Robust sound event detection in bioacoustic sensor networks.

Area of Science:

• Bioacoustics
• Machine Learning
• Computational Ecology

Background:

• Bioacoustic sensors (autonomous recording units/ARUs) offer scalable, non-invasive wildlife monitoring.
• Automated sound event detection (SED) is crucial for analyzing bioacoustic data but struggles with environmental noise variability.
• Convolutional neural networks (CNNs) are effective for SED but can be unreliable in heterogeneous acoustic conditions.

Purpose of the Study:

• To develop and benchmark machine listening techniques for robust sound event detection (SED) across varied acoustic environments.
• To improve the generalizability of automated systems for detecting avian flight calls amidst background noise.
• To introduce and combine noise adaptation methods for enhanced reliability of bioacoustic monitoring.

Main Methods:

• Developed and integrated two noise adaptation techniques: per-channel energy normalization (PCEN) for short-term context and a context-adaptive neural network (CA-NN) layer for long-term context.
• Applied these methods to a CNN-based SED model for detecting avian flight calls from field recordings.
• Benchmarked the combined techniques against a baseline CNN and artificial data augmentation.

Main Results:

• Per-channel energy normalization (PCEN) reduced temporal overfitting (dawn vs. dusk).
• Context adaptation on PCEN-based statistics reduced spatial overfitting (across sensor locations).
• The combined approach achieved state-of-the-art performance, surpassing artificial data augmentation.

Conclusions:

• The developed noise adaptation techniques significantly improve the generalizability and reliability of SED models in challenging acoustic environments.
• Combining PCEN and context adaptation offers a powerful strategy for enhancing bioacoustic monitoring systems.
• A pre-trained avian flight call detector, BirdVoxDetect, is released to facilitate field research.