While seabed characterization methods have often focused on estimating individual sediment parameters, deep learning suggests a class-based approach focusing on the overall acoustic effect. A deep learning classifier—trained on 1D synthetic waveforms from underwater explosive sources—can distinguish 13 seabed classes. These classes are distinct according to a proposed metric of acoustic similarity. When tested on seabeds not used in training, the classifier obtains 96% accuracy for matching such a seabed to one of the top-3 most acoustically similar classes from the 13 training seabeds. This approach quantifies the performance of a seabed classifier in the face of real seabed variability.

In the shallow ocean, seabed properties greatly affect how sound propagates. From a received sound, the signature of the ocean environment can be decoded using seabed characterization techniques. Many geoacoustic inversions,1–9 Bayesian optimizations,10–12 and neural network techniques13,14 have focused on estimating individual acoustic properties of the seabed. However, the acoustic impact of the individual properties is often coupled, which leads to non-unique solutions.15,16 Instead of searching for individual seabed properties, an alternative approach classifies a seabed type from the acoustic signals.

Machine and deep learning can be used for seabed classification. Three and four seabed classes were distinguished in Refs. 17 and 18, respectively, using scattering from side-scan sonar. Recent passive acoustic research into seabed classification has distinguished up to four seabeds of distinct sediment types using pressure time series from a single hydrophone,19 mid-frequency tones from a moving source,20 and spectrograms from passing ships.21 Reference 22 distinguishes four seabed types for multiple data types, both low-frequency pressure fields and high-frequency backscatter.

Building on this foundation, this paper presents a passive method for approximate seabed classification which is validated via a measure of acoustic similarity between seabeds. The main idea is to develop a catalog of representative seabed types that are acoustically distinguishable and to train a convolutional neural network (CNN) classifier on signals simulated with these representative seabed classes. The trained CNN can then be applied to acoustic signals from different seabeds, which it links to a similar representative seabed within the catalog. This process is essentially classification; however, the acoustic effect in the water column for each new seabed is unlikely to match any of the representative seabeds exactly, only approximately. The quality of this approximate classification can be validated by a measure of acoustic similarity between seabeds—since within this experimental setup we possess a geoacoustic model of the new seabed.

This method of linking a new seabed to an approximately similar representative class has several advantages over finding individual physical properties of the seabed. (1) Linking is more stable: A search for individual seabed properties can yield non-physical combinations of parameters and is dependent on parameter sensitivity and coupling. (2) Linking does not require a priori decisions on which parameters to search for. (3) Linking emphasizes the overall acoustic effect of the seabed and its impact on the propagation.

The key component to developing this seabed linking approach is a measure of seabed similarity. Specifically, our measure of acoustic similarity between seabeds is the Pearson correlation of transmission loss vs range for one-third-octave band center frequencies. This correlation measure provides a principled way to choose a catalog of distinct seabed types and also indicates the similarity of additional seabeds to those in the catalog.

The catalog of representative seabed types is used to generate synthetic pressure times series, which are then used to train a CNN classifier. Synthetic time series from additional seabeds are then passed through the classifier to test how well the classifier can link new seabeds to similar seabeds in the catalog. The performance is quantified using the Pearson correlation, and the median accuracy for linking a new seabed to one of the three most similar classes in the catalog is 96%.

One way to measure the acoustic similarity between two seabeds compares the resulting transmission loss (TL) as a function of range and frequency. This similarity in TL is quantified using the Pearson product-moment correlation.

The TL associated with each seabed is calculated for depths of water, source, and receiver, as well as ranges and frequency bands, that are similar to those used in the data generation (Sec. 2.3). In this work, ORCA, a normal-mode model,23 simulates TL for 14 one-third-octave band center frequencies from 25 to 500 Hz and for 250 source-receiver ranges from 1.5 to 12 km. A typical summer sound speed profile in the water column24 is used. For each seabed, the resulting TL matrix (over range and frequency) is then flattened to a 1D vector. The similarity between the characteristic TL vectors of two seabeds is measured using the Pearson product-moment correlation. For two TL vectors x and y, this is defined as

rxy=i=1n(xix¯)(yiy¯)i=1n(xix¯)2i=1n(yiy¯)2,thus1rxy+1.
(1)

The Pearson correlation is the normalized cross correlation with zero time offset and is selected because the elements of the vectors are aligned. Two example pairs of seabed characteristic TL vectors are shown in Fig. 1(a), with Pearson correlation values of 0.35 (upper) and 0.75 (lower).

Fig. 1.

(a) Transmission loss (TL) vectors (over one-third-octave band frequencies and ranges) for two pairs of seabeds. The Pearson correlations of these pairs of vectors are 0.35 (upper) and 0.75 (lower). (b) Pearson correlation values for pairs of TL vectors associated with the 30 seabeds listed in Table 1. The more acoustically distinct seabeds appear earlier in the numbering, preparing for the division between the catalog of representative seabeds (r1–r13) and the set of testing seabeds (t14–t30).

Fig. 1.

(a) Transmission loss (TL) vectors (over one-third-octave band frequencies and ranges) for two pairs of seabeds. The Pearson correlations of these pairs of vectors are 0.35 (upper) and 0.75 (lower). (b) Pearson correlation values for pairs of TL vectors associated with the 30 seabeds listed in Table 1. The more acoustically distinct seabeds appear earlier in the numbering, preparing for the division between the catalog of representative seabeds (r1–r13) and the set of testing seabeds (t14–t30).

Close modal

For this study, 30 seabed geoacoustic profiles were selected from 14 papers, as listed in Table 1. From fine sand to deep mud, these seabeds represent a variety of shallow water environments.

Table 1.

Seabeds used in this study come from peer-reviewed literature. Those labeled r1–r13 form the catalog of representative seabeds; t14–30 are used to test the CNN's ability to link them to a similar seabed type.

Seabed IdentifierRef.LocationSediment typeFrequency (source)
r1 25  NE Mud Patch mud over sand 25–275 Hz (Passive) 
r2 1  Mid -Atlantic Bight sandy silt 10–200 Hz (Passive) 
r3 2  Gulf of Mexico deep mud 75–500 Hz (Passive) 
r4 26  Multiple sand <1 kHz (Passive) 
r5-6 3  Gulf of Mexico sand-silt-clay 3.5 kHz (Sonar) 
r7,t16 4  New Jersey sand, clay 35–180 Hz (Passive) 
r8,t14,t21,t28 5  Jeongdongjin silt, sand, gravel 1 MHz (Core) 
r9 27  Yellow Sea “fine-grained” 1 MHz (Core) 
r10,t18-19,25,27,29-30 28  New Jersey sandy 35–265 Hz (Passive) 
r11-12,t15,t20 6  workshop sand, silt <500 Hz (Synthetic) 
r13,t24 7  Jinhae Bay mud, sand 4–12 kHz (Core) 
t17 8  Ulsan Coastal Area mud 1 MHz (Core) 
t22 9  Malta Plateau cobbles in silt 1.2 kHz (Passive) 
t23,t26 8  East China Sea sand, muddy sand 3.5 kHz (Core) 
Seabed IdentifierRef.LocationSediment typeFrequency (source)
r1 25  NE Mud Patch mud over sand 25–275 Hz (Passive) 
r2 1  Mid -Atlantic Bight sandy silt 10–200 Hz (Passive) 
r3 2  Gulf of Mexico deep mud 75–500 Hz (Passive) 
r4 26  Multiple sand <1 kHz (Passive) 
r5-6 3  Gulf of Mexico sand-silt-clay 3.5 kHz (Sonar) 
r7,t16 4  New Jersey sand, clay 35–180 Hz (Passive) 
r8,t14,t21,t28 5  Jeongdongjin silt, sand, gravel 1 MHz (Core) 
r9 27  Yellow Sea “fine-grained” 1 MHz (Core) 
r10,t18-19,25,27,29-30 28  New Jersey sandy 35–265 Hz (Passive) 
r11-12,t15,t20 6  workshop sand, silt <500 Hz (Synthetic) 
r13,t24 7  Jinhae Bay mud, sand 4–12 kHz (Core) 
t17 8  Ulsan Coastal Area mud 1 MHz (Core) 
t22 9  Malta Plateau cobbles in silt 1.2 kHz (Passive) 
t23,t26 8  East China Sea sand, muddy sand 3.5 kHz (Core) 

A catalog of representative seabed classes and a set of testing seabeds were chosen from these 30 seabeds using the Pearson correlation of the characteristic TL vectors. The Pearson correlation for each pair of these 30 seabeds is shown in Fig. 1(b) ordered by a heuristic algorithm ranking for acoustic dissimilarity: Starting with a seed set of seabeds—the four used in Ref. 19—the algorithm chose a fifth seabed for which the maximum Pearson correlation to any of the first four seabeds was lowest. Each subsequent seabed was ordered in the same way with reference to all the previous seabeds. The first 13 seabeds, named r1–r13, were selected as a catalog of representative seabed classes because the maximum Pearson correlation between them was no greater than among the seed set. These 13 seabeds were used to generate synthetic data for training the CNN and should span the space of acoustic variety of propagation effects from the 30 seabeds. The remaining 17 seabeds, named t14–t30, formed the testing set. A close-up of the correlations between the testing and representative seabeds is shown in Fig. 2(a) with a maximum value of 0.85.

Fig. 2.

(a) Pearson correlation between each of the 17 testing seabeds and the 13 representative acoustic classes according to their characteristic TL vectors, with the greatest correlation for each testing seabed marked with an X. (b) Linking matrix indicating how often the CNN classifier linked the 17 testing seabeds to each of the 13 representative seabeds (averaged over 92° data samples and five training instances). The X's carry over from (a).

Fig. 2.

(a) Pearson correlation between each of the 17 testing seabeds and the 13 representative acoustic classes according to their characteristic TL vectors, with the greatest correlation for each testing seabed marked with an X. (b) Linking matrix indicating how often the CNN classifier linked the 17 testing seabeds to each of the 13 representative seabeds (averaged over 92° data samples and five training instances). The X's carry over from (a).

Close modal

A great quantity of acoustic data from each seabed is required to train and test a CNN classifier such as ours. Since collecting that quantity of labeled acoustic data in situ is expensive, this proof-of-concept research uses ORCA23 to simulate the broadband frequency response of the ocean for each seabed—in combination with each of 23 sound speed profiles (SSPs), similar to those measured in the SW06 experiment,24,29 and a variety of source-receiver ranges. The physical configuration is an ocean depth of 80 m, a receiver depth of 0.5 m from the ocean floor, and a source depth of 18.3 m. The resulting Green's function is multiplied by the source spectral model, and an inverse fast Fourier transform is performed to obtain the synthetic time series. The source model selected for this study is for SUS charge explosions30,31 with a sampling frequency of 1000 Hz. For each seabed-SSP combination, 40 source-receiver ranges (2.00–11.75 km in increments of 0.25 km) are selected, and the resulting pressure time-series form the training and testing datasets. In both the training and testing datasets, the waveform peaks are aligned within a 1-s sample. There are 920 data samples for each seabed, with a total of 11 960 samples in the training dataset.

Before training or testing the CNN, the resulting 1-s synthetic time series were individually normalized, retaining no information on amplitude or absolute travel time (see Ref. 19). This works for seabed classification as the signature of the seabed is encoded in the interference pattern within the signal.

Convolutional neural networks (CNNs) have become popular in computer vision for their strength in pattern recognition for 2D images. For the single-dimensional scenario of pressure time series, a 1D CNN with a regression output layer was used in Ref. 19 to estimate source-receiver range and distinguish between four seabed types. The current CNN began with that same 1D implementation; however, the current network is designed to only classify the seabed and has an output neuron for each of the 13 representative seabed classes. We also increased the training epochs to 1000, switched to a cross-entropy loss function, and used a random search to tune the hyperparameters, in particular, increasing the kernel sizes and the number of channels. (This hyperparameter tuning was partly dependent on the training dataset but not on the testing dataset.) To validate that the CNN was learning to decode the seabed fingerprint from the pressure time series, fivefold cross-validation was performed on the training samples, obtaining an average 99.96% classification accuracy between the 13 representative seabed classes.

This paper addresses whether a CNN classifier can link an unknown seabed (one not used in training) to an acoustically similar seabed, as determined by the Pearson correlation of the characteristic TL vectors described in Sec. 2.1 and shown in Fig. 1(a).

Five randomly initialized and independently trained instances of the CNN were applied to the 920 waveform samples for each testing seabed. The results are shown as a linking matrix in Fig. 2(b). For each of the 17 testing seabeds (vertical axis), the color indicates the fraction of the times its 920 samples were classified as each of the 13 representative seabeds (horizontal axis), averaged over the five CNN instances. For example, the first row shows that a waveform from seabed t14 will probably be linked to either seabed r3 or r9, with roughly equal probability. Thus, each row may be interpreted as giving the probability a testing seabed is linked to each of the representative seabeds.

The quality of these testing results is evaluated by comparison with the Pearson correlation between seabeds, which gives insight into whether the CNN is decoding the broadband acoustic signature of the seabed. Comparing the correlation in Fig. 2(a) with the CNN output in Fig. 2(b) shows that many of the testing seabeds are linked to the most highly correlated representative seabed. For example, the last row of the linking matrix (t30) shows ideal agreement: The classifier almost always links t30 to representative seabed r10 in Fig. 2(b), which has the highest correlation in Fig. 2(a). All such highest-correlated seabed pairs are marked with an X in both figures. These X's show that 11 out of the 17 testing seabeds are most frequently linked to the most acoustically similar seabed. This validates the potential that a CNN classifier can link together seabeds that are acoustically similar.

Considering the six cases where the CNN did not link a testing seabed to the most highly correlated representative seabed class, these may indicate shortcomings of our metric of acoustic similarity for validating a classifier. For example, our metric indicates sub-ideal behavior for the CNN classifier on seabed t22 in the middle row of Fig. 2(b): While r9 is marked with an X as the most similar representative seabed class, the classifier instead chooses r10. The gestalt similarity of the characteristic TL vectors, however, tells a different story. A comparison of the characteristic TL vectors (broadband TL vs range) in Fig. 3(a) shows that the classifier has in fact made a reasonable choice for t22, since the TL vector of t22 looks more similar to r10 (lower) than to r9 (upper). This shortcoming of the Pearson correlation occurs because it measures proportionality but ignores absolute values. Alternative measures of acoustic similarity could potentially be constructed using other statistical correlations. Moreover, the characteristic TL vectors of Sec. 2.1 themselves could be replaced with an alternative physical measure, such as bottom loss, in order to avoid the dependence of TL on environmental configuration.

Fig. 3.

(a) Examples of the characteristic TL vector for seabed t22 combined with r9 (upper) and r10 (lower). (b) Cumulative distribution functions (CDFs) of seabed predictions relative to the representative seabeds used in training. The thick black line shows the median experimental results for how the CNN linked testing seabeds to representative seabeds, reordered based on correlation. The thick silver lines show the 10th and 90th percentile results. The dashed lines show the median results for testing seabeds with correlation to the representative seabeds greater than (blue) and less than (red) 0.75. Theoretically random results are plotted as a thin gray diagonal line.

Fig. 3.

(a) Examples of the characteristic TL vector for seabed t22 combined with r9 (upper) and r10 (lower). (b) Cumulative distribution functions (CDFs) of seabed predictions relative to the representative seabeds used in training. The thick black line shows the median experimental results for how the CNN linked testing seabeds to representative seabeds, reordered based on correlation. The thick silver lines show the 10th and 90th percentile results. The dashed lines show the median results for testing seabeds with correlation to the representative seabeds greater than (blue) and less than (red) 0.75. Theoretically random results are plotted as a thin gray diagonal line.

Close modal

While the linking matrix [Fig. 2(b)] shows the performance for each testing seabed in detail, a higher-level way to examine the classification results uses cumulative distribution functions (CDFs). For each testing seabed, the representative seabeds are reordered by the correlation values such that the CDFs reveal the probability that the CNN results agree with the correlations.

If the CNN classifier links a testing seabed to its most highly correlated representative seabed, it would have 100% top-1 accuracy. If that goal is not attained, the top-2 accuracy becomes interesting, that is, the accuracy for linking to either of the two most highly correlated representative seabeds. A CDF curve can plot this information effectively, starting at top-1 accuracy and climbing through top-3 accuracy all the way to top-13 accuracy (always 100%). The ideal CDF curve would have 100% top-1 accuracy, producing a horizontal line on the top of the graph. On the other hand, a random classifier would choose seabeds arbitrarily with no connection to acoustic similarity, leading to a diagonal line for the CDF with each of the 13 seabeds chosen as often as each other, i.e., the top-1 accuracy would be one thirteenth of 100%. This random classifier result—the null hypothesis—is plotted as a thin black line in Fig. 3(b) for comparison.

We consider the CDFs of each testing seabed independently and present the overall trends in Fig. 3(b). The results of the CNN classifier are significantly better than the null hypothesis. The median CDF result (thick black line) shows the overall accuracy metric for the 17 testing seabeds. The median CDF shows a top-1 accuracy of 66%. The top-3 accuracy jumps to 96% for linking to one of the top-3 most similar representative seabed types. This high accuracy shows that our CNN classifier is generally learning the broadband acoustic effect of a seabed in the water column and is capable of linking new seabeds to similar representative seabed classes.

These differences in correlation strength lead to substantial variation in the results. The 10th and 90th percentile results are shown as silver lines. Testing seabeds that have a high correlation with at least one of the representative seabed classes tend to show higher accuracy. Of the nine testing seabeds with a correlation of at least 0.75 to one of the representative seabeds, the median CDF (blue dashed line) shows better performance than the overall median, with a top-2 accuracy of 96%. On the other hand, the CNN classifier did not perform as well on the eight testing seabeds that had less than 0.75 correlation to any of the representative seabeds (red dashed line). Nevertheless, the median CDF shows a top-6 accuracy of 97%.

The overall results show that the CNN classifier can link acoustically similar seabeds together. The median top-3 accuracy of 96% shows that the CNN classifier is making physically grounded predictions as validated by our correlation metric.

This paper has shown that a deep learning classifier can learn the acoustic fingerprint of a set of representative seabeds and then link new seabeds to one of these representative seabeds based upon a similar acoustic signature. Our CNN classifier has attained a median 96% accuracy for linking a testing seabed to one of the top three most similar seabed classes in the representative set, using synthetic pressure time series. The most similar seabed classes are identified using the Pearson correlation of characteristic TL vectors for each seabed. We have shown how an example of lower accuracy can stem, not from the classifier itself, but rather from limitations of using this correlation metric to quantify similarity. Since the TL approach depends on source-receiver configuration, future work could use a more stable metric based on bottom loss to determine correlation. In summary, this paper has presented a method for choosing distinct seabed classes and also shown that a CNN classifier can link new seabeds to a similar seabed in a representative set. This work provides evidence that deep learning can characterize seabeds by acoustic type based on broadband effects of the sediment on acoustic propagation.

This research was supported in part by the National Science Foundation, REU Grant No. 1757998. The authors thank Mohsen Badiey for assistance with realistic sound speed profiles.

1.
G. R.
Potty
,
J. H.
Miller
, and
J. F.
Lynch
, “
Inversion for sediment geoacoustic properties at the New England Bight
,”
J. Acoust. Soc. Am.
114
(
4
),
1874
1887
(
2003
).
2.
D.
Knobles
,
R. A.
Koch
,
L.
Thompson
,
K.
Focke
, and
P.
Eisman
, “
Broadband sound propagation in shallow water and geoacoustic inversion
,”
J. Acoust. Soc. Am.
113
(
1
),
205
222
(
2003
).
3.
D.
Knobles
and
R. A.
Koch
, “
A time series analysis of sound propagation in a strongly multipath shallow water environment with an adiabatic normal mode approach
,”
IEEE J. Ocean. Eng.
21
(
1
),
1
13
(
1996
).
4.
Y.-M.
Jiang
,
N. R.
Chapman
, and
M.
Badiey
, “
Quantifying the uncertainty of geoacoustic parameter estimates for the New Jersey shelf by inverting air gun data
,”
J. Acoust. Soc. Am.
121
(
4
),
1879
1894
(
2007
).
5.
W.-H.
Ryang
,
S.-P.
Kim
,
D.-C.
Kim
, and
J.
Hahn
, “
Geoacoustic model of coastal bottom strata at Jeongdongjin in the Korean continental margin of the East Sea
,”
J. Korean Earth Sci. Soc.
37
,
200
210
(
2016
).
6.
N. R.
Chapman
,
S.
Chin-Bing
,
D.
King
, and
R. B.
Evans
, “
Benchmarking geoacoustic inversion methods for range-dependent waveguides
,”
IEEE J. Ocean. Eng.
28
(
3
),
320
330
(
2003
).
7.
H.
Kwon
,
J. W.
Choi
,
W.-H.
Ryang
,
S.-U.
Son
, and
S-k
Jung
, “
Measurements of mid-frequency bottom-interacting signals and geoacoustic inversion in Jinhae Bay, Southeast Korea
,”
J. Acoust. Soc. Am.
145
(
3
),
1205
1211
(
2019
).
8.
W.-H.
Ryang
,
S.-P.
Kim
,
S.
Kim
, and
D.-C.
Kim
, “
Geoacoustic model of the transverse acoustic variability experiment area in the northern East China Sea
,”
Geosci. J.
17
(
3
),
267
278
(
2013
).
9.
S. E.
Dosso
,
P. M.
Giles
,
G. H.
Brooke
,
D. F.
McCammon
,
S.
Pecknold
, and
P. C.
Hines
, “
Linear and nonlinear measures of ocean acoustic environmental sensitivity
,”
J. Acoust. Soc. Am.
121
(
1
),
42
45
(
2007
).
10.
S. E.
Dosso
and
M. J.
Wilmut
, “
Bayesian focalization: Quantifying source localization with environmental uncertainty
,”
J. Acoust. Soc. Am.
121
(
5
),
2567
2574
(
2007
).
11.
C. W.
Holland
and
S. E.
Dosso
, “
Mid frequency shallow water fine-grained sediment attenuation measurements
,”
J. Acoust. Soc. Am.
134
(
1
),
131
143
(
2013
).
12.
J.
Dettmer
,
S. E.
Dosso
, and
C. W.
Holland
, “
Trans-dimensional geoacoustic inversion
,”
J. Acoust. Soc. Am.
128
(
6
),
3393
3405
(
2010
).
13.
J.
Benson
,
N. R.
Chapman
, and
A.
Antoniou
, “
Geoacoustic model inversion using artificial neural networks
,”
Inv. Probl.
16
(
6
),
1627
(
2000
).
14.
J.
Piccolo
,
G.
Haramuniz
, and
Z.-H.
Michalopoulou
, “
Geoacoustic inversion with generalized additive models
,”
J. Acoust. Soc. Am.
145
(
6
),
EL463
EL468
(
2019
).
15.
M. D.
Collins
and
L.
Fishman
, “
Efficient navigation of parameter landscapes
,”
J. Acoust. Soc. Am.
98
(
3
),
1637
1644
(
1995
).
16.
T. B.
Neilsen
, “
An iterative implementation of rotated coordinates for inverse problems
,”
J. Acoust. Soc. Am.
113
(
5
),
2574
2586
(
2003
).
17.
Z.-H.
Michalopoulou
,
D.
Alexandrou
, and
C.
De Moustier
, “
Application of neural and statistical classifiers to the problem of seafloor characterization
,”
IEEE J. Ocean. Eng.
20
(
3
),
190
197
(
1995
).
18.
D.
Alexandrou
and
D.
Pantzartzis
, “
A methodology for acoustic seafloor classification
,”
IEEE J. Ocean. Eng.
18
(
2
),
81
86
(
1993
).
19.
D. F.
Van Komen
,
T. B.
Neilsen
,
K.
Howarth
,
D. P.
Knobles
, and
P. H.
Dahl
, “
Seabed and range estimation of impulsive time series using a convolutional neural network
,”
J. Acoust. Soc. Am.
147
(
5
),
EL403
EL408
(
2020
).
20.
T. B.
Neilsen
,
C. D.
Escobar-Amado
,
M. C.
Acree
,
W. S.
Hodgkiss
,
D. F.
Van Komen
,
D. P.
Knobles
,
M.
Badiey
, and
J.
Castro-Correa
, “
Learning location and seabed type from a moving mid-frequency source
,”
J. Acoust. Soc. Am.
149
,
692
705
(
2021
).
21.
D. F.
Van Komen
,
T. B.
Neilsen
,
D. B.
Mortenson
,
M. C.
Acree
,
D. P.
Knobles
,
M.
Badiey
, and
W. S.
Hodgkiss
, “
Seabed type and source parameters predictions using ship spectrograms in convolutional neural networks
,”
J. Acoust. Soc. Am.
149
,
1198
(
2021
).
22.
C.
Frederick
,
S.
Villar
, and
Z.-H.
Michalopoulou
, “
Seabed classification using physics-based modeling and machine learning
,”
J. Acoust. Soc. Am.
148
(
2
),
859
872
(
2020
).
23.
E. K.
Westwood
,
C. T.
Tindle
, and
N. R.
Chapman
, “
A normal mode model for acousto-elastic ocean environments
,”
J. Acoust. Soc. Am.
100
(
6
),
3631
3645
(
1996
).
24.
M.
Badiey
,
L.
Wan
, and
J. F.
Lynch
, “
Statistics of nonlinear internal waves during the shallow water 2006 experiment
,”
J. Atmos. Ocean. Technol.
33
(
4
),
839
846
(
2016
).
25.
D. P.
Knobles
,
P. S.
Wilson
,
J. A.
Goff
,
L.
Wan
,
M. J.
Buckingham
,
J. D.
Chaytor
, and
M.
Badiey
, “
Maximum entropy derived statistics of sound-speed structure in a fine-grained sediment inferred from sparse broadband acoustic measurements on the New England continental shelf
,”
IEEE J. Ocean. Eng.
45
(
1
),
161
173
(
2020
).
26.
J.-X.
Zhou
,
X.-Z.
Zhang
, and
D.
Knobles
, “
Low-frequency geoacoustic model for the effective properties of sandy seabottoms
,”
J. Acoust. Soc. Am.
125
(
5
),
2847
2866
(
2009
).
27.
W.-H.
Ryang
,
J.-H.
Jin
, and
J.
Hahn
, “
Geoacoustic model at the YSDP-105 Long-core Site in the mid-eastern Yellow Sea
,”
J. Korean Earth Sci. Soc.
40
(
1
),
24
36
(
2019
).
28.
D. P.
Knobles
,
J. A.
Goff
,
R. A.
Koch
,
P. S.
Wilson
, and
J. A.
Shooter
, “
Effect of inhomogeneous sub-bottom layering on broadband acoustic propagation
,”
IEEE J. Ocean. Eng.
35
(
4
),
732
743
(
2010
).
29.
D.
Tang
,
J. N.
Moum
,
J. F.
Lynch
,
P.
Abbot
,
R.
Chapman
,
P. H.
Dahl
,
T. F.
Duda
,
G.
Gawarkiewicz
,
S.
Glenn
,
J. A.
Goff
,
H.
Graber
,
J.
Kemp
,
A.
Maffei
,
J. D.
Nash
, and
A.
Newhall
, “
Shallow Water'06: A joint acoustic propagation/nonlinear internal wave physics experiment
,”
Oceanography
20
(
4
),
156
167
(
2007
).
30.
N. R.
Chapman
, “
Source levels of shallow explosive charges
,”
J. Acoust. Soc. Am.
84
(
2
),
697
702
(
1988
).
31.
P. S.
Wilson
,
D. P.
Knobles
,
P. H.
Dahl
,
A. R.
McNeese
, and
M. C.
Zeh
, “
Short-range signatures of explosive sounds in shallow water used for seabed characterization
,”
IEEE J. Ocean. Eng.
45
(
1
),
14
25
(
2020
).