固物研究生论坛 - focal depth by machine learning

Focal depth is one of the parameters difficult to obtain accurately in seismology all the time. And CNN (Convolutional Neural Network) has strong feature extraction and nonlinear mapping, which makes it possible to obtain focal depth range by CNN. In this study, earthquake depth is classified into shallow earthquake, intermediate earthquake and deep earthquake as the output classification. It can be acknowledged that seismic phase are sensitive to focal depth rather than epicentral distance by generating and comparing seismic data. For forward modeling data, waveform information of single and multiple stations is the input of CNN respectively. It turns out that the recognition effect of multiple stations is better than that of single station. The optimal combination of multiple station count range is 6~20 by testing different ones. At the same time, focal depth recognition accuracy can reach around 93% by optimizing CNN parameters. The seismic data in South America from 1990 to 2018 has been chosen for practically testing. On the basis of training parameters of the forward modeling data, the test accuracy of single station CNN training is about 85%, while as for multiple station training, it can reach around 99%. In addition, the time window of seismic data is within one minute after the first break with the stable performance.