Computational bioacoustics for the detection of rare acoustic events
| dc.contributor.author | McEwen, Ben | |
| dc.date.accessioned | 2024-10-16T03:16:12Z | |
| dc.date.available | 2024-10-16T03:16:12Z | |
| dc.date.issued | 2024 | |
| dc.description.abstract | The field of computational bioacoustics is a rapidly developing area of research. Tradi- tional bioacoustic monitoring has a long history of use particularly for monitoring of avian species. Species that were previously infeasible to monitor due to resource and time constraints are now a possibility. In New Zealand, bioacoustic monitoring has never been applied to the detection and surveillance of invasive terrestrial mammals. This dissertation presents the first research of its kind investigating the use of computational tools for the detection of invasive species. We evaluate the use of computational methods for the detection of common brushtail possums (Trichosurus vulpecula) a species that is invasive to New Zealand we also investigate the potential of acoustic monitoring of other challenging target species such as mustelids and rats. Low population densities (such as post-eradication efforts) encounter extremely sparse detections where it becomes challenging to reliably differentiate between species absence and non-detection. This issue of non-detection is accentuated by the use of passive acoustic monitoring technology at a landscape-scale. Thousands of hours of raw acoustic data can be collected weekly. Most of this data is empty of events of interest. It is infeasible for a human to manually analyse this data. The post-processing of acoustic data is a common challenge for passive acoustic monitoring applications. Rare feature detection (especially for challenging applications such as incursion detection and probability of absence testing) is currently at the border of feasibility using state- of-the-art computational bioacoustics tools. We present an active few-shot learning methodology that combines semi-supervised prototypical learning methods for efficient analysis of acoustic data with limited existing samples. We evaluate this methodology on an invasive species detection dataset demonstrating high performance at a range of data availability contexts. This methodology achieves a test accuracy of 98.4% (with fine- tuning) as well as 81.2% test accuracy using 2-shot, 2-way prototypical learning without fine-tuning, demonstrating high performance at varying data availability contexts. The development of improved methodologies capable of detecting rare acoustic features has clear benefits for other bioacoustic monitoring applications. This technology can be applied to other challenging applications such as monitoring of rare and at- risk species. These methods can be applied to scale the efficiency of data analysis of existing bioacoustic monitoring applications to achieve landscape-scale monitoring. Invasive species detection using bioacoustics represents a challenging but valuable area of research. We present the development of publicly available methodologies, tools and the first invasive species dataset containing 3500 labelled samples and over 1300 samples of brushtail possum vocalisations. The results of this work indicate the potential of computational bioacoustic methods for the detection of invasive species as well as rare and at-risk species that encounter similar monitoring challenges. In addition to improved bioacoustic detection methods, we also investigate the use of bioacoustic noise reduction methods evaluating both signal processing-based methods and deep audio enhancement methods. We investigate the efficacy of noise reduction for the performance of downstream segmentation and classification tasks identifying the limitations of common perceptual metric-based approaches. We find that noise reduction results in no improvement in segmentation precision and recall with an average AUC performance decrease of 19.2%. We also demonstrate no benefit in classification accuracy when tested using state-of-the-art time-domain and time-frequency-domain audio classification models with a marginal decrease in average validation accuracy of 0.41%. We contrast these findings with common perceptual metrics which demonstrate consistent increases in perceptual quality when noise reduction is applied with SnNR increases ranging from 14.0% to 41.3%. We also present initial work that investigates the use of visual detection methods. We develop predictive tracking methods used to improve population estimates of low- resolution and low-frame rate thermal cameras. We discuss the use of these tools within the context of predator-free New Zealand and the current pest management paradigm. | |
| dc.identifier.uri | https://hdl.handle.net/10092/107689 | |
| dc.language | English | |
| dc.language.iso | en | |
| dc.rights | All Right Reserved | |
| dc.rights.uri | https://canterbury.libguides.com/rights/theses | |
| dc.title | Computational bioacoustics for the detection of rare acoustic events | |
| dc.type | Theses / Dissertations | |
| thesis.degree.grantor | University of Canterbury | |
| thesis.degree.level | Doctoral | |
| thesis.degree.name | Doctor of Philosophy | |
| uc.college | Faculty of Engineering |
Files
Original bundle
1 - 1 of 1
No Thumbnail Available
- Name:
- McEwen, Benjamin (Ben)_ Final PhD Thesis.pdf
- Size:
- 27.6 MB
- Format:
- Adobe Portable Document Format
License bundle
1 - 1 of 1
No Thumbnail Available
- Name:
- license.txt
- Size:
- 1.71 KB
- Format:
- Item-specific license agreed upon to submission
- Description: