Source-linked AI summary
PhaseNet: A Deep-Neural-Network-Based Seismic Arrival Time Picking Method
Weiqiang Zhu, Gregory C. Beroza
TL;DR
Manual seismic phase picking is increasingly difficult as sensor networks grow, while existing automated methods remain vulnerable to noise and limited arrival-time accuracy. The paper presents PhaseNet, a deep-neural-network method for picking P and S arrivals, and reports substantially higher picking accuracy and recall than existing methods.
Problem
Growing seismic networks make comprehensive manual picking difficult, while traditional automated methods are susceptible to noise and low arrival-time accuracy.
Method
PhaseNet is a deep neural network algorithm for seismic phase picking that differentiates P and S waves beyond amplitude alone.
Results
Much higher picking accuracy and recall rate than existing methods.
Takeaways & Limitations
Improved S-wave picking is relevant because S-wave structure is important for strong ground motion prediction.
Takeaways & Limitations
Traditional automated picking remains susceptible to interference from multiple noise sources, constraining arrival-time accuracy.
Abstract
from arXiv · showhide
As the number of seismic sensors grows, it is becoming increasingly difficult for analysts to pick seismic phases manually and comprehensively, yet such efforts are fundamental to earthquake monitoring. Despite years of improvements in automatic phase picking, it is difficult to match the performance of experienced analysts. A more subtle issue is that different seismic analysts may pick phases differently, which can introduce bias into earthquake locations. We present a deep-neural-network-based arrival-time picking method called "PhaseNet" that picks the arrival times of both P and S waves. Deep neural networks have recently made rapid progress in feature learning, and with sufficient training, have achieved super-human performance in many applications. PhaseNet uses three-component seismic waveforms as input and generates probability distributions of P arrivals, S arrivals, and noise as output. We engineer PhaseNet such that peaks in probability provide accurate arrival times for both P and S waves, and have the potential to increase the number of S-wave observations dramatically over what is currently available. This will enable both improved locations and improved shear wave velocity models. PhaseNet is trained on the prodigious available data set provided by analyst-labeled P and S arrival times from the Northern California Earthquake Data Center. The dataset we use contains more than seven million waveform samples extracted from over thirty years of earthquake recordings. We demonstrate that PhaseNet achieves much higher picking accuracy and recall rate than existing methods.
SUMMARY
PhaseNet is trained on a large, long-term earthquake waveform dataset and achieves higher picking accuracy and recall than existing methods.
- More than seven million waveform samples were extracted from over thirty years of earthquake recordings.
- PhaseNet achieves much higher picking accuracy and recall rate than existing methods.
1 INTRODUCTION
Phase picking is increasingly difficult to perform manually and automatically as seismic networks grow, especially for noisy S waves. PhaseNet addresses these challenges with a deep neural network trained on analyst-labeled data to predict P and S arrivals.
- Motivation: Earthquake catalogs depend critically on both the number and accuracy of arrival-time measurements.
- Motivation: Growing sensor networks make comprehensive manual phase picking increasingly difficult, particularly across dense networks with thousands of sensors.
- Motivation: S arrivals are difficult to distinguish because they are later arrivals emerging from scattered P-wave coda.
- Limitations of existing methods: Traditional automatic methods can be noise-sensitive and have low arrival-time accuracy, particularly for shear waves.
- Limitations of existing methods: Complex waveforms and manually defined processing features limit traditional automated phase picking.
- PhaseNet: PhaseNet learns features from labeled data rather than using manually defined features, using analyst-picked P and S arrival times for training.
- PhaseNet: Unfiltered three-component waveforms are input to PhaseNet, which outputs P-wave, S-wave, and noise probability distributions.
- PhaseNet: Peaks in the P- and S-wave probability distributions correspond to predicted arrival times.
2 DATA
The study constructs a diverse Northern California waveform dataset from analyst-labeled P and S arrivals, including varied stations, instruments, and signal-to-noise conditions. The data are minimally preprocessed and arrival labels are represented probabilistically to account for uncertainty.
- Dataset construction: 779,514 three-component recordings with both P and S arrivals were retained from the Northern California Earthquake Data Center catalog.The data were split by station into 623,054 training, 77,866 validation, and 78,592 test samples.
- Dataset diversity: The dataset spans diverse waveform characteristics and includes recordings from 234,117 earthquakes and 889 seismic stations.The station and earthquake distributions are shown in Figure 1.
- Dataset diversity: Both high- and low-SNR recordings were selected, making automatic phase picking challenging while enabling more comprehensive evaluation.SNR is computed from the ratio of standard deviations in five-second windows following and preceding the P arrival.
- Preprocessing: Each waveform component was normalized by removing its mean and dividing by its standard deviation, and all data were sampled at 100 Hz.The 100-Hz sampling rate is the most common rate in the dataset.
- Label representation: P and S arrival times were converted into truncated Gaussian probability distributions with mean 0 s and standard deviation 0.1 s.Probabilistic labels represent arrival-time uncertainty and reduce the influence of labeling errors and biases.
- Dataset diversity: Large amounts of low-SNR data were deliberately included to improve arrival-time measurement for small events.
3 METHOD
PhaseNet is a modified U-net for three-component seismic time series that maps each time point to P-wave, S-wave, or noise probabilities. Arrival times are extracted from peaks in the predicted distributions, using softmax outputs and cross-entropy training.
- Architecture and outputs: The architecture modifies U-net for one-dimensional time series and uses four down-sampling and four up-sampling stages.Each stage applies convolution and rectified linear unit activation.
- Architecture and outputs: PhaseNet maps three-component seismograms of known earthquakes to probability distributions for P waves, S waves, and noise.
- Training objective: The final layer uses softmax probabilities for noise, P, and S categories, with z(x) denoting its unscaled values.
- Training objective: Training minimizes cross entropy between the true and predicted probability distributions, measuring their divergence.
- Architecture and outputs: Down-sampling extracts and compresses useful waveform information, while up-sampling converts it into per-time-point class probabilities.Skip connections concatenate shallower outputs with deeper layers.
- Arrival-time picking: P and S first arrivals are extracted from peaks in the corresponding output probability distributions.
4 EXPERIMENTS
The experiments evaluate PhaseNet’s picking accuracy, residuals, robustness across instruments, learned representations, and continuous-data detection. PhaseNet performs especially well for S arrivals, aligns predictions with analyst picks, and distinguishes P and S arrivals in continuous waveforms.
- Residual distributions for PhaseNet are narrower and less biased than those of the AR picker for both P and S arrivals.P-wave residuals are narrower than S-wave residuals, consistent with clearer P arrivals.
- PhaseNet remains robust across short-period, broad-band, high-gain, low-gain, accelerometer, and seismometer instruments without changing parameters or thresholds.The same model is trained on all instrument types and tested on instrument-specific subsets.
- Predicted probability peaks accurately align with true P and S picks, including ambiguous cases where abrupt waveform changes are absent.The model can still predict correct arrival times with high confidence in these ambiguous examples.
- Deepest-layer representations form distinct regions for P, S, and noise, indicating that the network extracts their characteristic features from raw waveforms.PCA is applied to condensed vectors produced from inputs containing P picks, S picks, or only noise.
- On continuous waveforms, PhaseNet produces detection spikes like STA/LTA while also distinguishing P from S arrivals, which may help reduce false detections.Events with both P and S picks are more likely to be true earthquakes than undifferentiated STA/LTA spikes.
5 DISCUSSION
PhaseNet improves P- and especially S-wave arrival picking over traditional methods, including noisy and clipped data, while reducing threshold-related timing bias. Its convolutional design supports waveform streams of arbitrary length and downstream earthquake-location and velocity-model applications.
- Picking performance: PhaseNet achieves F1 scores of 0.896 for P arrivals and 0.801 for S arrivals, compared with 0.558 for P arrivals using the AR picker.Evaluation used a strict true-positive criterion of Δt < 0.1 s, so relaxing the standard would yield higher F1 scores.
- Picking performance: PhaseNet’s improvement is greater for S waves than P waves, consistent with learned features outperforming manually defined features.STA/LTA relies on amplitude changes, while S arrivals are contaminated by P coda; PhaseNet learns additional features and distinguishes P, S, and noise.
- Robustness: PhaseNet picks S arrivals successfully in strongly clipped waveforms and accurately predicts P and S arrivals in low-SNR data.These conditions would make manual analyst picking difficult, yet the predictions remain accurate and high-probability.
- Thresholding: Unlike STA/LTA, PhaseNet’s probability threshold determines whether a pick is made, while the arrival time comes from the probability peak and is not strongly threshold-dependent.The authors use a 0.5 probability threshold for both P and S picks; tuning it can improve performance, but the effect is not significant.
- Applications: PhaseNet can process data of any length and generate running P- and S-arrival probabilities for earthquake detection when paired with an association algorithm.Accurate arrivals can also support absolute earthquake locations and seismic velocity models.
6 CONCLUSION
The paper introduces PhaseNet, trained on manually picked Northern California P and S arrivals, to learn waveform features and predict phase and noise probabilities. It improves existing methods particularly for S waves, although further testing against existing methods is still required before operational deployment.
- Conclusion: PhaseNet is a deep neural network trained on manually picked P and S arrivals from the Northern California Seismic Network catalog.It uses three-component waveform data to predict probability distributions for P waves, S waves, and noise, extracting arrival times from distribution peaks.
- Conclusion: Test results show significant improvement over existing methods, particularly for S-wave picking.PCA visualization indicates that condensed neural-network weights contain characteristics separating P waves, S waves, and noise.
- Conclusion: Further testing against existing methods is required before PhaseNet becomes operational.The authors state that the method is not far from operational capability.
- Implications: Reliable S-wave arrivals can substantially improve shear-wave velocity models and support strong-ground-motion path-effect prediction.More accurate P and S arrivals can also help extract information from growing waveform datasets for earthquake monitoring.
- Implications: PhaseNet can also be applied to other seismic phases when manually labeled training datasets are available.The authors identify large labeled datasets as an important ingredient for improving neural-network methods.