARA autocovariance processing (the PP estimation) do not take into
and IDRA processing including an additional noise clipping stage. consideration this specific feature of meteorological radar oper-
The tests have demonstrated high efficiency of the adaptive fil- ation (i.e., that the narrow-band signal is mixed with the wide-
tering estimator under low signal-to-noise ratio conditions. The band noise) and, therefore, are not optimal from the statistical
applicability of adaptive estimations in radar meteorology and the
limitations associated with a sample length are discussed. point of view. The FFT processing can be further improved by
applying noise clipping, i.e., suppressing the Doppler bins that
Index Terms—Adaptive estimation, atmospheric measure- are close to the noise power level. However, this clipping has
ments, Doppler radar, spectral moments.
two main drawbacks, i.e., it increases the complexity of the
algorithm and suppresses part of the useful spectrum, which
I. I NTRODUCTION
can lead to an underestimation of the reflectivity or, in the case
adaptive methods are rarely used in atmospheric radar systems The equation that establishes the relationship between the mea-
[13], [16]. sured and the useful signals in two consecutive time instants
In this paper, the performance of a simple adaptive algorithm can be written as
(adaptive PP algorithm) is formulated and implemented for
processing real atmospheric radar data. The algorithm is based Xt+1 = aZt + εt+1 + ηt+1 . (6)
on the assumption that the Doppler spectrum of backscattered
signals is symmetrical and unimodal. Some comments are needed concerning the application of
AR(1) model for description of atmospheric signals. First, since
parameter a is assumed to be constant, the model implies that
II. A DAPTIVE F ILTERING P ROCESSING signals are stationary within the measuring time interval. This
A. Signal and Noise Models assumption is more valid for wind profilers that use 0.5–5-s
intervals for measurements and less valid for continuously scan-
We assume that, after discretization, a backscattered signal ning radars with a 50–100-ms measuring interval. Second, the
from atmospheric objects Zt has the form of a complex au- AR(1) model introduces a simple one-mode symmetric spec-
toregressive series of the first order, with an unknown complex trum. Although, in most cases, atmospheric signals have such
coefficient a kind of spectrum, there are cases of more complicated spectra,
for example, if scatterers are hydrometeors of different types or
Zt+1 = aZt + εt+1 εt = 0 εi ε∗j = σε2 δij (1) when ground clutter exists. Third, if one is interested in measur-
ing finer characteristics of Doppler spectra such as asymmetry,
where t = 1, 2, . . . is the discrete time, εt is a normal white it is impossible to use AR(1). Therefore, the simplicity of the
random series with variance σε2 , a is a complex parameter spectrum shape may limit the application of the model.
characterizing the signal, symbol “∗ ” denotes conjugation, the
brackets mean averaging over the number of samples, and
δij is the Kronecker symbol. The spectrum of signal (1) is B. Adaptive Filter Estimator Design
symmetric relative to the average Doppler frequency and can
The adaptive filter estimator consists of two parts—the esti-
be represented as [18]
mation part and the filtering part. If both signals Xt and Zt are
σε2 σε2 known, the estimation of parameter a, which is optimal in terms
Fz (ω) = σz2 = (2) of maximum a posteriori probability, can be obtained using the
|1 − a exp(−iω)|2 1 − |a|2
following recursive equations [12, p. 65]:
where σz2 is the power (variance) of the signal Zt and ω is Zt∗ (Xt+1 − ât Zt )
the angular frequency which lies within the range [−π, π]. The ât+1 = ât + θt
σε2 + ση2 + θt |Zt |2
frequency of the discrete signal, which is equal to −π or π,
corresponds to the Nyquist frequency of the continuous signal |Zt |2
and determines the maximum unambiguous velocity. When θt+1 = θt − θt2 (7)
σε2 + ση2 + θt |Zt |2
|a| → 1, and is less than 1, the signal can be approximated
by a quasi-sinusoidal series with a random phase. As shown in where the symbol “∧ ” denotes estimation. θt can be interpreted
Appendix A, the mean frequency of the spectrum ω coincides as the variance of estimation of ât . If there is no white noise,
with the frequency of its maximum. The standard deviation of ση = 0, Zt = X t , and the solution
of (7) provides a regular PP
the spectrum (spectrum width Δω) depends solely on |a| estimation â = Zt∗ Zt+1 / |Zt |2 , the proposed method can
thus be interpreted as an adaptive PP algorithm.
ω = Λ = arg(a)
If we assume now that a is known, the optimal recursive
∞
π2 (−1)k+1 k filtering of the signal is described by the following difference
(Δω)2 = (ω − Λ)2 = −4 |a| . (3) equations representing a particular case of the Kalman filter
3 k2
k=1
([12, p. 67]):
To estimate the mean Doppler frequency and the Doppler 2
σε + |a|2 Rt (Xt+1 − aẐt )
spectrum width, it is enough to estimate the complex variable a. Ẑt+1 = aẐt +
We assume now that the useful atmospheric signal at the σε2 + ση2 + |a|2 Rt
estimator input Zt is measured while being mixed with an 2
uncorrelated additive normal complex white noise ηt (in the σε2 + |a|2 Rt
Rt+1 = σε2 2
+ |a| Rt − 2 (8)
bandwidth of ± Nyquist frequency) having variance ση2 . The σε + ση2 + |a|2 Rt
measured signal Xt is therefore
where Rt is an average square of the filtering error. The
Xt = Zt + ηt (4) equations in (8) actually represent a digital linear filter with one
complex pole and an amplification coefficient which is variable
where in time.
The mutual substitution of (7) and (8) leads to a system of
εt = ηt = 0 ηi ηj∗ = ση2 δij εi ηj = Zi ηj = 0. (5) equations for the adaptive filter estimator. The complete system
PINSKY et al.: APPLICATION OF A SIMPLE ADAPTIVE ESTIMATOR 117
Fig. 1. RMS errors in estimation of the (left column) mean Doppler velocity and (right column) Doppler spectrum width for different values of Doppler spectrum
widths Δω and SNRs. The errors are normalized by the Nyquist frequency ΩN .
namely, the standard PP and FFT (error analysis of the latter to the traditional PP algorithm and, particularly, as compared
two can be found in [14] and [10]). A 1024-sample-long series with the FFT spectrum integration approach. If the SNR is less
was used in the test. The mean Doppler frequency was com- than −8 dB, only the proposed adaptive algorithm provides
puted under different SNRs, ranging from 1 down to −12 dB. a satisfactory quality (error of ∼0.5–1.0 m/s) of the mean
The FFT estimation of the mean frequency was performed frequency estimation. The minimum errors achievable when
using the FFT algorithm without clipping and the appropriate processing the signal with the ideal theoretical algorithm, eval-
summation of the spectral components. uated according to the Cramer–Rao inequality that determines
The results of Test 2 are shown in Fig. 3. One can see the theoretical minimum estimation errors (see [15]), are shown
significant advantages of the adaptive algorithm as compared in Fig. 3 as well.
PINSKY et al.: APPLICATION OF A SIMPLE ADAPTIVE ESTIMATOR 119
shows the errors in the spectrum width. The left panels show the
bias, and the right panels show standard deviation (STD) of es-
timation. The method demonstrates the ability of measurement
of the mean frequency for samples typical of profilers (several
hundreds of samples) up to SNR = −10 dB. The bias of esti-
mation is negligible so that the errors can be attributed to STD.
For short samples typical of continuously scanning radars, the
satisfactory results are obtained only for SNR > 0 dB. The
errors in estimation of the spectrum width are substantially
bigger. The estimations have a large positive bias. Application
of short samples does not allow one to measure the spectrum
width at high enough accuracy. At long (∼1000) samples,
SNR > 15−20 dB are required for getting reasonable results.
Nevertheless, in our opinion, it would not be correct to state that
the method cannot be used for measurements of spectrum width
at all.
To compare the different methods, more realistic signals of
Fig. 2. Example of convergence of synthetic signal spectrum parameters to
their true values. SNR is equal to 0.5, the mean frequency (normalized by Test 3 were processed by the PP and the FFT-based methods.
Nyquist) is equal to 0.5, and the Doppler spectrum width (normalized) is equal The comparison is shown in Fig. 5 which presents dependences
to 0.4. of the bias and the STD on SNR for sample lengths equal to
512 (upper row) and 32 (lower row). The results show that
the bias introduces negligible errors in all the cases for all the
methods. The STD error of the adaptive method is smaller than
that of other methods, which is particularly pronounced for the
smallest values of SNR. This illustrates the main advantages
of the adaptive method. It is noteworthy that both the PP and
FFT-based methods actually indicate the same STD errors.
This result can be explained by the existing of a theoretical
dependence between the first spectral moment and the first
derivative of the correlation function. The comparison of Figs. 3
and 5 allows one to see different errors arising in the FFT-based
method. This can be attributed to the difference in summation.
In Test 2, standard summation was used, while in Test 3, we
implemented circular summation that brings about nonbiased
estimations.
Fig. 4. Errors in estimation of the (upper row) mean Doppler velocity and (lower row) Doppler spectrum width for Gaussian Doppler spectrum under different
SNRs. The left columns introduce bias, while the right columns introduce STD. The errors are normalized by the Nyquist frequency ΩN . The mean normalized
frequency is equal to 0.16, and the normalized spectrum width is equal to 0.3.
modulation-continuous wave (FM-CW) radars, unlike in pulse a priori from the measured signal power and then applying the
radars, the range information is contained in the frequency calibration factor.
domain. In addition, the mean of each row is removed, which is The implementation of the adaptive signal processing and
equivalent to suppressing the 0-Doppler bin and consequently the PP signal processing in IDRA is done off-line as well. It
suppressing the signals from static ground clutter. is only slightly different from the TARA processing. In the
The first step in the FFT-based algorithm is to perform an first place, the raw data in the raw data files in IDRA were
FFT on the rows in order to obtain the Doppler spectrum windowed with a 2-D Hamming window in order to decrease
and to calculate the power per a Doppler bin. Doppler cells the sidelobes when FFT is performed. Therefore, the first step
below the preestablished clipping level (usually 10 dB above of the adaptive processing is to multiply the rows by an inverse
the estimated noise threshold) are then eliminated in order to Hamming window to compensate the windowing. FFT is then
diminish the influence of noise spikes in the calculation of performed along the columns to get the range information
the Doppler parameters. The Doppler parameters are there- and the subtraction of the mean of the rows to suppress the
fore calculated using only the Doppler cells with the power 0-Doppler as in TARA. The rest of the processing is analogous
above the clipping level. As for the adaptive algorithm, the to that of TARA.
Doppler parameters are calculated directly from the time series As for FFT-based processing in IDRA, a well-known issue
using (9) and (10) developed in Section II. The reflectivity is is the presence of internal distortion lines at all ranges at a
calculated by subtracting the estimated noise power measured particular Doppler frequency [6]. This distortion lines are
PINSKY et al.: APPLICATION OF A SIMPLE ADAPTIVE ESTIMATOR 121
Fig. 5. Dependences of (left) the bias and (right) the STD of the estimated mean velocity on SNR. The normalized mean velocity and the spectrum width are
equal to 0.16 and 0.1, respectively. The sample lengths are equal to (upper row) 512 and (lower row) 32.
effectively removed using polarimetric Doppler filtering. How- the most demanding in terms of sensitivity. The measurement
ever, the adaptive Doppler processing algorithm cannot take was performed using the horizontal polarization in the high-
advantage of polarimetry. Therefore, to ensure a fair compar- resolution (3-m) mode. The maximum Doppler velocity was
ison, the polarimetric Doppler filtering was removed from the 22.7 m/s. In Fig. 7, a light-rain event measured by IDRA on
standard data processing. August 3, 2008 at 12:00 P. M . UTC is shown. This measurement
The number of real operations required to obtain the Doppler was performed using the standard IDRA resolution (30 m). The
parameters using the FFT method is roughly 6N log2 N , where maximum Doppler velocity was 9.5 m/s. The example offers an
N is the number of sweeps used. The number of operations interesting case of aliasing, i.e., the actual Doppler velocity of
required for the adaptive processing is just 41N . Therefore, the the atmospheric scatterers exceeded the unambiguous interval
adaptive processing is much faster, particularly when a large of the system. This showcases the performance of each algo-
number of sweeps are used to compute the parameters. This rithm in the case of aliasing.
represents a significant advantage in the case when real-time In the three algorithms, it is assumed that the atmospheric
processing is required. At the same time, the PP algorithm has signals are stationary and symmetrical with respect to the mean
the largest computational efficiency since it requires 12N real Doppler velocity. However, atmospheric signals can be consid-
operations only. ered stationary just for a limited period of time, after which
Data Analysis and Comparison of the Algorithms: Two statistical characteristics of a signal change. For example, in
sets of measurements are presented as an illustration of the the case of TARA which is an S-band radar, the signals are
algorithms’ capabilities. The data sets shown in Fig. 6 were considered to remain stationary for a period of approximately
obtained by TARA during a clear air measurement, which is half a second (512 samples with a sweep period of 1 ms).
122 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011
TABLE II
S YSTEM S PECIFICATIONS
That limitation constitutes a serious drawback for the adaptive where the variance of the scattering plot is much higher for
processing. Equation (3) shows that the mean Doppler velocity lower values of reflectivity than for higher ones.
depends only on the phase of parameter a while the Doppler Interestingly, the mean Doppler velocity in the low-
spectrum width depends on the modulus of a. While the phase reflectivity areas retrieved by the adaptive processing is 0
tends relatively fast to the asymptotic value and, therefore, a (Fig. 7), whereas the velocity retrieved by the FFT-based al-
good estimation of the mean Doppler velocity can be obtained, gorithm is the expected one. This could be due to the influence
this is not true for the modulus of a. Small errors in estima- of the aforementioned distortion lines present in the IDRA. If
tion of modulus lead to significant errors in estimation of the that were the case, it would mean that the adaptive processing is
Doppler spectrum width, particularly in the case of a narrow less robust to artifacts. At the same time, the Doppler spectrum
spectrum. As a result, in most cases, an overestimation of the width is correctly retrieved by the adaptive processing in areas
Doppler spectrum width was observed, particularly for narrow where there is aliasing, whereas it is not the case for the
spectra, as can be seen in Fig. 2. FFT-based processing. This is more clearly seen in Fig. 10,
Since the adaptive processor filters out most of the noise that where the differences between the two Doppler spectrum
contaminates the measurements, the performance of the adap- widths are plotted in a plan position indicator (PPI).
tive processor is substantially better than that of the FFT-based The comparison of the results of signal processing of IDRA
algorithm without clipping. However, if the latter algorithm is estimated by the proposed method, the standard FFT, and the
combined with clipping, the performances are similar, since the classical PP algorithm is shown in Fig 7. The analysis shows
clipping of the signal acts as efficiently as an adaptive filter. In that the difference between the fields is very small. The dif-
case the SNR is very low, or the spectrum is wide, i.e., all the ference between the mean velocities does not exceed 20 cm/s.
Doppler cells are below the clipping level, the estimation of the This is not surprising since the measurements were performed
Doppler parameters cannot be performed using the FFT-based for quite strong signals with SNR > 10−20 dB. At such SNR,
method, whereas the adaptive method is still able to estimate the all the algorithms provide close values. It is noteworthy that, at
mean Doppler velocity (see Table III and Fig. 8). The areas that very low SNR when the adaptive filter has no time to form,
were completely suppressed by the noise clipping are marked the values provided by the proposed and the PP algorithms
white in Figs. 6 and 7. Moreover, due to the suppression of the should be expected to lead to close results as follows from (9).
Doppler cells that are below the clipping level in the FFT-based However, when the spectrum width is estimated, the PP algo-
algorithm, there is an underestimation of the signal power and, rithm often shows larger nonrealistic values (up to 10 m/s) as
hence, of the reflectivity. This can be clearly seen in Fig. 9, compared with the algorithm proposed in this paper. This fact
PINSKY et al.: APPLICATION OF A SIMPLE ADAPTIVE ESTIMATOR 123
Fig. 6. Data sets obtained by the two algorithms: (left column) Adaptive processing and (right column) FFT-based processing. The white areas in the FFT-based
processing are areas where information is suppressed as a result of clipping.
reflects the advantages of the presented method over the PP sented. The essential feature of this method is the presence of
method. a feedback connection which defines the adaptation of the filter
(change of frequency characteristics and amplification coeffi-
cient) for a particular signal. The method provides estimation of
V. C ONCLUSION AND D ISCUSSION
the Doppler spectrum in the parametric form. The performance
A new method based on an adaptive procedure and aimed at of the new method was compared with the traditional PP and
calculation of the Doppler spectrum parameters has been pre- FFT-based methods. The comparison shows that including both
124 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 1, JANUARY 2011
Fig. 7. IDRA parameters measured during a light-rain event. The left column presents reflectivity, the mean Doppler velocity, and the Doppler spectrum width
retrieved by the adaptive processing algorithm. The middle column presents the values of the same parameters retrieved by the standard FFT processing algorithm.
The right column presents the same parameters retrieved by the PP algorithm.
TABLE III filter adaptation and noise filtering into the estimation proce-
P ROCESSING PARAMETERS ( THE SNR WAS S ET AT 0.09)
dure results in much smaller errors in the estimation of the mean
Doppler frequency and the spectrum width. This improvement
is particularly pronounced under small-SNR conditions. How-
ever, the presented method provides comparable performance
with the FFT-based algorithm combined with signal clipping
procedure.
The two main advantages of the proposed method were
found. First, the number of required operations is proportional
to the sample length N and is thus significantly smaller
compared to the FFT algorithms (where it proportional to
N log2 N ). It makes the new method faster which is important
in real-time applications. Second, the proposed method is
able to estimate the mean Doppler velocity with low bias
and variance even in the case of very low SNRs, where the
FFT method is not applicable for estimation since the signal
is below the clipping level. It is particularly important when
PINSKY et al.: APPLICATION OF A SIMPLE ADAPTIVE ESTIMATOR 125
Fig. 8. Synthetic signal spectrum with and without noise. The clipping level
used in the FFT-based processing was 10 dB above the noise floor. The
signal was completely suppressed by the FFT-based algorithm, whereas the
adaptive processing algorithm was still able to determine the mean Doppler
velocity.
Fig. 10. PPI with the differences in the Doppler spectrum widths obtained by
the adaptive processing and by the FFT-based processing.
To calculate the integrals in (A2), ω 2 can be expanded into a samples. For the stationary case solution, we obtain
Fourier series within the range [−π, π]
2
P∞ σz2 + σμ2
∞
|Ẑt | =
π2 (−1)k+1 cos(kω) ∞ 1 − (1 − P∞ )2 |a∞ |2
ω2 = −4 . (A3)
3 k2 2P∞ (1 − P∞ )|a||a∞ | ∗
k=1
+ Ẑt Zt . (B4)
1 − (1 − P∞ )2 |a∞ |2
The utilization of (A3) and the tabulated integral [4]
Using mutual substitutions (B2)–(B4) and the last equation
π from (B1), one can obtain |a∞ | = |a|. Since the estimation
cos(kω)dω π
2
= |a|k (A4) of the spectrum width is determined by the absolute value
1 + |a| − 2|a| cos ω 1 − |a|2
−π of the complex number a, the estimation is asymptotically
nonbiased.
leads to the equation for the spectrum variance
∞
π2 (−1)k+1 k ACKNOWLEDGMENT
(ω − Λ)2 = −4 |a| . (A5)
3 k2 The authors would like to thank IRCTR for providing
k=1
the data.
R EFERENCES
A PPENDIX B [1] J. M. B. Dias and J. M. N. Leitao, “Nonparametric estimation of mean
A SYMPTOTIC U NBIASEDNESS OF THE E STIMATION Doppler and spectral width,” IEEE Trans. Geosci. Remote Sens., vol. 38,
no. 1, pp. 271–282, Jan. 2000.
The equations in (9) in the limiting case, when t → ∞, can [2] J. M. B. Dias and J. M. N. Leitao, “Asymptotically efficient estimation
be rewritten as of spectral moments,” IEEE Trans. Signal Process., vol. 43, no. 9,
pp. 2222–2225, Sep. 1995.
a∞ a∞ [3] R. J. Doviak and D. S. Zrnic, Doppler Radar and Weather Observations.
= + Ẑt∗ Xt+1 New York: Academic, 1993.
γt+1 γt [4] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products. New York: Academic, 1980.
1 1 [5] F. Fabry and R. J. Keeler, ”Innovative signal utilization and processing,”
= + |Ẑt |2 in Radar and Atmospheric Science: A Collection of Essays in Honor of
γt+1 γt
David Atlas, R. M. Wakimoto and R. C. Srivastava, Eds., Amer. Meteorol.
Soc. Monograph, 2003, ch. 8.
Ẑt+1 = (1 − P∞ )a∞ Ẑt + P∞ Xt+1 [6] J. Figueras i Ventura and H. W. J. Russchenberg, “IDRA: A new instru-
ment for drizzle monitoring,” in Proc. IGARSS, Barcelona, Spain, 2007,
Q 1 − |a∞ |2 + P∞ |a∞ |2 pp. 3301–3304.
P∞ = (B1) [7] R. J. Keeler and L. J. Griffiths, “Acoustic Doppler extraction by adaptive-
1 + Q (1 − |a∞ |2 ) + P∞ |a∞ |2 prediction filtering,” J. Acoust. Soc. Amer., vol. 61, no. 5, pp. 1218–1227,
May 1977.
[8] R. J. Keeler, “Adaptive frequency estimation and new convergence proper-
where the symbol ∞ is used to define constant values when ties for the least mean square algorithm,” Nat. Tech. Inf. Service, Spring-
t → ∞. We will show that a∞ = a, and therefore, the Doppler field, VA, NOAA Tech. Memo. ERL WPL-49, 1980.
parameters are asymptotically nonbiased. Averaging the first [9] R. J. Keeler and R. E. Passarelli, “Signal processing for atmospheric
radars, (A90-39376 17-47),” in Radar in Meteorology. Boston, MA:
and the second equation from (B1) over a great number of Amer. Meteorol. Soc., 1990, pp. 199–229.
samples and substituting Xt from (6) when t → ∞ lead to [10] P. R. Mahapatra and D. S. Zrnic, “Practical algorithms for mean velocity
estimation in pulse Doppler weather radars using a small number of
samples,” IEEE Trans. Geosci. Remote Sens., vol. GE-21, no. 4, pp. 491–
Ẑt∗ Zt 501, Oct. 1983.
a∞ = a ∞ . (B2) [11] L. P. Ligthart and S. H. Heijnen, “Transportable multi-beam Doppler
|Ẑt |2 polarimetric radar,” in Proc. 14th Int. Conf. Microw., Radar Wireless
∞ Commun., (MIKON), May 20–22, 2002, vol. 2, pp. 334–340.
[12] R. S. Liptzer and A. N. Shiryaev, Statistics of Random Processes 2.
To evaluate the correlation Ẑt∗ Zt ∞ , we multiply the third New York: Springer-Verlag, 1977.
[13] M. B. Pinsky and A. B. Sterkin, “Personal computer processing of
equation in (B1) by the signal equation (6) and average over Doppler meteorological radar signals,” Soviet Meteorol. Hydrol., vol. 1,
the samples. The stationary solution has the following form: pp. 93–100, 1993.
[14] D. Sirmans and B. Bumgarner, “Numerical comparison of five mean
P∞ σz2 frequency estimators,” J. Appl. Meteorol., vol. 14, no. 6, pp. 991–1003,
Ẑt∗ Zt = . (B3) Sep. 1975.
∞ 1 − (1 − P∞ )aa∗∞ [15] M. Skolnik, Radar Handbook, 3rd ed. New York: McGraw-Hill, 2008.
[16] V. M. Vostrenkov, A. A. Ivanov, and M. B. Pinsky, “Utilization of adaptive
It follows from (B2) and (B3) that the ratio in the right part of filtering in Doppler Meteorological Radars,” Soviet Meteorol. Hydrol.,
vol. 10, pp. 114–118, 1989.
(B2) is a real value, and hence, arg a∞ = arg a. That means [17] L. C. Westphal, Handbook of Control Systems Engineering. Norwell,
that the estimation of the average frequency is asymptotically MA: Kluwer, 2001.
nonbiased. [18] A. M. Yaglom, “An Introduction to the Theory of Stationary Random
Functions,” in Dover Phoenix Editions. New York: Dover, 2004.
In order to calculate |Ẑt |2 , we multiply the third equation [19] D. S. Zrnic, “Simulation of weatherlike Doppler spectra and signals,”
from (B1) by the conjugated equation and average over the J. Appl. Meteorol., vol. 14, no. 4, pp. 619–620, Jun. 1975.
PINSKY et al.: APPLICATION OF A SIMPLE ADAPTIVE ESTIMATOR 127
Mark Pinsky received the M.S. degree from Alexander Sterkin received the M.Sc. degree from
the Moscow Telecommunication Institute, Moscow, the Moscow State Institute of Radio-Engineering,
Russia, in 1975 and the Ph.D. degree from the Electronics and Automation, Moscow, Russia, in
Central Aerological Observatory (CAO), Moscow, 1994, specializing in biomedical electronics and
in 1992. His Ph.D. thesis is entitled “Processing of equipment, and the Ph.D. degree in brain research
Doppler meteorological signals.” from the Weizmann Institute of Science, Rehovot,
From 1975 to 1992, he was a Scientific Researcher Israel, in 2002.
with the Departments of Upper Atmosphere and For several years, he was with the Central Aero-
Radar Meteorology, CAO. He is currently a Senior logical Observatory, Moscow, focusing his research
Scientist with the Institute of Earth Sciences, The on Doppler radar data analysis. For a number of
Hebrew University of Jerusalem, Jerusalem, Israel, years, he was collaborating with the Institute of Earth
where he has been working in the field of cloud physics and cloud modeling Sciences, The Hebrew University of Jerusalem, Jerusalem, Israel, developing
since 1992. His major research interests lie in the fields of cloud physics, cloud software for visualization and analysis of remote sensing information and cloud
modeling, microphysical processes in clouds, atmospheric turbulence, theory modeling results. He is currently a Senior Application Engineer with Intel
of turbulence, radar meteorology, signal processing, and continuous symmetry Corporation, Israel, working on multicore software optimization.
measure.