You are on page 1of 6

Weather Radar Spectral Estimation

Using Sparse Models


Manuel Redondo Garcı́a∗1 , Martı́n Hurtado∗2 , Sebastián Pazos∗3 and Carlos H. Muravchik4
∗ Departamento de Electrotecnia, Facultad de Ingenierı́a, UNLP
La Plata, Argentina
1
megredondo@gmail.com
2
martin.hurtado@ing.unlp.edu.ar
3
sebastian.pazos@ing.unlp.edu.ar
4
carlosm@ing.unlp.edu.ar

Abstract—In weather radar systems, the echoes from GMAP is based on the estimated power spectrum of the
the weather phenomena can be contaminated by ground received signal, this technique is severely affected by the
reflections (clutter). If this interference is not properly leakage of clutter power into the weather spectrum, as a
removed, it produces biased estimates of the meteorological
variables, or it can even hide the signal of interest. In this consequence of processing finite datasets. When clutter-to-
paper, we analyze different formulations of sparse models for signal ratio is low or moderate, window functions ame-
estimating the Doppler spectrum and spectral moments of liorate the power leakage at expenses of larger errors in
the radar signal. The performance of the sparse models is the estimation of weather spectral moments. On the other
compared with classical ground clutter filters using synthetic hand, the parametric time domain method (PTDM) estimates
data; and then validated using real data provided by INVAP
through their weather radar located in Bariloche. simultaneously clutter and weather spectral moments apply-
ing the maximum likelihood method [4]. Since there is no
Resumen— En sistemas de radar meteorológico, los ecos analytical solution of this optimization problem, it is solved
de los fenómenos meteorológicos pueden estar contaminados numerically by iterative procedures [5]. Hence, PTDM is
por reflexiones de la tierra, denominadas clutter. Si esta computationally intensive and requires properly select seed
interferencia no se remueve adecuadamente, produce
estimaciones sesgadas de las variables meteorológicas, o values to avoid local maximums.
incluso puede ocluir la señal de interés. En este artı́culo
analizamos diferentes formulaciones de modelos ralos para In this work, we attempt to improve the estimation of the
estimar el espectro Doppler y los momentos espectrales de weather signal parameters by computing a better estimate
una señal de radar. El desempeño de los modelos ralos es of the signal power spectrum. We propose using sparse
comparado con filtros de clutter terrestre clásicos utilizando
datos sintéticos. Adicionalmente se presenta un ejemplo de formulations and sparse inverse algorithms to address this
funcionamiento con datos reales provistos por INVAP a través problem, specially when clutter and weather spectrums
de su radar meteorológico ubicado en Bariloche. overlap. In the sparse representation, the observed signal
accepts a linear representation over a dictionary, which is
I. I NTRODUCTION a collection of known waveforms (atoms) [6]. The repre-
One of the main tasks of signal processing for weather sentation is sparse if only a few coefficients are significant.
radars is ground clutter suppression in order to improve Frequently, the number of available observations is smaller
the data quality and system performance. When it is not than the number of atoms in the dictionary. Then, this
properly removed, ground clutter introduces bias in the linear problem becomes under-determined. Nevertheless, it
estimates of the spectral moments of the weather signals is still possible to solve the inverse problem, even if its
such as power, mean Doppler frequency and spectral width. support is unknown [7], [8]. Several tractable methods have
Because ground clutter represents the backscattered field been proposed for retrieving signal representations in over-
from targets on the ground, its spectrum is located close to complete dictionaries. Here, we consider LASSO and fused
zero Doppler frequency. Simple solutions consist of applying LASSO, thoroughly studied in the field of compressive
high-pass filters to the radar signals, also named moving sensing, which are also computationally intensive, but they
target indicator (MTI) [1]. These techniques fail when the do not require seed values for initiating the algorithms.
spectrum of the signal and clutter overlap. An alternative
approach is the use of clutter maps that control where clutter The paper is organized as follows. In the next section we
filters should be applied [2]. The maps are created a priori present the fundamental ideas involved in this work. These
from data collected on clear days. However, this technique include the parameters of interest in the weather application,
can not incorporate new clutter areas that may exist when and the classic algorithms to eliminate ground clutter and
there is anomalous propagation. estimate those parameters. In Section III, we introduce
Developments in signal processing systems allow imple- sparse models to estimate the power spectrum and include
menting more complex and adaptive algorithms for weather this in a modified GMAP algorithm. We analyze simulation
radars. The Gaussian model adaptive processing (GMAP) results comparing these approaches in Section IV. Finally,
not only filters the clutter but also attempts to recover we summarize the results and discuss the future work in
the weather spectrum overlapping the clutter [3]. Because Section VI.
II. F UNDAMENTALS 3) Remove clutter points: Use the three central spectrum
A. Weather Spectral Moments Estimates points to fit a Gaussian with the required width in
m/s, and discard the points within the Gaussian clutter,
The three most important spectral moments estimates in above the noise level.
weather radar data processing are the weather signal power, 4) Replace clutter points: Using the signal components
the mean Doppler velocity and the spectrum width [9]. In left, fit a gaussian to fill-in the clutter points removed.
the present article we use the Pulse Pair Processing (PPP) 5) Recompute GMAP with optimal window based on the
algorithm to obtain these estimates [10]. PPP assumes that clutter to signal ratio.
the ground clutter has been removed, and that the Doppler
Finally, with the resulting clean spectrum we compute the
spectrum consists only of white noise and a single spectral
spectral moments via PPP.
peak due to backscatter from weather-related phenomena.
This weather spectral peak is often assumed to be ap- C. PTDM Algorithm
proximately Gaussian shaped, and is characterized by its
amplitude, mean and standard deviation. This algorithm exploits the fact that the received signal
For each range bin, the receiver generates a slow time data samples can be modeled as a multivariate Gaussian proba-
sequence x[l], l = 0, . . . , L obtained from L pulses sampled bility density function, using an assumption that the Doppler
at the pulse repetition frequency (PRF). The autocorrelation spectra of clutter and precipitation have a Gaussian shape.
and power spectrum of this sequence are respectively Using this probability density function one can write the
negative log-likelihood function as follows
L−k−1
X
sx [k] = x[l]x∗ [l + k], (1) L(θ) = ln (det R(θ)) + tr (R(θ)Rx ) (6)
l=0
Sx (ω) = F {sx } = |X(ω)|2 . (2) where R(θ) is the parametric representation of the covari-
ance matrix, θ is the vector of unknown parameters, Rx
Based on these definitions, we define the following param- is the sample covariance matrix, and tr() is the trace. The
eters of interest. algorithm finds spectral moments estimates of a precipitation
1) Power: The zeroth moment of the Doppler spectrum signal and clutter by solving the minimization problem,
can be related to liquid water content or precipitation rate
in the resolution volume. The PPP estimate for the power min L(θ). (7)
θ
P̂ is
L−1 L−1
III. S PARSE M ODELS + GMAP
X X
P̂ = sx [0] = x[l]x∗ [l] = |x[l]|2 . (3) Sparse models seek to represent a signal as a linear
l=0 l=0 combination of basic functions, called atoms, belonging to
2) Doppler Velocity: The first moment of the power- a dictionary [6].
normalized spectra, represents essentially the air motion in In the specific case of spectral estimation, this dictionary
a radial direction, towards or away from the radar. The PPP is defined by each frequency component,
estimate for the velocity v̂ is T
A = ej2πtf , (8)
−λ
v̂ = arg{sx [1]}, (4) where t = (−L/2P RF . . . L/2P RF )T are the sample
4πT
times, f = (−P RF/2 . . . P RF/2)T are the frequencies
where T = 1/P RF is the pulse repetition period and λ is
of interest. Assuming that there are very few non-zero
the wavelength of the transmitted radar signal.
components, the spectrum has a sparse representation.
3) Spectral Width: The square root of the second moment
Using these models, given a number of measurements y ∈
about the first of the normalized spectrum. This parameter
CL , the problem that the theory of compressive sensing (CS)
is a measure of the velocity dispersion (shear or turbulence
solves is to find x ∈ CN such that x is sparse and complies
within the range bin). The PPP estimate for the width σ̂v is
with the following noise formulation form
s  
−λ sx [1]
y = Ax + z, (9)
σ̂v = √ ln . (5)
8πT sx [0]
where z is the stochastic noise term.
B. GMAP Algorithm Compressive sensing theory states that, under this sparse
This technique is based on a frequency approach that representation, the number of samples and data necessary
assumes Gaussian shaped spectrum to model clutter and to solve the inverse problem can be significantly reduced
weather signals. The algorithm consist of the following without affecting the performance of the system [7], [8].
steps: These techniques have already been used in several areas
1) Window and DFT: Apply a Hamming window to of signal processing, including image processing, bioengi-
the time series x[l] and obtain the Doppler power neering, communications, and remote sensing. Furthermore,
spectrum through the DFT. from a deterministic point of view, the solution of the
2) Dynamic noise power: Re-organize the spectrum com- recovery problem turns out to be [11]
ponents in ascending order of intensity, thus defining x̂ = arg min kxkℓ0 s.t. ky − Axk2ℓ2 < ǫ, (10)
a signal/clutter region and a noise region. x
where ǫ is related to the statistics of the noise. Since the
20
minimization of the ℓ0 -norm functional is an NP hard prob- True spectrum
GMAP
lem the problem is relaxed by using the ℓ1 -norm functional 0
PTDM
LASSO+GMAPM
as a proxy, resulting in Basis Pursuit [6].

Power [dB]
FusedLASSO+GMAPM
-20
Alternative CS recovery algorithms, based on a Bayesian
perspective, have been studied resulting in the following -40

MAP estimate formulation of the compressive sensing re- -60


trieval [12],
-80
x̂ = arg max P(x|y) s.t. x ∼ P(x), (11) -15 -10 -5 0 5 10 15
x v [m/s]

where P(x) is a suitable prior used to enforce the sparsity (a)


of x.
20 True spectrum
Assuming a Laplacian prior, which encourages sparse GMAP
solutions, the MAP estimate is the solution to an ℓ1 - 0
PTDM
LASSO+GMAPM
norm regularized least-squares problem, known as the least

Power [dB]
FusedLASSO+GMAPM
-20
absolute shrinkage and selection operator (LASSO) [13],
-40
x̂LASSO (µ) = arg min ky − Axk2ℓ2 + µkxkℓ1 . (12)
x -60

The regularization parameter µ is an upper bound of


-80
the Fourier spectrum of the residual. Then, the algorithm -15 -10 -5 0 5 10 15
v [m/s]
neglects those frequency components whose corresponding
residual energy is smaller than µ. A simple way of tuning (b)
µ is provided in [14]. Fig. 1: Examples of two spectrums and their reconstructions
The ℓ1 -norm can be used to enforce more general struc- using GMAP, PTDM, LASSO and fused LASSO where
tural constraints on the solution by replacing the sparsity GMAP and PTDM fail to estimate (a) spectral width and
constraint kxkℓ1 with kDxkℓ1 for a structured matrix D. A (b) Doppler velocity.
particular choice for this matrix, defined as
 
−1 1 0 · · · 0 0
 ..  scenario, and examples of the reconstructed spectrums for
 . −1 1 · · · 0 0  each of the methods described. It can be appreciated that in
D= .
.. .. . . .. ..  ,
 (13) this case, the sparse models approach offers a better estimate
 .. . . . . .  of the spectral width, compared to GMAP and PTDM, which
0 0 0 · · · −1 1 provide a larger width. Figure 1(b) shows another spectrum
promotes block sparsity, not directly on the coefficients but where GMAP and PTDM cannot resolve the weather signal
rather on their successive differences. This is equivalent to a and the clutter introduces an excessive bias in the estimates
1D discrete Total Variation norm [14], which is essentially of the Doppler velocity, whereas the sparse models are able
ℓ1 norms of derivatives used to promote piecewise constant to correctly identify the weather signal.
solutions. IV. S IMULATIONS
The fused LASSO combines a LASSO prior with a
In this section we discuss the performance of the proposed
TV-norm prior to promote simultaneously sparsity in the
modification to GMAP through simulations. Weather-like
coefficients and sparsity in the difference of successive coef-
data with Gaussian shaped Doppler spectrum were generated
ficients that is few coefficients and flatness in the coefficients
using the procedure described in [15]. We generate L
profile.
samples, resulting in L atoms in the dictionary, equivalent
x̂F L (µ1 , µ2 ) = arg min ky − Axk2ℓ2 + (14) to L frequencies. We use a carrier frequency f0 = 5 GHz.
x
We use the bias as a measure of the performance in the
+µ1 kxkℓ1 + µ2 kDxkℓ1 ) . (15)
estimation of the Doppler velocity,
This combination makes fused LASSO suitable for applica- MC
1 X
tions that involve block sparsity or mixtures of spikes and Bv = v − v̂i , (16)
flat plateaus such as the weather spectrum. M C i=1
In this article we introduce a modification to the GMAP the root of the mean square error as a measure of the
algorithm. Instead of computing a DFT of the windowed performance in the estimation of the spectral width,
samples to estimate the power spectrum, we apply sparse v
u
u 1 X MC
models through LASSO and fused LASSO. This modifica-
RMSEσ = t ||σv − σ̂vi ||2ℓ2 , (17)
tion to step 1 allows us to improve the spectrum estimation. M C i=1
In this modified process we apply a Blackman window to
the time series, estimate its spectrum, remove the clutter, and the mean square error for the power estimation,
interpolate a Gaussian target in the removed points, if MC
1 X P̂i
necessary, and then estimate the spectral moments via PPP. MSEP = , (18)
M C i=1 P
Figure 1(a) shows a typical spectrum in the weather radar
2 2
GMAP
PTDM
0 1.5
LASSO+GMAPM
fusedLASSO+GMAPM
Bias in v [m/s]

Bias in v [m/s]
-2 1

-4 GMAP 0.5
PTDM
LASSO+GMAPM
-6 fusedLASSO+GMAPM 0

-8 -0.5
30 35 40 45 50 55 60 65 3 4 5 6 7 8 9 10 11 12 13
L v [m/s]

(a) (a)
2 1.4
GMAP GMAP
PTDM 1.2 PTDM
1.5 LASSO+GMAPM LASSO+GMAPM
fusedLASSO+GMAPM fusedLASSO+GMAPM
RMSE σ [m/s]

RMSE σ [m/s]
1

1 0.8

0.6
0.5
0.4

0 0.2
30 35 40 45 50 55 60 65 3 4 5 6 7 8 9 10 11 12 13
L v [m/s]

(b) (b)
50 25
GMAP
40 20
PTDM
LASSO+GMAPM
15 GMAP
30 fusedLASSO+GMAPM
MSE P [dB]

MSE P [dB]

PTDM
10 LASSO+GMAPM
20 fusedLASSO+GMAPM
5
10
0

0 -5

-10 -10
30 35 40 45 50 55 60 65 3 4 5 6 7 8 9 10 11 12 13
L v [m/s]

(c) (c)
Fig. 2: Performance of GMAP, PTDM and CS algorithms Fig. 3: Performance of GMAP, PTDM and CS algorithms
with modified GMAP, for different number of samples L with modified GMAP, for different values of target Doppler
when estimating (a) Doppler velocity, (b) spectral width (c) velocity v when estimating (a) Doppler velocity, (b) spectral
power. width (c) power.

where the variable M C is set to 1000 runs of Monte Carlo modified version of GMAP as previously described. It can
simulations. For each Monte Carlo realization, we randomly clearly be appreciated that we obtain a significant improve-
generate the noise using a fixed clutter spectrum width σc = ment in the Doppler velocity bias, for the whole range
0.3 m/s, target power Pt = 1; clutter power Pc = 104 ; noise of number of samples. As expected, GMAP has a poor
power Pn = 10−2 and clutter Doppler velocity vc = 0. We performance for small number of samples. Similarly, we
analyze the performance in terms of the number of samples obtain improvements in the estimates of the spectral width
L used (equivalent to the number of pulses), the overlap and the power.
between the clutter and the weather signal, and the spread
of the weather signal. B. Doppler Velocity
Second, we analyze the behavior as we vary the overlap
A. Number of Samples between the clutter and the weather signal, equivalent to
First, we analyze the behavior of the new techniques as changing the value of the target Doppler velocity v from 3
we decrease the number of samples L from 64 to 32. This to 13 m/s. In this case we use a favorable scenario for the
is critical since weather radars are not usually capable to classic algorithms, with L = 64 samples, and target spectral
process a large amount of pulses. The target radial velocity width σv = 1 m/s.
used in this simulation is v = 8 m/s, with a target spectral Figure 3 also shows the errors in the spectral estimates in
width σv = 1 m/s. this scenario. Again, we obtain a significant improvement in
Figure 2 shows the performance measures for the the estimation of the spectral width. When there is overlap
spectral moments estimates of interest, using GMAP, the performance is similar to the classic algorithms, but
GMAP+SMLE, and LASSO and fused LASSO with the increases as the signal recedes from the clutter. There is
0 L = 32 pulses per bin. The systems uses a carrier frequency
f0 = 5625 MHz, and a spectral width for the clutter σc = 0.5
-1 m/s.
Bias in v [m/s]

GMAP Figure 5(a) shows the doppler spectrum for each bin using
-2 PTDM
LASSO+GMAPM
the DFT. Figure 5(b) shows the reconstructed spectrum using
fusedLASSO+GMAPM GMAP, where it can be seen that fails to remove all the
-3
ground clutter interferences. Figures 5(c) and (d) show the
LASSO and fused LASSO reconstructed spectrums where it
-4
0.5 1 1.5 2 2.5 3 can be appreciated that they successfully remove the clutter
σv [m/s] around the bin at range of 30 km. We note that LASSO and
(a) fused LASSO produce similar results; however, the image
2 of fussed LASSO is smoother since this algorithm promotes
GMAP block sparse solutions.
PTDM
1.5 LASSO+GMAPM
VI. C ONCLUSIONS
RMSE σ [m/s]

fusedLASSO+GMAPM

1
In this article we introduced a new approach to improve
the estimates of the spectral moments in weather radar
0.5
applications. This new approach is based on sparse models
with compressive sensing. Applying these techniques allows
0
0.5 1 1.5 2 2.5 3 us to obtain a better estimate of the power spectrum of
σv [m/s] the signal, thus providing improved clutter and noise can-
(b) cellation. We compared this novel technique with classic
25 algorithms showing the improved performance. Resulting in
notable gains in the estimation of the Doppler velocity and
20
spectral width. We additionally illustrate the performance
when applied to real radar data, showing a better clutter
MSE P [dB]

15
rejection and noise cancellation.
10
GMAP The work discussed in this article is not concluded. We
5
PTDM
LASSO+GMAPM
are currently working on applying the new approach when
fusedLASSO+GMAPM staggered PRF are used to extend the Doppler velocity
0
0.5 1 1.5 2 2.5 3
unambiguous range. Sparse models can easily address the
σv [m/s] problem of processing non-uniform sampled data collected
(c) when the radar operates in this mode. Additionally, we are
analyzing the effect of the introduced technique for real
Fig. 4: Performance of GMAP, PTDM and CS algorithms data in complete scenarios with different weather conditions
with modified GMAP, for different values of target spectral (mild and heavy storms).
width σv when estimating (a) Doppler velocity, (b) spectral
width (c) power. ACKNOWLEDGMENT
This work was supported by CIN-CONICET PDTS 269,
ANPCyT PICT 2014-1232, PIP GI 2015-2017 and UNLP
no noticeable improvement in the Doppler velocity bias or
I-209. S. Pazos and M. Hurtado are supported by UNLP and
in the power estimates.
CONICET.
C. Spectral Width
R EFERENCES
Finally, we analyze the behavior as we change the spectral
width σv of the weather signal, varying its value from 0.5 to [1] H. L. Groginsky and K. M. Glover, “Weather radar canceller design,”
in 19th Conference on Radar Meteorology, 1980, pp. 192–198.
3 m/s. We use L = 64 samples, and a target Doppler velocity [2] P. Meischner, Weather Radar: Principles and Advanced Applications,
v = 5 m/s. ser. Physics of Earth and Space Environments. Springer Berlin
Figure 4 shows the results for this scenario. The estimates Heidelberg, 2005.
[3] A. D. Siggia and R. E. Passarelli, Jr., “Gaussian model adaptive
in the Doppler velocity are also improved by using sparse processing (GMAP) for improved ground clutter cancellation and
models, with a decrease in the performance as the spectral moment calculation,” in Third European Conference on Radar Mete-
width increases and overlaps with the clutter. In this case orology (ERAD), 2004, pp. 67–73.
[4] C. M. Nguyen, D. N. Moisseev, and V. Chandrasekar, “A parametric
we cannot improve on the performance of GMAP when time domain method for spectral moment estimation and clutter
estimating the spectral width or the power. mitigation for weather radars,” Journal of Atmospheric and Oceanic
Technology, vol. 25, no. 1, pp. 83–92, 2008.
V. R EAL DATA [5] E. Boyer, P. Larzabal, C. Adnet, and M. Petitdidier, “Parametric spec-
tral moments estimation for wind profiling radar,” IEEE Transactions
We illustrate the sparse model approach using real radar on Geoscience and Remote Sensing, vol. 41, no. 8, pp. 1859–1868,
data collected by the Weather Radar system at INVAP Aug 2003.
[6] S. S. Chen, D. L. Donoho, Michael, and A. Saunders, “Atomic de-
in Bariloche. The data corresponds to an azimuth angle composition by basis pursuit,” SIAM Journal on Scientific Computing,
of 259◦ , with 3500 range bins from 0 to 100 km and vol. 20, pp. 33–61, 1998.
0 0

100 -10 100 -10

90 90
-20 -20

80 80
-30 -30

70 70
Range [Km]

Range [Km]
-40 -40
60 60
-50 -50
50 50
-60 -60
40 40
-70 -70
30 30
-80 -80
20 20
-90 -90
10 10
-100 -100
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
Doppler Velocity [m/s] Doppler Vel [m/s]

(a) (b)
0 0

100 -10 100 -10

90 90
-20 -20

80 80
-30 -30

70 70
Range [Km]

Range [Km]
-40 -40
60 60
-50 -50
50 50
-60 -60
40 40
-70 -70
30 30
-80 -80
20 20
-90 -90
10 10
-100 -100
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
Doppler Vel [m/s] Doppler Vel [m/s]

(c) (d)
Fig. 5: Reconstructed Weather Radar image in range and doppler domain for (a) FFT, (b) GMAP, (c) LASSO+GMAPM,
and (d) fused LASSO+GMAPM.

[7] E. Candes, J. Romberg, and T. Tao, “Robust uncertainty principles:


Exact signal reconstruction from highly incomplete frequency infor-
mation,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509, Feb.
2006.
[8] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory,
vol. 52, no. 4, pp. 1289–1306, Apr. 2006.
[9] R. J. Doviak and D. S. Zrnic, Doppler Radar and Weather Observa-
tions. Academic Press, 1993.
[10] M. Richards, Fundamentals of Radar Signal Processing, ser. Profes-
sional Engineering. Mcgraw-hill, 2005.
[11] J. A. Tropp and S. J. Wright, “Computational methods for sparse
solution of linear inverse problems,” Proceedings of the IEEE, vol. 98,
no. 6, pp. 948–958, June 2010.
[12] S. Ji, Y. Xue, and L. Carin, “Bayesian compressive sensing,” Signal
Processing, IEEE Transactions on, vol. 56, no. 6, pp. 2346–2356,
June 2008.
[13] R. Tibshirani, “Regression shrinkage and selection via the lasso,”
Journal of the Royal Statistical Society. Series B (Methodological),
vol. 58, no. 1, pp. 267–288, 1996.
[14] L. I. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based
noise removal algorithms,” Physica D: Nonlinear Phenomena, vol. 60,
no. 1, pp. 259 – 268, 1992.
[15] D. S. Zrnic, “Simulation of weatherlike doppler spectra and signals,”
Journal of Applied Meteorology, vol. 14, no. 4, pp. 619–620, 1975.

You might also like