Martinez Balvanera, Santiago; (2025) Enabling Accessible and Adaptable AI for Bioacoustic Monitoring from Data Annotation to Edge Deployment. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Martinez Balvanera_10207196_Thesis.pdf Download (28MB) | Preview

Abstract

To address the critical challenge of biodiversity loss, it is essential to upscale monitoring efforts to help inform conservation actions. The growing application of AI for automated species detection and classification in audio streams offers a promising solution. However, the current application of AI in bioacoustics is limited in scope, often constrained by the lack of gold standard data, technical resource disparities, and a lack of accessible tools. In this thesis, I investigate techniques to facilitate accessible AI development and deployment in bioacoustics. Chapter 2 examines data annotation, the first stage of bioacoustic AI development, and whether current tools support collaborative and iterative improvement of AI models and datasets. I find that current bioacoustic annotation tools are insufficient for modern AI development and, in response, I develop whombat, an open-source tool designed for iterative data and model improvement. Chapter 3 investigates the type and quantity of annotations needed for effective bioacoustic classification. With an extensively annotated bat call dataset, I show that annotating spatio-temporal locations of calls substantially improves classification performance, especially in low-data scenarios. Chapter 4 investigates if standard AI techniques from computer vision are efficient for bioacoustic analysis. I show that models tailored to the temporal nature of bioacoustic data outperform previous approaches and adapt to small-scale bat call datasets from diverse regions. Chapter 5 examines deploying AI models on edge devices for bioacoustic monitoring, finding that while minimising maintenance and reliance on extensive data infrastructure, tailoring solutions to specific monitoring goals can require advanced coding skills. To address this challenge, I develop acoupi, an open-source framework that simplifies the creation and deployment of such tailored solutions, with its effectiveness demonstrated through a month-long field deployment of a novel bat detection model. This research helps overcome challenges limiting AI-powered bioacoustics, paving the way for its broader use in conservation.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item