Source-linked AI summary
Convolutional Neural Network for Earthquake Detection and Location
Thibaut Perol, Michaël Gharbi, Marine Denolle
TL;DR
Induced, low-magnitude earthquakes are difficult to detect reliably in noisy, expanding seismic records, motivating more exhaustive catalogs. ConvNetQuake uses a supervised convolutional network to detect and locate earthquakes from a single waveform. In Oklahoma, it combines high detection and location accuracy with substantially faster processing than established methods, while requiring cataloged earthquakes for training.
Problem
Existing detection methods often miss low-magnitude induced earthquakes in seismic noise, while waveform autocorrelation is computationally intensive and does not scale to long time series.
Method
ConvNetQuake is a supervised convolutional network that classifies a three-channel waveform window as noise or an earthquake associated with one of six geographic clusters.
Results
ConvNetQuake achieved 100% event-detection accuracy, 99.9% accuracy on noise windows, 74.5% six-class location accuracy, and approximately 13,500-times-faster processing than autocorrelation.
Takeaways & Limitations
The method detects previously uncataloged and non-repeating events and supports scalable monitoring of continuous seismic records from a single signal.
Takeaways & Limitations
ConvNetQuake requires a pre-existing history of cataloged earthquakes for training, limiting its suitability in areas of low seismicity or recent instrumentation.
Abstract
from arXiv · showhide
The recent evolution of induced seismicity in Central United States calls for exhaustive catalogs to improve seismic hazard assessment. Over the last decades, the volume of seismic data has increased exponentially, creating a need for efficient algorithms to reliably detect and locate earthquakes. Today's most elaborate methods scan through the plethora of continuous seismic records, searching for repeating seismic signals. In this work, we leverage the recent advances in artificial intelligence and present ConvNetQuake, a highly scalable convolutional neural network for earthquake detection and location from a single waveform. We apply our technique to study the induced seismicity in Oklahoma (USA). We detect 20 times more earthquakes than previously cataloged by the Oklahoma Geological Survey. Our algorithm is orders of magnitude faster than established methods.
Results
ConvNetQuake detects and locates earthquakes from single three-channel waveform windows using a convolutional network trained on labeled data. It achieves high detection and location accuracy, generalizes beyond repeating events, scales substantially faster than established methods, and reveals a key dependence on catalog history and discrete geographic classes.
- Method: ConvNetQuake predicts noise or one of six geographic earthquake clusters from a three-channel waveform window.The network is trained with augmented event examples recorded at two stations and Gaussian-noise perturbations.
- Detection: 100% detection accuracy was achieved on cataloged events in the independent test set.Detection accuracy treats the six earthquake labels as one event class and evaluates event-versus-noise classification.
- Detection: 99.9% accuracy was obtained on noise windows, with 116 false detections among 129,954 correctly classified noise windows.The test set contained 131,072 noise windows overall.
- Location: 74.5% location accuracy exceeded the 16.7% six-class chance baseline, while 50 clusters yielded 22.5% accuracy versus 2% chance.The larger-cluster result was attributed to only 40 training samples per class on average.
- Generalization: ConvNetQuake detects previously uncataloged and non-repeating events that template matching can miss, including all inserted events above 7 dB SNR.On one month of continuous Oklahoma data, it detected 4,225 previously uncataloged events, of which autocorrelation confirmed 3,949.
- Scalability: The method processed one week of continuous data approximately 13,500 times faster than autocorrelation and 48 times faster than FAST.It processed one month in 4 minutes 51 seconds, while FAST required 4 hours 20 minutes; its parameter storage remained constant at 500 kB.
- Limitations: The approach requires a pre-existing catalog of earthquakes for training and is limited here to local events and fixed discrete location categories.The authors identify low-seismicity or recently instrumented areas as poorly suited settings and suggest continuous probabilistic maps as future work.
Latitude
The study evaluates ConvNetQuake using synthetic seismic records, waveform classification, and comparisons with template-based detection methods. The experiments examine generalization to unseen events, detection coherence, and computational scaling.
- Synthetic-data evaluation: Synthetic day-long records insert scaled earthquake templates into Gaussian noise at SNRs from -1 to 8 dB.Training uses template T1, while testing uses both T1 and an unseen template T2.
- Detection performance: For SNRs higher than 7 dB, ConvNetQuake detects all inserted seismic events.The comparison uses synthetic data with ground-truth event locations and insertions.
- Generalization: ConvNetQuake detects most unseen inserted templates, whereas template matching misses almost all of them.Both methods detect events from the seen template T1, but only ConvNetQuake generalizes effectively to T2.
- Detection validation: Autocorrelation analysis of 4225 detected windows shows a nonzero correlation distribution, indicating that many detections contain waveform coherence.A threshold of 0.2 retains coherent waveforms visible on at least two of the three components, yielding 2129 waveforms in cluster 3.
Latitude
The supplementary figures visualize waveform correlations and cluster-specific waveform structure in the Oklahoma dataset. They compare correlation thresholds and show how waveform components vary across detected events.
- Correlation analysis: The cross-correlation coefficient distribution is plotted for windows classified as events by ConvNetQuake.The figure summarizes correlation structure among detected event windows after autocorrelation.
- Correlation-threshold analysis: 2271 Cluster 3 waveforms are shown at a correlation coefficient threshold of 0.1 across North, East, and vertical components.Waveforms are ordered by increasing absolute correlation with a reference event and aligned or flipped according to correlation sign.
- Correlation-threshold analysis: 2129 Cluster 3 waveforms are shown at the main-manuscript threshold of 0.2 across North, East, and vertical components.S-waves dominate both horizontal components, while vertical-component dominance differs between highly correlated and most other events.
- Correlation-threshold analysis: 845 Cluster 3 waveforms are shown at a correlation coefficient threshold of 0.3 across the three waveform components.The figure provides the most selective waveform set among the three displayed thresholds.