Windy events detection in big bioacoustics datasets using a pre-trained Convolutional Neural Network.

Area of Science:

• Bioacoustics and Ecoacoustics
• Machine Learning for Environmental Monitoring

Background:

• Passive Acoustic Monitoring (PAM) generates large datasets for wildlife studies.
• Wind noise is a significant challenge, causing misclassifications and masking biological sounds in acoustic data.
• Reliable analysis of wildlife vocalizations requires effective noise reduction techniques.

Purpose of the Study:

• To develop a straightforward approach for identifying wind-affected recordings in PAM datasets.
• To pre-process acoustic data, ensuring the integrity and reliability of wildlife sound analysis.
• To evaluate the effectiveness of a pre-trained convolutional neural network (YAMNet) for wind noise detection.

Main Methods:

• Utilized YAMNet, a deep learning model for sound classification.
• Applied YAMNet in a Transfer Learning scenario with annotated data from the Stony Point Penguin Colony.
• Compared the performance of the pre-trained YAMNet against a fine-tuned model.

Main Results:

• YAMNet achieved a precision of 0.71 and recall of 0.66 in its pre-trained state.
• Transfer learning significantly improved performance, reaching a precision of 0.91 and recall of 0.92 (>28% relative increment).
• Demonstrated the model's capability to effectively identify wind-induced noise in acoustic recordings.

Conclusions:

• The fine-tuned YAMNet model offers a robust solution for detecting wind noise in bioacoustic data.
• This pre-processing step is crucial for ensuring the reliability of Passive Acoustic Monitoring data analysis.
• The open-access code and efficient model performance enable broad application on standard hardware for wind-noise-free acoustic research.