Highlight

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 the accuracy and efficiency. However, accuracy of those methods rely on 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 limited number of reliable labels, can we use them to generate more synthetic labels; (2) Can we use those synthetic lables to improve the detectability? Among all the existing generative models, 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 develop a generator with three-pipeline structure and a discriminator using features from both high and low frequency components of the seismic time series.

The full paper can be found on arxiv:

EarthquakeGen: Earthquake Simulation Using Generative Adversarial Networks

What is EarthquakeGen?

EarthquakeGen consists of two components: Generator and Discriminator, which are showing below.

Generator: The structure of our generator reflects the characteristics of seismic time series. Particularly, through previous sections we learned that real seismic samples contain 3 channels of 1D time series length of 1,600. Correspondingly, we design our generator to comprise of three pipelines to synthesize each channel of the data individually. All three pipelines share an identical network structure as shown below. Each pipeline is a four-layer convolutional network. There is no interaction among the three pipelines except that they share the same input.

Discriminator: The discriminator will be used to evaluate the quality of input samples (real or synthetic). The discriminator first learns features representative to seismic signals including both earthquake and non-earthquake events and further provides critics based on features learned. According to the domain knowledge, raw seismic time series consists of earthquake events which reside in high-frequency band and non-earthquake events which exist in the low-frequency band. Therefore, we design our discriminator to include two sequential modules, namely, ``feature extraction'' and ``sample critic''. The module of feature extraction learns high-quality feature vector, which will be passed onto the module of sample critic that evaluates the quality of the samples.

The following are synthetic earthquake and non-earthquake events generated by our EarthquakeGen.