A Semisupervised Seismic Events Classifier Based on Generative Adversarial Network

Abstract This article presents a semisupervised generative deep learning seismic event classifier that uses a generative adversarial network to extract the features of unlabeled seismic events and subsequently creates a supervised seismic event classification through a synchronous weight sharing mechanism, thereby achieving a high-accuracy seismic event classification with only a small amount of labeled seismic data. In this study, we used records from 59 broadband seismic stations in the Henan region in the China Seismic Network and conducted independent testing using the data not involved in training. The results show that the classifier can classify seismic events related to natural earthquakes, blasting, and collapse using 1-min seismic waveforms. We compared the latent space dimensions, different numbers of labels, and generalization capabilities of different methods and in different regions of the classifier. The test results demonstrate the robustness and reliability of our developed method. The results show that our method can achieve a classification performance of >92.5% compared with traditional deep learning methods when using only 30% of the labeled samples from the entire dataset. The results of this study indicate that generative artificial intelligence methods can effectively extract the features of unlabeled seismic waveforms, and it has broad potential applications in seismology.