An artificial neural network approach for broadband seismic phase picking

Abstract

Abstract This article presents a method for picking broadband seismic phases by using backpropagation neural networks (BPNNs) as detectors. By combining the results from three BPNN detectors—long term, mid-term, and short term—the method combines the features of short term's higher accuracy and long term's lower false alarm rate. We demonstrate that proper pre- and postprocessing of the data can help to improve the system's performance. The determination of the architecture and parameters for BPNNs is also discussed in this article. The devised BPNN detector is applied to 1254 broadband seismograms of the IRIS network to determine the first arrival, which is expected to be used in tomographic studies of the mantle structure. The results show that the first arrival can be identified for more than 95% of the 1254 seismograms. The automatically picked travel times have a reasonable accuracy; more than 85% have an error of less than 1 sec, and about 80% have an error of less than 0.5 sec.