Odontocete Echolocation Clicks Research Articles

Denoising odontocete echolocation clicks using a hybrid model with convolutional neural network and long short-term memory network.

Ocean noise negatively influences the recording of odontocete echolocation clicks. In this study, a hybrid model based on the convolutional neural network (CNN) and long short-term memory (LSTM) network-called a hybrid CNN-LSTM model-was proposed to denoise echolocation clicks. To learn the model parameters, the echolocation clicks were partially corrupted by adding ocean noise, and the model was trained to recover the original echolocation clicks. It can be difficult to collect large numbers of echolocation clicks free of ambient sea noise for training networks. Data augmentation and transfer learning were employed to address this problem. Based on Gabor functions, simulated echolocation clicks were generated to pre-train the network models, and the parameters of the networks were then fine-tuned using odontocete echolocation clicks. Finally, the performance of the proposed model was evaluated using synthetic data. The experimental results demonstrated the effectiveness of the proposed model for denoising two typical echolocation clicks-namely, narrowband high-frequency and broadband echolocation clicks. The denoising performance of hybrid models with the different number of convolution and LSTM layers was evaluated. Consequently, hybrid models with one convolutional layer and multiple LSTM layers are recommended, which can be adopted for denoising both types of echolocation clicks.

Monitoring cetacean occurrence and variability in ambient sound in Scottish offshore waters

The characterisation of marine soundscapes allows observation of spatiotemporal distribution of vocalising species and human activities, which can inform an assessment of their interactions. Such data are important for monitoring the ecological status of marine habitats. The Scottish Atlantic Frontier is an important habitat for a variety of cetacean species. Historic whaling has heavily impacted several species inhabiting these waters and current comprehensive information about seasonal occurrence and distribution is lacking for all species. This study presents year-round passive acoustic monitoring data from ten sites in this understudied region. The three most offshore sites were examined for baleen whale vocalisations, and four species were regularly detected. Fin whale detections peaked from October to January and were at their lowest during May and June. Humpback whale song was detected as early as January but showed a strong seasonal peak in March and April. In contrast, minke whales were detected regularly throughout the year but with a peak in detections from October to November, when sei whales were also detected. All monitoring sites showed frequent occurrence of odontocete echolocation clicks and whistles. Comprised mainly of delphinid vocalisations, whistles and clicks were detected on an almost-daily basis among the offshore sites, with a slight reduction in detections from May to July particularly among the more inshore sites. Ambient sound levels (root-mean-square sounds pressure level; SPL) varied by site and season in relation to species presence, anthropogenic contributions, and environmental conditions. Monthly median SPL across the array varied up to 18 dB within 1/3-octave bands. Throughout the year, variability in median SPL was lowest in the higher frequency bands (>10 kHz), while highest variability was found between January to July in specific lower frequency bands (<1 kHz). Results from this study demonstrate the value of passive acoustic data in providing novel baseline information about cetacean occurrence and distribution in Scottish offshore habitats where data are limited and outdated. The results will feed into statutory reporting on underwater noise, support the identification and designation of future marine protected areas for cetaceans, and help guide management of future human-marine mammal interactions in Scottish offshore waters.

Discriminating and classifying odontocete echolocation clicks in the Hawaiian Islands using machine learning methods.

Passive acoustic monitoring (PAM) has proven a powerful tool for the study of marine mammals, allowing for documentation of biologically relevant factors such as movement patterns or animal behaviors while remaining largely non-invasive and cost effective. From 2008–2019, a set of PAM recordings covering the frequency band of most toothed whale (odontocete) echolocation clicks were collected at sites off the islands of Hawaiʻi, Kauaʻi, and Pearl and Hermes Reef. However, due to the size of this dataset and the complexity of species-level acoustic classification, multi-year, multi-species analyses had not yet been completed. This study shows how a machine learning toolkit can effectively mitigate this problem by detecting and classifying echolocation clicks using a combination of unsupervised clustering methods and human-mediated analyses. Using these methods, it was possible to distill ten unique echolocation click ‘types’ attributable to regional odontocetes at the genus or species level. In one case, auxiliary sightings and recordings were used to attribute a new click type to the rough-toothed dolphin, Steno bredanensis. Types defined by clustering were then used as input classes in a neural-network based classifier, which was trained, tested, and evaluated on 5-minute binned data segments. Network precision was variable, with lower precision occurring most notably for false killer whales, Pseudorca crassidens, across all sites (35–76%). However, accuracy and recall were high (>96% and >75%, respectively) in all cases except for one type of short-finned pilot whale, Globicephala macrorhynchus, call class at Kauaʻi and Pearl and Hermes Reef (recall >66%). These results emphasize the utility of machine learning in analysis of large PAM datasets. The classifier and timeseries developed here will facilitate further analyses of spatiotemporal patterns of included toothed whales. Broader application of these methods may improve the efficiency of global multi-species PAM data processing for echolocation clicks, which is needed as these datasets continue to grow.

Identification of western North Atlantic odontocete echolocation click types using machine learning and spatiotemporal correlates.

A combination of machine learning and expert analyst review was used to detect odontocete echolocation clicks, identify dominant click types, and classify clicks in 32 years of acoustic data collected at 11 autonomous monitoring sites in the western North Atlantic between 2016 and 2019. Previously-described click types for eight known odontocete species or genera were identified in this data set: Blainville’s beaked whales (Mesoplodon densirostris), Cuvier’s beaked whales (Ziphius cavirostris), Gervais’ beaked whales (Mesoplodon europaeus), Sowerby’s beaked whales (Mesoplodon bidens), and True’s beaked whales (Mesoplodon mirus), Kogia spp., Risso’s dolphin (Grampus griseus), and sperm whales (Physeter macrocephalus). Six novel delphinid echolocation click types were identified and named according to their median peak frequencies. Consideration of the spatiotemporal distribution of these unidentified click types, and comparison to historical sighting data, enabled assignment of the probable species identity to three of the six types, and group identity to a fourth type. UD36, UD26, and UD28 were attributed to Risso’s dolphin (G. griseus), short-finned pilot whale (G. macrorhynchus), and short-beaked common dolphin (D. delphis), respectively, based on similar regional distributions and seasonal presence patterns. UD19 was attributed to one or more species in the subfamily Globicephalinae based on spectral content and signal timing. UD47 and UD38 represent distinct types for which no clear spatiotemporal match was apparent. This approach leveraged the power of big acoustic and big visual data to add to the catalog of known species-specific acoustic signals and yield new inferences about odontocete spatiotemporal distribution patterns. The tools and call types described here can be used for efficient analysis of other existing and future passive acoustic data sets from this region.

Cetaceans and seismic surveys in the Southern Adriatic Sea

The first long-term, broadband, passive acoustic monitoring effort was conducted in the south Adriatic Sea from October 2018 to December 2019 using a High-frequency Acoustic Recording Package (HARP). The recording effort coincided with seismic surveys that took place off Montenegro in the fall and winter 2018–2019. These surveys occurred at distances 70 to 125 km from the recording site, at water depths between 70 and 800 m. Recordings were manually reviewed with start and end times of air-gun encounters and ship passages marked. A generic energy detector and unsupervised clustering procedure were used to extract odontocete echolocation clicks. During the surveys, sound pressure levels (SPL) in the 15 to 200 Hz band increased by up to 20 dB during fall and up to 10 dB during winter survey, with an average increase of 7.5 dB. Cuvier's beaked whale echolocation clicks were very commonly detected in the recordings, but sperm whale clicks were only sporadically present. At least three additional echolocation click types were also detected and their attribution to particular species is underway. Although mortality related to seismic surveys has not been recorded, high levels of noise have the potential to cause long term disruption to cetaceans in this region.

Classification of odontocete echolocation clicks using convolutional neural network

A method based on a convolutional neural network for the automatic classification of odontocete echolocation clicks is presented. The proposed convolutional neural network comprises six layers: three one-dimensional convolutional layers, two fully connected layers, and a softmax classification layer. Rectified linear units were chosen as the activation function for each convolutional layer. The input to the first convolutional layer is the raw time signal of an echolocation click. Species prediction was performed for groups of m clicks, and two strategies for species label prediction were explored: the majority vote and maximum posterior. Two datasets were used to evaluate the classification performance of the proposed algorithm. Experiments showed that the convolutional neural network can model odontocete species from the raw time signal of echolocation clicks. With the increase in m, the classification accuracy of the proposed method improved. The proposed method can be employed in passive acoustic monitoring to classify different delphinid species and facilitate future studies on odontocetes.

Developing a click type “library” for Hawaiian odontocetes using machine learning methods

Though odontocete echolocation clicks are highly useful in passive acoustic monitoring, their acoustic features vary depending on behavior, orientation, and location relative to the receiver. Using unsupervised machine learning, regional sets of echolocation click types can be identified that allow for automated labeling of bioacoustic data while accounting for within-type signal variability. This methodology derives labels for clusters of click types based on parameters including spectral shape, inter-click interval, and click duration. Here we present classification results of echolocation clicks produced by Hawaiian odontocetes. Unsupervised methods were used to establish a set of click types for 10 years of recordings (200 kHz) collected from a bottom-mounted hydrophone off the coast of Kona, HI. These automated labels were compared with manual labels for species with established and recognizable clicks, which allowed for ground-truthing of some click types identified by the unsupervised machine learning method. Towed array data (500 kHz) was used to assign click types to species with visual verification, expanding the set of known species-specific click types for the region. The click types and method established here will be useful in future analyses of Hawaiian odontocete density estimation, abundance, and distribution patterns using acoustic data without requiring visual confirmation.

Convolutional neural network for detecting odontocete echolocation clicks.

In this work, a convolutional neural network based method is proposed to automatically detect odontocetes echolocation clicks by analyzing acoustic data recordings from a passive acoustic monitoring system. The neural network was trained to distinguish between click and non-click clips and was subsequently converted to a full-convolutional network. The performance of the proposed network was evaluated using synthetic data and real audio recordings. The experimental results indicate that the proposed method works stably with echolocation clicks of different species.

The use of context in machine learning for bioacoustics

Biological acoustic signals are often produced in context. Contextual information includes things such as timing between calls, conspecific or interspecies cues, and physical environmental cues such as sunrise or sunset. We show how some forms of contextual information can be used to improve the results of detection and classification tasks for biological acoustic signals. We examine how context can be used to improve labeled data, resulting in more-accurate classification results, as well as how learners can exploit context directly. We demonstrate these improvements via two bioacoustic detection/classification tasks. The first algorithm detects odontocete echolocation clicks. We used a decision support system that allowed analysts to label echolocation clicks using between call timing cues as well as other measurements and found that deep learners trained with these high quality data are able to detect clicks in adverse environments. The second algorithm applies contextual information surrounding North Atlantic right whale upcalls to improve precision and recall. [This work was supported by ONR Grant Nos. N00014-17-1-2867 and N00014-15-1-2299.]

Using detection received level as a guide for odontocete density estimation

Odontocete echolocation clicks have been used for density estimation using cue-counting. We present an approach for assessment of the appropriate threshold for click detection based on the click received sound-pressure-level distribution. Under the assumption of random spatial distribution of echolocating animals with respect to the sensor, the number of detections should decrease with increasing received level. This pattern is verified by simulation modeling, even after incorporating diving behavior and sound production directionality. When measured received level distributions show peaks at values above the detection threshold, errors in the detection process are the most likely explanation. Potential remedies involve setting a higher detection threshold and/or fully characterizing both missed and false positive rates.Odontocete echolocation clicks have been used for density estimation using cue-counting. We present an approach for assessment of the appropriate threshold for click detection based on the click received sound-pressure-level distribution. Under the assumption of random spatial distribution of echolocating animals with respect to the sensor, the number of detections should decrease with increasing received level. This pattern is verified by simulation modeling, even after incorporating diving behavior and sound production directionality. When measured received level distributions show peaks at values above the detection threshold, errors in the detection process are the most likely explanation. Potential remedies involve setting a higher detection threshold and/or fully characterizing both missed and false positive rates.

Categorizing click trains to increase taxonomic precision in echolocation click loggers.

Passive acoustic monitoring is an efficient way to study acoustically active animals but species identification remains a major challenge. C-PODs are popular logging devices that automatically detect odontocete echolocation clicks. However, the accompanying analysis software does not distinguish between delphinid species. Click train features logged by C-PODs were compared to frequency spectra from adjacently deployed continuous recorders. A generalized additive model was then used to categorize C-POD click trains into three groups: broadband click trains, produced by bottlenose dolphin (Tursiops truncatus) or common dolphin (Delphinus delphis), frequency-banded click trains, produced by Risso's (Grampus griseus) or white beaked dolphins (Lagenorhynchus albirostris), and unknown click trains. Incorrect categorization rates for broadband and frequency banded clicks were 0.02 (SD 0.01), but only 30% of the click trains met the categorization threshold. To increase the proportion of categorized click trains, model predictions were pooled within acoustic encounters and a likelihood ratio threshold was used to categorize encounters. This increased the proportion of the click trains meeting either the broadband or frequency banded categorization threshold to 98%. Predicted species distribution at the 30 study sites matched well to visual sighting records from the region.

Automatic detection method for monitoring odontocete echolocation clicks

Passive acoustic monitoring records large amounts of acoustic data, and thus an efficient detection method is essential to analyse these data. A novel approach is proposed to automatically detect odontocete echolocation clicks. A time–frequency filter detects echolocation clicks in the time–frequency domain, and then the Teager–Kaiser energy operator and Gabor curve-fitting method determine the start and end points of each echolocation click precisely. Detector performance is assessed by using synthetic data. The experimental results showed that the recall rates were higher than 90% when the signal-to-noise ratio was above 10 dB.

Single-sensor, cue-counting population density estimation: Average probability of detection of broadband clicks.

Odontocete echolocation clicks have been used as a preferred cue for density estimation using single-sensor data sets, requiring estimation of detection probability as a function of range. Many such clicks can be very broadband in nature, with 10-dB bandwidths of 20-40 kHz or more. Detection distances are not readily obtained from single-sensor data. Here, the average detection probability is estimated in a Monte Carlo simulation using the passive sonar equation along with transmission loss calculations to estimate the signal-to-noise ratio (SNR) of tens of thousands of click realizations. Continuous-wave (CW) analysis, i.e., single-frequency analysis, is inherent to basic forms of the passive sonar equation. Using CW analysis with the click's center frequency while disregarding its bandwidth has been shown to introduce bias into detection probabilities and hence to density estimates. In this study, the effects of highly broadband clicks on density estimates are further examined. The usage of transmission loss as an appropriate measure for calculating click SNR is also discussed. The main contributions from this research are (1) an alternative approach to estimate the average probability of detection of broadband clicks, and (2) understanding the effects of multipath clicks on population density estimates.

Single-sensor density estimation of highly broadband marine mammal calls

Odontocete echolocation clicks have been used as a preferred cue for density estimation studies from single-sensor data sets. Such sounds are broadband in nature, with 10-dB bandwidths of 20 to 40 kHz or more. Estimating their detection probability is one of the main requirements of density estimation studies. For single-sensor data, detection probability is estimated using the sonar equation to simulate received signal-to-noise ratio of thousands of click realizations. A major problem with such an approach is that the passive sonar equation is a continuous-wave (CW) analysis tool (single-frequencies). Using CW analysis with a click’s center frequency while disregarding its bandwidth has been shown to introduce bias to detection probabilities and hence to population estimates. In this study, the methodology used to estimate detection probabilities is re-evaluated, and the bias in sonar equation density estimates is quantified by using a synthetic data set. A new approach based on the calculation of arrivals and subsequent convolution with a click source function is also presented. Application of the new approach to the synthetic data set showed accurate results. Further complexities of density estimation studies are illustrated with a data set containing highly broadband false killer whale (Pseudorca crassidens) clicks. [Work supported by ONR.]

Morphology of the Nasal Apparatus in Pygmy (Kogia Breviceps) and Dwarf (K. Sima) Sperm Whales.

Odontocete echolocation clicks are generated by pneumatically driven phonic lips within the nasal passage, and propagated through specialized structures within the forehead. This study investigated the highly derived echolocation structures of the pygmy (Kogia breviceps) and dwarf (K. sima) sperm whales through careful dissections (N = 18 K. breviceps, 6 K. sima) and histological examinations (N = 5 K. breviceps). This study is the first to show that the entire kogiid sound production and transmission pathway is acted upon by complex facial muscles (likely derivations of the m. maxillonasolabialis). Muscles appear capable of tensing and separating the solitary pair of phonic lips, which would control echolocation click frequencies. The phonic lips are enveloped by the "vocal cap," a morphologically complex, connective tissue structure unique to kogiids. Extensive facial muscles appear to control the position of this structure and its spatial relationship to the phonic lips. The vocal cap's numerous air crypts suggest that it may reflect sounds. Muscles encircling the connective tissue case that surrounds the spermaceti organ may change its shape and/or internal pressure. These actions may influence the acoustic energy transmitted from the phonic lips, through this lipid body, to the melon. Facial and rostral muscles act upon the length of the melon, suggesting that the sound "beam" can be focused as it travels through the melon and into the environment. This study suggests that the kogiid echolocation system is highly tunable. Future acoustic studies are required to test these hypotheses and gain further insight into the kogiid echolocation system.

Single-sensor, cue-counting density estimation of highly broadband marine mammal calls

Odontocete echolocation clicks have been used as a preferred cue for density estimation studies from single-sensor data sets, studies that require estimating detection probability as a function of range. Many such clicks can be very broadband in nature, with 10-dB bandwidths of 20 to 40 kHz or more. Because detection distances are not realizable from single-sensor data, the detection probability is estimated in a Monte Carlo simulation using the sonar equation along with transmission loss calculations to estimate the received signal-to-noise ratio of tens of thousands of click realizations. Continuous-wave (CW) analysis, that is, single-frequency analysis, is inherent to basic forms of the passive sonar equation. Considering transmission loss by using CW analysis with the click’s center frequency while disregarding its bandwidth has recently been shown to introduce bias to detection probabilities and hence to population estimates. In this study, false killer whale (Pseudorca crassidens) clicks recorded off the Kona coast of Hawai’i are used to quantify the bias in sonar equation density estimates caused by the center-frequency approach. A different approach to analyze data sets with highly broadband calls and to correctly model such signals is also presented and evaluated. [Work supported by ONR.]

The energy ratio mapping algorithm: A tool to improve the energy-based detection of odontocete echolocation clicks

The energy ratio mapping algorithm (ERMA) was developed to improve the performance of energy-based detection of odontocete echolocation clicks, especially for application in environments with limited computational power and energy such as acoustic gliders. ERMA systematically evaluates many frequency bands for energy ratio-based detection of echolocation clicks produced by a target species in the presence of the species mix in a given geographic area. To evaluate the performance of ERMA, a Teager-Kaiser energy operator was applied to the series of energy ratios as derived by ERMA. A noise-adaptive threshold was then applied to the Teager-Kaiser function to identify clicks in data sets. The method was tested for detecting clicks of Blainville's beaked whales while rejecting echolocation clicks of Risso's dolphins and pilot whales. Results showed that the ERMA-based detector correctly identified 81.6% of the beaked whale clicks in an extended evaluation data set. Average false-positive detection rate was 6.3% (3.4% for Risso's dolphins and 2.9% for pilot whales).

Passive‐acoustic monitoring of odontocetes using a Seaglider: First results of a field test in Hawaiian waters.

In fall 2009 the University of Washington, Applied Physics Laboratory conducted in collaboration with the Oregon State University, a comprehensive field test of a passive‐acoustic Seaglider along the western shelf‐break of the island of Hawaii. During the 3 week mission, a total of approximately 170 h of broadband acoustic data [194 kHz sampling rate] were collected. The recordings were manually analyzed by an experienced analyst for beaked whale (Ziphiidae), dolphin (Delphinidae), and sperm whale (Physeter macrocephalus) echolocation clicks as well as echo sounder pings emitted by boats in the area. Here we present and discuss first results of these data analysis, which revealed that more than 50% of the recorded files (each of 1‐minute duration) contain bioacoustic signals. Furthermore the recorded data and the results of the manual analysis are used to validate and optimize an automated classifier for odontocete echolocation clicks, which was developed in a collaborative effort with San Diego State University. The algorithm is intended to be implemented on the Seaglider to enable species identification by classifying detected echolocation clicks in (near) real‐time during sea trials. [This work is funded by the Office of Naval Research and the U.S. Navy’s Environmental Readiness Division.]

A real-time detection system for odontocete echolocation clicks in the low-energy processing environment of an acoustic glider.

This presentation will introduce a real-time system, the energy ratio mapping algorithm (ERMA), for detection of odontocete echolocation clicks. The system has been developed to run continuously over extended periods of time in low-energy processing environments such as acoustic gliders. The detector operates in the time domain and takes advantage of the species-specific differences in the power spectra and interclick interval of echolocation clicks to minimize the number of false positive detections. We will demonstrate the operation of ERMA using the clicks of beaked whales, family Ziphiidae, as example target species. To get information on the performance of ERMA, a comparison with more sophisticated and already established detection systems was conducted. Preliminary results indicate that the developed method is a promising tool to reliably detect echolocation clicks emitted by beaked whales in real time. On-going work is focused on implementation of the detector in the acoustic seaglider developed by the Applied Physics Laboratory of the University of Washington (APL-UW) and the use of this platform for autonomous passive acoustic monitoring and mitigation of beaked whales and other odontocete species.

Comparison of feature extraction methods for the identification of odontocete species based upon echolocation clicks

Recent work by several groups has shown that odontocete echolocation clicks contain information that can be used to detect or identify specific species. In this study, we compare the relative performance of cepstral and wavelet features on various Pacific Ocean species of odontocetes. Comparison of features within different systems is often complicated by the large number of variables unrelated to feature extraction that change between systems. By experimenting within a proven state of the art classification framework, it is possible to make meaningful comparisons of feature extraction performance with respect to common machine learning algorithms such as neural networks, support vector machines, and Gaussian mixture models.