Deep learning categorization of infrasound array data

Area of Science:

• Geophysics
• Signal Processing
• Machine Learning

Background:

• Infrasound monitoring is crucial for detecting natural and anthropogenic events.
• Traditional methods for infrasound signal detection and categorization can be challenging due to signal complexity and noise.
• Automated analysis is needed to improve efficiency and reliability in processing large infrasound datasets.

Purpose of the Study:

• To develop and evaluate a deep learning methodology for infrasound detection and categorization.
• To improve the reliability of signal detection by utilizing features from beamforming analysis.
• To assess the performance of the deep learning model in distinguishing various infrasound signal types.

Main Methods:

• Convolutional neural networks (CNNs) with self-attention layers were employed for signal processing.
• Features extracted from coherence and direction-of-arrival (DOA) from beamforming were used as model inputs.
• An analyst-reviewed dataset from three International Monitoring System stations was utilized for training and validation.
• Models were evaluated using k-fold cross-validation, comparison with state-of-the-art methods, and transportability analysis.

Main Results:

• The deep learning model demonstrated high accuracy in detecting signals-of-interest (SOI), with f-scores consistently above 0.96 for noise vs. non-noise and SOI categorization.
• The model showed improved signal detection capabilities compared to using raw infrasound waveform data.
• Performance in distinguishing between short-duration stationary and non-stationary signals was mixed.
• Transportability analysis indicated the model's potential for application across different infrasound arrays.

Conclusions:

• Deep learning, particularly CNNs with self-attention, offers a highly accurate approach for infrasound signal categorization.
• Feature extraction from beamforming significantly enhances the reliability of infrasound detection.
• The developed methodology provides a robust tool for identifying signals-of-interest in complex infrasound data, supporting global monitoring efforts.