Spectrogram-derived graphs and inductive learning for multi-label avian vocalization detection in field recordings.

Area of Science:

• Bioacoustics
• Machine Learning
• Deep Learning

Background:

• Accurate detection of avian vocalizations is crucial for biodiversity monitoring and ecological research.
• Traditional methods often struggle with complex acoustic environments and overlapping sounds.
• Deep learning offers potential for improved automated sound event detection.

Purpose of the Study:

• To develop and evaluate an inductive spatial geometric deep learning framework for multi-label avian vocalization detection.
• To compare the performance of GraphSAGE and Graph Attention Network (GAT) models in this task.
• To assess the impact of data augmentation (SpecAugment) on model robustness.

Main Methods:

• Constructing graphs from Mel-spectrograms using a Deep Convolutional Neural Network (Deep CNN).
• Employing spatial inductive graph neural networks (GraphSAGE and GAT) for node-feature graph processing.
• Utilizing SpecAugment for data augmentation to enhance model generalization.

Main Results:

• The proposed framework achieved high performance on the Xeno-canto dataset.
• GraphSAGE achieved a macro F1-score of 0.90, and GAT achieved 0.92.
• Replacing Deep CNN with AudioProtoPNet-20 and using GAT resulted in a macro F1-score of 0.93.

Conclusions:

• Inductive spatial graph-based deep learning models show superior performance for avian vocalization detection.
• GAT demonstrates strong potential for accurate bird sound identification.
• The methodology offers a robust and generalizable approach for bioacoustic analysis.