Advancements in preprocessing, detection and classification techniques for ecoacoustic data: A comprehensive review for large-scale Passive Acoustic Monitoring

Expert Systems with Applications, Journal Year: 2024, Volume and Issue: 252, P. 124220 - 124220

Published: May 16, 2024

Computational ecoacoustics has seen significant growth in recent decades, facilitated by the reduced costs of digital sound recording devices and data storage. This progress enabled continuous monitoring vocal fauna through Passive Acoustic Monitoring (PAM), a technique used to record analyse environmental sounds study animal behaviours their habitats. While collection ecoacoustic become more accessible, effective analysis this information understand monitor populations remains major challenge. survey paper presents state-of-the-art approaches, with focus on applicability large-scale PAM. We emphasise importance PAM, as it enables extensive geographical coverage monitoring, crucial for comprehensive biodiversity assessment understanding ecological dynamics over wide areas diverse approach is particularly vital face rapid changes, provides insights into effects these changes broad array species ecosystems. As such, we outline most challenging tasks, including pre-processing, visualisation, labelling, detection, classification. Each evaluated according its strengths, weaknesses overall suitability recommendations are made future research directions.

Language: Английский

---

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

Heliyon, Journal Year: 2023, Volume and Issue: 9(10), P. e20275 - e20275

Published: Sept. 22, 2023

Soundscape ecology is a promising area that studies landscape patterns based on their acoustic composition. It focuses the distribution of biotic and abiotic sounds at different frequencies attribute relationship said with ecosystem health metrics indicators (e.g., species richness, biodiversity, vectors structural change, gradients vegetation cover, connectivity, temporal spatial characteristics). To conduct such studies, researchers analyze recordings from Acoustic Recording Units (ARUs). The increasing use ARUs capacity to record hours audio for months time have created need automatic processing methods reduce consumption, correlate variables implicit in recordings, extract features, characterize sound related attributes. Consequently, traditional machine learning been commonly used process data characteristics soundscapes, mainly presence–absence species. In addition, it has employed call segmentation, identification, source clustering. However, some authors highlight importance new approaches unsupervised deep improve results diversify assessed this paper, we present systematic review field ecoacoustics processing. includes recent trends, as semi-supervised methods. Moreover, maintains format found reviewed papers. First, describe papers analyzed, configuration, study sites where datasets were collected. Then, provide an ecological justification relates monitoring features. Subsequently, explain followed assess various show trend towards label-free can large volumes gathered years. Finally, discuss adopt approach other biological dimensions landscapes.

Language: Английский

---

The effect of soundscape composition on bird vocalization classification in a citizen science biodiversity monitoring project

Ecological Informatics, Journal Year: 2023, Volume and Issue: 75, P. 102065 - 102065

Published: March 13, 2023

There is a need for monitoring biodiversity at multiple spatial and temporal scales to aid conservation efforts. Autonomous recording units (ARUs) can provide cost-effective, long-term systematic species data sound-producing wildlife, including birds, amphibians, insects mammals over large areas. Modern deep learning efficiently automate the detection of occurrences in these sound with high accuracy. Further, citizen science be leveraged scale up deployment ARUs collect reference vocalizations needed training validating models. In this study we develop convolutional neural network (CNN) acoustic classification pipeline detecting 54 bird Sonoma County, California USA, vocalization collected by scientists within Soundscapes Landscapes project (www.soundscapes2landscapes.org). We trained three ImageNet-based CNN architectures (MobileNetv2, ResNet50v2, ResNet100v2), which function as Mixture Experts (MoE), evaluate usefulness several methods enhance model Specifically, we: 1) quantify accuracy fully-labeled 1-min soundscapes an assessment real-world conditions; 2) assess effect on precision recall additional pre-training external archive (xeno-canto) prior fine-tuning from our domain; and, 3) how detections errors are influenced presence coincident biotic non-biotic sounds (i.e., soundscape components). evaluating (n = 37 species) across probability thresholds models, found followed improved average 10.3% relative no pre-training, although there was small 0.8% reduction recall. selecting optimal architecture each based maximum F(β 0.5), MoE approach had total 84.5% 85.1%. Our exhibit issues arising applying county scale, relatively low fidelity recordings background noise overlapping vocalizations. particular, human significantly associated more incorrect (false positives, decreased precision), while physical interference (e.g., recorder hit branch) geophony wind) classifier missing negatives, recall). process surmounted obstacles, final predictions allowed us demonstrate applied low-cost paired valuable diversity

Language: Английский

---

Soundscape Characterization Using Autoencoders and Unsupervised Learning

Sensors, Journal Year: 2024, Volume and Issue: 24(8), P. 2597 - 2597

Published: April 18, 2024

Passive acoustic monitoring (PAM) through recorder units (ARUs) shows promise in detecting early landscape changes linked to functional and structural patterns, including species richness, diversity, community interactions, human-induced threats. However, current approaches primarily rely on supervised methods, which require prior knowledge of collected datasets. This reliance poses challenges due the large volumes ARU data. In this work, we propose a non-supervised framework using autoencoders extract soundscape features. We applied dataset from Colombian landscapes captured by 31 audiomoth recorders. Our method generates clusters based autoencoder features represents cluster information with prototype spectrograms centroid decoder part neural network. analysis provides valuable insights into distribution temporal patterns various sound compositions within study area. By utilizing autoencoders, identify significant characterized recurring intense types across multiple frequency ranges. comprehensive understanding area's allows us pinpoint crucial sources gain deeper its environment. results encourage further exploration unsupervised algorithms as promising alternative path for environmental changes.

Language: Английский

---

Deep learning bird song recognition based on MFF-ScSEnet

Ecological Indicators, Journal Year: 2023, Volume and Issue: 154, P. 110844 - 110844

Published: Aug. 30, 2023

Bird diversity plays an important role in ecological balance, and bird song identification is of great practical significance. The spectrum generated by feature extraction shows good performance on classification. However, the information extracted filter process spectrogram generation can cause loss, which limits learning ability birdsong recognition. This study proposes a fusion network (MFF-ScSEnet) to solve this problem. audios Mel-spectrogram with low-frequency advantage Mel-filter, Sinc-spectrogram timbral Sincnet-filter, respectively, perform early strategy. ScSEnet attention module introduced into backbone ResNet18 enhance sound ripple spectrogram, reduce influence noise recognition improve network. Based MFF-ScSEnet paper, accuracy experimental results self-built dataset (Huabei_dataset), public datasets Urbansound8K Birdsdata reached 96.28%, 98.34%, 96.66%, respectively. indicated that method proposed paper superior recent latest method.

Language: Английский

---

Classification of Complicated Urban Forest Acoustic Scenes with Deep Learning Models

Forests, Journal Year: 2023, Volume and Issue: 14(2), P. 206 - 206

Published: Jan. 20, 2023

The use of passive acoustic monitoring (PAM) can compensate for the shortcomings traditional survey methods on spatial and temporal scales achieve all-weather wide-scale assessment prediction environmental dynamics. Assessing impact human activities biodiversity by analyzing characteristics scenes in environment is a frontier hotspot urban forestry. However, with accumulation data, selection parameter setting deep learning model greatly affect content efficiency sound scene classification. This study compared evaluated performance different models classification based recorded data from Guangzhou forest. There are seven categories classification: sound, insect bird bird–human insect–human bird–insect silence. A dataset containing was constructed, 1000 samples each scene. requirements training volume epochs were through several sets comparison experiments, it found that able to satisfactory accuracy when sample single category 600 100. To evaluate generalization new small test multiple trained used make predictions dataset. All experimental results showed DenseNet_BC_34 performs best among models, an overall 93.81% validation provides practical experience application techniques perspectives technical support further exploring relationship between biodiversity.

Language: Английский

---

Graph-Based Audio Classification Using Pre-Trained Models and Graph Neural Networks

Sensors, Journal Year: 2024, Volume and Issue: 24(7), P. 2106 - 2106

Published: March 26, 2024

Sound classification plays a crucial role in enhancing the interpretation, analysis, and use of acoustic data, leading to wide range practical applications, which environmental sound analysis is one most important. In this paper, we explore representation audio data as graphs context classification. We propose methodology that leverages pre-trained models extract deep features from files, are then employed node information build graphs. Subsequently, train various graph neural networks (GNNs), specifically convolutional (GCNs), GraphSAGE, attention (GATs), solve multi-class problems. Our findings underscore effectiveness employing represent data. Moreover, they highlight competitive performance GNNs endeavors, with GAT model emerging top performer, achieving mean accuracy 83% classifying sounds 91% identifying land cover site based on its recording. conclusion, study provides novel insights into potential learning techniques for analyzing

Language: Английский

---

Exploring emergent soundscape profiles from crowdsourced audio data

Computers Environment and Urban Systems, Journal Year: 2024, Volume and Issue: 110, P. 102112 - 102112

Published: April 8, 2024

The key component of designing sustainable, enriching, and inclusive cities is public participation. soundscape an integral part immersive environment in cities, it should be considered as a resource that creates the acoustic image for urban environment. For planning professionals, this requires understanding constituents citizens' emergent experience. goal study to present systematic method analyzing crowdsensed data with unsupervised machine learning methods. This applies sound- scape experience collection low threshold aim analyze using methods give insights into perception quality. purpose, qualitative raw audio were collected from 111 participants Helsinki, Finland, then clustered further analyzed. We conclude analysis combined accessible, mobile crowdsensing enable results can applied track hidden experiential phenomena soundscape.

Language: Английский

---

Sonotope patterns within a mountain beech forest of Northern Italy: a methodological and empirical approach

Frontiers in Ecology and Evolution, Journal Year: 2024, Volume and Issue: 12

Published: March 7, 2024

According to the Sonotope Hypothesis, heterogenous nature of acoustically sensed, but not yet interpreted, environmental sounds (i.e., sonoscape) is created by spatial and temporal conformation sonic patches (sonotopes) as recently been described in a Mediterranean rural landscape. We investigated Hypothesis mountain beech forest Northern Apennines, Italy that notoriously poor soniferous species. Our aim was test whether sonotopes were temporally distinct over seasonal astronomical timeframes spatially configured relation vegetation variables. used Acoustic Complexity Index (ACI tf ) analyze heterogeneity information gathered from an array 11 sound recorders deployed within lattice eleven 4-ha hexagonal sample sites distributed throughout 48-ha managed forest. visualized patterns ACI between seasons (May–June July–August 2021), across six periods (Night I, Morning Twilight, Morning, Afternoon, Evening Night II), according two aggregated frequency classes (≤2000 >2000 Hz). introduced Spectral Sonic Signature (SSS) calculated sequence values along bins descriptor dynamic production scales. Mean Dissimilarity compare SSS sites. identified grouping similar for each site generated cluster analyses their arrangements. Frequencies ≤2000 Hz (mainly geophonies wind rain) more prevalent than frequencies biophonies songbirds). Despite there being no strong relationship variables minimal biophony anthropophony, still emerged every period. This suggests sonoscape expresses sonotope configurations associated with geophysical events generate animal or anthropogenic occurrences. A new strategy based on reintroduction indigenous trees shrubs clearings should be considered enhancing local biodiversity conservation ecoacoustic monitoring Hypothesis.

Language: Английский

---

Temporal patterns in Malaysian rainforest soundscapes demonstrated using acoustic indices and deep embeddings trained on time-of-day estimation

The Journal of the Acoustical Society of America, Journal Year: 2025, Volume and Issue: 157(1), P. 1 - 16

Published: Jan. 1, 2025

Rapid urban development impacts the integrity of tropical ecosystems on broad spatiotemporal scales. However, sustained long-term monitoring poses significant challenges, particularly in regions. In this context, ecoacoustics emerges as a promising approach to address gap. Yet, harnessing insights from extensive acoustic datasets presents its own set such time and expertise needed label species information recordings. Here, study an investigating soundscapes: use deep neural network trained time-of-day estimation. This research endeavors (1) provide qualitative analysis temporal variation (daily monthly) soundscape using conventional ecoacoustic indices embeddings, (2) compare predictive power both methods for estimation, (3) performance supervised classification unsupervised clustering specific recording site, habitat type, season. The study's findings reveal that proposed embeddings exhibit overall comparable performance. article concludes by discussing potential avenues further refinement method, which will contribute understanding across space.

Language: Английский

---