Abstract
Detecting weak seismic events from noisy sensors is a difficult perceptual task. We formulate this task as Bayesian inference and propose a generative model of seismic events and signals across a network of stations. Our approach, SIGVISA (Signal-based Vertically Integrated Seismic Analysis), directly models raw waveforms, allowing it to leverage weak signals that fall below traditional detection thresholds. By combining physical models of wave propagation with Gaussian Processes to capture local signal nuances, SIGVISA triples the detection rate of weak events and reduces location errors by 75% compared to current international standards.
Introduction
Seismic monitoring is a critical component of global security, particularly for the International Monitoring System (IMS) which listens for clandestine nuclear tests. Traditional monitoring systems often struggle to distinguish weak signals from background noise, primarily due to their reliance on a "bottom-up" approach that thresholds waveforms into discrete detections before network-level association.
The Problem with Traditional Monitoring
In conventional systems, a station thresholds a waveform to create a detection (e.g., a P-wave arrival). This process is inherently lossy: small events may never cross the threshold, while noise can trigger false positives. Furthermore, the Earth's complex geology distorts signals in ways that simple physics models cannot fully predict, leading to systematic errors in event location and characterization.
Methodology: The SIGVISA Approach
SIGVISA shifts the paradigm to a "top-down" generative model. Instead of associating detections, it asks: "What set of events most likely generated these raw waveforms?"
The Generative Model
SIGVISA defines a joint probability over event histories (location, time, magnitude) and the resulting signals. The signal is modeled in two parts:
- Parametric Envelope: Captures the general shape of the energy using parameters for arrival time, rise time, amplitude, and decay.
- Modulation Process: Captures detailed fluctuations using Daubechies wavelets.
Integrating Physics and Machine Learning
To predict signal characteristics, SIGVISA uses standard geophysical models (like IASPEI-91) augmented by Gaussian Processes (GPs). The GPs model the residuals—the difference between simple physics and observed waveforms—allowing the system to learn regional geological quirks over time. This enables "waveform matching," where the system recognizes similar signals from known faults.
Inference via RJMCMC
Since the number of events is unknown, the authors employ Reversible Jump Markov Chain Monte Carlo (RJMCMC). To maintain computational tractability, they use a "collapsed" model that marginalizes out high-dimensional wavelet coefficients using Kalman filtering.
Evaluation and Results
The system was evaluated on a two-week dataset from the Western United States following a magnitude 6.0 earthquake near Wells, Nevada. SIGVISA was compared against the automated system used by the CTBTO (SEL3) and a previous detection-based Bayesian system (NETVISA).
- Higher Recall: SIGVISA recovered 19.3% of events at the same precision level where SEL3 recovered only 6.4%—a 3x improvement.
- Superior Accuracy: Mean location errors were reduced by a factor of four.
- Sensitivity: The system excelled at catching low-magnitude events (below mb 2.5) that traditional systems missed entirely.
- Robustness: Even in "de novo" regions with no historical data, SIGVISA matched or exceeded existing performance by falling back on its underlying physics models.
Conclusion and Future Work
SIGVISA demonstrates that signal-based Bayesian inference significantly outperforms traditional detection-based methods. While computationally demanding—requiring significant CPU resources for processing—the gains in sensitivity and accuracy are substantial. Future work will focus on scaling the system for real-time global monitoring, incorporating multi-axis data, and exploring unsupervised learning from raw seismic data.
Summary updated