Discussion
Managing the wellbeing of ecosystems requires identifying when and where
human activities are impacting species’ occurrence, movement, and
behaviour. PAM is a useful approach for the detection of both large- and
small-scale changes in urban and wild environments, as it allows for
continuous and prolonged ecosystem monitoring. Challenges in employing
PAM as a standard monitoring tool arise after data collection, when
researchers and practitioners need to quickly extract useful information
from large acoustic datasets, to understand when and where management
actions are needed to preserve the well-being of ecosystems. The
relatively new field of ecoacoustics provides the theoretical background
for linking specific characteristics of the acoustic environment to
biodiversity and ecosystem health. However, identifying a common
analytical approach has been an obstacle to the broad application of
ecoacoustics theory so far, and most studies employing ecoacoustics
indices are not suited for replicability and comparison.
We addressed these problems by linking marine ecoacoustics assessment to
the realms of machine learning and dimensionality reduction. We applied
a deep-learning approach to characterize the biological and
anthropogenic components of marine acoustic environments, and we
illustrated how acoustic features derived from a pre-trained
Convolutional Neural Network capture both the coarse and fine-grained
structure of large PAM datasets. These methods can be applied to a broad
range of marine and terrestrial systems.
Our analyses revealed several applications for inferring population- and
location-specific information from acoustic datasets. When datasets are
already labelled and focused on a specific taxon, such as the WMD, we
found that acoustic features were particularly suited for the
discrimination of marine mammal vocalizations. Understanding the
evolution of vocal diversity and the role of vocalizations in the
ecology of a species is one of the key objectives of bioacoustics
research (Luís et al., 2021). Full acoustic repertoires are not
available for most species, as building comprehensive lists of
vocalizations requires considerable research effort. Here we show how a
general acoustic classification model (VGGish) used as a feature
extractor allows us to detect differences and similarities among marine
mammal species, without requiring prior knowledge on the species’ vocal
repertoires. Our results for orcas are of particular interest, as they
provide insights on the vocal similarities and differences between
distinct populations of the same species. A large number of orca call
samples labelled as EN Pacific were classified as WN
Atlantic whales using the methodology in this study. Orcas show both
genetic divergence and differences in call frequency that are more
pronounced for sympatric ecotypes than whales found in different ocean
basins (Filatova et al., 2015). Although we cannot consider the
artefactual conflation of EN Pacific orcas with NW
Atlantic orcas in the WMD as definitive evidence of convergence in
vocal behaviour, we suggest that this aspect should be further
investigated, perhaps using more recent recordings of these different
orca populations.
More than 60 different ecoacoustic indices are being employed as
descriptors of terrestrial soundscapes (Bradfer-Lawrence et al., 2019),
making the search for indices that are successfully measuring
biodiversity across widely variable environments very challenging
(Minello et al., 2021). So far, ecoacoustic indices have been applied to
marine environments with little success (Bohnenstiehl et al., 2018). Due
to higher sound propagation efficiency, marine acoustic environments can
receive acoustic energy from many sources with some that are hundreds of
kilometres distant, making them more complex to study than terrestrial
environments. Accordingly, the biases shown by acoustic indices
measuring terrestrial species diversity (Eldridge et al., 2018;
Fairbrass et al., 2017; Heath et al., 2021) are amplified when
transferred to the study of marine environments (Bohnenstiehl et al.,
2018; Dimoff et al., 2021; Minello et al., 2021).
Machine learned acoustic features are a promising alternative to the use
of ecoacoustics indices for monitoring terrestrial biodiversity (Heath
et al., 2021; Sethi et al., 2020). In this study, we show how this
approach can also be extended to the study of marine soundscapes. The
derived acoustic features were successful in discriminating between two
different marine environments that differed in type and intensity of
anthropic activity: recordings collected in Burin were dominated by
distant seismic airgun pulses in the low frequency range, and the Red
Island hydrophone recordings were characterized by frequent ship noise.
Both sites yielded recordings of humpback whale vocalizations, and our
results show that machine-learned acoustic features can be employed for
detecting marine mammal sounds across different acoustic contexts.
Machine-learned acoustic features respond to multiple marine sound
sources, and can be employed successfully for investigating both the
biological and anthropic components of marine soundscapes.
Reducing acoustic features to two UMAP dimensions, however, resulted in
poorly performing classifiers for three sets of labels: airgun noise
presence, ship presence, and humpback whale presence. In all three
cases, repeating the analysis on a larger set of 128 features improved
model performance at the cost of increased processing time. The best
models used as little as two features, and as many as 64, whereas
classifiers based on the full 128 features were selected as best models
for all iterations of the humpback whale classifier (Appendix S1). This
indicates that the number of acoustic features could be significantly
reduced in some instances, thus reducing processing time and virtual
memory requirements. The poor performance observed in the UMAP ship
presence classifiers could be partly due to the approach adopted for
labelling presences and to the fact that ship noise was almost
ubiquitous in the Red Island recordings. Most samples collected at the
Red Island deployment location were more than 3 dB higher than the full
dataset median, but only a fraction of such samples contributed to the
broadband SPL (Appendix S2.2), indicating that ship presence may have
been underestimated. As an alternative, using records of vessel
positions obtained from the Automatic Identification System (AIS) as an
indicator of ship presence may improve model performance, at the cost of
underestimating the presence of small vessels, which are rarely equipped
with AIS.
Acoustic features have been shown to overcome many of the limitations of
ecoacoustics indices; for example, acoustic features outperform common
ecoacoustic indices in discriminating different environmental
characteristics (Sethi et al., 2020). Furthermore, acoustic features are
resilient to audio file compression and reduction of Nyquist frequency,
and provide results that are independent from type of recorders deployed
and choices relative to the temporal fragmentation of acoustic datasets
(Heath et al., 2021; Sethi et al., 2020). Here, we show that acoustic
features and UMAP dimensions allow for the comprehensive exploration of
marine PAM datasets. Features can be used to train classification models
focusing on biological and anthropogenic sound sources and allow for
fine-grain comparison of marine mammal vocalizations.
Two limitations persist. VGGish, the CNN used to extract the acoustic
features, is pre-trained on audio files with a sampling rate of 16 kHz,
resulting in a Nyquist frequency of 8 kHz. This is sufficient to capture
low frequency vocalizations but reduces its ability to discriminate
high-frequency sounds. Nonetheless, we were able to correctly classify
both high- and low-frequency vocalizations in the WMD examples,
including Phocoenidae sounds, a family that includes species that
can produce sounds up to 150 kHz. A second limitation is that acoustic
features are not a plug and play product, as establishing links between
features and relevant ecological variables requires additional analyses,
while ecoacoustic indices are designed as measures of specific
environmental characteristics.
By presenting a set of examples focused on marine mammals, we have
demonstrated the benefits and challenges of implementing acoustic
features as descriptors of marine acoustic environments. Our future
research will extend feature extraction and testing to full PAM datasets
spanning several years and inclusive of multiple hydrophone deployment
locations. Other aspects warranting further investigation are how
acoustic features perform when the objective is discriminating
vocalizations of individuals belonging to the same species or
population, as well as their performance in identifying samples with
multiple active sound sources.
Acoustic features are abstract representations of PAM recordings which
preserve the original structure and underlying relationships between the
original samples, and, at the same time, are a broadly applicable set of
metrices that can be used to answer ecoacoustics, ecology, and
conservation questions. As such, they can help us understand how natural
systems interact with, and respond to, anthropogenic pressures across
multiple environments. Lastly, the universal nature of acoustic features
analysis could help bridge the gap between terrestrial and marine
soundscape research. This approach could deepen our understanding of
natural systems by enabling multi-system environmental assessments,
allowing researchers to investigate and monitor, for example, how
stressor-induced changes in one system may manifest in another. And
these benefits accrue from an approach that is more objective than
manual analyses and requires far less human effort.
ACKNOWLEDGEMENTS
This project was funded by the Species at Risk, Oceans Protection Plan,
and Marine Ecosystem Quality programmes of the Department of Fisheries
and Oceans Canada, Newfoundland and Labrador Region, by Memorial
University of Newfoundland and Labrador, and by the Ph.D. program in
Evolutionary Biology and Ecology (University of Parma, agreement with
University of Ferrara and University of Firenze). Simone Cominelli was
also supported by the TD Bank Bursary for Environmental Studies. Nicolo’
Bellin was also supported by the ‘COMP-HUB’ Initiative, funded by the
‘Departments of Excellence’ Project of the Italian Ministry for
Education, University and Research (MIUR). We would like to express our
gratitude to the curators of the Watkins Marine Mammal Sound Database as
we believe that open access databases are incredibly relevant to the
development of global monitoring of natural systems. We thank all the
graduate students of the Northern EDGE Lab for their support, and their
effort in creating a welcoming and inclusive work environment. Lastly,
we would like to thank Madelyn Swackhamer, Sean Comeau, Lee Sheppard,
Greg Furey, and Andrew Murphy from DFO’s Marine Mammal Section for
collecting the data used in this study, providing detailed information
about the hydrophone deployments, and for their help and support with
accessing DFO’s PAM databases.
CONFLICT OF INTEREST STATEMENT
The authors declare that there are no conflicts of interest.
AUTHORS’ CONTRIBUTIONS
Simone Cominelli and Nicolo’ Bellin developed the concepts and
methodology described here, acquired the necessary databases, ran the
analysis, and prepared the first draft of the manuscript. Dr. Carissa
Brown and Dr. Valeria Rossi supervised the two main authors (Simone
Cominelli and Nicolo’ Bellin) throughout the preparation of the
manuscript, and provided space and equipment for conducting the
research. Dr. Jack Lawson provided access to DFO’s PAM database,
provided input during the development of the methodology, and reviewed
analysis results. All authors contributed critically to the drafts and
final submission, and gave approval for publication.
DATA AVAILABILITY
Scripts to reproduce the images and analysis described here, as well as
sample wav files, and tables containing all acoustic features and their
labels are available for download as Jupiter Notebooks through Dryad:
Link_to_repository_here