Research on seismic wave first arrival Picking Based on improved U-Net

For the current problem of low accuracy of seismic wave phase pickup, this paper proposes a method based on the improved U-Net seismic wave P-wave and S-wave first-to-arrival pickup. Firstly, in order to reduce the impact of pre-processing operations such as filtering, the 20,000 three-component seismic waveform data will be normalized as input. Secondly, in order to reduce the loss of information during the seismic waveform extraction and increase the amount of information of effective features at the starting point, the Residual Unit is added to the encoder part of U-Net only, instead of adding it to the decoder part. In order to reduce the loss of high frequency information of seismic waveform, convolution operation is used instead of pooling operation. Finally, the probability distributions of P-wave, S-wave and noise are used as the output. The results show that the accuracy of the first-to-arrival pickup of P-wave is 93% and the accuracy of the first-to-arrival pickup of S-wave is 90% of the test set. The improved U-Net increased 30.9% and 6.2% in P-wave accuracy, 7.6% and 2.8% in S-wave accuracy, 39.6% and 4.0% in the P-wave recall, and 41.5% and 4.7% in the S-wave recall, respectively, compared with the conventional STA/LTA and U-Net. The improved U-Net outperforms the other two models in both evaluation metrics of recall and accuracy. The research results of this paper are of great practical significance in the localization of earthquakes, the interpretation of the earthquake gestation mechanism, and the prevention and mitigation of earthquakes. (paper)