Automating the detection of marine sound sources from passive acoustic monitoring (PAM) data

Abstract

Passively listening to the marine environment allows us to characterise the combination of acoustic signals present over space and time, collectively known as the soundscape. A regional soundscape provides a reliable indicator of habitat quality, providing an observation of the distribution of vocalising species and monitors human activities which may influence the ecological health of an environment. In this thesis I use the COMPASS array, a suite of ten acoustic monitoring locations in the coastal region of western Scotland, to identify the complexities within the shallow water soundscape over small spatial scales. I describe how the cacophony of heterogenous acoustic sources present shape the temporal, spatial, and spectral patterns of each specific environment across the array, with variability in reported sound level metrics across space and time. To automate the detection and classification of spatial and temporal variations in ecosystem health and species presence from passive acoustic monitoring (PAM) data, algorithms must be capable of discrimination between marine species and the presence of anthropogenic and environmental noise, and be robust to fluctuating ambient noise levels. Within this thesis, I present a ‘small-scale’ convolutional neural network (CNN), trained to detect biological and anthropogenic signals from broadband acoustic data (0 – 48 kHz). I demonstrate that a computationally low-cost deep learning model can be used to mine extensive marine acoustic datasets for more than a single signal type, providing information on temporal scales relevant to the management of marine protected areas and the conservation of vulnerable species. For the CNN to be of use to the wider bioacoustic community, it must be robust within ambient noise fields, which vary to that of the original training set. The adaptability of the CNN to acoustic environments is demonstrated by fine-tuning the model with training sets of differing sizes curated from data collected in the Gulf of Mexico. The network reports an increase in accuracy of 30% after fine-tuning on a small amount of site-specific data (2.5 minutes), increasing by a further 5% with 25 minutes of data. Before CNNs can become commonplace in the bioacoustic toolbox, they must demonstrate performance benefits above that of currently used automated approaches to species monitoring. The final element of this thesis empirically evaluates the performance of the CNN, as a dolphin monitoring tool, in comparison to that of a widely popular data logger, the C-POD. Acoustic recordings (2136 hours) collected at two locations within the COMPASS array were analysed by both the CNN, the C-POD, and also manually to ground-truthed the detection positive hours (dph). The CNN achieves an overall accuracy of 0.84, and a false-positive rate of 0.05, outperforming the C-POD for all analysis periods. The CNN accurately describes seasonal and diel patterns in dolphin presence around each of the moorings, even when sensitivity is hindered by adverse weather periods. This thesis presents to the research community a multi sound source CNN for monitoring marine signals in diverse acoustic environments, which can provide information to management and stakeholders on an appropriate timeframe for conserving vulnerable species and habitats.

More information

Identifiers

ORCID for Paul White: orcid.org/0000-0002-4787-8713

ORCID for Jonathan Bull: orcid.org/0000-0003-3373-5807

Catalogue record

Export record