Source-linked AI summary
Generalized Seismic Phase Detection with Deep Learning
Zachary E. Ross, Men-Andrin Meier, Egill Hauksson, Thomas H. Heaton
TL;DR
Earthquake catalogs need sensitive detection but template matching is blind to events without suitable templates. This paper trains a ConvNet on hand-labeled Southern California phases for generalized phase detection, showing portability across earthquake sizes and settings while retaining template-matching-level sensitivity. The framework could improve catalogs and earthquake early warning.
Problem
Template-matching catalogs are biased because they cannot detect earthquakes for which no suitable waveform templates exist, including events in quiet regions or very large earthquakes.
Method
A convolutional neural network scans continuous seismic data in 4-second windows and classifies each window as P-wave, S-wave, or noise using millions of hand-labeled Southern California records.
Results
The GPD method detects P- and S-waves from very small to very large earthquakes, including data outside the training coverage, with sensitivity that rivals template matching.
Takeaways & Limitations
GPD may improve seismicity catalogs and provide a more stable basis for earthquake early warning by identifying signals reminiscent of P-waves.
Takeaways & Limitations
The model is trained specifically for direct P- and S-phases at local distances and may require retraining or added classes for larger distances and additional signals.
Abstract
from arXiv · showhide
To optimally monitor earthquake-generating processes, seismologists have sought to lower detection sensitivities ever since instrumental seismic networks were started about a century ago. Recently, it has become possible to search continuous waveform archives for replicas of previously recorded events (template matching), which has led to at least an order of magnitude increase in the number of detected earthquakes and greatly sharpened our view of geological structures. Earthquake catalogs produced in this fashion, however, are heavily biased in that they are completely blind to events for which no templates are available, such as in previously quiet regions or for very large magnitude events. Here we show that with deep learning we can overcome such biases without sacrificing detection sensitivity. We trained a convolutional neural network (ConvNet) on the vast hand-labeled data archives of the Southern California Seismic Network to detect seismic body wave phases. We show that the ConvNet is extremely sensitive and robust in detecting phases, even when masked by high background noise, and when the ConvNet is applied to new data that is not represented in the training set (in particular, very large magnitude events). This generalized phase detection (GPD) framework will significantly improve earthquake monitoring and catalogs, which form the underlying basis for a wide range of basic and applied seismological research.
1 Introduction
Seismic monitoring remains limited by tiny earthquakes and pervasive nuisance signals, while template matching improves sensitivity but misses events unlike available templates. The paper introduces deep learning as a way to detect seismic phases across broader circumstances without those template-related biases.
- Most earthquakes are very small, and nuisance signals continually interfere with detecting every earthquake.
- Southern California’s magnitude of completeness is now approximately M1.8 in most areas after improvements in instruments, networks, and detection methods.
- Template matching correlates cataloged-event waveforms with continuous data and can detect an order of magnitude more events than routine methods.
- Similarity-based methods require stringent waveform and network-pattern matches, leaving catalogs blind to events in inactive regions or very large earthquakes.
- The authors develop a ConvNet-based generalized phase detector trained on hand-labeled Southern California data to detect and classify seismic body-wave phases.
2 Methods and Results
The GPD ConvNet learns phase representations from millions of labeled records, classifies 4-second windows, and generalizes across noise, frequencies, instruments, regions, and earthquake magnitudes.
- Neural Networks and Deep Learning: The ConvNet automatically extracts features through convolution, decimation, and activation layers, avoiding manually selected input features.Its architecture uses four convolution layers and two fully connected layers.
- A ConvNet for Generalized Seismic Phase Detection: The model classifies each 4-second, three-component window as P-wave, S-wave, or noise using millions of hand-labeled Southern California records.The training data comprised 4.5 million records, evenly divided among 1.5 million P-wave, S-wave, and noise windows.
- Generalization: The method remained effective across a broad range of filter frequencies and detected P- and S-waves in the Mw 7.0 Kumamoto earthquake using acceleration traces filtered at 0.1 Hz.Training used a 2-Hz high-pass filter, but identical filtering was not required for new data.
- Classification Performance on the Validation Set: Precision exceeded 99% for both phases across almost all thresholds, while recall remained 96–99% for most threshold choices and was nearly as strong for S-waves as P-waves.The results indicate that noise was rarely labeled as seismic phase, although some phases were classified as noise.
- Classification Performance on the Validation Set: Probability thresholds trade false positives against false negatives and can therefore be selected according to catalog-generation or earthquake-early-warning priorities.Lower thresholds can increase recall when additional measures reduce false positives.
- Application to the 2016 Bombay Beach, California Swarm: In a two-minute Bombay Beach swarm segment, the method detected 13 earthquakes, identified both phases in 12, and resolved overlapping events with consistent P-before-S arrivals.Six detected events significantly overlapped in time, and the detections were consistent with real swarm earthquakes.
- Application to the 2016 Mw 7.0 Kumamoto, Japan Earthquake: Template-based methods cannot detect damaging earthquakes unlike their prior examples, whereas GPD detected P-waves at every tested Kumamoto station and S-waves at nearly every station.The model was trained only on Southern California data and earthquakes below M5, yet detected phases in Japan on different instruments.
3 Discussion
GPD extends sensitive seismic phase detection beyond template-matching biases, including large earthquakes and unfamiliar regions. The framework may support earthquake early warning, phase association, and characterization of seismicity across scales, while requiring adaptation for broader settings and signal classes.
- Detection framework: GPD’s sensitivity rivals template matching without inheriting similarity-based detection biases.The discussion attributes this to generalized representations of millions of P-waves, S-waves, and noise records.
- Limitations: Broader applications may require retraining with larger-distance recordings or adding signal and phase classes that are poorly represented in the current training data.The presented model specifically targets direct P- and S-phases at local distances, while transient noise and non-direct phases can complicate classification.
- Applications: GPD could provide a more stable foundation for earthquake early warning by distinguishing genuine earthquake phases from spurious signals.Phase labels could also exclude S-wave observations from P-wave-based magnitude estimation and help assess whether a rupture is still growing.
- Applications: Reliable phase labels may help phase association combine stations and disentangle simultaneous, overlapping events during earthquake swarms.The paper presents this as a suggested use of the method’s individual-station sensitivity and P-versus-S classification.
- Implications: GPD may sharpen characterization of earthquake evolution from barely detectable micro-earthquakes to destructive large-magnitude sequences.The stated scope concerns earthquake evolution across spatial and temporal scales and the study of seismogenic processes.
4 Data and Resources
The study uses a large Southern California waveform archive containing manually picked earthquake phases from 273,882 earthquakes recorded across 692 stations between 2000 and 2017.
- Earthquake archive: The dataset contains 273,882 Southern California earthquakes recorded from 2000–2017 at 692 broadband and short-period three-component stations.The earthquake magnitudes span −0.81 < M < 5.7.
- Labels and continuous data: SCSN analysts manually determined 1.5 million P-wave picks and 1.5 million S-wave picks for the waveform data.The study also used 24 hours of continuous data from station CI BOM on 2016-09-26.