32 MARCH 2022 • WORLD AQUACULTURE • WWW.WA S .ORG learning to detect acoustic signatures of marine mammals from hydrophone recordings or in real time from connected hydrophone systems. The acoustic detection model uses a multi-step process to characterize ocean sounds and determine if a recording includes a vocalization from a species of interest. First, a sample of environmental background noise is collected and used to subtract consistent features of the audio signal from the recording. Next, two sets of features are generated from the audio signal to form the inputs of the classification algorithm. The first consists of a set of ten features based on the auditory characteristics of the recording. The spectrogram of the recording is then translated into an image from which additional features are generated. The resulting feature set is then passed through a neural network, a style of machine learning algorithm inspired by the wiring of neurons in the human brain used to perform a wide range of artificial intelligence tasks. To fully classify a given acoustic recording (Fig. 2), the derived feature set is passed through a series of two neural networks. The first of these classifies an audio signature into one of three taxonomic groups: baleen whales, toothed whales or pinnipeds (a taxonomic clade including animals such as seals, walruses and sea lions). These functional groups represent the bulk of marine mammals of interest for entanglement concerns and tend to group species with similar acoustic signatures. When this classification is performed, the recording is passed through a group-specific classifier to attempt to further identify the species producing the acoustic signal. The utility of this two-step classification process is two-fold. First, a significant increase in species-level accuracy was observed on the initial set of testing data when using the two-step model as opposed to a single species-level classifier. Second, this approach has potential benefits where species level data is scarce, as identifying the species type can still provide useful data for facility operators, and in some instances is practically equivalent to a species determination. A pertinent example is the Gulf of Mexico, where the local population of the Bryde’s whale is an emphasis of conservation efforts. The Bryde’s whale is the sole species of baleen whale in the Gulf, therefore a classifier that can accurately identify baleen whale calls will functionally serve as a species-level determination in that region. A similar scenario exists for the Hawaiian monk seal, an endangered species that also represents the sole pinniped in the waters of the Hawaiian archipelago. The first generation of the classifier was trained on an initial set of 25 species of marine mammals (Table 1). The model was trained on sound clips obtained frommultiple sources, including Discovery of Sound in the Ocean (dosits.org), the International Workshop On Detection, Classification, Localization, and Density Estimation of Marine Mammals Using Passive Acoustics, NOAA researchers, and various other sources. With the initial testing set, the model was 97 percent accurate in predicting taxonomic group and 91 percent accurate in predicting species overall. This performance was based on a relatively small training and testing set, which is currently being expanded to include the Watkins Marine Mammal Sound Database curated by the Woods Hole Oceanographic Institution and the New Bedford Whaling Museum. The described models can be applied to discrete sound clips of limited duration and are best applied to individual instances of mammal vocalizations. To adapt these models to use with long-form hydrophone recordings or streaming audio, the model is applied to a sliding 10-sec window of audio. The software then reports identification occurrences along with a timestamp to locate the call within a larger audio sample. This methodology was used to scan long-form hydrophone recordings obtained by Salem State University at their offshore mussel farm, which had previously been analyzed by bioacoustic experts at CSA Ocean Sciences. The classifier identified several instances of humpback whale calls within the data (though some calls that had been identified by thirdparty analyzers were missed). Additionally, the model identified several instances of fin whale calls that were not identified by the third-party analysis but were observed at the site. Performance of the model is expected to improve with inclusion of the Watkins database that will represent an order of magnitude increase in the size of the dataset. However, the true test of the acoustic classifier will be its performance in real-world conditions. Current deployments will help to guide development and refinement moving forward (Table 1). Looking Ahead There are several opportunities for future development in the application of acoustic detection to entanglement prevention. Beyond measuring proximity and population distributions, researchers and aquaculture stakeholders are seeking to better understand how the behavior of marine mammals are impacted by offshore facilities and how they might interact with them. Acoustic detection is positioned to meet this need and specific call types have been associated with different behaviors such as foraging, traveling and mating in cetacean species. By training machine learning models to differentiate between individual call types, these tools can help to understand the behavior of nearby marine mammals in addition to species determination. One limitation of detection using a single hydrophone approach is the lack of information regarding location. The extensive detection range of underwater acoustic signals that makes the use of hydrophones so effective means that any observed signal could be originating anywhere within a large region surrounding the hydrophone. This problem can be addressed through use of a hydrophone array, using a technique known as multilateration. This involves measuring the time of arrival for a given sound signal at each individual sensor in an array. Using the position of each hydrophone and differences in arrival times, the position of the source object can be located. There are limitations to the use of multilateration and accurate results require hydrophones to be spaced roughly on the order of the distance to the signal. For best accuracy, the source object should be within the array of hydrophones. While coordination of a large-scale hydrophone array presents significant challenges, understanding proximity to or location within an aquaculture facility could provide vital information to operators and increasing understanding of mammal behaviors. Managing how aquaculture facilities interact with larger ecosystems and at-risk species will be a critical step in the expansion of aquaculture to the offshore environment and bring new capabilities to existing markets. The application of novel ( C O N T I N U E D O N P A G E 7 2 )
RkJQdWJsaXNoZXIy MjExNDY=