PARTICLE FILTERING FOR CHANNEL PARAMETER TRACKING IN A NOISY
SHALLOW OCEAN ENVIRONMENT USING A VERTICAL ARRAY
X. Zhong , V. N. Hari , W. Wang , H. Wang# and X. Shen#
Innitus, School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, Singapore.
Acoustic Research Lab, National University of Singapore, 639798, Singapore.

CVSSP, Department of Electronic Engineering, University of Surrey, UK, GU2 7XH.
#
School of Marin Science and Technology, Northwestern Polytechnical University, ShaanXi, China.
xhzhong@ntu.edu.sg,
#
and
ABSTRACT
Acoustic signals in a shallow ocean environment are severely
distorted due to the time-varying and inhomogeneous nature
of the propagation channel. In this paper, a state-space
model is introduced to characterize the uncertainties of
the shallow ocean and a Rao-Blackwellized particle lter
(RBPF) is developed to estimate the model parameters. Since
both modal functions and horizontal wave numbers of the
channel are assumed unknown, the state-space model has
a high nonlinearity and high dimensionality. As the modal
functions are linear with the measurements conditioning on
the horizontal wave numbers, a Kalman ltering (KF) is
employed to marginalize out the modal functions. Hence
only the horizontal wave numbers need to be estimated by
using a PF. Simulation results show that the proposed RBPF
algorithm signicantly outperforms the existing approaches.
Index Terms Shallow ocean acoustic model, wave
number, modal function, Rao-Blackwellized particle lter.
I. INTRODUCTION
Modeling acoustic wave propagation in a shallow ocean
environment is an important topic and lies at the heart
of many underwater signal processing applications [1][3].
It is a challenging problem as uncertainties arise due to
the time-varying and inhomogeneous nature of the ocean
environment, and the received signal is seriously distorted
due to multiple reections from ocean boundaries.
Traditionally, matched-eld processor (MFP) that com-
pares the measured pressure-eld to that predicted by a prop-
agation model has been employed for parameter estimation
[4]. To implement an MFP, the source location is required
and is obtained by computing model predictions of the eld
at the array for various assumed source positions. However,
the shallow ocean acoustic channel varies with space and
time. In [5], a state-space model was formulated to charac-
terize all these dynamics and accordingly, extended Kalman
lter (EKF) and unscented Kalman lter (UKF) approaches
were introduced to estimate the states that characterize the
978-1-4673-7297-8/15/$31.00 2015
c IEEE

harivishnu@gmail.com, w.wang@surrey.ac.uk,
{h.wang, x.shen}@nwpu.edu.cn.
channel. Such approaches are categorized as model-based
methods in that the acoustic propagation and measurement
uncertainties are modeled and included into the estimation
algorithms. Very recently, the particle ltering (PF) approach
which is more appropriate for nonlinear and non-Gaussian
systems has been employed for parameter estimation of the
shallow ocean model [5][7] and shown to achieve better
estimation performance.
In a shallow ocean model, the horizontal wave numbers
are considered unknown parameters in addition to the modal
functions due to inhomogeneous time-varying nature of the
environment. However, if all these parameters are blindly
encapsulated into the estimated state, the performance of PF
algorithm suffers due to the high dimensionality of the state.
In this paper, a Rao-Blackwellized particle ltering (RBPF)
method is developed to reduce the dimensionality of the state
to be processed by PF. In essence, in this method the system
model is reorganized in a linear fashion conditioned on the
horizontal wave numbers. The Kalman lter (KF) is then
employed to provide an optimal solution for this linear part.
Following this, the horizontal wave numbers are the only
state to be estimated by the PF. Hence, the PF can achieve
better estimation accuracy with the same number of particles
because the dimensionality is reduced. Our contribution here
is incorporating a Rao-Blackwellization technique to obtain
an analytical solution for part of the state, and therefore
reduce the dimensionality of the state for PF algorithm.
Simulations are organized to demonstrate the superiority of
the proposed algorithm over the existing EKF and PF.
II. SIGNAL MODEL
For modeling acoustic wave propagation in a shallow
ocean environment, we assume a horizontally stratied ocean
of depth h with a known horizontal source range rs and
depth zs , and the acoustic wave propagation is governed
by the Helmholtz equation [8]. The normal-mode acoustic
pressure propagation model can be written as
M

s(rs , z) = m (rs , zs )m (z) (1)
m=1
450
  
where m (rs , zs ) is the modal coefcient dened as where C x (z) = [1 (), 0, 2 (), 0, . . . , M (), 0], and
w(z) N (0, 2 ) is the measurement noise process. Equa-
er (m)rs jr (m)rs tions (9), (10) and (11) give a full state space model that
m (rs , zs ) = qm (zs )  e (2)
r (m)rs describes the dynamics and uncertainties of the shallow
ocean characteristics such as horizontal wave numbers and
where q is the source amplitude, and r (m), r (m) and modal functions. In next section, an RBPF approach will be
m () are the modal attenuation, the horizontal wave number developed to estimate these characteristics.
and the modal function associated with the mth mode respec-
tively. The modal function holds an eigenvalue equation in
z, given as [9] III. RAO-BLACKWELLIZED PARTICLE
FILTERING IMPLEMENTATION
d2
m (z) + 2z (m)m (z) = 0 (3) Assume that z is the depth of the
th sensor and z1: =
dz 2 [z1 , . . . , z ]. Given a measurement sequence y(z1: )) =
for m = 1, . . . , M . The eigen-value set {m (z)} are the [y(z1 ), . . . , y(z )], the task is to estimate the posterior dis-
modal functions and z is the wave number in the depth tribution p(x(z)|y(z1: )). Such a task can be achieved by
direction. The solutions of (3) depend on the sound speed using a Bayesian recursive estimation, given as
prole c(z), the boundary conditions and the corresponding
Predict :
dispersion relation given by
     
2 p x(z ) y(z1:1 ) = p x(z )x(z1

2 = = 2r (m) + 2z (m) (4)   
c2 (z) p x(z1 )y(z1:1 ) dx(z1 ); (12)
where r (m) is the mth horizontal wave number and Update :
     
is the harmonic source frequency. Let m (z) = p x(z )y(z1: ) p y(z )x(z )
[m1 (z), m2 (z)] T   
 and dene m1 (z) = m (z) and p x(z )y(z1:1 ) . (13)
m2 (z) = d/dz m (z) . The eigenvalue equation in state-
  
space form is In thisrecursion,
 p y(z )x(z ) is the likelihood of the state
d and p x(z )y(z1:1 ) is the probability density function
m (z) = Am (z)m (z) (5) (PDF) of a prior distribution. The Bayesian recursion states
dz
that given the transition density and likelihood, the posterior
where Am (z) is the coefcient matrix given as distribution of the state can be recursively estimated.
 
0 1 The measurement function is nonlinear and therefore, the
Am (z) = (6)
2z (m) 0 EKF approach [9] and PF approach [7] have been employed
to estimate the posterior distribution. It has been shown
that the PF approach is more appropriate under such a
The horizontal wave number r (m) can be roughly
highly nonlinear scenario [10]. However, taking all states
estimated using wavenumber spectrum estimation methods.
into account is cumbersome for the PF approach due to the
However, these are unable to account for the uctuating
curse of the dimensionality. Recall the state space model,
and time-varying nature of the environment. Thus r (m)
the state of modal function x (z) holds a linear relationship
is considered as an unknown environmental parameter to
with the measurement conditioning on the state x (z). This
be estimated and is included into the state vector. Let
means given the estimation of x (z), an analytical solution
m (z) = r (m). The whole state vector can be constructed
for p x (z)x (z), y(z1: ) can be obtained. Hence, it is
as x(z) = [x (z)T , x (z)T ]T , where
possible to exploit a Kalman lter to marginalize out the
x (z) = [T1 (z), T2 (z), . . . , TM (z)]T R2M 1 (7) modal functions. Consequently, only the horizontal wave
numbers need to be handled by using the PF. Such a
x (z) = [1 (z), 2 (z), . . . , M (z)]T RM 1 (8) technique is referred to as Rao-Blackwellization and widely
The state process can be written as used for the state estimation where part of state space model
is linear and Gaussian [10]. Using Bayesian theorem, the
d posterior distribution can be decomposed as
x (z) = A(z)x (z) + v (z) (9)      
dz
d p x(z )x(z1 ), y(z ) = p x (z )y(z )
x (z) = 0 + v (z) (10)


dz   PF

  p x (z )x (z ), x (z1 ), y(z ) , (14)
where A(z) = diag A1 (z), . . . , AM (z) , and v (z) and


v (z) are the zero-mean Gaussian processes given by KF
v (z) N (0, ) and v (z) N (0, ) respectively. The   
in which p x (z 
measurement process can be written as  x (z
 )   ), x (z1 ), y(z ) is analytically
  tractable and p x (z )y(z ) can be estimated by PF ap-
y(z) = s(rs , z) = C x (z) x (z) + w(z) (11) proximation. Since part of the state can be estimated by using

2015 IEEE 5th Asia-Pacific Conference on Synthetic Aperture Radar(APSAR) 451

a KF, the dimension of the state to be processed by the Algorithm 1: RBPF for model parameter estimation.
PF can be reduced. Consequently, the Rao-Blackwellization
based PF is able to provide better estimates than the standard Initialisation: for i = 1, . . . , N , draw particles
(i)
PF when the same number of particles is used. according to (26); set the initial weight w0 = 1/N ;
for
1 to L do
The core idea of PF is that it uses a set of particles for i 1 to N do
and importance weights of these particles to approximate 1) draw samples according to equation (15);
the posterior distribution. Assuming that N particles are 2) KF marginalization from (16) to (21);
used
 to approximate
  the above Bayesian recursion, the PDF 3) compute the likelihood according to (23);
 (i) (i)
p x (z ) y(z ) is represented by {x (z ), wk }N i=1 . The 4) calculate the weight according to (22);
entire procedure of PF processing can be summarized as end
following. At each time step, the particles are sampled
according to the state dynamic model (10), given as (i) (i) N (i)
5) normalise the weight wk = wk / i=1 wk ;
 (i)  (i) 
x (z ) p x (z )x (z1 ) .
(i) 6) resample the particles according to the weights;
(15)
7) output the estimates.
These particles are then employed in the KF steps to end
marginalize out the modal functions. Assume that at the
previous time step, the state and covariance estimates are
(i) (i)
x (z1 ) and P (z1 ) respectively. The predictions are: (i)
where () is a Dirac-delta function, and wk is a normal-
(i) (i) ized weight. In practical implementation, the state can be
x (z |z1 ) =A(i) (z )x (z1 ) (16)
initialized by estimates via the MFP method. Assume that
(i) (i) (i)
P (z |z1 ) =A (z )P (z1 )A(i) (z )T + (17) the initial state is x0 . The initial distribution can be given as
(i)
The Kalman gain is then calculated as x0 N (x0 , 0 ) (26)
 (i)  (i)  (i) 
S(i) =C x (z ) P (z |z1 )CT x (z ) + 2 (18) where 0 is the variance of initial distribution that charac-
(i)  (i)  1 terizes the error of the MFP estimates. All implementation
K(i) =P (z |z1 )CT x (z ) S(i) (19) steps of the RBPF approach are summarized in Algorithm
The dependency on z in (18) and (19) is ignored to simplify 1. The proposed approach differs from traditional PF [7] for
the expression. The state and covariance are updated as model parameter estimation in that the modal functions are
(i) (i)
analytically estimated by using KF, and only the horizontal
x (z ) =x (z |z1 ) + K(i) wave numbers are estimated by using PF.
 
 (i)  (i)
y(z ) C x (z ) x (z |z1 ) (20) IV. SIMULATIONS
(i) (i)
P (z ) =P (z |z1 ) K(i) In this section, simulations are provided to demonstrate
 (i)  (i) the tracking performance. The performance is compared with
C x (z ) P (z |z1 ) (21) that of existing EKF approach [9] and PF approach [7].
 (i)  A noisy shallow ocean channel with a depth of 100 m is
The ltered distribution is N x (z ), P(i) (z ) . The im- simulated. The signal source is located at a depth of 36
portance weights of the particles are then evaluated by m and a horizontal range of 5 km. The center frequency
  (i) 
wk = wk1 p y(z )x (z ) ,
(i) (i) of the source signal is 100 Hz. This leads to the signal
(22)
eld propagating with six normal modes. Considering the
The likelihood of the particles are then calculated as uctuations in the ocean, the wave numbers are initialized
  (i)  1 with a bias of 1.0 104 . This accounts for the uncertainty
p y(z )x (z ) = present in describing the shallow ocean environment. Other
 2 2 parameters are set as following: N = 200, 2 = 1.0 103 ,
 2 
1  (i)  (i) = 1.0 106 and = 1.0 108 . These parameters
exp 2 y(z ) C x (z ) x (z |z1 ) are selected based on extensive experimental study and are
2
(23) found be able to provide good estimates.

After the resampling scheme, the posterior distribution of Figure 1 presents a comparison of estimation using the
the state is thus approximated by RBPF against the EKF and traditional PF at an SNR of
10 dB. The ground truth and corresponding estimates of
N
      modal functions 1, 2, 4 and 5 are plotted on the top
p x (z )y(z )
(i)
wk x(i) (z ) x (z ) , (24) of the gure. Signal eld and wave numbers are also

i=1 presented. The estimation results show that the proposed
N
   RBPF algorithm provides better accuracy than traditional
p x (z )y(z )
(i)
wk N (x(i) (z ), P(i) (z )), (25) EKF and PF approaches. RBPF converges faster to the
i=1 true value of the wave numbers despite the initial bias. Its

452 2015 IEEE 5th Asia-Pacific Conference on Synthetic Aperture Radar(APSAR)

 3 3
x 10 x 10
10 10
wave num,RBPF modal fun,RBPF signal,RBPF
100 100 100 100 wave num,PF modal fun,PF signal,PF
9 wave num,EKF modal fun,EKF signal,EKF
9
80 80 80 80 0.25
8
60 60 60 60 8
depth

depth

depth

depth
7 0.2
40 40 40 40
6 7

MSE
20 20 20 20 0.15
5 6
0 0 0 0
0 0.1 0.2 0.2 0 0.2 0.2 0 0.2 0.2 0 0.2
modal function modal function modal function modal function 4 0.1
5

10 3
0.05
Pure Signal 4
8 2
amp. of Signal

RBPF
6 PF
1 3 0
EKF 0 10 20 30 0 10 20 30 0 10 20 30
SNR/dB SNR/dB SNR/dB
4

0 Fig. 2. RMSE versus different SNRs for wave numbers (left),

0 10 20 30 40 50 60 70 80 90 100 modal functions (center) and signal eld (right)
depth

100 100 100 100

80 80 80 80 Hence, the modal functions are analytically estimated by us-

60 60 60 60
ing a KF and only the wave numbers need to be estimated by
depth

depth

depth

depth

the PF. The estimation accuracy is thus improved by the pro-

40 40 40 40
posed approach. Simulations show that the proposed RBPF
20 20 20 20 algorithm signicantly outperforms the existing approaches
in estimating these parameters. Future work includes modal
0.4175 0.418
wavenumber
0.4145 0.415
wavenumber
0.402 0.4025
wavenumber
0.392 0.3925
wavenumber
order detection and real underwater data applications.

VI. REFERENCES

Fig. 1. Results of estimation and ground truth under SNR = [1] X. Zhong, V. N. Hari, A. Premkumar, and A. S. Mad-
10 dB for: modes 1, 2, 4 and 5 (top), pressure eld (middle) hukumar, Particle ltering with enhanced likelihood
and corresponding wave numbers (bottom). model for underwater acoustic source DOA tracking,
in OCEANS, 2011 IEEE - Spain, June 2011, pp. 16.
[2] X. Zhong, A. Premkumar, and W. Wang, Direction of
accuracy in estimating modal functions is signicantly better. arrival tracking of an underwater acoustic source using
Consequently, the signal recovery is signicantly improved particle ltering: real data experiments, in TENCON
in the noisy environment. Spring Conference, 2013 IEEE, April 2013, pp. 420
424.
Multiple Monte Carlo (MC) simulations are also orga- [3] X. Zhong, A. B. Premkumar, and A. S. Madhukumar,
nized to study the performance of the proposed algorithm. Particle ltering for acoustic source tracking in im-
The normalized mean square error (MSE) over 50 MC pulsive noise with alpha-stable process, IEEE Sensor
runs for the modal function, wave number and signal eld J., vol. 13, no. 2, pp. 589600, 2013.
estimation is presented in Fig. 2. Different noisy data from [4] A. B. Baggeroer, W. A. Kuperman, and P. N.
0dB to 30dB with a 5dB increment are generated. The results Mikhalevsky, An overview of matched eld methods
further demonstrate the superiority of the proposed algorithm in ocean acoustics, IEEE J. Ocean. Eng., vol. 18, no. 4,
over other existing approaches. For all estimations, the MSE pp. 401424, 1993.
is signicantly lower than that of other existing approaches. [5] C. Yardim, P. Gerstoft, and W. S. Hodgkiss, Tracking
Due to Rao-Blackwellization, the state dimensionality to of geoacoustic parameters using kalman and particle
be processed by the PF is reduced. Consequently, RBPF lters, J. Acoust. Soc. Amer., vol. 125, no. 2, pp. 746
performs much better than the traditional PF approach. 760, 2009.
[6] J. Candy, Particle ltering for signal enhancement in
V. CONCLUSION a noisy shallow ocean environment, in OCEANS 2010
IEEE, Sep. 2010, pp. 16.
An RBPF approach is introduced in this paper to estimate [7] , An adaptive particle ltering approach to track-
the modal functions and wavenumbers of a shallow ocean ing modes in a varying shallow ocean environment, in
channel. Conditioning on the horizontal wave numbers, the OCEANS, 2011 IEEE, Santander, Spain, Jun. 2011, pp.
modal functions are linearly dependent on the measurements. 1 7.

2015 IEEE 5th Asia-Pacific Conference on Synthetic Aperture Radar(APSAR) 453

 [8] L. Bjorno, Ed., Underwater Acoustics and Signal Pro-
cessing. Dordrecht: Reidel, 1981.
[9] J. Candy and E. Sullivan, Model-based identication:
an adaptive approach to ocean-acoustic processing,
Ocean. Eng., IEEE J., vol. 21, no. 3, pp. 273 289,
jul 1996.
[10] B. Ristic, S. Arulampalam, and N. Gordon, Beyond the
Kalman Filter: Particle Filters for Tracking Applica-
tions. Artech House, 2004.

454 2015 IEEE 5th Asia-Pacific Conference on Synthetic Aperture Radar(APSAR)