Earthquake event detection in seismic time series data is an important and challenging problem. The current state-of-the-art machine-learning based detection methods mostly belong to supervised learning category and the detection accuracy highly relies on the quality and quantity of the labeled data. How- ever, acquisition of high-quality training set can be technical challenging and expensive in that it requires intensive training on domain knowledge. Therefore, expanding the dataset with artificially generated labeled data can be extremely useful. In this paper, we develop an earthquake generator, EarthquakeGen, by using generative adversarial networks (GAN). GAN is a recently raised generative model based on neural net- work. By training in a min max game process, GAN is able to produce samples looks similar but actually different with the real ones. To verify the performance of our EarthquakeGen, we apply it to seismic field data acquired at Oklahoma, where induced earthquake events have been reported. Through our numerical results, we show that our EarthquakeGen can yield high-quality artificial earthquakes. More importantly, we show that the earthquake detection accuracy can be significantly improved by using augmented training sets combining both artificial and real samples.
Presentation Date: Wednesday, September 18, 2019
Session Start Time: 10:35 AM
Presentation Time: 10:35 AM
Location: Poster Station 13
Presentation Type: Poster
