Santiago Martinez Balvanera: Data-driven bat monitoring: leveraging machine learning

There is a growing need for innovative and effective tools for large-scale biodiversity monitoring to address the biodiversity crisis. Machine Learning (ML) has emerged as a powerful tool in this pursuit, as exemplified by the promising application of machine listening to analyze environmental audio data. However, acoustic monitoring presents significant challenges due to the wide variety of deployment environments, the inherent complexities of target sounds (often requiring subtle differentiations), and limited data availability in terms of taxonomic and geographic coverage. Despite these difficulties, they offer compelling research problems in the intersection of ML research and its application within complex ecological contexts.

In this talk I will explore key challenges encountered in developing machine learning (ML) tools for bat acoustic identification, focusing on how my research addresses these challenges. A critical obstacle lies in the efficient generation and curation of high-quality training data for bioacoustic ML tasks. My research tackles this by investigating best practices in annotation and developing collaborative annotation tools to facilitate the iterative refinement of both ML models and datasets. Furthermore, the limited availability of data necessitates sample-efficient ML training strategies. This can involve incorporating biologically relevant inductive biases into model architectures and maximizing the use of existing data resources. Finally, the presentation concludes by exploring the complexities of transitioning these ML tools into accessible solutions readily available to both researchers and decision-makers.

Santiago Martinez Balvanera is a PhD Student at University College London, investigating machine learning methods for extracting ecological insights from audio data. His primary focus lies in developing sound event detection models for bioacoustics, particularly in the few-shot regime. In collaboration with Kate Jones and Oisin Mac Aodha, he specializes in generalizable models for bat acoustic identification. Additionally, he contributes to open-source software tooling for computational bioacoustics and advocates for data standardization and sharing practices.