Improving Regional and Teleseismic Detection for Single-Trace Waveforms Using a Deep Temporal Convolutional Neural...

Area of Science:

• Seismology and Geophysics
• Artificial Intelligence and Machine Learning

Background:

• Seismic event detection is crucial for Nuclear Treaty Monitoring.
• Traditional methods require expensive multi-instrument seismic arrays.

Purpose of the Study:

• To introduce DeepPick, a novel algorithm for seismic detection using single seismic traces.
• To demonstrate array-like detection performance from single-trace data.

Main Methods:

• Constructed a high-fidelity dataset pairing array-beam catalog arrival times with single-trace waveforms.
• Created an idealized characteristic function with exponential peaks.
• Employed a deep temporal convolutional neural network (CNN) to learn waveform transformations.

Main Results:

• DeepPick achieved 56% recall at a 0.001 type-I error rate, significantly outperforming existing methods.
• The algorithm demonstrated over twice the detections of STA/LTA and a 35% improvement over kurtosis-based detectors.
• DeepPick offers a 4 dB sensitivity improvement and is an order of magnitude faster computationally.

Conclusions:

• DeepPick offers a cost-effective and efficient solution for seismic event detection.
• The algorithm shows strong generalization and transportability to new seismic stations.
• DeepPick has the potential to significantly enhance global treaty monitoring network effectiveness.