Traditional ecological survey methods are expensive and time-consuming as they typically involve a domain expert in the field. As a result of this, continuous audio recordings are playing an ever more critical role in conservation and biodiversity monitoring. However, listening to these recordings is often infeasible, as they can be thousands of hours long. The knowledge of domain experts can be efficiently leveraged using visualization. Traditionally, spectrograms are used to visualize audio data. However, there is a limit to the duration of audio that can fit on a computer screen as a spectrogram. Several techniques have been adapted to overcome this. Each of these techniques has its own set of advantages and disadvantages that make it ideal for some situations but unsuitable for others. In this paper, we propose a novel visualization based on embeddings produced by Frequency Preserving Autoencoders. We evaluate by visually comparing with traditional spectrogram and spectral index false-color spectrogram, using human-generated annotations as a baseline. We find that calls from some species are more consistently visible in our autoencoder-based visualization, while others are more clearly visible in the alternative visualizations. This novel visualization presents a new opportunity to assist ecologists in identifying bird calls in long-duration audio, if calls are difficult to identify using other methods.

1. Acoustic monitoring is an effective and scalable way to assess the health of important bioindicators like bats in the wild. However, the large amounts of resulting noisy data requires accurate tools for automatically determining the presence of different species of interest. Machine learning-based solutions offer the potential to reliably perform this task, but can require expertise in order to train and deploy. 2. We propose BatDetect2, a novel deep learning-based pipeline for jointly detecting and classifying bat species from acoustic data. Distinct from existing deep learning-based acoustic methods, BatDetect2’s outputs are interpretable as they directly indicate at what time and frequency a predicted echolocation call occurs. BatDetect2 also makes use of surrounding temporal information in order to improve its predictions, while still remaining computationally efficient at deployment time. 3. We present experiments on five challenging datasets, from four distinct geographical regions (UK, Mexico, Australia, and Brazil). BatDetect2 results in a mean average precision of 0.88 for a dataset containing 17 bat species from the UK. This is significantly better than the 0.71 obtained by a traditional call parameter extraction baseline method. 4. We show that the same pipeline, without any modifications, can be applied to acoustic data from different regions with different species compositions. The data annotation, model training, and evaluation tools proposed will enable practitioners to easily develop and deploy their own models. BatDetect2 lowers the barrier to entry preventing researchers from availing of effective deep learning bat acoustic classifiers. Open source software is provided at: ### Competing Interest Statement The authors have declared no competing interest.

1. Forests on private land have a wide range of uses that span activities such as recreation, primary production and nature conservation. Traditionally, it has been difficult for researchers to access private land to undertake systematic surveys. We used mini-acoustic sensors (Audiomoth) mailed via the postal service to overcome landholder concerns about researchers accessing private property, with a focus on properties used for private native forestry. 2. We surveyed koalas, an iconic threatened marsupial, in north-east New South Wales, Australia using passive acoustics, with repeat surveys over consecutive nights to account for imperfect detection in an occupancy modelling framework. 3. Over 3 years, we surveyed 128 sites and recorded 2,560 male bellows. Detection probability over seven nights was high (>0.79), but varied substantially between years, due to use of different sensors, housings and weather conditions. After accounting for detection probability, modelling revealed that koalas commonly occupied private native forests of the study region (probability of occupancy = 0.58 +/- 0.08). 4. Occupancy was modelled against several covariates and it varied with the landscape extent of sealed roads (.ve), NDVI (.ve) and a habitat suitability model (+ve, but minor). There was no support for occupancy in private forests to be related to a range of other factors including extent of surrounding cleared land, timber harvesting history, fire and other measured habitat features. 5. Synthesis and applications. We conclude that mini--acoustic recorders mailed to landholders were a highly effective method for assessing koala occupancy, after accounting for variable detection, and the approach could be deployed more widely for a range of species. Private native forests in partly cleared landscapes are commonly occupied by koalas, highlighting that this tenure is crucial for koala conservation and that practices seeking to balance conservation and production should be encouraged. In addition to sensitive habitat management in private forests, sealed road density is a major threat needing to be addressed.

Context It is notoriously difficult to estimate the size of animal populations, especially for cryptic or threatened species that occur in low numbers. Recent advances with acoustic sensors make the detection of animal populations cost effective when coupled with software that can recognise species-specific calls. Aims We assess the potential for acoustic sensors to estimate koala, Phascolarctos cinereus, density, when individuals are not identified, using spatial count models. Sites were selected where previous independent estimates of density were available. Methods We established acoustic arrays at each of five sites representing different environments and densities of koalas in New South Wales. To assess reliability, we compared male koala density estimates derived from spatial count modelling to independently derived estimates for each site. Key results A total 11 312 koala bellows were verified across our five arrays. Koalas were detected at most of our sample locations (96–100% of sensors; n = 130), compared with low detection rates from rapid scat searches at trees near each sensor (scats at <2% of trees searched, n = 889, except one site where scats were present at 69% of trees, n = 129). Independent estimates of koala density at our study areas varied from a minimum of 0.02 male koalas ha−1 to 0.32 ha−1. Acoustic arrays and the spatial count method yielded plausible estimates of male koala density, which, when converted to total koalas (assuming 1:1 sex ratio), were mostly equivalent to independent estimates previously derived for each site. The greatest discrepancy occurred where the acoustic estimate was larger (although within the bounds of uncertainty) than the independent mark–recapture estimate at a fragmented, high koala-density site. Conclusions Spatial count modelling of acoustic data from arrays provides plausible and reliable estimates of koala density and, importantly, associated measures of uncertainty as well as an ability to model spatial variations in density across an array. Caution is needed when applying models to higher-density populations where home ranges overlap extensively and calls are evenly spread across the array. Implications The results add to the opportunities of acoustic methods for wildlife, especially where monitoring of density requires cost-effective repeat surveys.

The compatibility of forestry and koala conservation is a controversial issue. We used a BACIPS design to assess change in koala density after selective harvesting with regulations to protect environmental values. We also assessed additional sites heavily harvested 5–10 years previously, now dominated by young regeneration. We used replicate arrays of acoustic sensors and spatial count modelling of male bellowing to estimate male koala density over 3600 ha. Paired sites in nearby National Parks served as controls. Naïve occupancy was close to 100% before and after harvesting, indicating koalas were widespread across all arrays. Average density was higher than expected for forests in NSW, varying between arrays from 0.03–0.08 males ha −1 . There was no significant effect of selective harvesting on density and little change evident between years. Density 5–10 years after previous heavy harvesting was equivalent to controls, with one harvested array supporting the second highest density in the study. Within arrays, density was similar between areas mapped as selectively harvested or excluded from harvest. Density was also high in young regeneration 5–10 years after heavy harvesting. We conclude that native forestry regulations provided sufficient habitat for koalas to maintain their density, both immediately after selective harvesting and 5–10 years after heavy harvesting.

Automatically detecting the calls of species of interest in audio recordings is a common but often challenging exercise in ecoacoustics. This challenge is increasingly being tackled with deep neural networks that generally require a rich set of training data. Often, the available training data might not be from the same geographical region as the study area and so may contain important differences. This mismatch in training and deployment datasets can impact the accuracy at deployment, mainly due to confusing sounds absent from the training data generating false positives, as well as some variation in call types. We have developed a multiclass convolutional neural network classifier for seven target bird species to track presence absence of these species over time in cotton growing regions. We started with no training data from cotton regions but we did have an unbalanced library of calls from other locations. Due to the relative scarcity of calls in recordings from cotton regions, manually scanning and labeling the recordings was prohibitively time consuming. In this paper we describe our process of overcoming this data mismatch to develop a recognizer that performs well on the cotton recordings for most classes. The recognizer was trained on recordings from outside the cotton regions and then applied to unlabeled cotton recordings. Based on the resulting outputs a verification set was chosen to be manually tagged and incorporated in the training set. By iterating this process, we were gradually able to build the training set of cotton audio examples. Through this process, we were able to increase the average class F1 score (the harmonic mean of precision and recall) of the recognizer on target recordings from 0.45 in the first iteration to 0.74.

Continuous audio recordings are playing an ever more important role in conservation and biodiversity monitoring, however, listening to these recordings is often infeasible, as they can be thousands of hours long. Automating analysis using machine learning is in high demand. However, these algorithms require a feature representation. Several methods for generating feature representations for these data have been developed, using techniques such as domain-specific features and deep learning. However, domain-specific features are unlikely to be an ideal representation of the data and deep learning methods often require extensively labeled data.In this paper, we propose a method for generating a frequency-preserving autoencoder-based feature representation for unlabeled ecological audio. We evaluate multiple frequency-preserving autoencoder-based feature representations using a hierarchical clustering sample task. We compare this to a basic autoencoder feature representation, MFCC, and spectral acoustic indices. Experimental results show that some of these non-square autoencoder architectures compare well to these existing feature representations.This novel method for generating a feature representation for unlabeled ecological audio will offer a fast, general way for ecologists to generate a feature representation of their audio, which does not require extensively labeled data.

Abstract Fauna surveys are traditionally manual, and hence limited in scale, expensive and labour‐intensive. Low‐cost hardware and storage mean that acoustic recording now has the potential to efficiently build scale in terrestrial fauna surveys, both spatially and temporally. With this aim, we have constructed the Australian Acoustic Observatory. It provides a direct and permanent record of terrestrial soundscapes through continuous recording across Australian ecoregions, including those periodically subject to fire and flood, when manual surveys are dangerous or impossible. The observatory comprises 360 permanent listening stations deployed across Australia. Groups of four sensors are deployed at each of 90 sites, placed strategically across ecoregions, to provide representative datasets of soundscapes. Each station continuously records sound, resulting in year‐round data collection. All data are made freely available under an open access licence. The Australian Acoustic Observatory is the world's first terrestrial acoustic observatory of this size. It provides continental‐scale environmental monitoring of unparalleled spatial extent, temporal resolution and archival stability. It enables new approaches to understanding ecosystems, long‐term environmental change, data visualization and acoustic science that will only increase in scientific value over time, particularly as others replicate the design in other parts of the world.

Continuous audio recordings are playing an ever more important role in conservation and biodiversity monitoring, however, listening to these recordings is often infeasible, as they can be thousands of hours long. Automating analysis using machine learning algorithms requires a feature representation. In this paper we propose a technique for learning a general feature representation from unlabelled audio using auto-encoders, which can be used for analysing environmental audio on a small timescale. We start by segmenting the audio data into non-overlapping 1-s long chunks and generating audio spectrograms. These audio spectrograms are then used to train a basic auto-encoder, with the output of the encoder network being used to generate the feature representation. We have found that at a 1-s timescale, our feature representation offers marginal improvements over ?acoustic indices?, a common representation for analysing environmental audio.

This thesis contributes to the field of acoustic environmental monitoring by developing novel semiautomated methods of processing long audio recordings to conduct species richness surveys efficiently. These methods allow a machine to select rich subset of the recordings though estimations of acoustic variety, which can then be presented to the human listener for species identifications. This work represents a step towards more effective biodiversity monitoring of vocal species that can be performed at a larger scale than is possible with traditional methods.

Acoustic recordings have been shown to be an effective way to conduct avian species surveys, whereby a trained expert listens to the audio and records observations, a task that can be very time consuming. In practice, most identification of species are first made by visual inspection of the spectrogram, with listening then performed for verification. This paper presents an approach for a surveyor to rapidly scan long duration recordings of environmental audio by automatically filtering parts with low activity and repetitions of the same call types. Recordings are segmented into fixed-length one-second non-overlapping clips. A classifier filters segments of low activity using features that are robust to different levels of background noise. The non-silent segments are then clustered using a feature representation derived from Time-domain Cepstral Coefficients, calculated from the discrete Fourier transform of downsampled spectrogram rows. This time-invariant feature representation allows for arbitrary segmentation, which is advantageous because segmentation of complex audio soundscapes into individual events is difficult and prone to errors. A visualization tool displays a representative segment from each cluster, grouped hierarchically, allowing an ecological researcher to very rapidly visually scan through the entire variety of audio events that occurred throughout the long recording, without wasting time on silent portions of the recording or repetitions of the same call-type. This tool provides functionality missing from both time-consuming audio players and black-box pattern recognizers, allowing conservation scientists to visually explore the entirety of their recordings.

In experimental research using computation, a workflow is a sequence of steps involving some data processing or analysis where the output of one step may be used as the input of another. The processing steps may involve user-supplied parameters, that when modified, result in a new version of input to the downstream steps, in turn generating new versions of their own output. As more experimentation is done, the results of these various steps can become numerous. It is important to keep track of which data output is dependent on which other generated data, and which parameters were used. In many situations, scientific workflow management systems solve this problem, but these systems are best suited to collaborative, distributed experiments using a variety of services, possibly batch processing parameter sweeps. This paper presents an R package for managing and navigating a network of interdependent data. It is intended as a lightweight tool that provides some visual data provenance information to the experimenter to allow them to manage their generated data as they run experiments within their familiar scripting environment, where it may not be desirable to commit to a fully-blown comprehensive workflow manager. The package consists of wrapper functions for writing and reading output data that can be called from within the R analysis scripts, as well as a visualization of the data-output dependency graph rendered within the R-studio console. Thus, it offers benefit to the experimenter while requiring minimal commitment for integration in their existing working environment.

In experimental research using computation, a workflow is a sequence of steps involving some data processing or analysis where the output of one step may be used as the input of another.The processing steps may involve user-supplied parameters, that when modified, result in a new version of input to the downstream steps, in turn generating new versions of their own output.As more experimentation is done, the results of these various steps can become numerous.It is important to keep track of which data output is dependent on which other generated data, and which parameters were used.In many situations, scientific workflow management systems solve this problem, but these systems are best suited to collaborative, distributed experiments using a variety of services, possibly batch processing parameter sweeps.This paper presents an R package for managing and navigating a network of interdependent data.It is intended as a lightweight tool that provides some visual data provenance information to the experimenter to allow them to manage their generated data as they run experiments within their familiar scripting environment, where it may not be desirable to commit to a fully-blown comprehensive workflow manager.The package consists of wrapper functions for writing and reading output data that can be called from within the R analysis scripts, as well as a visualization of the data-output dependency graph rendered within the R-studio console.Thus, it offers benefit to the experimenter while requiring minimal commitment for integration in their existing working environment.

Avian species richness surveys, which measure the total number of unique avian species, can be conducted via remote acoustic sensors. An immense quantity of data can be collected, which, although rich in useful information, places a great workload on the scientists who manually inspect the audio. To deal with this big data problem, we calculated acoustic indices from audio data at a one-minute resolution and used them to classify one-minute recordings into five classes. By filtering out the non-avian minutes, we can reduce the amount of data by about 50% and improve the efficiency of determining avian species richness. The experimental results show that, given 60 one-minute samples, our approach enables to direct ecologists to find about 10% more avian species.

Environmental acoustic recordings can be used to perform avian species richness surveys, whereby a trained ornithologist can observe the species present by listening to the recording. This could be made more efficient by using computational methods for iteratively selecting the richest parts of a long recording for the human observer to listen to, a process known as "smart sampling". This allows scaling up to much larger ecological datasets.In this paper we explore computational approaches based on information and diversity of selected samples. We propose to use an event detection algorithm to estimate the amount of information present in each sample. We further propose to cluster the detected events for a better estimate of this amount of information. Additionally, we present a time dispersal approach to estimating diversity between iteratively selected samples.Combinations of approaches were evaluated on seven 24-hour recordings that have been manually labeled by bird watchers. The results show that on average all the methods we have explored would allow annotators to observe more new species in fewer minutes compared to a baseline of random sampling at dawn.

Acoustic classification of anurans (frogs) has received increasing attention for its promising application in biological and environment studies. In this study, a novel feature extraction method for frog call classification is presented based on the analysis of spectrograms. The frog calls are first automatically segmented into syllables. Then, spectral peak tracks are extracted to separate desired signal (frog calls) from background noise. The spectral peak tracks are used to extract various syllable features, including: syllable duration, dominant frequency, oscillation rate, frequency modulation, and energy modulation. Finally, a k-nearest neighbor classifier is used for classifying frog calls based on the results of principal component analysis. The experiment results show that syllable features can achieve an average classification accuracy of 90.5% which outperforms Mel-frequency cepstral coefficients features (79.0%).

Frog protection has become increasingly essential due to the rapid decline of its biodiversity. Therefore, it is valuable to develop new methods for studying this biodiversity. In this paper, a novel feature extraction method is proposed based on perceptual wavelet packet decomposition for classifying frog calls in noisy environments. Pre-processing and syllable segmentation are first applied to the frog call. Then, a spectral peak track is extracted from each syllable if possible. Track duration, dominant frequency and oscillation rate are directly extracted from the track. With k-means clustering algorithm, the calculated dominant frequency of all frog species is clustered into k parts, which produce a frequency scale for wavelet packet decomposition. Based on the adaptive frequency scale, wavelet packet decomposition is applied to the frog calls. Using the wavelet packet decomposition coefficients, a new feature set named perceptual wavelet packet decomposition sub-band cepstral coefficients is extracted. Finally, a k-nearest neighbour (k-NN) classifier is used for the classification. The experiment results show that the proposed features can achieve an average classification accuracy of 97.45% which outperforms syllable features (86.87%) and Mel-frequency cepstral coefficients (MFCCs) feature (90.80%).

Acoustic recordings of the environment provide an effective means to monitor bird species diversity. To facilitate exploration of acoustic recordings, we describe a content-based birdcall retrieval algorithm. A query birdcall is a region of spectrogram bounded by frequency and time. Retrieval depends on a similarity measure derived from the orientation and distribution of spectral ridges. The spectral ridge detection method caters for a broad range of birdcall structures. In this paper, we extend previous work by incorporating a spectrogram scaling step in order to improve the detection of spectral ridges. Compared to an existing approach based on MFCC features, our feature representation achieves better retrieval performance for multiple bird species in noisy recordings.

In our large library of annotated environmental recordings of animal vocalizations, searching annotations by label can return thousands of results. We propose a heat map of aggregated annotation time and frequency bounds, maintaining the shape of the annotations as they appear on the spectrogram. This compactly displays the distribution of annotation bounds for the user's query, and allows them to easily identify unusual annotations. Key to this is allowing zero values on the map to be differentiated from areas where there are single annotations.