Is there an accurate and generalisable way to use soundscapes to monitor biodiversity?

Abstract

Acoustic monitoring has the potential to deliver biodiversity insight on vast scales. Whilst autonomous recording networks are being deployed across the world, existing analytical techniques struggle with generalisability. This limits the insight that can be derived from audio recordings in regions without ground-truth calibration data. By calculating 128 learned features and 60 soundscape indices of audio recorded during 8,023 avifaunal point counts from diverse ecosystems, we investigated the generalisability of soundscape approaches to biodiversity monitoring. Within each dataset, we found univariate correlations between several acoustic features and avian species richness, but features behaved unpredictably across datasets. Training a machine learning model on compound indices, we could predict species richness within datasets. However, models were uninformative when applied to datasets not used for training. We found that changes in soundscape features were correlated with changes in avian communities across all datasets. However, there were cases where avian communities changed without an associated shift in soundscapes. Our results suggest that there are no common hallmarks of biodiverse soundscapes across ecosystems. Therefore, soundscape monitoring should only be used when high quality ground-truth data exists for the region of interest, and in conjunction with more targeted and accurate in-person ecological surveys. By better understanding how to use interpret data reliably, we hope to unlock the scale at which acoustic monitoring can be used to deliver true impact for land managers and scientists monitoring biodiversity around the world.

Using acoustic indices in ecology: Guidance on study design, analyses and interpretation

“…However, there are contradictory patterns reported in the literature. There do not appear to be any acoustic indices that hold a consistent relationship with species richness or density across regions or taxonomic communities, limiting their potential as biodiversity proxies (Alcocer et al, 2022; Sethi et al, 2023). Some contradictory findings may arise from issues in research design (see considerations below), but other inconsistencies reflect soundscape differences among species, acoustic communities, biomes and ecosystems (Barbaro et al, 2022; Buxton et al, 2018; Eldridge et al, 2018).…”

“…As with all ecological data, exploratory analyses are necessary to sense‐check apparent trends against the patterns that acoustic indices may reflect (Table S1 and the references therein). This is particularly important when associating indices with biodiversity metrics; our current understanding is that there is no widely generalisable relationship between acoustic diversity and biodiversity (Alcocer et al, 2022; Sethi et al, 2023). Moreover, ‘biodiversity’ includes a broad range of different metrics, which themselves should be carefully selected for each ecological question.…”

“…These approaches are powerful, but there are some drawbacks. Firstly, existing methods are prone to over‐fitting; it is not clear how well representations learned from one soundscape generalise to novel contexts (Sethi et al, 2023). Secondly, these methods are notoriously opaque, making interpretation difficult (Rudin, 2019); future research should take inspiration from the field of machine vision (e.g.…”

Towards interpretable learned representations for Ecoacoustics using variational auto-encoding

“…The success of deep learning (DL) in real world audio data such as human speech recognition [30, 16, 40] and music generation [19, 36, 4] means many are now looking to deep neural networks for automated soundscape analysis [11]. Convolutional neural networks (CNNs) have proven successful at bioacoustic segmentation and classification tasks [20, 26] and early applications to generate whole soundscape descriptors show some promise [8, 41, 44, 42]. Audio CNNs learn a hierarchy of feature maps to embed latent representations that capture key characteristics of acoustic events.…”

“…Several recent papers have trialled using a CNN model VGGish pre-trained by supervision to extract features from PAM audio spectrograms [41, 44, 42, 43]. However, VGGish’s training data and pre-processing pipeline hard-code prior assumptions about the feature recognition task that may not necessarily be suitable for ecoacoustics.…”