Search

Acoustic monitoring has proven to be an effective tool for monitoring biotic soundscapes in the marine, terrestrial, and aquatic realms. Recently it has been suggested that it could also be an effective method for monitoring soil soundscapes, but has been used in very few studies, primarily in temperate and polar regions. We present the first study of soil soundscapes using passive acoustic monitoring in tropical forests, using a novel analytical pipeline allowing for the use of in-situ recording of soundscapes with minimal soil disturbance. We found significant differences in acoustic index values between burnt and unburnt forests and the first indications of a diel cycle in soil soundscapes. These promising results and methodological advances highlight the potential of passive acoustic monitoring for large-scale and long-term monitoring of soil biodiversity. We use the results to discuss research priorities, including relating soil biophony to community structure and ecosystem function, and the use of appropriate hardware and analytical techniques.

Ecoacoustics, the study of environmental sound, is a rapidly growing discipline offering ecological insights at scales ranging from individual organisms to whole ecosystems. Substantial methodological developments over the last 15 years have streamlined extraction of ecological information from audio recordings. One widely used set of methods are acoustic indices, which offer numerical summaries of the spectral, temporal and amplitude patterns in audio recordings. Currently, the specifics of each index's background, methodology and the soundscape patterns they are designed to summarise, are spread across multiple sources. Critically, details of index calculation are sometimes scarce, making it challenging for users to understand how index values are generated. Discrepancies in understanding can lead to misuse of acoustic indices or reporting of spurious results. This hinders ecological inference, replicability and discourages adoption of these tools for conservation and ecosystem monitoring, where they might otherwise provide useful insight. Here we present the Acoustic Index User's Guide—an interactive RShiny web app that defines and deconstructs eight of the most commonly used acoustic indices to facilitate consistent application across the discipline. We break the acoustic indices calculations down into easy‐to‐follow steps to better enable practical application and critical interpretation of acoustic indices. We demonstrate typical soundscape patterns using a suite of 91 example audio recordings: 66 real‐world soundscapes from terrestrial, aquatic and subterranean systems around the world, and 25 synthetic files demonstrating archetypal soundscape patterns. Our interpretation figures signpost specific soundscape patterns likely to be reflected in acoustic indices' values. This RShiny app is a living resource; additional acoustic indices will be added in the future through collaboration with authors of pre‐existing and new indices. The app also serves as a best‐p

A significant proportion of the world’s Common Scoters Melanitta nigra spend the non-breeding season in the coastal waters around Britain and Ireland. Overland movements of scoters across Britain were first documented over 140 years ago but the prevailing assumption in the ornithological literature has been that their primary migration routes are largely coastal. With the increased uptake of nocturnal migration monitoring, it has become apparent that Common Scoters are a seasonally predictable feature of the night-time soundscape across Britain, even at inland locations. We crowdsourced records of nocturnally migrating Common Scoters from England in the spring of 2020 to try and assess the migration routes taken. We combined this data with 20 years of reported sightings of inland Common Scoters to assess the phenology of these events. Our results suggest that Common Scoters from a number of geographically disjunct wintering areas move overland across England in spring using several potentially discrete migration routes, and that the timing of overland migration differs between the north and south of England. Across the whole of England, adult male Common Scoters formed the majority of reports in spring (62%) and summer (67%). However, in the late autumn period, the proportions were reversed, with 67% of birds concerning females or immatures. These trends were consistent across the three regions studied – the north, the midlands and the south – with the largest disparity between the sexes occurring during summer in the north of England, where 70% of the reports concerned drakes.

Estimation of avian biodiversity is a cornerstone measure of ecosystem condition. Surveys conducted using autonomous recorders are often more efficient at estimating diversity than traditional point-count surveys. However, there is limited research into the optimal temporal resolution for sampling—the trade-off between the number of samples and sample duration when sampling a survey window with a fixed survey effort—despite autonomous recorders allowing easy repeat sampling compared to traditional survey methods. We assess whether the additional temporal coverage from high temporal resolution (HTR) sampling, consisting of 240 15-s samples spread randomly across a survey window detects higher alpha and gamma diversity than low temporal resolution (LTR) sampling of four 15-min samples at the same locations. We do so using an acoustic dataset collected from 29 locations in a region of very high avian biodiversity—the eastern Brazilian Amazon. We find HTR sampling outperforms LTR sampling in every metric considered, with HTR sampling predicted to detect approximately 50% higher alpha diversity, and 10% higher gamma diversity. This effect is primarily driven by increased coverage of variation in detectability across the morning, with the earliest period containing a distinct community that is often under sampled using LTR sampling. LTR sampling produced almost four times as many false absences for species presence. Additionally, LTR sampling incorrectly found 70 species (34%) at only a single forest type when they were in fact present in multiple forest types, while the use of HTR sampling reduced this to just two species (0.9%). When considering multiple independent detections of species, HTR sampling detected three times more uncommon species than LTR sampling. We conclude that high temporal resolution sampling of passive-acoustic monitoring-based surveys should be considered the primary method for estimating the species richness of bird communities in tropical forests

Passive acoustic monitoring can be an effective method for monitoring species, allowing the assembly of large audio datasets, removing logistical constraints in data collection and reducing anthropogenic monitoring disturbances. However, the analysis of large acoustic datasets is challenging and fully automated machine learning processes are rarely developed or implemented in ecological field studies. One of the greatest uncertainties hindering the development of these methods is spatial generalisability—can an algorithm trained on data from one place be used elsewhere? We demonstrate that heterogeneity of error across space is a problem that could go undetected using common classification accuracy metrics. Second, we develop a method to assess the extent of heterogeneity of error in a random forest classification model for six Amazonian bird species. Finally, we propose two complementary ways to reduce heterogeneity of error, by (i) accounting for it in the thresholding process and (ii) using a secondary classifier that uses contextual data. We found that using a thresholding approach that accounted for heterogeneity of precision error reduced the coefficient of variation of the precision score from a mean of 0.61 ± 0.17 (SD) to 0.41 ± 0.25 in comparison to the initial classification with threshold selection based on F-score. The use of a secondary, contextual classification with thresholding selection accounting for heterogeneity of precision reduced it further still, to 0.16 ± 0.13, and was significantly lower than the initial classification in all but one species. Mean average precision scores increased, from 0.66 ± 0.4 for the initial classification, to 0.95 ± 0.19, a significant improvement for all species. We recommend assessing—and if necessary correcting for—heterogeneity of precision error when using automated classification on acoustic data to quantify species presence as a function of an environmental, spatial or temporal predictor variable.

Passive acoustic monitoring can be an effective method for monitoring species, allowing the assembly of large audio datasets, removing logistical constraints in data collection and reducing anthropogenic monitoring disturbances. However, the analysis of large acoustic datasets is challenging and fully automated machine learning processes are rarely developed or implemented in ecological field studies. One of the greatest uncertainties hindering the development of these methods is spatial generalisability—can an algorithm trained on data from one place be used elsewhere? We demonstrate that heterogeneity of error across space is a problem that could go undetected using common classification accuracy metrics. Second, we develop a method to assess the extent of heterogeneity of error in a random forest classification model for six Amazonian bird species. Finally, we propose two complementary ways to reduce heterogeneity of error, by (i) accounting for it in the thresholding process and (ii) using a secondary classifier that uses contextual data. We found that using a thresholding approach that accounted for heterogeneity of precision error reduced the coefficient of variation of the precision score from a mean of 0.61 ± 0.17 (SD) to 0.41 ± 0.25 in comparison to the initial classification with threshold selection based on F‐score. The use of a secondary, contextual classification with thresholding selection accounting for heterogeneity of precision reduced it further still, to 0.16 ± 0.13, and was significantly lower than the initial classification in all but one species. Mean average precision scores increased, from 0.66 ± 0.4 for the initial classification, to 0.95 ± 0.19, a significant improvement for all species. We recommend assessing—and if necessary correcting for—heterogeneity of precision error when using automated classification on acoustic data to quantify species presence as a function of an environmental, spatial or temporal predictor variable.

Acoustic indices are increasingly employed in the analysis of soundscapes to ascertain biodiversity value. However, conflicting results and lack of consensus on best practices for their usage has hindered their application in conservation and land‐use management contexts. Here we propose that the sensitivity of acoustic indices to ecological change and fidelity of acoustic indices to ecological communities are negatively impacted by signal masking. Signal masking can occur when acoustic responses of taxa sensitive to the effect of interest are masked by less sensitive acoustic groups, or target taxa sonification is masked by non‐target noise. We argue that by calculating acoustic indices at ecologically appropriate time and frequency bins, masking effects can be reduced and the efficacy of indices increased. We test this on a large acoustic dataset collected in Eastern Amazonia spanning a disturbance gradient of undisturbed, logged, burned, logged‐and‐burned, and secondary forests. We calculated values for two acoustic indices: the Acoustic Complexity Index and the Bioacoustic Index, across the entire frequency spectrum (0‐22.1 kHz), and four narrower subsets of the frequency spectrum; at dawn, day, dusk and night. We show that signal masking has a large impact on the sensitivity of acoustic indices to forest disturbance classes. Calculating acoustic indices at a range of narrower time‐frequency bins substantially increases the classification accuracy of forest classes by random forest models. Furthermore, signal masking led to misleading correlations, including spurious inverse correlations, between biodiversity indicator metrics and acoustic index values compared to correlations derived from manual sampling of the audio data. Consequently, we recommend that acoustic indices are calculated either at a range of time and frequency bins, or at a single narrow bin, predetermined by a priori ecological understanding of the soundscape.

The increasing demand for cost-efficient biodiversity data at large spatiotemporal scales has led to an increase in the collection of large ecoacoustic datasets. Whilst the ease of collection and storage of audio data has rapidly increased and costs fallen, methods for robust analysis of the data have not developed so quickly. Identification and classification of audio signals to species level is extremely desirable, but reliability can be highly affected by non-target noise, especially rainfall. Despite this demand, there are few easily applicable pre-processing methods available for rainfall detection for conservation practitioners and ecologists. Here, we use threshold values of two simple measures, Power Spectrum Density (amplitude) and Signal-to-Noise Ratio at two frequency bands, to differentiate between the presence and absence of heavy rainfall. We assess the effect of using different threshold values on Accuracy and Specificity. We apply the method to four datasets from both tropical and temperate regions, and find that it has up to 99% accuracy on tropical datasets (e.g. from the Brazilian Amazon), but performs less well in temperate environments. This is likely due to the intensity of rainfall in tropical forests and its falling on dense, broadleaf vegetation amplifying the sound. We show that by choosing between different threshold values, informed trade-offs can be made between Accuracy and Specificity, thus allowing the exclusion of large amounts of audio data containing rainfall in all locations without the loss of data not containing rain. We assess the impact of using different sample sizes of audio data to set threshold values, and find that 200 15 s audio files represents an optimal trade-off between effort, accuracy and specificity in most scenarios. This methodology and accompanying R package ‘hardRain’ is the first automated rainfall detection tool for pre-processing large acoustic datasets without the need for any additional rain gauge data.

The increasing demand for cost-efficient biodiversity data at large spatiotemporal scales has led to an increase in the collection of large ecoacoustic datasets. Whilst the ease of collection and storage of audio data has rapidly increased and costs fallen, methods for robust analysis of the data have not developed so quickly. Identification and classification of audio signals to species level is extremely desirable, but reliability can be highly affected by non-target noise, especially rainfall. Despite this demand, there are few easily applicable pre-processing methods available for rainfall detection for conservation practitioners and ecologists. Here, we use threshold values of two simple measures, Power Spectrum Density (amplitude) and Signal-to-Noise Ratio at two frequency bands, to differentiate between the presence and absence of heavy rainfall. We assess the effect of using different threshold values on Accuracy and Specificity. We apply the method to four datasets from both tropical and temperate regions, and find that it has up to 99% accuracy on tropical datasets (e.g. from the Brazilian Amazon), but performs less well in temperate environments. This is likely due to the intensity of rainfall in tropical forests and its falling on dense, broadleaf vegetation amplifying the sound. We show that by choosing between different threshold values, informed trade-offs can be made between Accuracy and Specificity, thus allowing the exclusion of large amounts of audio data containing rainfall in all locations without the loss of data not containing rain. We assess the impact of using different sample sizes of audio data to set threshold values, and find that 200 15 s audio files represents an optimal trade-off between effort, accuracy and specificity in most scenarios. This methodology and accompanying R package ‘hardRain’ is the first automated rainfall detection tool for pre-processing large acoustic datasets without the need for any additional rain gauge data.