TL;DR
This paper introduces SIGVISA, a novel Bayesian system for seismic monitoring that directly models raw seismic waveforms. By incorporating the physics of signal generation and using Gaussian Processes to capture waveform similarities, SIGVISA significantly outperforms existing methods in detecting weak seismic events and accurately locating them, even in regions with no historical seismic data.
What Problem Does This Paper Address?
Detecting and locating seismic events, especially weak ones, from noisy sensor data is a challenging perceptual task. Traditional seismic monitoring systems often rely on a "bottom-up" approach, where signals are first processed into discrete "detections" which are then associated into events. This process can miss subtle signals and is sensitive to false detections. Furthermore, existing methods struggle to detect events in regions with little to no historical seismic activity, which is a critical limitation for applications like nuclear test ban treaties. The paper aims to bridge this gap by developing a more robust and sensitive seismic monitoring system that directly leverages the rich information contained within raw seismic waveforms.
Methods
The core of SIGVISA is a generative probabilistic model that directly models seismic waveforms. Unlike previous systems that relied on pre-processed "detections," SIGVISA models the entire waveform, allowing it to incorporate detailed physical information about signal generation. Key methodological contributions include:
- Generative Waveform Model: SIGVISA models seismic events and the signals they produce at a network of stations. It uses a Poisson process for event occurrence and models each event's location, depth, origin time, and magnitude.
- Physics-Informed Signal Modeling: The system models the waveform of each seismic phase using a parametric envelope (linear onset followed by poly-exponential decay) and a modulation process. This allows it to capture path-dependent signal characteristics and predictable travel times.
- Gaussian Processes (GPs) for Repeatability: GPs are used to model the signal envelope parameters and wavelet coefficients. This enables SIGVISA to learn repeatable waveform characteristics from historical events, allowing it to detect new events by matching their signals to known patterns, even from a single station.
- Collapsed Signal Model: To improve inference efficiency, the model analytically marginalizes out the wavelet coefficients, resulting in a multivariate Gaussian distribution for the observed signals. This is then efficiently handled using Kalman filtering.
- Reversible Jump MCMC Inference: The system employs reversible jump Markov Chain Monte Carlo (MCMC) for inference, allowing it to dynamically propose, accept, and reject new events, and to handle unassociated arrivals as a mechanism for event birth and death.
Datasets
- Training Data: The system was trained on one year of historical seismic data from the western United States (January to December 2008), with a special focus on the first six hours following a major earthquake near Wells, NV, to capture aftershock correlations.
- Test Data: The test period comprised two weeks of seismic data, starting twelve hours after the Wells mainshock.
- Reference Bulletin: A reference bulletin was constructed by combining events from regional networks (NEIC), an analysis of Wells aftershocks, and data from the transportable US Array and temporary instruments. This served as a "ground truth" for evaluating the IMS-based systems.
Key Insights and Ideas
- Waveforms Hold Rich Information: The paper strongly demonstrates that directly modeling raw seismic waveforms, rather than relying on pre-processed detections, unlocks significantly more information for event detection and localization.
- Physics and Statistics Synergy: By combining physics-based models of seismic wave propagation with statistical inference (Bayesian methods and GPs), SIGVISA achieves a more accurate and sensitive understanding of seismic activity.
- Repeatability as a Detection Mechanism: The use of GPs to model repeatable waveform characteristics is a powerful way to detect events, especially weak ones, by recognizing their similarity to known patterns, even when only observed by a single station.
- Unified Inference for Detection and Location: SIGVISA's Bayesian framework naturally integrates event detection, association, and location into a single inference problem, overcoming limitations of traditional, separate processing steps.
- Handling "De Novo" Events: The system's ability to perform well even in regions with no historical seismicity is crucial for applications like nuclear monitoring, where detecting novel events is paramount.
Key Results
- Increased Event Recall: SIGVISA recovered three times as many events as the SEL3 baseline system at the same precision level.
- Reduced Location Errors: Mean location errors were reduced by a factor of four compared to previous work.
- Enhanced Sensitivity: The system showed greatly increased sensitivity to low-magnitude events.
- Competitive Performance on De Novo Events: SIGVISA matched or exceeded other systems in detecting "de novo" events (those with no nearby historical seismicity).
- Human-Level Performance: The human-reviewed LEB achieved near-perfect precision but only 9% recall, while SIGVISA achieved 33% recall at 14% precision, suggesting it recovers events not obvious to human analysts.
Limitations and Future Work
- Assumptions on Training Data: The current training relies on a bulletin of historical event locations as ground truth. The authors note that incorporating noisy training locations or fully unsupervised training would be valuable future work.
- Model Improvements: Future directions include incorporating more sophisticated tomographic travel-time models, directional information from seismic arrays, explicit modeling of earthquake vs. explosion sources, and better quantification of uncertainty and detection limits.
- Scalability: The authors aim to scale SIGVISA to global seismic data by exploiting parallelism and refining inference moves and implementation.
- Accessibility: They hope that successful applications of complex Bayesian models will inspire advances in probabilistic programming systems to make generative modeling more accessible.
Summary updated