Deep learning for seismic event detection of earthquake aftershocks

We design a system based on a convolutional neural network (CNN) for event detection and event time picking, and apply it to continuous recordings during aftershocks of the 2008 Mw7.9 Wenchuan Earthquake. The system detects 97% of the labeled seismic phases in a standard catalog. Off-line training of the CNN remains successful at 95% accuracy after reducing the training set size by a factor of ten to a few thousand training samples. When applied to a continuous aftershock dataset, the CNN-based method achieves accurate detection of manually labeled phases, precise arrival times of picked phases, as well as discovering many weak events not listed in the given seismic catalog. On-line processing of a 31-day recording takes less than 12 hours on a single GPU.