DuSAFNet: A Multi-Path Feature Fusion and Spectral-Temporal Attention-Based Model for Bird Audio Classification.

Area of Science:

• Bioacoustics
• Machine Learning
• Computational Ecology

Background:

• Fine-grained bird audio classification is crucial for biodiversity monitoring.
• Existing methods often struggle with complex spectro-temporal patterns and require significant computational resources.
• Automated acoustic monitoring offers a scalable solution for ecological assessments.

Purpose of the Study:

• To develop a lightweight deep neural network, DuSAFNet, for accurate fine-grained bird audio classification.
• To enhance the model's ability to capture both local spectral textures and long-range temporal dependencies.
• To improve inter-class separability across different frequency bands for robust classification.

Main Methods:

• DuSAFNet employs dual-path feature fusion and spectral-temporal attention mechanisms.
• A multi-band ArcMarginProduct classifier is utilized to boost inter-class separability.
• The model processes Mel-spectrograms, integrating local and global spectro-temporal cues.

Main Results:

• DuSAFNet achieved 96.88% accuracy and 96.83% F1 score on a dataset of 17,653 recordings across 18 species.
• The model demonstrates high efficiency with only 6.77 million parameters and 2.275 GFLOPs.
• Cross-dataset evaluation on Birdsdata yielded 93.74% accuracy, indicating strong generalization.

Conclusions:

• DuSAFNet offers a high-performance, lightweight solution for bird audio classification.
• Its efficiency makes it suitable for edge-device deployment and real-time alerts for species monitoring.
• This research supports scalable automated acoustic monitoring for biodiversity assessment and conservation planning.