Systematic review of machine learning methods applied to ecoacoustics and soundscape monitoring

Unlocking the soundscape of coral reefs with artificial intelligence: pretrained networks and unsupervised learning win out

“…Previous work using automated techniques has predominantly used individual acoustic indices, which have revealed relationships between the soundscape and key ecological attributes on reefs (10,(33)(34)(35)(36). However, comparing individual indices provides only a weak signal and often fails in new contexts (2,25,37,38); this finding is supported by our findings (Fig. 4, Fig.…”

“…Our findings support this in the context of soundscapes given the broadly similar performance exhibited at the scale of data used here. Furthermore, datasets of this scale are easily achieved using raw soundscape data 23 . Alternatively, future work could identify the minimum sampling regime necessary, enabling efficient handling of equipment resources and field time by soundscape ecologists.…”

“…Our findings support this in the context of soundscapes given the broadly similar performance exhibited at the scale of the data used here. Furthermore, datasets of this scale are easily achieved using raw soundscape data, one of the key advantages of whole soundscape analysis over bioacoustic signal analysis which requires extensive human annotation to construct labelled datasets (25). Further accuracy improvements to supervised classifiers are more likely obtainable by considering recording periods longer than the one-minute samples used here.…”

“…We identified three approaches broadly representative of the techniques that can be used to extract feature embeddings from soundscape data for input into ML algorithms 23 . The first approach used a 'compound index', generated by calculating multiple acoustic indices from recordings and combining these into a vector, which can be used to train computationally cheap ML algorithms.…”

“…Both supervised and unsupervised ML algorithms leverage feature embeddings, compressed representations of the data either calculated using handcrafted features designed by experts or learned by the algorithm during the process. We identified three approaches broadly representative of the techniques that can be used to extract feature embeddings from soundscape data, which remain largely unbenchmarked against one another (25). The first approach uses a 'compound index', generated by calculating multiple acoustic indices from recordings and combining these indices into a vector.…”