BirdNET: A deep learning solution for avian diversity monitoring

• DOI: 10.1016/j.ecoinf.2021.101236
• Publication Date: 2021-01-27T14:56:52Z
• Authors: Stefan Kahl; Connor M. Wood; Maximilian Eibl; Holger Klinck
• Keywords: Soundscape; Bioacoustics; Citizen Science

Variation in avian diversity space and time is commonly used as a metric to assess environmental changes. Conventionally, such data were collected by expert observers, but passively acoustic rapidly emerging an alternative survey technique. However, efficiently extracting accurate species richness from large audio datasets has proven challenging. Recent advances deep artificial neural networks (DNNs) have transformed the field of machine learning, frequently outperforming traditional signal processing techniques domain event detection classification. We developed DNN, called BirdNET, capable identifying 984 North American European bird sound. Our task-specific model architecture was derived family residual (ResNets), consisted 157 layers with more than 27 million parameters, trained using extensive pre-processing, augmentation, mixup. tested against three independent datasets: (a) 22,960 single-species recordings; (b) 286 h fully annotated soundscape array autonomous recording units design analogous what researchers might use measure setting; (c) 33,670 single high-quality omnidirectional microphone deployed near four eBird hotspots frequented birders. found that domain-specific augmentation key build models are robust high ambient noise levels can cope overlapping vocalizations. Task-specific designs training regimes for recognition perform on-par very complex architectures other domains (e.g., object images). also temporal resolution input spectrograms (short FFT window length) improves classification performance sounds. In summary, BirdNET achieved mean average precision 0.791 recordings, F0.5 score 0.414 soundscapes, correlation 0.251 hotspot observation across 121 4 years data. By enabling efficient extraction vocalizations many hundreds potentially vast amounts data, similar tools potential add tremendous value existing future may transform ecology conservation.