Source-linked AI summary
P-wave arrival picking and first-motion polarity determination with deep learning
Zachary E. Ross, Men-Andrin Meier, Egill Hauksson
TL;DR
Precisely measured P-wave arrivals and first-motion polarities are needed for earthquake hypocenters and focal mechanisms, but automated estimates have lagged human analysis. This paper trains two CNNs directly on seismograms to estimate both quantities. The networks produce results comparable to or better than human expert measurements, with high reported pick precision and polarity classification precision.
Problem
P-wave arrival and first-motion measurements are essential for hypocenters and focal mechanisms, while automated methods have been less accurate than human experts.
Method
The study trains two convolutional neural networks on expert-labeled Southern California seismograms, using the waveforms directly without feature extraction.
Results
95% precision was achieved for first-motion polarity classification, while the networks produced P-wave picks comparable to or better than human measurements.
Takeaways & Limitations
The trained networks can automate P-wave picking and first-motion classification at a quality level comparable to or better than human experts.
Takeaways & Limitations
CNN design choices were selected partly through trial and error, although 2–6-second waveform windows produced similar results.
Abstract
from arXiv · showhide
Determining earthquake hypocenters and focal mechanisms requires precisely measured P-wave arrival times and first-motion polarities. Automated algorithms for estimating these quantities have been less accurate than estimates by human experts, which is problematic for processing large data volumes. Here, we train convolutional neural networks to measure both quantities, which learn directly from seismograms without the need for feature extraction. The networks are trained on 18.2 million manually picked seismograms for the southern California region. Through cross-validation on 1.2 million independent seismograms, the differences between the automated and manual picks have a standard deviation of 0.023 seconds. The polarities determined by the classifier have a precision of 95% when compared with analyst-determined polarities. We show that the classifier picks more polarities overall than the analysts, without sacrificing quality, resulting in almost double the number of focal mechanisms. The remarkable precision of the trained networks indicates that they can perform as well, or better, than expert seismologists.
1 Introduction
Precise P-wave arrival times and first-motion polarities are needed for earthquake hypocenters and focal mechanisms, motivating automated methods that can handle growing seismic data volumes. The study applies CNNs directly to seismograms and reports results comparable to or better than human measurements.
- P-wave arrival times and first-motion polarities support determining earthquake hypocenters and focal mechanisms.
- Automated analysis is increasingly necessary as seismic networks grow and is required for real-time seismology.Automated procedures also provide consistency and repeatability across analyses.
- CNNs learn features directly from input seismograms rather than relying on separately extracted features.Their convolution and pooling layers distill relevant information before passing it to a fully connected network.
- The study develops two CNNs for automated P-wave picking and first-motion classification using millions of Southern California expert annotations.The networks were trained and validated on SCSN measurements from 2000–2017.
- CNN outputs are reported as extraordinarily precise and comparable to, or better than, human expert decisions.
2 Data
The study uses a large Southern California Seismic Network archive containing manually labeled earthquake seismograms, P-wave picks, and first-motion polarities.
- 273,882 earthquakes were recorded by the SCSN from 2000–2017 at 692 stations.
- 4,847,248 manually determined P-wave picks and 2,530,857 analyst-assigned first-motion polarities are associated with the seismograms.
- The dataset includes HHZ, HNZ, and EHZ channels, restricted to stations within 120 km epicentral distance.
3 Methods and Results
The study uses separate convolutional neural networks to pick P-wave arrivals and classify first-motion polarities directly from seismograms. Across validation data, the networks achieve precise arrival picks, high polarity precision, and improved focal-mechanism catalogs.
- Methods: Separate CNNs locate P-wave onsets with a regressor and classify first-motion signs with a second network.The arrival-time network estimates onset time, while the polarity network assigns up, down, or unknown labels.
- Methods: The waveform inputs are preprocessed by down-sampling to 100 Hz, detrending, and applying a causal 1–20 Hz Butterworth filter.The models use the vertical sensor component and 4 s windows containing 400 waveform samples.
- P-wave picking: 6.1 million validation samples yield a 0.023 s standard deviation in arrival-pick differences, with 75% of picks within 0.028 s of analyst picks.The mean difference is -3x10^-3 s, and the 90th percentile absolute difference is 0.074 s.
- P-wave picking: Pick error increases rapidly for SNR < 5, while the network still mimics analyst behavior with high precision under low-SNR conditions.Error also increases weakly with epicentral distance.
- First-motion polarity classification: Up and down polarity classifications achieve average precision above 95%, while recall is 0.80 for up and 0.81 for down.The network often assigns unknown when uncertain, preserving precision at the expense of recall.
- Focal mechanism comparisons: CNN-derived polarities produce 6003 quality A–D focal mechanisms versus 4613 from manual polarities, with fewer high-misfit events and nearly twice as many events below 10% misfit.The number of A, B, C, and D mechanisms increases by 84%, 90%, 51%, and 18%, respectively.
4 Discussion
The paper treats phase measurement as direct seismogram recognition rather than feature-engineered triggering, using CNNs trained on extensive expert-labeled archives. The discussion emphasizes broad measurement improvements while acknowledging design choices and noisy onsets as important constraints.
- Scientific relevance: Improved phase measurements can propagate into earthquake locations, magnitudes, and source-property estimates.
- Methodological perspective: CNNs learn P-wave onset characteristics directly from minimally pre-processed seismograms, without extracting hand-designed features.This frames picking as an image-recognition problem analogous to human waveform interpretation.
- Design choices: The 4-second waveform window tolerates velocity-model uncertainty, while windows from 2 to 6 seconds produce similar results.
- Methodological perspective: CNN decision boundaries are systematically optimized from seismograms, whereas traditional pickers commonly rely on difficult-to-tune hard thresholds.The paper presents this difference as a reason CNN approaches can outperform traditional methods.
- Caveats: Noisy waveform onsets can be intrinsically ambiguous, limiting the accuracy of both human labels and automated measurements.The true onset may lie beneath the noise level, making precise ground truth unavailable for some seismograms.
- Scientific relevance: Training on nearly two decades of expert-labeled records allows the networks to encode collective SCSN analyst experience for future automated picks.The discussion states that this knowledge base can expand as more data become available.