Fully convolutional network for P-wave arrival detection and time picking using strong-motion data in China

Abstract

Over the past decade, the amount of available strong-motion records in China mainland has increased significantly, prompting the need for automatic processing to extract the vast amount of information from such data sets. Meanwhile, the need for seismic analysis and earthquake disaster estimation during quaking demands even more reliable P-wave detection and accurate P-wave arrival time picking. Here, we present a fully convolutional network (FCN) for on-set P-wave detection and time picking for China strong-motion Network. The FCN uses three-component seismic acceleration waveforms as input and generates probability distributions of P-wave arrival time as output. The FCN is trained and tested on the datasets (1708 events) selected from China Earthquake Network Center Catalog (CENC, 2017). The trained network is tested against the short-term average/long-term average united Akaike Information Criterion (STA/LTA-AIC) method, the FCN has an advantage over the STA/LTA method, achieves 99.3% probability in P-wave arrival detection and 98.3% probability in P-wave arrival time picking within the residual error of 0.1s.