Deep Learning for Seismic Template Recognition

Detecting earthquakes is one of the most fundamental tasks in seismology. As continuous seismic recordings grow and become readily available, they offer opportunities for identifying weak seismic signals (e.g., microseismicity, nonvolcanic tremor). Mining this source of data recordings is now considered a substantial method to study various geophysical phenomena. Widely used methods, such as the template matching algorithm (TMA), use waveforms of confirmed seismic signal to slide through corresponding continuous recordings searching similarities. These methods present several weaknesses, including sensitivity to the choice of thresholds, signal-to-noise ratios (SNRs), and bias. In this work, we present a solution for seismic signal recognition using emerging deep learning techniques. We have designed a Convolutional Neural Network (CNN) that recognizes seismic signals using features extracted by multiple filter kernels. This software is designed for training with data from multiple stations and the SNRs of template waveforms. We demonstrate the accuracy of our approach using seismic waveform recordings from multiple stations during the 2016 ML 6.6 Meinong earthquake sequence in Taiwan.