DISTRIBUTION STATEMENT A. Approved for public release; distribution is unlimited.

Sequential Geoacoustic Filtering and Utilizing Ambient

Noise for Geoacoustic Inversion

Peter Gerstoft, Caglar Yardim, Ravi Menon, and Bill Hodgkiss

Marine Physical Laboratory
Scripps Institution of Oceanography
La Jolla, CA 92093-0701
phone: / (858) 534-7768 / (858) 822-4865 /(858) 534-1798
email: [email protected] / [email protected] /[email protected]

Award Numbers: N00014-11-1-0320 and N00014-11-1-0321

http://www.mpl.ucsd.edu/people/pgerstoft/

LONG-TERM GOALS

The development of new geoacoustic inversion methods, their use in the analysis of shallow water
experimental data, and evaluation of geoacoustic model and parameter uncertainties including the
mapping of these uncertainties through to system performance uncertainties.

OBJECTIVES

Analysis of geoacoustic inversion data collected from various experiments. Of specific technical
interest are: (1) development of methods to track the environmental parameters using sequential
filtering, (2) use of ambient noise for estimation of seafloor structure parameters, and (3) the
development of new inversion methods for use into the kHz frequency regime. In an ONR Graduate
Traineeship Awards we address using Random Matrix Theory in ocean acoustics.

APPROACH

1. Sequential filtering

A common feature of inverse problems in ocean acoustics is that estimates of underlying physical
parameters are extracted from measured acoustic data. Geoacoustic inversion has been approached in
the same framework, estimating, in addition to source location, ocean environment parameters and
their uncertainty. Often, those parameters evolve in time or space, with acoustic data arriving at
consecutive steps. Information on parameter values and uncertainty at preceding steps can be
invaluable for the determination of future estimates but is often ignored.

Sequential Bayesian filtering, tying together information on parameter evolution, a physical model
relating acoustic field measurements to the unknown quantities, and a statistical model describing
random perturbations in the field observations, offers a framework for the solution of such problems.

1
 Form Approved
Report Documentation Page OMB No. 0704-0188

Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it
does not display a currently valid OMB control number.

1. REPORT DATE 2. REPORT TYPE 3. DATES COVERED

2012 N/A -
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Sequential Geoacoustic Filtering and Utilizing Ambient Noise for 5b. GRANT NUMBER
Geoacoustic Inversion
5c. PROGRAM ELEMENT NUMBER

6. AUTHOR(S) 5d. PROJECT NUMBER

5e. TASK NUMBER

5f. WORK UNIT NUMBER

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION

REPORT NUMBER
Marine Physical Laboratory Scripps Institution of Oceanography La
Jolla, CA 92093-0701
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)

11. SPONSOR/MONITOR’S REPORT

NUMBER(S)

12. DISTRIBUTION/AVAILABILITY STATEMENT

Approved for public release, distribution unlimited
13. SUPPLEMENTARY NOTES
The original document contains color images.
14. ABSTRACT

15. SUBJECT TERMS

16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON
a. REPORT b. ABSTRACT c. THIS PAGE
SAR 9
unclassified unclassified unclassified

Standard Form 298 (Rev. 8-98)

Prescribed by ANSI Std Z39-18
2. Extracting information from noise cross-correlations

We have focused extensively on extracting information from noise in ocean acoustics with both
theoretical work as well as experimental work. The passive fathometer is based on relating the down-
and up-going signals on the array and can be implemented in the time or frequency domain. Here, we
are exploring the passive fathometer by aligning arrivals using phase information from the fathometer
(currently only the magnitude is used). We have evidence that the vertical fathometer array moves with
the waves on the sea surface. Thus if we can correct for this movement it will be possible to align the
reflections better and then average the reflection time series with phase as opposed to just using the
envelope. This should give sharper definition of the seafloor and sub bottom reflections and enables
estimating environmental geoacoustic parameters in addition to depths of reflecting layers. We also are
exploring accelerating convergence for the noise cross-correlation by various signal processing
strategies, e.g. averaging, rejecting interference dominated time series, eigenvalue/eigenvector
decomposition, and focusing on specific arrivals using beamforming.

WORK COMPLETED

One application of passive estimation of the time-domain Green’s function is in the use of cross-
correlations of upward and downward pointing vertical line array beams observing ambient noise to
extract seabed layer structure (i.e. a passive fathometer) [Traer et al., 2011, 2012]. This so called
passive fathometer technique exploits the naturally occurring acoustic sounds generated on the sea-
surface, primarily from breaking waves. The method is based on the cross-correlation of noise from the
ocean surface with its echo from the seabed, which recovers travel times to significant seabed
reflectors. To limit averaging time and make this practical, beamforming is used with a vertical array of
hydrophones to reduce interference from horizontally propagating noise. The initial development used
conventional beamforming, but significant improvements have been realized using adaptive
techniques. An analytical model is presented in [Traer et al., 2011] for the passive fathometer response
to ocean surface noise, interfering discrete noise sources, and locally uncorrelated noise in an ideal
waveguide. The leading order term from the ocean surface noise produces the cross-correlation of
vertical multipaths, yielding the depth of sub-bottom reflectors.

We have explored incorporating Kalman and particle filter tracking techniques into the geoacoustic
inversion problem [Yardim 2011a, 2011b, 2012, Michalopoulou 2012]. This enables spatial and
temporal tracking of environmental parameters and their underlying probability densities, making
geoacoustic tracking a natural extension to geoacoustic inversion techniques.

RESULTS

In many cases, it is of interest to estimate geoacoustic parameters over a larger spatial region rather
than just the parameters characterizing propagation between a fixed source and receiver (or receiving
array) location. Data might be available at a moored vertical receiving array from a towed acoustic
sound source or a source might be received by a towed horizontal array. In both cases, the typical
approach would be to treat each record of data independently of the others and carry out a full
geoacoustic inversion for every record resulting in a sequence of geoacoustic parameter estimates and,
in some cases, posteriori probability densities of the environmental parameters. The latter enables the
environmental uncertainty to be projected into other waveguide characterizations such as propagation
loss and its uncertainty.

2
In a review paper we have studied the basis and use of sequential filtering in ocean acoustics [Yardim
2011]. Sequential filtering provides a consistent framework for estimating and updating the unknown
parameters of a system as data become available, see Figs. 1-2. Despite significant progress in the
general theory and implementation, sequential Bayesian filters have been sparsely applied to ocean
acoustics. The foundations of sequential Bayesian filtering with emphasis on practical issues are first
presented covering both Kalman and particle filter approaches. Filtering becomes a powerful
estimation tool, employing prediction from previous estimates and updates stemming from physical
and statistical models that relate acoustic measurements to the unknown parameters. Ocean acoustic
applications are then reviewed focusing on the estimation of environmental parameters evolving in
time or space. Some possible scenarios for geoacoustic inversion are shown in Fig. 3.

Random Matrix Theory

A new direction is taken in Menon [2012a, 2012b, 2012c] and Gerstoft [2012] where random matrix
theory is used to analyze noise cross-spectral density matrices. Isotropic noise fields are often used to
model environmental noise surrounding an array of sensors. For a line array of equidistant sensors in
such a noise field, the true covariance matrix of the observations in the frequency domain is a
symmetric Toeplitz sinc matrix. In Menon [2012a], we derive the eigenvalues of the true covariance
matrix as the size of the matrix approaches infinity. For arrays spaced at less than half a wavelength
apart, the covariance matrix is shown to be rank deficient and this has implications in techniques such
as adaptive beamforming, which require the inverse covariance matrix. The zero eigenvalues are
related to classical array processing concepts such as the invisible region in frequency-wavenumber
space (region where there is no propagating energy, but a spectrum can be calculated). Using random
matrix theory, we derive the eigenvalue density of the sample noise covariance matrix, whose
knowledge is useful in reliable signal detection. An example of such processing is seen in Figure 4.

Often the ocean acoustic data sample covariance matrix (SCM, or cross-spectral matrix) is assumed to
consist of a few large signal-plus-noise eigenvalues followed by a set of equal-value noise-only
eigenvalues representing uncorrelated noise. However, it is well-known that the SCM from real data
observations is characterized by steadily decaying noise-only eigenvalues. In array processing, a
common rule of thumb is that the SCM is “well-estimated” when the number of snapshots is 2 to 3
times the array dimension. This depends on the type of noise and application under consideration.
Often, the number of snapshots available for forming the SCM is less than this, especially for large
arrays. Using Random Matrix Theory (RMT) to model the statistical properties of the SCM, the
eigenvalue distributions are more informative than using the expectation alone. A random matrix is a
matrix-valued random variable, i.e., the elements are stochastic variables. RMT can be used to study
the distribution of eigenvalues under asymptotic assumptions. [Menon 2012a, 2012b, 2012c; Gerstoft
2012]. Here we will use the SCM to extract eigenvalue distributions for representative noise scenarios
and this will quantify the information content in the data.

The excellent matching of SCM eigenvalues using RMT is demonstrated using data from a towed
horizontal array during the long range acoustic communications (LRAC) experiment from 10:00-11:00
UTC on 16 September 2010 in the NE Pacific in 5-km water depth. Other data periods yield similar
results to those shown here. The array was towed at 3.5 knots at a depth of 200 m. The data were
sampled at 2000 Hz using a nested array with each configuration having 64 channels

Figure 5 shows the eigenvalues of the SCM at selected values β, the element spacing to wavelength
ratio, for the four arrays. Due to the low sampling frequency (2000 Hz), the HF array only can be used

3
up to β=1/4 (1000 Hz). All eigenvalues are based on one-hour observations, meaning that for M=64 the
eigenvalues are averaged over 13 SCM eigenvalues. The first few eigenvalues for each SCM likely are
due to the distant transiting ships and noise from the towship, as seen in the beamform power vs angle
time series. The eigenvalues drop sharply above 2β(N-1)+1 (vertical dotted line) as predicted by theory,
and indicates that the coherent noise is stronger than the incoherent noise. The eigenvalues of the SCM
of the LF and ULF arrays show a similar behavior as the MF and HF arrays though with less strong
transition between the two eigenvalue regimes.

Figure 5 shows that the eigenvalues depend on β. As β increases, all the eigenvalue spectra become
more extended and at β=0.5 (half wavelength spacing) the SCM ideally should become diagonal with
eigenvalues that approximately are all equal. Comparing the four arrays at β=1/8, first column in Fig. 5
shows that the higher eigenvalue numbers (containing mostly incoherent noise) are relatively larger at
low frequencies. At half-wavelength spacing β=1/2, last column in Fig. 5, all eigenvalues remain large
for the three arrays, except when using a relatively small number of snapshots (M=N).

The observed and modeled noise eigenvalues are compared in Fig. 6. It is important to realize that
there is towship radiated noise as well as broadband signatures from several distant ships arriving at the
array, especially at low frequencies. These “signals” are among the largest eigenvalues extracted from
the data.

IMPACT / APPLICATIONS

Geoacoustic inversion techniques are of general interest for the estimation of waveguide parameters
thus facilitating system performance prediction in shallow water. Natural transition paths for these
results will be the PEO-C4I Battlespace Awareness and Information Operations Program Office
(PMW-120) and the Naval Oceanographic Office.

RELATED PROJECTS

None.

PUBLICATIONS

2012:
Gerstoft, Peter, Ravishankar Menon, William S. Hodgkiss, Christoph F. Mecklenbräuker, Eigenvalues
of the sample covariance matrix for a towed array, J Acoust. Soc. Am., 132, 2388-2396.
doi:10.1121/1.4746024. [published, refereed]

Menon, Ravishankar, Peter Gerstoft, and William S. Hodgkiss (2012a), Asymptotic Eigenvalue
Density of Noise Covariance Matrices, IEEE Signal processing. 60, 3415-3424,
doi:10.1109/TSP.2012.2193573. [published, refereed]

Menon, Ravishankar, Peter Gerstoft, William S. Hodgkiss (2012b), Cross-correlations of diffuse noise
in an ocean environment using eigenvalue based statistical inference, J Acoust. Soc. Am., 132,
DOI:10.1121/1.4754558, Nov 2012. [published, refereed]
Menon, Ravi, Peter Gerstoft and William S. Hodgkiss (2012c) Passive acoustic monitoring using
random matrix theory, IEEE Statistical Signal Processing workshop. [published, refereed]

4
Michalopoulou, Zoi-Heleni, Caglar Yardim, and Peter Gerstoft (2012), Passive Fathometer Tracking, J
Acoust. Soc. Am EL.131, EL74-EL80, doi:10.1121/1.3670004. [published, refereed]

Yardim, Caglar, Peter Gerstoft and WS Hodgkiss (2012), Sequential geoacoustic inversion at the
continental shelfbreak, J Acoust. Soc. Am, 131, 1722-1732, doi:10.1121/1.3666012. [published,
refereed]

Yardim, Caglar, Peter Gerstoft and WS Hodgkiss (2012), Particle smoothers in sequential
geoacoustical inversion, J Acoust. Soc. Am [Submitted]

2011:
Gerstoft, Peter and WS Hodgkiss (2011), Improving beampatterns of 2D random arrays using convex
optimization, J Acoust. Soc. Am EL, 129, EL135-140 DOI:10.1121/1.3556896, April 2011.
[published, refereed]

Park, Cheolsoo, Woojae Seong, Peter Gerstoft, and WS Hodgkiss (2011), Fluctuating arrivals of short-
range acoustic data, J Acoust. Soc. Am., 129, 98–103, doi:10.1121/1.3514505. [published,
refereed]

Traer, James, Peter Gerstoft, and W. S. Hodgkiss (2011), Ocean bottom profiling with ambient noise: a
model for the passive fathometer, J. Acoust. Soc. Am., 129, 1825-1836, doi:10.1121/1.3552871,
April 2011. [published, refereed]

Traer, James and Peter Gerstoft, 2012, Coherent averaging of the passive fathometer response using
short correlation time, J Acoust. Soc. Am, 131, 130, 3633-3641, Doi:10.1121/1.3654026
DOI: 10.1121/1.3654026 [Published, refereed]

Yardim, Caglar, Zoi-Heleni Michalopoulou, and Peter Gerstoft (2011a), An overview of sequential
Bayesian filtering in ocean acoustics, IEEE Oceanic Eng 36(1), 73-91,
Doi:10.1109/JOE.2010.2098810. [published, refereed]

Yardim, Caglar, Peter Gerstoft and WS Hodgkiss (2011b), Range Aliasing in Frequency Coherent
Geoacoustic Inversion, J Acoust. Soc. Am EL, 130, EL154-EL160, Doi:10.1121/1.3630021.
[published, refereed]

5
 Figure 1. Sequential Bayesian filtering. From state xk−1, state xk is first predicted via the state
equation, providing xk|k−1 . As data yk becomes available, the observation equation is employed to
update state xk|k-1 , providing xk|k .

Figure 2. A quick guide to filter selection leading to the Kalman filter (KF), extended Kalman filter
(EKF), unscented Kalman filter (UKF), and particle filter (PF).

6
 Figure 3. Geoacoustic environmental tracking: (a) Temporal tracking of the ocean sound speed
profile for a fixed-receiver and a fixed-source and (b) tracking of the changing environment
between the receiver and a moving source. Here shown for a vertical line array (VLA) of receivers.

Figure 4. Asymptotic eigenvalue density (solid line) and the empirical

eigenvalue density, with N = 100 array elements, ν = 1/4 (ratio of number of sensors to number of
snapshots) and spacing to wavelength ratios β of (a) 1/4 (b) 1/2 and (c) 3/4. (d,e,f): Same as in
(a,b,c) except with N = 20. The dotted lines show the locations of the distinct non-zero true
eigenvalues. For more information see Menon [2012a].

7
 Figure 5. Eigenvalues of the towed array SCM for HF array (1st row, β =1/8,1/4), MF array (2nd
row, β =1/8, 1/4, 1/2), LF array (3rd row, β =1/8, 1/4, 1/2), and ULF array (4th row, β =1/8, 1/4, 1/2).
The eigenvalues are based on 64 (ν=1, dashed) and 864 (ν=13, solid) snapshots. The eigenvalues
are normalized with the largest eigenvalue and the vertical dotted line indicates the edge of the
visible region, β is the element spacing to wavelength ratio.
For more information see Gerstoft [2012].

8
Figure 6. Comparison of observed and modeled eigenvalues. Eigenvalues of the towed array SCM
(solid) and modeled eigenvalues 3D (dashed) and 2D (dash-dotted) for HF array (1st row, β=1/8,
1/4) and LF array (2nd row, β=1/8, 1/4, 1/2). The observed eigenvalues are based on 864 (ν=13,
solid) snapshots, the modeled eigenvalues use a coherent/incoherent noise ratio 700 for HF array
and coherent/incoherent noise ratio 100 for LF array. The eigenvalues are normalized with the
largest eigenvalue and the vertical dotted line indicates the edge of the visible region, β is element
spacing to wavelength ratio. For more information see Gerstoft [2012].