[Paper Review] SeismoGen: Seismic Waveform Synthesis Using Generative Adversarial Networks

Tiantong Wang, Daniel T. Trugman+1 more|arXiv (Cornell University)|Nov 10, 2019

Seismology and Earthquake Studies56 references5 citations

TL;DR

This paper proposes SeismoGen, a conditional generative adversarial network (GAN) that synthesizes realistic 3-component seismic waveforms for both earthquake and noise classes. Trained on limited real data from Oklahoma, the GAN-generated waveforms significantly improve earthquake detection accuracy when used to augment small real training sets, demonstrating the potential of synthetic data for data-scarce seismological applications.

ABSTRACT

Detecting earthquake events from seismic time series has proved itself a challenging task. Manual detection can be expensive and tedious due to the intensive labor and large scale data set. In recent years, automatic detection methods based on machine learning have been developed to improve accuracy and efficiency. However, the accuracy of those methods relies on a sufficient amount of high-quality training data, which itself can be expensive to obtain due to the requirement of domain knowledge and subject matter expertise. This paper is to resolve this dilemma by answering two questions: (1) provided with a limited number of reliable labels, can we use them to generate more synthetic labels; (2) Can we use those synthetic labels to improve the detectability? Among all the existing generative models, the generative adversarial network (GAN) shows its supreme capability in generating high-quality synthetic samples in multiple domains. We designed our model based on GAN. In particular, we studied several different network structures. By comparing the generated results, our GAN-based generative model yields the highest quality. We further combine the dataset with synthetic samples generated by our generative model and show that the detectability of our earthquake classification model is significantly improved than the one trained without augmenting the training set.

Motivation & Objective

• To address the challenge of limited labeled seismic data for training machine learning models in earthquake detection.
• To develop a generative model capable of producing realistic synthetic seismic waveforms that mimic both event and noise classes.
• To evaluate whether GAN-generated synthetic waveforms can improve the performance of earthquake detection algorithms when used to augment real training data.
• To validate the realism and utility of synthetic waveforms through qualitative and quantitative benchmarks using a trained classifier.

Proposed method

• A conditional GAN architecture is employed, with a generator network mapping random noise to 3-component seismic waveforms conditioned on class labels (event or noise).
• The discriminator network is trained to distinguish between real seismic waveforms and those generated by the generator, using adversarial training to improve realism.
• The model is trained on a limited real dataset from two seismic stations (V34A and V35A) in Oklahoma, focusing on waveform patterns specific to that region.
• Synthetic waveforms are generated in both event and noise classes and combined with real data to form augmented training sets for downstream detection models.
• Performance is evaluated using a machine learning-based earthquake classifier, comparing detection accuracy on real-only vs. real+synthetic training sets.
• Both visual inspection and quantitative metrics (e.g., F1-score, AUC) are used to validate the realism and utility of the synthetic waveforms.

Experimental results

Research questions

• RQ1 Can a GAN model generate high-fidelity synthetic seismic waveforms that are indistinguishable from real waveforms in both event and noise classes?
• RQ2 Can synthetic waveforms generated from a limited real dataset improve the performance of earthquake detection models when used for data augmentation?
• RQ3 How does the performance of detection models trained on real-only data compare to those trained on real-plus-synthetic data in low-data regimes?
• RQ4 To what extent do the synthetic waveforms preserve the temporal and channel-wise correlations present in real seismic data?

Key findings

• The GAN-generated waveforms are visually and quantitatively indistinguishable from real waveforms in both event and noise classes, as confirmed by a trained classifier.
• Detection models trained on augmented datasets (real + synthetic) achieve significantly higher F1-scores and AUC values compared to models trained on real data alone, especially when real data is scarce.
• The model demonstrates robustness in generating realistic waveforms even with only two seismic stations’ data, suggesting strong generalization within the local geologic context.
• The synthetic waveforms preserve realistic channel-to-channel correlations and temporal patterns, indicating fidelity in waveform structure.
• The approach effectively mitigates data scarcity issues, enabling high-performance detection with minimal labeled training data.