Ecoacoustic Data AnalyticsCase study
Advancing AI methods to determine ecosystem composition from acoustic recordings, distinguishing species, geophonic & anthropogenic sounds in soundscapes as well as flagging unusual or unanticipated sounds.
Ecological monitoring using acoustic data of the whole ecosystem (the soundscape) from passive acoustic sensors and AI is now a common approach for understanding impacts of environmental change on biodiversity and habitat health. However, scaling AI across different habitats, consistently separating the individual biotic, anthropogenic and geophonic sounds and classifying species specific sounds remains a significant challenge. The main bottleneck is limited amount of publicly available labelled data from many regions or underrepresented taxonomic groups to train the AI to classify target sounds, which is prohibitively time consuming and expensive to generate, requiring manual annotation of reference recordings. This is where unsupervised AI methods, such as self-supervised learning, would be an advantage, allowing us to learn representations and find hidden structure in unlabelled data.
Motivated by the recent success of AI large language models, we are exploring self-supervised learning, to learn the ‘grammar’ of soundscapes from unlabelled data. These methods generate internal representations of soundscapes (the AI ‘embeddings’) that separate and cluster individual species from other sounds, sub-cluster their behaviours and isolate unanticipated ‘anomalous’ sounds from unlabelled data. The AI models can be extended to applications when some labelled data is available, for example, for species classification. As a head-start we are adapting and refining models developed for bird sound to orthoptera, geophonic and anthropogenic sounds.
A key driver of our self-supervised approach is the use of ecoacoustics for soil health assessment, where limited knowledge of many below ground sounds constrain the applicability of the passive acoustic monitoring approach. Labelled soil sounds data is currently sparse, yet the self-supervised approach could circumvent the need for extensive reference soil sounds collection.
For information on the different habitats under investigation please see our associated project ‘Ecoacoustics for assessing ecosystem health and function, from air to soil’.
