Classification of dispersive gunshot calls using a convolutional neural network

Goldwater, Mark; Bonnel, Julien; Cammareri, Alejandro; Wright, Dana; Zitterbart, Daniel P.

doi:10.1121/10.0006718

Cited by 4 publications

(1 citation statement)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A common approach is to first transform acoustic data into the TF domain using the STFT to create a spectrogram representation of the signal. Then a large collection of such labeled synthetic or experimental data is used to train a convolutional neural network (CNN) to distinguish between signals of interest, a method called supervised learning [12], [96], [97]. Preparing the data using an STFT is considered a form of preprocessing or manual feature extraction.…”

Section: Classificationmentioning

confidence: 99%

Automatic baleen whale detection and 2D localization using a network of unsynchonized passive acoustic sensors

Goldwater

View full text Add to dashboard Cite

Underwater acoustics is a powerful tool for learning about the ocean’s soniferous marine life. However, most modern acoustic sensing systems consist of expensive arrays of timesynchronized recorders which require a crewed research vessel and significant expertise to deploy, operate, and recover. Recently, there has been a growing corpus of research related to algorithms for low-cost and accessible acoustic hardware. Deep learning methods have shown great promise when applied to underwater acoustics inverse problems. While many signal processing or physics-based algorithms exhibit long run times and require manual labor to extract signals of interest, tune parameters, as well as visually verify the results, an appropriately trained neural network can quickly process data with no human supervision. Both low-cost passive acoustic monitoring (PAM) sensing platforms and algorithms that can analyze massive amounts of raw data are critical to accessible and scalable approaches in ocean acoustic monitoring. This thesis presents a method for detection and 2D (latitude-longitude) localization of underwater acoustic sources without requiring synchronized sensors. The signals of interest here are the dispersive low-frequency impulsive gunshot vocalizations of North Pacific and North Atlantic right whales (NPRWs, NARWs). In shallow-water channels, the timefrequency representation of the received signal is strongly dependent on source-receiver range, making these impulses ideal candidates for range-based localization. The first step in the localization pipeline uses a temporal convolutional network (TCN) to simultaneously detect gunshot vocalizations and predict their ranges. Trained on spectrograms of synthetic data simulated in a variety of environments, the TCN is applied to PAM data from moorings in the Bering Sea. Gunshots are detected with high precision, and the range estimates are comparable to those estimated using traditional physics-based processing. Both methods use a minimal set of a priori environmental information including water column depth, sound speed, and density. Depending on the sensor layout, the TCN may need to scan large windows of data, so the number of unique acoustic sources is unknown. To automatically associate and localize range measurements, the proposed method seeks subsets of measurements across unique sensors which are internally consistent. For every considered measurement subset, locations are estimated with single constituent measurements left out and checked to be sufficiently close to the excluded measurement’s set of potential locations. If a measurement subset is entirely consistent in this manner, the measurements are added as neighboring nodes in a graph-based representation, and strongly connected components are used to determine data associations and calculate the final source location estimates. Informed by the methods developed in this thesis, an array of low-cost TOSSIT moorings was deployed in Cape Cod Bay and used to collect experimental PAM data. The localization results are comparable to another similar physics-based inversion approach. Overall, this thesis aims to fill a gap in acoustic data processing methods where data from a low-cost network of unsynchronized acoustic sensors are fused to localize acoustic sources. The presented methods and data processing pipeline demonstrate the great potential of low-cost acoustic sensing systems.

show abstract

Section: Classificationmentioning

confidence: 99%

Automatic baleen whale detection and 2D localization using a network of unsynchonized passive acoustic sensors

Goldwater

View full text Add to dashboard Cite

show abstract

Machine-learning-based simultaneous detection and ranging of impulsive baleen whale vocalizations using a single hydrophone

Goldwater

Zitterbart

Wright

et al. 2023

The Journal of the Acoustical Society of America

View full text Add to dashboard Cite

The low-frequency impulsive gunshot vocalizations of baleen whales exhibit dispersive propagation in shallow-water channels which is well-modeled by normal mode theory. Typically, underwater acoustic source range estimation requires multiple time-synchronized hydrophone arrays which can be difficult and expensive to achieve. However, single-hydrophone modal dispersion has been used to range baleen whale vocalizations and estimate shallow-water geoacoustic properties. Although convenient when compared to sensor arrays, these algorithms require preliminary signal detection and human labor to estimate the modal dispersion. In this paper, we apply a temporal convolutional network (TCN) to spectrograms from single-hydrophone acoustic data for simultaneous gunshot detection and ranging. The TCN learns ranging and detection jointly using gunshots simulated across multiple environments and ranges along with experimental noise. The synthetic data are informed by only the water column depth, sound speed, and density of the experimental environment, while other parameters span empirically observed bounds. The method is experimentally verified on North Pacific right whale gunshot data collected in the Bering Sea. To do so, 50 dispersive gunshots were manually ranged using the state-of-the-art time-warping inversion method. The TCN detected these gunshots among 50 noise-only examples with high precision and estimated ranges which closely matched those of the physics-based approach.

show abstract

Approximation of modal wavenumbers and group speeds in an oceanic waveguide using a neural network

Varón

Mars

Bonnel

2023

JASA Express Letters

View full text Add to dashboard Cite

Underwater acoustic propagation is influenced not only by the property of the water column, but also by the seabed property. Modeling this propagation using normal mode simulation can be computationally intensive, especially for wideband signals. To address this challenge, a Deep Neural Network is used to predict modal horizontal wavenumbers and group velocities. Predicted wavenumbers are then used to compute modal depth functions and transmission losses, reducing computational cost without significant loss in accuracy. This is illustrated on a simulated Shallow Water 2006 inversion scenario.

show abstract

Classification of dispersive gunshot calls using a convolutional neural network

Cited by 4 publications

References 13 publications

Automatic baleen whale detection and 2D localization using a network of unsynchonized passive acoustic sensors

Automatic baleen whale detection and 2D localization using a network of unsynchonized passive acoustic sensors

Machine-learning-based simultaneous detection and ranging of impulsive baleen whale vocalizations using a single hydrophone

Approximation of modal wavenumbers and group speeds in an oceanic waveguide using a neural network

Contact Info

Product

Resources

About