Bayesian Convolutional Neural Networks for Uncertainty-Aware Classification of Infrasound Events

Area of Science:

• Geophysics
• Signal Processing
• Machine Learning

Background:

• Accurate infrasound signal classification is crucial for nuclear-test verification, natural-hazard warning, and geophysical monitoring.
• Conventional convolutional neural networks (CNNs) struggle with small, imbalanced datasets and quantifying predictive uncertainty.

Purpose of the Study:

• To develop a Bayesian CNN framework to address limitations of conventional CNNs in infrasound signal classification.
• To quantify epistemic and aleatoric uncertainty in infrasound signal classification models.

Main Methods:

• Implemented a Bayesian CNN framework using variational approximation, converting standard networks (LeNet-5, AlexNet, 4Conv3Fc) into Bayesian counterparts.
• Utilized short-time Fourier transform (STFT) to generate time-frequency spectrograms as input for the models.
• Evaluated models on a highly imbalanced dataset including nuclear tests, explosions, volcanic eruptions, rocket launches, earthquakes, and lightning.

Main Results:

• Bayes 4Conv3Fc achieved 99.14% accuracy on an imbalanced dataset without data augmentation.
• Demonstrated significant improvements in precision (up to 6.91%), recall (up to 7.12%), F1-score (up to 7.20%), and Cohen's Kappa (up to 8.86%) compared to deterministic baselines.
• Bayesian models showed lower Brier scores than class-weighted cross-entropy, indicating benefits from uncertainty modeling.

Conclusions:

• The Bayesian CNN framework effectively classifies infrasound signals, outperforming conventional methods on imbalanced datasets.
• Quantified and validated reliable epistemic and aleatoric uncertainty estimates, enhancing model interpretability and reliability.
• The approach offers a promising direction for uncertainty-aware infrasound monitoring and event detection.