Efficient Masked Autoencoder for Birdsong Representation with Applications on Wild Bird Species Classification.

Area of Science:

• Ecology
• Bioacoustics
• Machine Learning

Background:

• Birds are vital indicators of biodiversity and ecological health.
• Monitoring avian populations non-invasively is crucial but challenging due to environmental noise and the need for extensive data annotation in traditional methods.
• Sensor technology for bird song identification offers a promising, eco-friendly approach.

Purpose of the Study:

• To develop an advanced bird song recognition system, CResMAE-BirdNET, that effectively extracts features from unlabeled acoustic data.
• To overcome limitations of existing methods, including environmental noise interference and reliance on manual data annotation.
• To enhance the accuracy and robustness of avian diversity monitoring.

Main Methods:

• Proposed CResMAE-BirdNET, integrating contrastive learning with a masked autoencoder framework.
• Incorporated audio enhancement techniques and a time-frequency self-calibration fusion module (TFSC) to mitigate noise and leverage spectral ripple features.
• Utilized residual attention in the encoder and a residual multi-layer perceptron in the decoder for superior local and global feature representation.

Main Results:

• Achieved high recognition accuracies of 99.35% on the Bird40Song dataset and 98.43% on the Birdsdata dataset.
• Attained F1-scores of 99.34% and 98.28% on the respective datasets, demonstrating exceptional performance.
• Validated the effectiveness of CResMAE-BirdNET in handling noisy acoustic environments and extracting meaningful features from unlabeled data.

Conclusions:

• CResMAE-BirdNET significantly advances bird song recognition capabilities.
• The proposed method offers a robust and efficient solution for large-scale ecological monitoring and biodiversity research.
• Autonomous feature extraction from unlabeled acoustic data holds great potential for bioacoustics and conservation efforts.