Biodiversity loss is a global environmental issue (Cardinale et al., 2012), and it is now imperative to develop methods to efficiently monitor wildlife, accounting for spatial and temporal coverage (Joppa et al., 2016). Remote sensing techniques are being used to fill this gap, as they can be applied over large geographic areas where access may be difficult, allowing for some degree of unattended monitoring (Kerr and Ostrovsky, 2003). Remote sensing techniques include a range of technologies, like satellite imaging (Bonthoux et al., 2018), camera traps (Fontúrbel et al., 2021), Unmanned Aerial Vehicles (UAVs) (Nowak et al., 2019), and passive acoustic monitoring (PAM) (Froidevaux et al., 2014; Wrege et al., 2017).

Passive acoustic monitoring is now routinely used in terrestrial environments to monitor biodiversity (Gibb et al., 2019) with several purposes, such as understanding acoustic community composition of frog choruses (Ulloa et al., 2019), investigating acoustic species diversity of different taxonomic groups (Aide et al., 2017), and bird species recognition based on syllable recognition (Petrusková et al., 2016). Long-term recording can enable detection of species responses to important environmental impacts like climate change (Krause and Farina, 2016), and species recovery following extreme weather events (Duarte et al., 2021). However, recordings comprise large datasets which can be challenging to store, access and analyze (Ulloa et al., 2018). Subsampling is one way of dealing with these constraints, but it can limit the temporal and/or spatial scale of monitoring, therefore methods to analyze and filter recordings are necessary.

Currently, analysis of acoustic recordings still heavily relies on manual listening and inspection of recordings: this greatly limits the applicability of PAM. One alternative to that is to summarize acoustic information using acoustic indices, which mathematically represent different aspects of sound (e.g., frequency, intensity, etc.) (Sueur et al., 2014). Acoustic indices have, in some cases, been inspired by ecological indices. For example, the acoustic diversity index (Villanueva-Rivera et al., 2011) is based on the Shannon diversity index (Shannon and Weaver, 1964). NDSI (Gage and Axel, 2014) measures the ratio between biophony (biological sounds) and anthrophony (human and technological sounds) and is derived from NDVI, an index used in the remote sensing analysis of vegetation (Pettorelli, 2013). Acoustic indices have been used in different contexts such as to evaluate the differences in faunal beta-diversity between forests and plantations (Hayashi et al., 2020), to detect rainfall in acoustic recordings (Sánchez-Giraldo et al., 2020), to examine differences among indices representing taxonomic groups (e.g., birds, anurans, mammals and insects) (Ferreira et al., 2018), to relate indices with bird diversity (Tucker et al., 2014), and to identify frog species (Brodie et al., 2020).

Although there are numerous acoustic indices to choose from, different indices represent different acoustic phenomena in terrestrial environments, and the translation of acoustic into ecological information may vary depending on the context (Machado et al., 2017; Jorge et al., 2018; Bradfer-Lawrence et al., 2020). While there is no consensus on linking one index to one taxa, research has shown that combining indices can provide a good representation of different soundscapes (i.e., sounds in the landscape), especially across varying environments (Towsey et al., 2018), and can even be used to recognize different species (Brodie et al., 2020). Visualization tools such as false-color spectrograms (FCS) successfully combine three acoustic indices [Acoustic Complexity Index (Pieretti et al., 2011), Temporal Entropy (Sueur et al., 2008) and Events per Second (Towsey, 2018)] allowing different sound sources to be identified. The FCS and its combination of indices have been shown to provide a good representation of soundscapes in different contexts (e.g., Brodie et al., 2020; Znidersic et al., 2020). While visual representations of soundscapes are useful for scanning recordings for different phenomena (like rain, wind, or a frog chorus, for example), there is currently no available tool to statistically analyze these images. The underlying index data used to create the FCS can be retrieved and analyzed, but the mathematical interpretation of multiple indices remains a challenge, and therefore the statistical analysis of single indices is currently the favored approach. If mass deployments are required, [e.g., Australian Acoustic Observatory—(Roe et al., 2021)], we need to develop reliable, reproducible analysis methods with some degree of automation.

Furthermore, most statistical methods used for continuous recordings require an approach that accounts for temporal autocorrelation of the data (i.e., most statistical tests applied in ecology require independence of data). This means that each minute is not independent of the previous one in a recording, and this is often ignored in ecoacoustic studies. While spatial autocorrelation can be dealt with through experimental design, temporal correlation exists even when data are non-continuous (e.g., subsampled for example 1 min every 15 min) or arbitrarily split into time periods (e.g., day/night). Standard statistical approaches which assume independence of data cannot be applied for autocorrelated data.

Aiming to address the different challenges faced by researchers when analyzing recordings, we present a novel workflow for analyzing multi-index acoustic data. Our goal was to provide a tool that can be used by ecologists in a rapid assessment of terrestrial acoustic recordings. By having such a tool, ecologists can forward recordings of interest (i.e., for species identification) to specialists more efficiently, but also have quick metrics to compare ecosystems and/or recordings from different points in time. To deal with autocorrelated data capturing repeated patterns in acoustic indices is an alternative. Using the Hierarchical Based Motif Enumeration (HIME) (Gao and Lin, 2017) of acoustic indices, repetitive patterns of the data were detected (here referred to as motifs) in continuous recordings. As the algorithm searches for repetition of patterns in the time-series (Zolhavarieh et al., 2014) it was expected that noisy minutes (i.e., non-signal) would be excluded from the results as they tend to be random and not have a structure that repeats across time. Here we outline a semi-supervised method to classify the motifs according to dominant sounds. We demonstrate the transferability of the analysis in different environments and timescales with two case studies using data from two distinct ecosystems and recorded with different sampling schemes and devices.

Different lengths and minimum window sizes were tested for the two datasets and the minimum length selected was 30 min for both datasets. From an ecological perspective, 30 min of recording provides good resolution of fine-scale phenomena (e.g., a single species calling). Moreover, it can reveal soundscape changes throughout a day as the HIME progressively increases the window size. Having the same window length for both datasets is an advantage as it allows future comparisons to be made between results. All the selected motifs were labeled for both datasets so that the model accuracy could be measured, and sample size could be correctly estimated. The Bowra dataset had 549 selected motifs with a mean distance of 3.88 ± 1.24 and a mean length (in minutes) of 35.14 ± 3.88. The SERF dataset had 789 selected motifs with a mean distance of 4.28 ± 0.81 and a mean length (in minutes) of 36.02 ± 2.83. SERF dataset had 10% more hours than Bowra (542 and 494, respectively) and 43% more selected motifs.

The approach proposed here using time-series motif discovery and random forest classification represents a significant improvement in how acoustic indices are currently analyzed for terrestrial soundscapes. It resolves some major challenges and constraints associated with acoustic data analysis including: (1) accounts for temporal autocorrelation of acoustic data, which violates most statistical test assumptions; (2) combines more than one index to assign soundscape components; and (3) performs in different contexts, as demonstrated by the finding that the same set of indices identified the same soundscape components in different ecosystems surveyed at different times, using varying recording schemes.

Acoustic recording and indices are now routinely used to monitor biodiversity (Doohan et al., 2019; Moreno-Gómez et al., 2019), but statistical analysis of recordings is problematic due to temporal autocorrelation. By using sub-sequence time-series search, we were able to group sequences of minutes with repetitive patterns across the recordings, reducing the number of consecutive minutes analyzed as independent samples.

Acoustic data analysis approaches are varied and include linear mixed models (Francomano et al., 2020), mean differences (Carruthers-Jones et al., 2019), manual inspection and tagging species or groups of interest (Ferreira et al., 2018) or using non-index based metrics (like amplitude and frequency) direct from sound files (Furumo and Aide, 2019). Despite the variety of ways to analyze sound data, single index approach is still one of the most common approaches. As previously stated, single index data can be problematic because they cannot be consistently interpreted across different taxonomic groups or environments. For example, studies using ACI have shown that this index was positively correlated with bird species (Jorge et al., 2018; Mitchell et al., 2020), but also rain and wind (Duarte et al., 2015). Acoustic entropy has been found to have higher values in biodiversity rich habitats (Sueur et al., 2008), although higher values in quiet recordings and lower values in recordings dominated by insects have also been documented (Bradfer-Lawrence et al., 2019). Other studies have also tried to find a direct relationship between one index and one taxonomic group (Brown et al., 2019; Indraswari et al., 2020) but this relationship often does not hold across environments. From these findings we can conclude that a single index provides only a crude or obscure representation of biodiversity and is context-dependant. However, in our study, we have developed, validated and tested a new workflow that can be used in different terrestrial environments. This is particularly important because with recent advances in development of cost-effective ecoacoustic technology, passive acoustic recording is becoming a commonplace ecological survey approach worldwide. It is important that analytical tools are developed to meet this need, and are transferable across environments, providing standardized outputs for comparison or benchmarking.

An alternative to analyzing single index data is to combine indices, but this has been rarely attempted. One study used clusters to combine indices and classify major soundscape components (Phillips et al., 2018). However, their method still relies on listening to many recording minutes, which is extremely time-consuming and usually not feasible for large datasets. Another study combined indices using RF models to predict avian species richness (Buxton et al., 2018) and revealed that acoustic entropy and ACI were among the best predictors of avian biodiversity. But still, their aim was to link indices to a specific taxonomic group. Our approach is different because it shifts the focus from the index itself, to instead examine what is being captured by it. While often the focus of an ecological study is a target species or taxonomic group, soundscapes can provide valuable insights on processes (such as geophony and anthropophony) that may influence biodiversity. Until now, no analytical approach exists that can efficiently extract soundscape components in a semi-supervised and transferable way.

Using index-based spectrograms for visual inspection of recordings, we were able to accurately assign sound labels to motifs, extrapolating these labels to the whole data. Although a certain level of generalization was required when using multiple indices and automated classification techniques, this approach represents a progression from single indices and the manual identification of sounds or calls. Along with the generalization required, there were also issues with misclassifications by the algorithm. Nevertheless, inspection of misclassified motifs showed that, for example, some of them that were not classified as wind, did have wind present. The labeling process was based on the most predominant sound, which does not exclude the possibility of having more than one sound present at a given motif. In fact, the presence of more than one soundscape component is quite common, and for the SERF data here presented an additional label had to be created to address multiple dominant sounds in one motif. Ecosystems are complex and biodiversity is subject to a variety of influencing factors that will change according to geography and its features (Gaston, 2000). This variation challenges the use of automatic analyses, and it also makes it harder to compare different contexts. Nevertheless, it is necessary to establish a baseline for analysis so recordings can be effectively used for environmental and temporal comparisons. Moreover, it highlights the importance of the label process that provides the researcher with the opportunity to adjust the method to the context.

As the two study sites were in different ecoregions (Bowra is classified under the Temperate Grasslands, Savannas and Shrublands while SERF is Temperate Broadleaf and Mixed Forest (Environment Australia, 2000), it was expected that their soundscapes would vary due to the distinct biodiversity, ecology, and environmental conditions. Besides expected differences, the SERF dataset had 10% more minutes than Bowra, but 43% more motifs. Potential explanations include that SERF has a more complex soundscape, or more likely that the increase may be attributed to the lack of wind at SERF relative to Bowra, resulting in more minutes with signal and less noise.

Although soundscapes are known to vary between different environments, major soundtopes (Farina, 2014) were still expected to be found in both ecosystems (e.g., dawn and dusk choruses). Daily cycles were evident across the month at SERF, although variation could still be detected. This reflects environmental processes which also vary naturally across days. For example, areas near urban settlements, such as SERF, traffic noise can exhibit differences between weekdays and weekends. Biophony is also expected to change in response to temperature, rainfall, sunlight, and many other environmental factors that influence animal behavior (Pijanowski et al., 2011). The differences found here among and within ecosystems emphasizes again the importance of labeling motifs by a researcher before running the algorithm. Each recording will have distinct features that need to be addressed before data analysis. Furthermore, it provides an opportunity for the researcher to understand patterns and to become acquainted with the data specific to the site. It is also important to keep in mind that like other sampling methods, acoustic surveys are a snapshot of the moment in which the recordings were made. In order to track changes and effectively use this method as a biodiversity monitoring tool, it is important to establish sampling schemes that can capture different moments in time so that natural variation can be examined (Metcalf et al., 2020), as well as man-made impacts.

The labels in this study were generalized, however, future research could attempt to create more specific categories. At the same time, we argue that keeping upper levels of categories is important for model optimization, but it also might be useful when comparing results across datasets and studies. For example, every environment might have different species assemblages but similar patterns of biophony. For this reason, we believe the method presented here will help standardize analyses in ecoacoustics research. Another improvement that can be done is to assign more than one soundscape category per motif, creating a rank of sound presence. In this way, the misclassifications could be measured more accurately and potentially improved.

Ecoacoustics is a promising tool which is widely used to monitor biodiversity, and it has increased even more with the advent of acoustic indices. Nevertheless, until now there has been no consensus on how to transform acoustic indices into broad, transferrable ecological information, especially when combining indices. It is crucial to have an approach that standardizes and enables rapid assessment of terrestrial soundscapes. Although the analysis presented here treats indices separately as independent time series, there is no distinction between them for classification. This is important because it addresses the narrow assumption that each index serves as a good proxy for measuring specific taxonomic groups, and that these relationships will hold in different contexts. By combining different analysis techniques (time-series motif discovery and RF model classification), we were able to label grouped minutes of recordings translating acoustic indices into important components of the soundscape. We tested this approach on two datasets acquired using different recording devices and from different environments, providing strong evidence that this method can capture important temporal patterns in insect and bird biodiversity, as well as environmental geophonic sounds across environments. Given the global biodiversity loss that we are currently facing in the Anthropocene (Johnson et al., 2017), it is even more important that monitoring and methods of analysis are developed allowing to track changes in biodiversity.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: https://doi.org/10.5281/zenodo.4784758.

MS, SF, and BL conceived the ideas. MS and DT collected the data. MS and BL designed the methodology and analyzed the data. All authors contributed to data interpretation, drafts, critical revision and gave final approval for submission.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.