Source-linked AI summary
A Machine-Learning Approach for Earthquake Magnitude Estimation
S. Mostafa Mousavi, Gregory C. Beroza
TL;DR
Rapid earthquake magnitude estimation is challenging because amplitude information is important for some magnitude scales, while early warning requires fast broadcasts. This study presents an end-to-end, single-station neural-network approach and finds that site-specific learning can improve performance, with potential applications in monitoring and early warning.
Problem
Earthquake magnitude estimation is challenging because amplitude information plays a key role for scales such as local magnitude, while warning systems require rapid estimates.
Method
The study presents a fast, reliable end-to-end method for estimating earthquake magnitudes from single-station waveforms, including waveform displacement after instrument-response correction.
Results
Site-specific learning can improve performance, whereas region-based training might be less important; borehole stations also produced better results than surface stations.
Takeaways & Limitations
The approach may support applications ranging from routine earthquake monitoring to early warning systems.
Takeaways & Limitations
The network requires more training data to prevent overfitting, and the better borehole-station results cannot be attributed to site-amplification effects alone.
Abstract
from arXiv · showhide
In this study we develop a single-station deep-learning approach for fast and reliable estimation of earthquake magnitude directly from raw waveforms. We design a regressor composed of convolutional and recurrent neural networks that is not sensitive to the data normalization, hence waveform amplitude information can be utilized during the training. Our network can predict earthquake magnitudes with an average error close to zero and standard deviation of ~0.2 based on single-station waveforms without instrument response correction. We test the network for both local and duration magnitude scales and show a station-based learning can be an effective approach for improving the performance. The proposed approach has a variety of potential applications from routine earthquake monitoring to early warning systems.
signals. Here, we introduce the first successful deep-learning approach to estimate directly
The study presents a deep-learning approach for estimating earthquake magnitude directly from raw seismic signals recorded at a single station. Its convolutional-recurrent network estimates both local and duration magnitudes.
- A deep-learning approach estimates earthquake magnitude from raw seismic signals recorded on a single station.
- The network combines convolutional and recurrent neural networks.
- It estimates both ML and Md directly from raw seismograms recorded on a single station.
2 Introduction
Earthquake magnitude is a fundamental parameter, but different empirical scales measure distinct wave properties and can diverge across earthquake sizes and source conditions. Standard estimation involves instrument-response correction, propagation correction, station-level estimates, and averaging, motivating rapid single-station estimation for warning applications.
- Earthquake magnitude is a fundamental parameter for characterizing earthquakes and represents source strength.
- Different magnitude scales measure different seismic-wave properties and suit different earthquake sizes and epicentral-distance ranges.
- Magnitude scales can differ by more than one magnitude unit for very large or small earthquakes and some source classes.
- Typical estimation corrects raw seismograms for instrument response and propagation effects, estimates magnitude at each station, then averages the results.
- Rapid, reliable preliminary magnitude estimates from immediately available data are important for earthquake and tsunami early-warning systems.
- The study presents a fast, reliable end-to-end method for estimating magnitude from raw single-station seismograms and reports it as the first successful study of its kind.
3 Method
The method addresses the challenge that un-normalized inputs hinder magnitude estimation even though amplitude information is important. It uses convolutional layers for dimensionality reduction and feature extraction followed by recurrent processing and regression.
- Un-normalized inputs make magnitude estimation challenging because waveform amplitude information is important, especially for local magnitude.
- The network mainly uses convolutional and recurrent layers, with convolutional layers serving dimensionality reduction and feature extraction without activation functions.
- LSTM units retain temporal dependencies among input elements and are useful for sequential earthquake signals.
- The model takes three-channel, 30-second seismograms containing 3000 samples as input.
- Two convolutional layers precede dropout and maxpooling, which reduce input dimensions by a factor of 4 before a bidirectional LSTM with 100 units.
- A one-neuron fully connected layer with linear activation estimates magnitude, using dropout rate 0.2 and mean square error minimization during training.
4 Results
The network estimates earthquake magnitude from single-station waveforms, including local and duration scales, with near-zero mean error and ~0.2 standard deviation. Results also examine regional and site-based learning, noise sensitivity, and agreement among station estimates.
- Overall performance: ~0.2 standard deviation and mean error close to zero characterize the network’s earthquake-magnitude predictions.The evaluation reports this overall regression result, while performance is worst near magnitude bounds with fewer samples.
- Magnitude scales: The network predicts both local and duration magnitudes from single-station waveforms.The results indicate that the model learns attenuation-related amplitude information and duration directly from the input waveform.
- Regional learning: A southern-California model shows only a small improvement in standard deviation despite a slight decrease in determination coefficient.The regional model uses a smaller training set, limiting the observed improvement.
- Site effects and noise: Surface-station performance deteriorates relative to borehole-station performance, but the difference cannot be attributed to site amplification alone.The passages identify higher surface-data noise as another factor with a direct impact on waveforms.
- Site-based learning: The station-based model obtains the best result, suggesting sites affect learning performance more strongly than regions.This model uses globally distributed stations with more than 1000 observations each.
- Cross-station consistency: Single-station predictions vary very little across stations for most of 311 events and generally remain within the ground-truth range.The comparison uses events with at least four available station observations and reports predicted station values, averages, and ground truth.
5 Conclusions
The study shows that neural networks can estimate earthquake magnitude directly from raw single-station waveforms across different magnitude types. Results also suggest site-specific learning can improve performance, although borehole-versus-surface differences cannot be attributed to site amplification alone because noise level also matters.
- Neural networks can learn general relations for estimating earthquake magnitudes directly from raw single-station waveforms.
- The approach supports different magnitude types when sufficient training data are available.
- Site-specific learning can be an effective strategy for improving performance, whereas region-based training may be less important.
- Borehole stations produced better results than surface stations, but site amplification alone cannot explain the difference.
- Noise level also shows a strong effect on the borehole-versus-surface comparison.
- The method provides fast earthquake-magnitude estimation from raw single-station seismograms, with potential applications in routine monitoring and early warning systems.