800 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 27, NO.

4, OCTOBER 2002

Novel Physical Interpretations of

K-Distributed Reverberation
Douglas A. Abraham and Anthony P. Lyons

Abstract—Interest in describing and modeling envelope the probability of false alarm ( ). The Rayleigh distribution
distributions of sea-floor backscatter has increased recently, is obtained by assuming that there are enough scatterers con-
particularly with regard to high-resolution active sonar systems. tributing to the reverberation in any given resolution cell (after
Sea-floor scattering that results in heavy-tailed-matched-filter-en-
velope probability distribution functions (i.e., non-Rayleigh beamforming and matched filtering) so that the central limit the-
distributions exemplified by the K, Weibull, Rayleigh mixture, orem (CLT) holds and the received signal is Gaussian, resulting
or log-normal distributions) is often the limiting factor in the in a Rayleigh distributed envelope. Non-Rayleigh reverberation,
performance of these types of sonar systems and in this context is in particular reverberation with tails heavier than Rayleigh, can
referred to as reverberation or acoustic clutter analogous to radar occur when the conditions of the CLT are violated. For example,
clutter. Modeling of reverberation has traditionally entailed fitting
various candidate distributions to time samples of the envelope there may be too few scatterers in the resolution cell or the scat-
of the scattered sonar (or radar) returns. This type of descriptive terers may not be identically distributed.
analysis and the asymptotic (infinite number of scatterers) analysis Statistical analysis of the reverberation envelope usually en-
defining the K-distribution yield little insight into the environ- tails fitting the observed data to a set of models to determine
mental mechanisms responsible for heavy-tailed distributions which provides the best representation [1]–[5]. Researchers are
(e.g., distributions and clustering of discrete scatterers, patchiness
in geo-acoustic properties, scattering strength of scatterers, etc.) usually forced into such descriptive analyses because there are
and do not allow evaluation of the effect of changing sonar system very few physics-based models that link a tenable description
parameters such as bandwidth and beamwidth. In contrast, we of the environment and sonar system to the probability density
derive the envelope distribution for the scattered returns starting function (PDF) of the reverberation induced matched filter enve-
from simple physical descriptions of the environment with a finite lope output. Among them are Crowther’s model [6] for sea-bed
number of scatterers. It is shown that plausible descriptions of the
environment can lead to K-distributed reverberation. This result backscatter when the sea-bed is comprised of patches of differing
explains, at least partially, the success of the K-distribution in the types and Middleton’s KA-distribution [7] which was developed
modeling of radar clutter and sonar reverberation at a variety of using counting functionals in what he terms a “physical-statis-
frequencies and scales. The finite-number-of-scatterers model is tical” approach. The KA-distribution is a generalization of the
then used to predict how the shape parameter of the K-distribution well-known K-distribution [8], [9], which is itself derived in the
will change as the beamwidth of a towed-array receiver is varied.
Analysis of reverberation data from a low-frequency (450–700 Hz) limit by assuming that the number of scatterers in a resolution
active sonar system illustrates that, within our ability to estimate cell follows a negative binomial distribution with a mean that
it, the shape parameter of the K-distribution is proportional to tends to infinity. The K-distribution has also been described as
the beamwidth of the towed-array receiver, a result important a compound process arising from the modulation of Rayleigh
for sonar simulation and performance prediction models. These backscatter by an intensity that varies slowly in time and is well
results should prove useful in developing methods for modeling,
predicting and mitigating reverberation on high-resolution sonar modeled by the gamma distribution [10]. Owing to their com-
systems. plicated nature or limited applicability, the models of Crowther
Index Terms—Clutter, K-distribution, non-Rayleigh, reverbera-
and Middleton are rarely employed. The K-distribution, however,
tion, scattering. is a standard model for radar clutter [11] and has been shown to
represent sonar reverberation well [5], [12]–[14]. Unfortunately,
neither the asymptotic derivation nor the compound representa-
I. INTRODUCTION tion (which is phenomenological) of the K-distribution retain the
pertinent description of the environment that would allow predic-
H IGH-resolution active sonar systems have been developed
to improve detection performance by lowering the power
of the interfering reverberation in a given resolution cell. Un-
tion of the PDF of the matched filter envelope and how it changes
as a function of sonar system parameters or is affected by varying
fortunately, this can have the adverse effect of producing rever- environmental scattering mechanisms.
beration with a statistical distribution that is significantly more In this paper, we start with a plausible physical description of
heavy-tailed than the traditionally assumed Rayleigh distribu- the sea-floor scattering environment and derive the PDF for the
tion at the matched filter envelope, leading to an increase in reverberation envelope while retaining the environmental char-
acterization necessary for predictive analysis such as changes in
Manuscript received October 5, 2002; revised July 3, 2002. This work was
supported by the Office of Naval Research under Grants N00014-02-1-0115 PDF as a function of system or environmental parameters. In Sec-
and N00014-01-1-0352. tion III-A, it will be shown that reverberation arising from a finite
The authors are with the Applied Research Laboratory, Pennsyl- number of discrete scattering objects having an exponential size
vania State University, State College, PA 16804-0030 USA (e-mail:
[email protected]). distribution results in a K-distributed matched filter envelope. The
Digital Object Identifier 10.1109/JOE.2002.804324 shape parameter of the K-distribution is related to the number of
0364-9059/02$17.00 © 2002 IEEE
ABRAHAM AND LYONS: NOVEL PHYSICAL INTERPRETATIONS OF K-DISTRIBUTED REVERBERATION 801

scatterers within each resolution cell of the sonar system and the where now represents time delay and is the ambient noise
scaleparameterisrelatedtotheaveragesizeofthescatterers.Scat- after matched filtering. The effect of matched filtering is to re-
tering from a sea floor comprised of a finite number of patches duce the number of scatterers that contribute to at time .
of differing sizes is considered in Section III-B. Interface scat- If is the width of the autocorrelation function of , then,
tering from each individual patch produces a Gaussian reverber- only scatterers with delays such that will con-
ation response; however, when the patches are assumed to have tribute to at time . Scatterers that are farther away in time
an exponentially distributed size, the reverberation envelope is are attenuated by the term in (3). is also taken
seen to be K-distributed. These models that are based on a finite as the size of the resolution cell in range (delay) and is usually
number of scatterers allow the prediction of the PDF of the rever- well approximated by one over the bandwidth of the transmit
beration induced matched filter envelope and how it changes as a waveform. It should be noted that matched filtering is not nec-
function of sonar system parameters such as transmit waveform essarily optimal when the interfering signals are non-Gaussian
bandwidth and receiving array beamwidth. In Section IV, exper- or correlated in time; however, it is often employed as it can
imental data from a low-frequency active sonar system are used in many cases significantly increase SNR. Additionally, the op-
to show that the K-distribution shape parameter is, as predicted timal processor at this stage is not known when the interferences
by the model of Section III-A, proportional to the beamwidth of are non-Gaussian.
the towed-array receiver. After matched filtering, and assuming little or no target
Doppler, the beam output may be modeled as a bandpass
II. BACKGROUND random process with energy predominantly in the transmission
band. As such, the envelope of the process is a sufficient statistic
Active sonar systems transmit signals and process the re-
for the detection of fluctuating targets and nonfluctuating tar-
sulting echoes for the purposes of detecting, classifying and lo-
gets with random phase when the interfering components are
calizing targets. The transmitted pulse, say , reflects off of a
Gaussian [15]. The importance of sufficiency in this application
potential target and a multitude of inhomogeneities and irregu-
is that the optimal likelihood ratio detector then only requires
larities (i.e., scatterers) in the ocean medium and is then received
access to the sufficient statistic and not the complete data. Thus,
by an array of hydrophones. The response of each scatterer is
the design and performance analysis of a detection algorithm
modeled as a time-delayed and amplitude-scaled version of the
may be determined solely from the statistical description of the
transmitted pulse. The hydrophone array is designed to allow
envelope. Owing to the one-to-one correspondence between
spatial filtering (i.e., beamforming) of the array data which im-
the matched filter envelope and intensity, the latter may be
proves the signal-to-noise power ratio (SNR) by rejecting en-
equivalently used as a sufficient statistic. The envelope is also
ergy coming from directions away from the main response axis
of interest when the interfering components are not Gaussian.
(MRA) while accepting energy from a limited region in angle
As seen in Appendix A, the envelope is a sufficient statistic
about the MRA. The received signal after beamforming may be
for the more general case where the complex envelope of
described as
the interfering components has a circularly symmetric PDF.
(1) Reverberation that can be described as arising from the product
between a Gaussian component and some positive modulating
where represents the angle of the MRA, is the transmitted random variable has a circularly symmetric complex envelope
pulse, and are the target echo amplitude and arrival time, PDF. The K-distribution, Rayleigh mixture, and Crowther’s
and represent the amplitude and arrival time of the rever- model satisfy these requirements.
beration component from the th scatterer, is the ambient A convenient intermediate step in forming the envelope of
noise, and is the total number of scatterers contributing to is to first form the complex envelope by basebanding the
reverberation on this particular beam. Note that, in this paper, data (i.e., frequency shifting with a complex exponential at the
random variables are generally denoted by upper-case letters center frequency of the band, ) followed by low-pass filtering
and constants by lower-case ones. It will be initially assumed to remove out-of-band components
that there is no acoustic wave propagation with multiple
boundary interaction (e.g., multipath propagation). This case
will be discussed in Section III-C.
If the interfering signals (the reverberation or clutter and
(4)
ambient noise) are assumed to be a white Gaussian random
process, then, the optimal detection processing [15] requires
matched filtering the beam data by correlating them with the where is the basebanded autocorrelation function of the
transmitted waveform. If the autocorrelation function of the transmit waveform and is the complex envelope of the am-
transmitted waveform is bient noise after matched filtering. In this paper, variables with
a tilde are complex. The envelope is then simply the modulus
(2) of the complex envelope, . When the reverber-
then, the matched filter output has the form ation dominates the ambient noise and no target is present, the
complex envelope simplifies to the middle term of (4) and may
(3) be approximated by assuming that the autocorrelation function
is binary and either fully passes the individual scatterer echoes
802 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 27, NO. 4, OCTOBER 2002

when they are within the resolution cell, or fully rejects them they require some simplifying assumptions, yet they provide the
when they are not desired link between the number of scatterers and the statistical
distribution of the detection statistic.

III. SEA-FLOOR SCATTERING MODELS

(5) Central to the novel physical interpretation of K-distributed
where is the number of scatterers that contribute to the reverberation arising from a finite number of scatterers is an
complex envelope at time delay and . For no- assumption of a specific functional form for describing the
tational simplicity, it is assumed that these scatterers are size frequency distribution of discrete scatterers or scattering
indexed by . The time delays are assumed to patches. In nature, distributions of size often exhibit the prop-
be uniformly random over an integer number of periods of the erty that there are far fewer large objects or features than small.
center frequency so may be equivalently assumed to be This is true for the size distributions of large geologic sea-floor
uniformly distributed from 0 to 2 . It should be noted that if the features such as seamounts or bathymetry [18] as well as for
transmit waveform is a short rectangular continuous wave pulse smaller features such as rock outcrops and fragments or even
and matched filtering is not employed [i.e., gas bubbles in sediment [19]. Various probability distribution
for with an integer multiple of ] then the re- functions have been used to fit data that display this character-
sult of (5) is exact rather than approximate. istic including the exponential and the power law functions.
Traditionally, it has been assumed that is nonrandom The exponential function, in particular, has often been found
and large enough that the CLT [16] applies and is a zero to be useful for describing the size distribution of large-scale
mean complex Gaussian random process, leading to a Rayleigh geologic sea-floor features including fault displacements
distributed envelope and lengths [20] and sea-floor slopes [21]. Additionally, for
many sites, both terrestrial and extraterrestrial, the exponential
(6) distribution has often provided the best fit to rock size data
[22], [23]. In the case of discrete objects such as boulders and
where is the power. The K-distribution has been shown to rocks, exponential size distributions are justified based on the
arise by first assuming that is a random integer following a theory of fracture and fragmentation [23].
negative binomial distribution and then letting the average value Given the validity of an exponential function for describing
of tend to infinity [9]. The PDF for a K-distributed rever- the size distribution of many types of natural features, in the fol-
beration envelope is lowing analysis, we assume that the size of a randomly occur-
ring scatterer is described by an exponential distribution func-
(7) tion. Although numerical methods could be used to obtain the
PDF of the matched filter envelope for a number of other distri-
bution functions, the exponential form has the advantage of re-
where is the Basset function (i.e., a modified Bessel func- sulting in an analytical solution. In the following sections, both
tion of the third kind) [17], is a scale parameter, is the shape reverberation arising from discrete scatterers and from sea-floor
parameter, and the power is . The Rayleigh dis- patches of different scattering strength are considered.
tribution is a submember of the -distribution that is obtained
when , while . A. Exponentially Sized Discrete Scatterers
The significance of these two results lies in their validity irre-
spective of the PDF of the amplitudes , as long as the distribu- Assuming that reverberation arises from a finite number of
tions satisfy some minimal regularity conditions such as inde- scattering objects that have an exponentially distributed size and
pendence and finite variance. However, owing to their asymp- that the amplitude of the scattering return is proportional to its
totic nature, these results do not allow the analysis of a sonar size (e.g., as in geometrical scattering [24]), then in (5) would
system as a function of the beamwidth or bandwidth; that is, be exponentially distributed with PDF
when the number of scatterers within a resolution cell is fi-
nite and is allowed to change with system parameters (e.g., in- (8)
creasing bandwidth) or environmental conditions. In particular,
a sonar system with a large receiving array or a wide band- where is the average size of the scatterers. Assuming that the
width transmitter may reduce the resolution cell size so that amplitude and delay of each scatterer are independent of each
not enough scatterers are present for the CLT to hold or for other and of those of other scatterers, the distribution of
the K-distribution to be accurate as an asymptotic approxima- may be found by forming the joint characteristic function (CF)
tion. In this situation, the distribution of the summation in (5) of its in-phase (real) and quadrature (imaginary) components
will depend highly on the PDFs of the scatterer echo ampli-
tudes and delays. The following section considers two plausible (9)
physical characterizations of sea-floor scattering with a finite
number of scatterers or a finite number of patches, both leading The joint CF of the sum in (5) is the product of the joint CFs of
to the K-distribution for the matched filter envelope PDF. The the components of the sum or, as the summands are statistically
derivations may not be rigorous with regard to the physics as identical, the joint CF of raised to the power . Let
ABRAHAM AND LYONS: NOVEL PHYSICAL INTERPRETATIONS OF K-DISTRIBUTED REVERBERATION 803

, , , , and where the CF of a zero-mean Gaussian random variable with

. The joint CF of is then variance is [26]
(14)
Clearly, (10) and (13) have the same form, illustrating that ex-
ponentially sized scatterers produce K-distributed reverberation
with the shape parameter of the K-distribution equal to half the
number of scatterers

(15)

and the scale parameter related to the average scatterer size

through
(10)
(16)

where and The importance of the last two equations lies in the fact that they
link the parameters of the K-distribution to a physical model of
(11) the environment, making prediction of the reverberation PDF
possible for different sonar system configurations.
is a Bessel function of the first kind and order zero [17], and sim- It is worthwhile to note that the previous derivation does not
plification of the last line comes from Gradshteyn and Ryzhik imply the following converse statement: if the matched filter en-
[25, p. 729, 6.611-1]. velope is K-distributed, then, the scatterer amplitude in (5) must
Inversion of would show that the exponential scat- be exponentially distributed. In fact, there are other statistical
tering model in fact results in K-distributed reverberation. It is, distributions that may be used to describe the scatterer ampli-
however, easier to derive the joint CF for K-distributed rever- tude that will result in a K-distributed matched filter envelope.
beration and exploit the uniqueness property of CFs than to in- An example of this may be found in the work of Lord [27] where
vert (10). In order to accomplish this it is convenient to use the the joint PDF of an -dimensional spherically invariant random
product description of K-distributed reverberation. Ward [10] vector is described as having an exponential form in terms of
noted that it was possible to describe K-distributed reverbera- the vector length. For the two-dimensional case that represents
tion as the product between a Rayleigh speckle component and a the complex data model of (5), this results in a gamma dis-
slowly fluctuating variable following a square-root gamma dis- tributed vector length (i.e., scatterer amplitude) with a shape pa-
tribution. The complex envelope of such reverberation would rameter equal to 2 and is seen to result in a K-distributed vector
have the form length (i.e., the matched filter envelope) when such vectors
are summed.
(12)
B. Exponentially Sized Sea-Floor Patches
where is a zero-mean, low-pass, complex, Gaussian A theoretical model for reverberation resulting from interface
random process with variance and is a gamma dis- scattering off of a sea floor comprised of patches was devel-
tributed random process with unit scale, shape parameter oped by Crowther [6] where it was assumed that the patches
and a low-pass spectrum. A more realistic model would were characterized by a two-dimensional Markov model. Sup-
allow to be nonstationary with a time-varying variance. pose instead that the patches dominate the background and are
However, for the purposes of this paper, such a generality is characterized as having random size; for example, an exponen-
not necessary, so the variance is assumed constant in order to tially distributed area
simplify notation. If , then and
are independent, zero-mean, real, Gaussian random processes (17)
with variance . The joint CF of in (12) is found to be
[with , , , and ] where is the area for the th patch and is the average patch
area. Diffuse scattering from the interface produces a Gaussian
distributed return from each patch by virtue of the CLT, with
power proportional to the patch area and backscattering coeffi-
cient. Now assume that there are such patches within a res-
olution cell of the sonar system. The reverberation component
of the complex envelope of the received signal in this resolution
cell (i.e., after beamforming and matched filtering) may then be
represented as

(13) (18)
804 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 27, NO. 4, OCTOBER 2002

where is a zero-mean, complex, Gaussian random variable PDFs between samples from two different times) have not yet
with variance . Such a characterization of the sea floor can been established and thus limit the physical interpretation of the
not be described by the model (5) of the previous section be- random processes of the product formulation of K-distributed
cause the time delay of the individual scatterers is not uniformly reverberation, and .
distributed.
Noting that an exponential random variable is a gamma
IV. BEAMWIDTH ANALYSIS OF EXPERIMENTAL DATA
random variable with unit shape parameter, the joint CF of one
term in the summation of (18) is found from (13) by setting A potential use for the models described in the previous
and . Assuming that the responses from each section lies in improving the fidelity of models that simulate
patch are statistically independent then leads to the joint CF of sonar systems or predict their performance. In the sonar com-
the summation munity, the Rayleigh distribution and the associated receiver
operating characteristic curves of probability of detection
(19) versus probability of false alarm, as may be found in Urick [29]
or Burdic [30], are typically used regardless of the goodness
Clearly, this is the same form as that of (13) for the joint CF of of fit of the Rayleigh distribution to observed reverberation
the K-distribution with , the number of patches in a reso- data. The radar community has accepted statistical models that
lution cell, and , the average area of each patch times are more representative of their heavy-tailed data including the
the backscattering power per unit area. It is interesting to note Weibull and K-distributions [11], though these models tend to
that even though each patch produces Gaussian reverberation, be used with measured parameters rather than ones derived
the resultant sum is non-Gaussian owing to the random power from physics based models [11], [31].
of each component. This result also illuminates the fact that the As a motivating example, consider a sonar system operating
K-distribution, at the bandpass or complex envelope stage, is in an environment where the scatterer density and propagation
closed under addition when the summands have the same-scale are such that there are scatterers in a particular range-
parameter ( ). Thus, if the sea-bed is comprised of bearing resolution cell with each scatterer having average power
patches of differing types the result will still be K-distributed so that (using the results of Section III-A) . The
if the scale parameters are equal; i.e., . When the reverberation power at the matched filter output is then
scale parameters are not equal, the envelope will not be K-dis- where . The probability of false alarm
tributed; however, if one patch type dominates the other in the for K-distributed reverberation and the Rayleigh distribution of
sense of then, the K-distribution may be a good the same power are shown in Fig. 1, as Case 1. At ,
approximation. the K-distribution predicts a threshold level 1.7-dB higher than
that predicted by the Rayleigh distribution, illustrating how it
C. Multipath Propagation accounts for the heavier tails of the non-Rayleigh reverberation
As previously admitted, the models and methods used in the and will, in turn, result in more accurate predictions of the prob-
previous sections to link a description of the environment in ability of detection when the reverberation is not Rayleigh dis-
terms of the number of scatterers within a resolution cell of a tributed. In the future, it may be possible to obtain values for
sonar system to the statistical distribution of the matched filter and from geophysical measurements of scatterer size and
envelope are at times coarse approximations. Perhaps, the most density (i.e., frequency of occurrence). For now, however, these
egregious of the omissions has revolved around the implicit as- must be estimated from observed reverberation data, as is done
sumptions about acoustic propagation between the source array, in the analysis of the following sections.
the scatterers, and the receiving array. No mention has been Now, suppose that a sonar system were employed with a
made of any particular model; however, it has been unrealisti- towed-array receiver twice the size of the previous example.
cally assumed that no multipath propagation occurs. Were it to The larger array provides beamwidths half the size of the
occur, the effect would be to make the reverberation at any given smaller array and would therefore yield half as many scatterers
time more Rayleigh-like, owing to scatterers arriving from dif- in the range-bearing resolution cell under consideration,
ferent parts of the sea-floor through alternative acoustic prop- resulting in and the expected 3-dB reduction in the
agation paths. That the K-distribution is closed under addition reverberation power level. The curves for this situation are
indicates that the envelope may still be K-distributed in the pres- shown in Fig. 1, as Case 2, and illustrate a 2.7-dB difference in
ence of multipath if the path amplitudes do not vary signifi- the predicted threshold levels for the Rayleigh and K-distribu-
cantly. If the path amplitudes do differ significantly, the joint tional models when . The disparity in prediction
CF of the complex envelope would be a product of terms sim- is worse when , and, therefore, , is smaller and better when
ilar to (13) and the envelope distribution would have to be found is larger and the K-distribution more closely resembles the
computationally through either Padé approximations [28] or a Rayleigh distribution.
numerical Hankel transform. Under both distributional models of this example (Rayleigh
Under multipath propagation conditions the reverberation and K), the reverberation power level decreases by a factor of
would also become more correlated in time, which exposes a two when the array size is doubled. However, under the K-distri-
second weakness of the current derivation. The link between bution model, the shape parameter also decreases by a factor of
exponential scattering and the K-distribution has been estab- two. It is this aspect of the finite-number-of-scatterers model de-
lished for first-order PDFs. The second–order PDFs (i.e., joint veloped in Section III-A that we wish to examine using real data,
ABRAHAM AND LYONS: NOVEL PHYSICAL INTERPRETATIONS OF K-DISTRIBUTED REVERBERATION 805

Fig. 1. Probability of false alarm under the Rayleigh and K-distributional models. When P = 10 , the Rayleigh model predicts a 1.7 dB smaller threshold
than the K-distribution for Case 1 and a 2.7-dB smaller threshold for Case 2. Case 2 differs from Case 1 in that the receiving towed-array is twice as large yielding
a beamwidth that is twice as small.

specifically to see if the estimated shape parameter changes the latter produced Rayleigh distributed reverberation and the
linearly with the beamwidth of a sonar receiving array as is former produced heavier-tailed non-Rayleigh reverberation.
predicted by the model. The following sections describe the The towed-array receiver was comprised of 128 hydrophones
sea-trial from which data are analyzed, illustrate that they are with a 0.5-m spacing, resulting in a design frequency of 1500
well described by the K-distribution, and finally examine the Hz. The array data were beamformed with a hanning window
dependence of the shape parameter on the beamwidth of the such that the beampatterns of adjacent beams overlapped at their
towed-array receiver. 3-dB down points at 900 Hz, resulting in 54 beams equally
spaced in wavenumber, spanning from forward to aft along the
A. Sea-Trial Description array. The signal processing applied to the beamformed data
prior to analysis of the reverberation statistics included base-
The data to be analyzed were taken during the SCARAB ’97 banding, match filtering and normalization. The normalizer is
sea-trial (Chief Scientist, Charles Holland) sponsored by the of the cell-averaging type [11] with a split-window and gap, as
SACLANT Undersea Research Centre, La Spezia, Italy. The described in [34]. The leading and lagging windows each con-
trial occurred off the coast of Italy in June, 1997 in the Capraia tained 100 samples at a sampling rate of 250 Hz,1 so the rever-
Basin which is north of the Island of Elba and contained several beration power estimate is formed by averaging 0.8 s of rever-
components related to scattering and reverberation [32]. The beration data. The gap on each side of the sample being normal-
data analyzed in this paper were taken on June 2–3 using ized was 5-samples long.
the Centre’s low-frequency towed hydrophone-array and the
towed vertically directive source (TVDS), which was config- B. Fit of the Rayleigh and K-Distributions
ured to send a linear frequency modulated pulse with a 2-s
duration spanning 450–700 Hz. The subsequent reverberation The shape parameter of the K-distribution is estimated from
was recorded from the towed-array for 50 s following each the beamformed, matched filtered, and normalized data. Before
transmission. Forty-nine such pings of data were collected evaluating the shape parameter and how it changes with array
at the points shown in Fig. 2. Owing to the shallow water beamwidth, it is necessary to determine how well fit the data
environment and downward refracting sound-speed profiles are by the K-distribution. Fig. 3 contains the probability of false
that were measured during the sonar data acquisition, the alarm estimated from a window of normalized matched filter
reverberation data are dominated by bottom scattering. As data 2000 samples long (8 s). The ping from which these data are
described in [33], bottom scattering from the western portion obtained is noted in Fig. 2. The obtained from the Rayleigh
of the Capraia Basin seems to be dominated by magmatic rock and K-distributions whose parameters are estimated from this
outcrops while the central and eastern parts of the basin are data are also shown where the K-distribution clearly provides a
dominated by scattering from sub-bottom sediment volume 1As the sampling rate is equal to the transmit waveform bandwidth, consec-
inhomogeneities. The analysis presented in [33] illustrated how utive samples are approximately uncorrelated.
806 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 27, NO. 4, OCTOBER 2002

Fig. 2. Locations of each ping of data analyzed on a bathymetric map of the Capraia Basin. Each circle indicates the location of the ship. The trailing line from
each circle illustrates where the towed-array was located relative to the ship (i.e., ship heading minus 180 ). Bathymetry is in meters.

Fig. 3. Probability of false alarm estimated from normalized matched filter data along with Rayleigh and K-distribution models with parameters estimated from
the data.

better fit to the data. This particular segment of data was chosen In order to quantify how much of the data are well fit by the
to illustrate that the Rayleigh distribution is not always a good fit Rayleigh and K-distributions, the Kolmogorov–Smirnov (KS)
to observed reverberation data. A significant portion of the data test [35] is applied to the normalized matched filter data to test
collected and analyzed, however, were well fit by the Rayleigh the ability of the models to represent the observed data. The KS
distribution. As the K-distribution has the Rayleigh distribution test evaluates the maximum difference between the sample cu-
as a submember, the K-distribution fit these data well too. mulative distribution function (CDF) generated by the data and
ABRAHAM AND LYONS: NOVEL PHYSICAL INTERPRETATIONS OF K-DISTRIBUTED REVERBERATION 807

TABLE I bearing resolution cell. As such, if the beamwidth of the sonar is

PERCENT OF DATA ACCEPTED AS BEING WELL FIT BY THE RAYLEIGH OR doubled, then should also double, regardless of the density of
K-DISTRIBUTIONS BASED ON THE K–S TEST AT VARIOUS LEVELS OF
FALSE REJECTION scatterers as long as they are uniformly placed within the range-
bearing resolution cell. This latter point implies that the effect
should occur for all the data observed, regardless of whether
they are non-Rayleigh or Rayleigh-like.
The beamwidth of a sonar approximately doubles when the
array size is halved. The same effect may be accomplished by
coherently summing adjacent beams. As previously mentioned,
the towed-array data have been spatially filtered into 54 beams
that overlap at the 3-dB down points of their beampatterns
at 900 Hz. The transmit waveform being analyzed ranged
from 450 to 700 Hz, at which frequencies the beampatterns of
a test CDF which is, in this case, either the Rayleigh or K-dis- adjacent beams overlap, respectively, at the 0.75- and 1.8-dB
tribution with their parameters estimated from the data being down points. A beampattern having width equal to the com-
tested. The Rayleigh distribution only depends on its power, bined width of the individual beams may be formed without
which is estimated by the sample intensity (i.e., the average of significant destruction of the mainlobe by coherently summing
the matched filter intensity over the window being tested). As every other beam at the spacing described above. This does
the data have already been normalized to have unit power, this result in a ripple in the mainlobe with less than a one decibel
should be near one. Estimation of the K-distribution parameters height at 450 Hz and less than a two decibel height at 700 Hz.
is more involved. The method of moments estimator (MME) has To illustrate this beam-summing, the beampatterns of three of
been employed and is described in detail in Appendix B. the original beams near broadside to the array and the summed
For the data under consideration, windows 1000-samples beam they form are shown in Fig. 4. From this figure, it may be
long (4 s of data) with 50% overlap are used to estimate the seen that the width of the broadside beam at 700 Hz is 2.6 . At
model parameters and then form the KS test statistic. Using 450 Hz, the broadside beam is 4.1 wide.
the asymptotic -value2 of the KS test statistic [35], the data The shape parameter of the K-distribution is estimated from
are either accepted as being well fit by the Rayleigh or K-dis- either the original beam data or the summed beam data after
tribution or rejected. Table I contains the results for various matched filtering and normalization. For each beam, the data
-values. For example, at the level, 68% of the data are separated into windows that are 500-samples long (at a sam-
are well fit by the Rayleigh distribution and 97.4% are well pling rate of 250 Hz this is 2 s of data) with an 80% overlap. For
fit by the K-distribution. Note that only data between 2 and the th window on the th beam of the th ping, call this es-
40 s following the end of signal transmission on each ping are timate where the index indicates the span of how
utilized in this and subsequent data analysis. Over these times, many original beams the beam-sum is formed over (beam-spans
the data are reverberation limited except on the few beams of 1, 3, 5, 7, and 9 are analyzed). The MME, as described
where nearby shipping dominated the reverberation or where in Appendix B, is used to estimate the K-distribution shape pa-
geological features precluded acoustic propagation (e.g., the rameter. The estimates of the shape parameter from the single
Island of Capraia causing shadow zones). ping of data mentioned in the previous section are displayed in
As described in Appendix B, the method of moments does not Fig. 5 in the form of histograms for each of the beam-spans.
always yield a solution for the K-distribution parameters. When As expected, the data exhibit a trend toward higher values as
a solution is not available, the shape parameter of the K-dis- the beamwidth increases. On the figure, this is evident from
tribution is arbitrarily set to and the scale parameter the median, mean (which was trimmed by removing the largest
is obtained using the moment equation described in (39). That and smallest 0.5 percent of the values), and the quantity of es-
these cases tend to be when the data are nearly Rayleigh is evi- timates that exceed 50. It is also clear that there is wide range
dent from the change in acceptance percentages seen in Table I of values observed on this single ping, indicating that the den-
when only those data yielding an MME solution are considered. sity of scatterers (i.e., frequency of occurrence) varies within
In this case, a slightly higher percentage of the data are well fit the geographic region represented by this ping. To remove this
by the K-distribution and slightly smaller percentage are well fit variability, the ratio of the estimated shape parameters on the
by the Rayleigh distribution. Based on the acceptance percent- summed beams to that estimated on the individual beams is
ages shown in Table I, the K-distribution is accepted as a good formed for each individual data window
model for these data and it is noted that a significant portion of
the data are Rayleigh-like. (20)
C. Real Data Analysis
where represent the original beams that
Based on the model developed in Section III-A, it may be span the beamwidth of summed beam . An average value is
hypothesized that the effective number of scatterers parameter formed for each ping and beam sum
( ) of the K-distribution is proportional to the area of the range-
2The p-value is the probability that a data sample would be rejected when it (21)
should be accepted as being well fit by the model under consideration.
808 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 27, NO. 4, OCTOBER 2002

Fig. 4. Beampatterns of three original beams near broadside to the array and the summed beam that spans them at 700 Hz. The ripple in the main-lobe of the
summed beam, approximately 2 dB, is at its worst at this frequency.

Fig. 5. Histograms of the estimates of the K-distribution shape parameter from a single ping for various beamwidths. The expected trend toward higher values of
 as beamwidth increases is evident in the mean, median and number of estimates exceeding 50.

where the summation is over the cases of the indices ( ) is large, it is difficult to accurately estimate its value even when
such that the method moments estimators provide shape param- an efficient estimator is used and the MME is not likely to be
eter estimates for both the individual and summed beams and the efficient.
estimates satisfy and . Ex- The change in the shape parameter predicted by the model
tremely large estimates of are discarded because of their high is exactly the change in beamwidth. This is formed in a similar
variability. As shown in Appendix B, the Cramér–Rao lower manner to how the change in the shape parameter is estimated
bound (CRLB) for the variance of any unbiased estimator of is from the data, by taking the average of the change in beamwidth
well approximated by , especially for large . Thus, when between the summed beams and each of the individual beams
ABRAHAM AND LYONS: NOVEL PHYSICAL INTERPRETATIONS OF K-DISTRIBUTED REVERBERATION 809

Fig. 6. Change in the shape parameter of the K-distribution between summed and individual beams as a function of the span of the beams summed (i.e.,
beamwidth). The + marks are the average values from each ping. The change as estimated from the data is very similar to that predicted by the change in
beamwidth with the exception of an explainable upward bias.

for all of the summed beams. Despite the fact that beamwidth number of scatterers with the shape parameter tied to the
varies with arrival angle, the ratios of the beamwidths of the number of scatterers in a sonar range-bearing resolution cell
summed and individual beams do not vary significantly. These can provide more realistic simulations or predictions of sonar
average ratios are computed for the edges of the frequency band system performance as a function of system parameters such
of the transmit waveform (450 and 700 Hz) and also for the peak as beamwidth (shown here) and possibly bandwidth when the
power frequency of the source (600 Hz) and shown with the av- reverberation data are known to be non-Rayleigh.
erage change in the estimated shape parameter of the K-distribu-
tion in Fig. 6. In the figure, it is seen that the estimated changes
V. CONCLUSION
fall very close to the predictions and that the average
value over all pings has a slope similar to that predicted for the In this paper, the statistical distribution of reverberation
450 Hz case, but is biased high. This may be explained by first resulting from two plausible scenarios for sea-floor scattering
noting that the expected change (20) is formed from the ratio of has been examined. Sea floors composed of exponentially
two random variables, say . Assuming that both sized individual scatterers were considered as well as those
and are positive random variables, it can be shown through the comprised of homogeneous patches that have exponential size
use of Jensen’s inequality [36] that . This and are dominated by interface scattering. In both cases, the
result also requires that the two random variables in the ratio be matched filter envelope was found to be K-distributed with
independent, which is not the case with the real data analyzed. the shape parameter related to either the number of scatterers
However, for the minimal dependence that is expected, the re- or the number of patches and the scale parameter related to
sult should be approximate and does proffer an explanation for the average size of the scatterer or patch and backscattering
the bias seen in the data. coefficient. The models were developed using a finite number
The data shown in Fig. 6 are also seen to increase in variability of scatterers or patches, in contrast to the more traditional
as the beam-span increases. This may be the result of variability asymptotic derivation of the K-distribution that requires an
induced by nonstationarity in the frequency of occurrence of infinite number of scatterers. The importance of this lies in the
scatterers on the bottom over wider angles or may arise from ability to link the reverberation envelope distribution to measur-
the increased variability of shape parameter estimates when able geo-acoustic properties in conjunction with the lens with
is large. The latter of which is expected to occur more often which they are viewed (i.e., the sonar system). To illustrate this,
for larger beamwidths (i.e., larger range-bearing resolution cell low-frequency active sonar reverberation data were analyzed
sizes) and when the data are nearly Rayleigh distributed (which in terms of the shape parameter of the K-distribution and the
is known to occur frequently in this data set). beamwidth of the towed-array receiver. The analysis confirmed
The close similarity between the estimated and predicted that, as predicted by the finite-number-of-scatterers model, the
change in the K-distribution shape parameter indicates that, shape parameter of the K-distribution is proportional to the
within our ability to estimate it, the shape parameter is pro- array beamwidth, even when the data were well described by
portional to the beamwidth of the towed-array receiver. This the Rayleigh distribution (i.e., a large number of scatterers in
also implies that a K-distribution model assuming a finite each range-bearing resolution cell).
810 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 27, NO. 4, OCTOBER 2002

The relationships presented in this paper provide the founda- the envelope . Thus, can be factored
tion necessary for solving several important problems related so that the envelope is seen to be a sufficient statistic
to the detection of targets in non-Rayleigh reverberation in- for target parameters required to define and interference
cluding both environmentally driven concerns and signal pro- parameters required to define .
cessing issues. Some of the former include prediction of the By making varying assumptions about the distribution on the
reverberation induced envelope probability distribution func- target amplitude, the common Swerling target models [38] may
tions from measured geo-acoustic data as well as consideration be obtained. When , then the target has a
of more complicated sea-floor scattering and multipath prop- constant amplitude and the nonfluctuating or Swerling
agation scenarios. Signal processing issues that may now be 0 target model arises. When is Rayleigh distributed the target
addressed include the development of optimal and suboptimal component is a zero-mean complex Gaussian random vari-
detectors and system performance evaluations that account for able, which produces the fluctuating or Swerling 1 target model.
non-Rayleigh reverberation such as evaluating the trade-offs be- Sufficiency of the envelope may also be shown for the one-dom-
tween improved SNR derived from an increase in the transmit inant-plus-Rayleigh Swerling 3 target through a similar proce-
signal bandwidth and the concomitant increase in false alarms dure.
resulting from heavier reverberation distribution tails.
APPENDIX B
APPENDIX A ESTIMATION OF THE K-DISTRIBUTION SHAPE PARAMETER
SUFFICIENCY OF THE ENVELOPE FOR CIRCULARLY
SYMMETRIC REVERBERATION Considering the PDF of (7), estimation of and for the
K-distribution certainly seems to be a formidable problem.
Let the complex envelope of an active sonar return be com- Ward [31] provided an empirical formula for based on
posed of a target component and interference (reverberation or the grazing angle and cross-range resolution of a vertically
clutter, and ambient noise) polarized radar. Raghavan [39] proposed an estimator based on
(22) approximating the K-distribution by a gamma distribution. The
estimators resulted in a form depending on the arithmetic and
where and are the amplitude and phase of the target echo
geometric sample means that perform well when the K-distri-
and is the complex envelope of the interference. Equation
bution is distinctly non-Rayleigh. Joughin et al. [40] presented
(22) may be obtained from (4) by setting and grouping
an analysis of maximum likelihood estimators (MLE) for the
the latter terms into , which is assumed to have a circularly
parameters obtained through a numerical maximization of the
symmetric PDF; that is, if then
likelihood function. They compared the MLE to Raghavan’s
(23) method [39] and method of moment estimators based on either
Sufficiency is typically shown using the Neyman factoriza- the first and second or second and fourth moments. Simulation
tion theorem [37] where the PDF is factored into two terms, the analysis showed that the MLE performed best when was
first one depending on the statistic to be proven sufficient and small with Raghavan’s method being nearly equivalent and that
the parameters for which it is sufficient and the second term when is large the MME using the first two moments performs
only depending on the data. The random variables in (22) are best with the MLE being nearly equivalent.
, and . Conditioning on and Because the maximum likelihood techniques require numer-
, the joint PDF of is ical optimization routines and evaluation of Bessel functions,
they are computationally intensive and therefore inappropriate
for evaluation of large data sets. The MMEs are relatively quick
but suffer from a nonzero probability of not returning a solution.
Despite this failing they are used in the data analysis presented
in Section IV because they require significantly less computa-
(24) tional power than the maximum likelihood based techniques. In
this section, the MME that is employed is described and an ap-
where . Now, if is assumed to be indepen- proximation to the CRLB is derived for the shape parameter and
dent of and uniformly random over then seen to be well approximated by .

A. MME
Based on the results of Joughin et al. [40] MME using the
first two moments is recommended. This estimator provides an
acceptable trade-off between performance and cost of imple-
mentation. Let the matched filter envelope data used to estimate
the K-distribution shape parameter be . Define the
sample moments of the envelope as
(25)
where . Clearly the joint PDF of (26)
can be described as only depending on through
ABRAHAM AND LYONS: NOVEL PHYSICAL INTERPRETATIONS OF K-DISTRIBUTED REVERBERATION 811

and A reasonable initial estimate of is obtained by using the ap-

proximation of (37), without the third multiplicative term, in the
(27) natural logarithm of (31). Applying a second order approxima-
tion to then results in the
The moments of the matched filter envelope following a K-dis- starting point
tribution are easily obtained from the product form of the K-dis-
tribution [10] (38)

(28) After the Newton–Raphson iteration has converged, is es-

timated according to (29)
Matching the second moment results in choosing
(39)
(29)
Examination of the function reveals that it approaches
which, when solved for and substituted into the relationship infinity as and monotonically decreases to as
for the first moment results in . Thus, in order for there to be a positive solution to (31),
. Unfortunately, there is a nonzero probability
(30) that this inequality is not met and therefore no solution exists.
As might be expected, this happens more frequently when
or, equivalently is high and/or the number of samples used to estimate the mo-
ments is small. The problem is surmounted either by obtaining
(31) the numerically more intensive maximum likelihood estimate
or by accepting either a large (e.g., ) or the Rayleigh
distribution as an approximate fit.
Owing to the gamma functions, no closed form solution for is
obtainable. Defining the function B. CRLB for K-Distribution Shape Parameter
The CRLB provides a means for establishing the best per-
(32)
formance a parameter estimator may achieve [37]. Specifically,
it provides a lower bound on the variance of an unbiased esti-
it can be seen that (31) will be satisfied when in the mator. If the matched filter envelope data are K-distributed
function with PDF as described in (7) but with the dependence
on made explicit, then the Fisher information of with re-
(33) spect to , which is inversely proportional to the CRLB, is

The function is chosen because it is numerically stable for (40)

a wide range of . The method of moments estimate of may
then be obtained through use of a Newton–Raphson iteration and the CRLB is
[41] to find the root of . The iteration is
(41)
(34)
where is an unbiased estimator of the shape parameter (i.e.,
which requires the derivative ) and is the number of independent samples of
that are used to form .
(35) Noting that is a scale parameter, it is straightforward to see
that is invariant to . Thus, the Fisher information need
where only be determined for , which leads to the PDF

(42)
(36)
the log-likelihood function
is the digamma function [17]. Note that for large most algo-
rithms evaluating the gamma function will have problems. Ap- (43)
plication of Stirling’s approximation [41] to the gamma function
and its derivative with respect to
results in

(37)

(44)
which should be used for .
812 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 27, NO. 4, OCTOBER 2002

where is the digamma function [17]. The third term in (44) [2] M. Gensane, “A statistical study of acoustic signals backscattered from
may be obtained by approximating when is large the sea bottom,” IEEE J. Ocean. Eng., vol. 14, pp. 84–93, Jan. 1989.
[3] S. Stanic and E. G. Kennedy, “Fluctuations of high-frequency shallow-
and positive [17] water sea-floor reverberation,” J. Acoust. Soc. Amer., vol. 91, no. 4, pp.
1967–1973, 1992.
[4] W. K. Stewart, D. Chu, S. Malik, S. Lerner, and H. Singh, “Quantitative
(45) sea-floor characterization using a bathymetric sidescan sonar,” IEEE J.
Oceanic Eng., vol. 19, pp. 599–610, Oct. 1994.
The partial derivative of (45) with respect to , normalized by [5] A. P. Lyons and D. A. Abraham, “Statistical characterization of high-
, is then frequency shallow-water sea-floor backscatter,” J. Acoust. Soc. Amer.,
vol. 106, no. 3, pp. 1307–1315, 1999.
[6] P. A. Crowther, Fluctuation Statistics of Sea-Bed Acoustic
Backscatter. New York: Plenum, 1980, pp. 609–622.
[7] D. Middleton, “New physical-statistical methods and models for clutter
and reverberation: The KA-distribution and related probability struc-
tures,” IEEE J. Oceanic Eng., vol. 24, pp. 261–284, July 1999.
[8] E. Jakeman and P. N. Pusey, “A model for non-Rayleigh sea echo,” IEEE
Trans. Antennas Propagat., vol. AP-24, pp. 806–814, Nov. 1976.
[9] E. Jakeman and R. J. A. Tough, “Non-Gaussian models for the statistics
of scattered waves,” Adv. Phys., vol. 37, no. 5, pp. 471–529, 1988.
(46) [10] K. D. Ward, “Compound representation of high resolution sea clutter,”
Electron. Lett., vol. 17, no. 16, pp. 561–563, 1981.
where the final result is obtained by assuming that , [11] P. Z. Peebles, Jr., Radar Principles. New York: Wiley, 1998.
, and for small . [12] H. Griffiths, J. Dunlop, and R. Voles, “Texture analysis of sidescan sonar
imagery using statistical scattering models,” in High Frequency Acous-
Inserting (46) into (44) and then (40) and noting that tics in Shallow Water, N. G. Pace, E. Pouliquen, O. Bergem, and A. P.
for large , the Fisher information is seen to Lyons, Eds. La Spezia, Italy: NATO SACLANT Undersea Research
be approximated by Centre, pp. 187–194.
[13] S. Dugelay, N. G. Pace, G. J. Heald, and R. J. Brothers, “Statistical anal-
ysis of high frequency acoustic scatter: What makes a statistical distribu-
tion?,” in Proc. Fifth Eur. Conf. Underwater Acoustics, ECUA’00, Lyon,
France, July 2000, pp. 269–274.
[14] M. Gu and D. A. Abraham, “Using McDaniel’s model to represent non-
Rayleigh reverberation,” IEEE J. Oceanic Eng., vol. 26, pp. 348–357,
July 2001.
(47) [15] H. L. Van Trees, Detection, Estimation, and Modulation Theory: Part
III. New York: Wiley, 1971.
where is the scintillation index (i.e., the variance of the [16] A. Papoulis, Probability, Random Variables, and Stochastic Processes,
matched filter intensity divided by the square of the average 3rd ed. New York: McGraw-Hill, 1991.
[17] J. Spanier and K. B. Oldham, An Atlas of Functions. Washington, DC:
intensity). For the K-distribution, the scintillation index is Hemisphere, 1987.
[18] C. G. Fox and D. E. Hayes, “Quantitave methods for analyzing the
(48) roughness of the sea floor,” Rev. Geophys., vol. 23, pp. 1–48, 1985.
[19] A. L. Anderson, F. Abegg, J. A. Hawkins, M. E. Duncan, and A. P.
which is approximately one for large (i.e., when the K-distri- Lyons, “Bubble populations and acoustic interaction with the gassy floor
of Eckernfoerde Bay,” Continental Shelf Res., vol. 18, no. 14–15, pp.
bution is near the Rayleigh distribution). 1807–1838, 1998.
The CRLB is then approximately [20] J. Escartin, P. A. Cowie, R. C. Searle, S. Allerton, N. C. Mitchell, C. J.
MacLeod, and P. Sloogweg, “Quantifying tectonic and magmatic strain
at a slow-spreading ridge segment (mid-Atlantic ridge, 29 N),” J. Geo-
(49) phys. Res., vol. 105, no. B5, pp. 10 421–10 427, 1999.
[21] N. C. Mitchell, M. A. Tivey, and P. Gente, “Sea-floor slopes at mid-
for large which implies that it is harder to estimate when ocean ridges from submersible observations and implications for inter-
preting geology from sea-floor topography,” Earth Planetary Sci. Lett.,
the K-distribution is nearly Rayleigh (i.e., when the two param- vol. 183, pp. 543–545, 2000.
eter model is well described by a one-parameter model). Numer- [22] S. W. Anderson, E. R. Stofan, J. J. Plaut, and D. A. Crown, “Block size
ical evaluation of the CRLB through Monte-Carlo integration of distributions on silicic lava flow surfaces: Implications for emplacement
conditions,” Geol. Soc. Amer. Bull., vol. 110, no. 10, pp. 1258–1267,
(40) with a first order difference approximation to the derivative 1998.
with respect to confirmed the above approximation as accu- [23] M. Golombek and D. Rapp, “Size-frequency distributions of rocks on
rate for large . mars and earth analog sites: Implications for future landed missions,” J.
Geophys. Res., vol. 102, pp. 4117–4129, 1997.
[24] H. Medwin and C. S. Clay, Fundamentals of Acoustical Oceanog-
ACKNOWLEDGMENT raphy. Boston, MA: Academic, 1998.
[25] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Prod-
The authors wish to thank K. Fisher and B. La Cour for ucts, fifth ed, A. Jeffrey, Ed. San Diego, CA: Academic, 1994.
pointing out the work of Lord [27], and C. Holland for many [26] M. Evans, N. Hastings, and B. Peacock, Statistical Distributions, 3rd
helpful discussions and his efforts in organizing and leading ed. New York: Wiley, 2000.
[27] R. D. Lord, “The use of the Hankel transform in statistics I. General
the SCARAB sea trial. theory and examples,” Biometrika, vol. 41, no. 1/2, pp. 44–55, 1954.
[28] D. M. Drumheller, “Padé approximations to matched filter amplitude
REFERENCES probability functions,” IEEE Trans. Aerosp. Electron. Syst., vol. 35, pp.
1033–1045, July 1999.
[1] N. P. Chotiros et al., “Acoustic backscattering at low grazing angles [29] R. J. Urick, Principles of Underwater Sound. New York: McGraw-
from the ocean bottom. Part II. Statistical characteristics of bottom Hill, 1983.
backscatter at a shallow water site,” J. Acoust. Soc. Amer., vol. 77, no. [30] W. S. Burdic, Underwater Acoustic System Analysis. Englewood
3, pp. 975–982, 1985. Cliffs, NJ: Prentice-Hall, 1984.
ABRAHAM AND LYONS: NOVEL PHYSICAL INTERPRETATIONS OF K-DISTRIBUTED REVERBERATION 813

[31] K. D. Ward, “Compound representation of high resolution sea clutter,” Douglas A. Abraham received the B.S., M.S., and Ph.D. degrees in electrical
in Proc. IEE Conf. Publ. 216 (Radar-82), 1982, pp. 203–207. engineering, and the M.S. degree in statistics, all from the University of Con-
[32] C. W. Holland and J. Osler, “High-resolution geoacoustic inversion necticut, Storrs, in 1988, 1990, 1993, and 1994, respectively.
in shallow water: A joint time- and frequency-domain technique,” J. He is currently a Senior Research Associate with the Applied Research Labo-
Acoust. Soc. Amer., vol. 107, no. 3, pp. 1263–1279, 2000. ratory, Pennsylvania State University. During 1989–1995, he was with the Naval
[33] D. A. Abraham and C. W. Holland, “Statistical analysis of low-fre- Undersea Warfare Center, New London, CT. During 1995–1998, he was a Se-
quency active sonar reverberation in shallow water,” in Proc. 4th Eur. nior Scientist at the NATO SACLANT Undersea Research Centre, La Spezia,
Conf. Underwater Acoustics, Rome, Italy, Sept. 1998. Italy. During 1998–2000, he held a Visiting Faculty Position at the University
[34] D. A. Abraham and P. K. Willett, “Active sonar detection in shallow of Connecticut. His work is primarily in the area of statistical signal processing
water using the Page test,” IEEE J. Oceanic Eng., vol. 27, pp. 35–46, applied to underwater acoustic applications. His current interests are in repre-
Jan. 2002. senting and accounting for non-Rayleigh active sonar reverberation in signal
[35] M. Fisz, Probability Theory and Mathematical Statistics, 3rd ed. New processing algorithms for detection, classification and localization.
York: Wiley, 1963.
[36] E. B. Manoukian, Modern Concepts and Theorems of Mathematical Sta-
tistics. New York: Springer Verlag, 1985.
[37] R. V. Hogg and A. T. Craig, Introduction to Mathematical Statistics, 4th Anthony P. Lyons received the B.S. degree in physics from Henderson State
ed. New York: Macmillan, 1978. University, Arkadelphia, AR, in 1988 and the M.S. and Ph.D. degrees in
[38] J. V. DiFranco and W. L. Rubin, Radar Detection. Dedham, MA: oceanography from Texas A&M University, College Station, in 1991 and 1995,
Artech House, 1980. respectively.
[39] R. S. Raghavan, “A method for estimating parameters of K-distributed He worked as a Scientist at the SACLANT Undersea Research Center, La
clutter,” IEEE Trans. Aerosp. Electron. Syst., vol. 27, pp. 238–246, Mar. Spezia, Italy, from 1995 to 2000. where he was involved in a variety of projects
1991. in the area of environmental acoustics. He is currently a Senior Research Asso-
[40] I. R. Joughin, D. B. Percival, and D. P. Winebrenner, “Maximum like- ciate at the Applied Research Laboratory, Pennsylvania State University, State
lihood estimation of K distribution parameters for SAR data,” IEEE College, where he is engaged in studies of high-frequency shallow-water prop-
Trans. Geosci. Remote Sensing, vol. 31, pp. 989–999, Sept. 1993. agation, acoustic interaction with the sea floor, and high-resolution characteri-
[41] L. Råde and B. Westergren, Beta  Mathematics Handbook, 2nd zation of sea-floor sediments.
ed. Boca Raton, FL: CRC, 1990. Dr. Lyons is a member of the Acoustical Society of America.