Source-linked AI summary
CRED: A Deep Residual Network of Convolutional and Recurrent Units for Earthquake Signal Detection
S. Mostafa Mousavi, Weiqiang Zhu, Yixiao Sheng, Gregory C. Beroza
TL;DR
Large and growing seismic datasets require earthquake detectors that are sensitive, robust, and efficient. The paper introduces CRED, a residual network combining convolutional and recurrent units to learn time-frequency characteristics from three-component seismograms. CRED achieves high test performance, generalizes to Central Arkansas, and processes continuous data efficiently, while its training data have limited parameter diversity.
Problem
Growing seismic data volumes require efficient earthquake detection that remains sensitive to small events and robust to noise, non-earthquake signals, and varied waveforms.
Method
CRED combines convolutional and bidirectional LSTM units in a residual-learning architecture for sequence-to-sequence detection from three-component seismogram spectrograms.
Results
CRED generalized from Northern California training data to Central Arkansas, detecting three orders of magnitude more events than STA/LTA, while using no templates and achieving an F-score of 99.95.
Takeaways & Limitations
CRED provides fast, scalable event detection with reported robustness to background-noise levels and applicability to real-time monitoring and large seismic datasets.
Takeaways & Limitations
Training used only 250k earthquake waveforms with limited parameter diversity, so broader training data could improve performance.
Abstract
from arXiv · showhide
Earthquake signal detection is at the core of observational seismology. A good detection algorithm should be sensitive to small and weak events with a variety of waveform shapes, robust to background noise and non-earthquake signals, and efficient for processing large data volumes. Here, we introduce the Cnn-Rnn Earthquake Detector (CRED), a detector based on deep neural networks. The network uses a combination of convolutional layers and bi-directional long-short-term memory units in a residual structure. It learns the time-frequency characteristics of the dominant phases in an earthquake signal from three component data recorded on a single station. We train the network using 500,000 seismograms (250k associated with tectonic earthquakes and 250k identified as noise) recorded in Northern California and tested it with an F-score of 99.95. The robustness of the trained model with respect to the noise level and non-earthquake signals is shown by applying it to a set of semi-synthetic signals. The model is applied to one month of continuous data recorded at Central Arkansas to demonstrate its efficiency, generalization, and sensitivity. Our model is able to detect more than 700 microearthquakes as small as -1.3 ML induced during hydraulic fracturing far away than the training region. The performance of the model is compared with STA/LTA, template matching, and FAST algorithms. Our results indicate an efficient and reliable performance of CRED. This framework holds great promise in lowering the detection threshold while minimizing false positive detection rates.
1. Introduction
Seismological data volumes are growing rapidly, creating a need for efficient processing tools. Earthquake detection remains difficult because existing methods can be sensitive to noise, miss small events, produce false positives, or require prior templates.
- More than 50 terabytes of seismic data are archived annually at IRIS, motivating efficient extraction of useful information from large datasets.Machine learning approaches may help address this processing challenge.
- Continuous-data processing for earthquake-signal detection is an early stage in extracting information from large seismological datasets.
- STA/LTA is generalized and efficient but can be sensitive to time-varying background noise, miss small events, generate false positives, and struggle with closely spaced events.
- Template matching improves detection thresholds through cross-correlation but requires prior knowledge of templates.
- Deep CNN studies have reported promising robust and efficient earthquake-signal detection, while earlier neural networks used sparse feature vectors transformed through fully connected layers.
- CRED formulates detection as sequence-to-sequence learning, using bidirectional LSTM units and residual learning to model temporal relations and enable deeper learning.
2. Methodology
The methodology combines recurrent, convolutional, and residual-learning components to model seismic sequences and train a deep end-to-end detector. LSTM memory and bidirectional processing capture temporal patterns, while residual shortcuts support deeper architectures.
- Sequential Learning: RNNs retain relations among sequential inputs through internal states, enabling temporal modeling across seismogram positions.
- LSTM Networks: LSTM units use memory cells and gates to reduce vanishing or exploding gradients and control which states are retained.Forget, update, and output gates regulate memory and state transitions.
- Bidirectional LSTM: Bidirectional LSTM layers combine forward and backward processing to provide information from both directions at each time step.
- Residual Learning: Residual blocks learn F(x) instead of the original mapping H(x), using shortcut addition without extra parameters or computational complexity.The residual formulation is y = F(x, {Wi}) + x.
- Network Architecture: CRED uses spectrograms from three-component seismograms and combines convolutional, recurrent, and fully connected layers in a residual structure.The architecture includes convolutional feature extraction, bidirectional recurrent processing, and a final unidirectional LSTM for time-increasing acquisition data.
- Network Architecture: The proposed network has 12 layers and 256,000 trainable parameters in an end-to-end structure designed for precise seismic characterization.
3. Data
The study builds its dataset from Northern California three-component seismograms, balancing earthquake and noise examples. Signals are cleaned, standardized, transformed into spectrograms, and paired with binary labels covering dominant P- and S-wave energy.
- 550,000 30-second three-component seismograms from 889 Northern California stations support network training and validation.Half represent earthquakes from January 1987 to December 2017, while the other half contain ambient and non-ambient seismic noise.
- Noise traces were screened to remove uncataloged events and residual earthquake spectral features.A detection algorithm and de-signaling procedure reduced contamination by small events hidden beneath background noise.
- All traces were detrended, band-pass filtered from 1 to 45 Hz, resampled at 100 Hz, and normalized.Short Time Fourier Transform generated spectrograms for the model inputs.
- Binary label vectors mark samples from the P arrival through P + 3d, where d = S-P, as earthquake-signal intervals.This interval captures dominant P- and shear-wave spectral energy in the time-frequency domain.
- Figure 7 contrasts vertical-component seismograms, label vectors, and spectrograms for two noises and two earthquake samples.
4. Training and Testing of the Model
The model is trained and evaluated on separate dataset partitions using pointwise detection probabilities and precision, recall, and F-score metrics. It achieves high accuracy and strong detection performance across threshold choices, while test examples show broad signal generalization.
- Training: The dataset was split into training, evaluation, and test sets of 80%, 10%, and 10%, respectively.Training used binary cross-entropy, ADAM optimization, and 62 epochs on one Tesla V100 GPU.
- Training: 99.33% training accuracy and 99.24% validation accuracy were accompanied by losses of 0.02 and 0.03.Mean absolute errors were 0.01 for training and 0.009 for validation, based on point-by-point comparisons with ground truth.
- Testing: Detection performance was evaluated on 50,000 test samples using precision, recall, and F-score.F-score combines precision and recall to reduce effects from class imbalance.
- Testing: Precision exceeded 96% and recall exceeded 99% across thresholds, with a maximum F-score of 99.95% between thresholds 0.1 and 0.3.The confusion matrix was reported at threshold 0.11.
- Testing: Figure 8 displays the precision-recall curve and F-score as functions of threshold values.
- Generalization: Figure 9 shows high-precision predictions across large and small, local and teleseismic, noisy, damaged-instrument, and non-earthquake examples.Samplewise predictions allow detection without requiring the full signal length, supporting real-time processing.
5. Sensitivity Test
A semi-synthetic experiment tests CRED under varying Gaussian noise and compares it with STA/LTA and template matching. CRED detects more low-SNR events while maintaining zero false positives in the reported conditions.
- Experimental design: The experiment assembled 8.4 hours of continuous semi-synthetic data from 500 high-SNR earthquakes, 500 Ricker-wavelet non-earthquake signals, and 23 Gaussian-noise levels.
- Experimental design: CRED, STA/LTA, and template matching were applied with detection thresholds tuned to maximize precision.Two templates were used for template matching.
- Results: 100% of events with SNR >= 12 dB were detected by CRED with 0 false positives for all SNRs.
- Results: At 7 dB SNR, CRED detected 80% of events, compared with 27% for STA/LTA and 3% for template matching.
- Results: CRED showed superior performance to STA/LTA and template matching in the semi-synthetic comparison.The authors associate its noisy-condition performance with reliance on signal spectral contents rather than waveform shape.
6. Application to Central U.S.
CRED was applied to one month of continuous Arkansas data to test efficient detection and generalization beyond its training region. It identified cataloged and previously unrecognized events across a broad magnitude range, while lowering the threshold increased false positives.
- Continuous-data application: 1 hour and 9 minutes of processing detected 1102 events in one month of continuous data from station WHAR.The processing included three-channel data conversion into STFT and model application on a laptop.
- Detection outcomes: 680 detections matched the existing catalog, 77 were verified as new events, and 345 were false positives, yielding ~69 % precision.The remaining detections were visually checked before classification.
- Magnitude range: CRED detected events from magnitude -1.2 to 2.6, extending into low-magnitude signals.The magnitude-frequency distribution is shown for detections from different algorithms.
- Threshold trade-off: Lowering the detection threshold can recover more cataloged events but also produces more false positives.
- Waveform examples: Almost all events larger than ML 1.0 were detected, alongside lower-magnitude events spanning varied SNRs and waveforms.Newly detected events were estimated at 0.1 < ML < 0.4, but their magnitudes were not directly calculated.
7. Discussion
The discussion presents CRED as a residual convolutional-recurrent detector that generalizes from Northern California training data to substantially different Arkansas seismicity. Its efficiency, sensitivity, and scalability support applications from real-time monitoring to archived-data reprocessing, while training-data coverage remains a limitation.
- Architecture: CRED combines convolutional and recurrent units to learn sparse time-frequency features and sequential characteristics of earthquake signals.The architecture uses LSTM recurrent units within a deep residual-learning framework.
- Generalization: The model generalized from Northern California tectonic events to much smaller, shallower Central Arkansas microseismic events in a different crustal setting.The Arkansas test involved substantially different crustal structure, lower magnitudes, shorter distances, and shallower depths.
- Comparative performance: CRED detected three orders of magnitude more events than STA/LTA, ran more than 100 times faster than non-parallel FAST, and used no templates during training.Its detection rate was ~30 % of template matching and FAST, while newly detected events exposed limitations of similarity-search methods.
- Limitations: Training used only 250k earthquake waveforms with limited parameter distributions, so broader and more variable training data could improve performance.The authors also suggest recursive retraining, more reliable labels, and additional non-earthquake signals.
- Applications: The framework is described as suitable for real-time monitoring, microseismic monitoring, earthquake early warning, and scalable processing across many sensors or long durations.Archived-data reprocessing could support new insights into earthquake phenomena.
8. Conclusion
The conclusion presents CRED as a fast, precise, and generalizable detector combining residual convolutional and recurrent learning of seismic time-frequency structure. It reports real-time applicability, scalable architecture, and low false-positive sensitivity attributed to high-resolution spectral modeling.
- Contribution: CRED combines convolutional and recurrent units for deep residual learning of earthquake-signal time-frequency characteristics.
- Performance: The learned model is reported to have low sensitivity to background noise, generalize across regions, and outperform STA/LTA in sensitivity and template matching in efficiency.
- Operational use: The method is fast enough for application to seismic streams in real time after training, and its architecture can scale easily.
- False positives: False positive rates are described as minimal because earthquake signals are modeled at high resolution through their spectral structure.