Abstract:Rapid increase of seismic data poses challenges to seismic data processing, in which phase-picking is a crucial procedure and accordingly attracts attention from seismologists. Among the breakthroughs in deep learning methods for seismic data processing, seismologists have especially proposed many seismic phase-picking methods based on deep learning, which have shown much better accuracy than traditional automatic phase-picking methods. However, further improvements are needed to meet the requirements for processing continuous seismic records. Here we propose a new approach combining seismic array scheme and U-Net, an end-to-end neural network developed in image segmentation. Previous studies have shown that U-Net can classify waveforms and determine the onset of seismic phases accurately with the help of contracting and expanding paths. Based on U-Net, we first design a deep learning phase picker (PP) model with a long (40 s) time window. To avoid difficulties in time window selection and interference between P- and S-phase identifications, the model involves two networks with the same structure, which pick the P and S phases separately. To improve model training, we employ data transformation to expand the sample set, add typical noises to the sample set for better resemblance of real data, and slightly increase the weight of the seismic signal containing seismic records for a better balance between precision and recall. We also adopt a more effective array-assisted phase picker (APP) model which correlates the picking results from different stations to minimize misidentification. Using data from Hi-net and Alibaba Aftershock AI Capture Contest, we design training and test sets of different sizes to examine the performance and generalization ability of the model. Results show that the APP can effectively pick the body-wave phases on continuous records with a good generalization ability. The relationship between the model performance and training set is such that the model performance is directly related to the data quality rather than the size of the training set, and there is no serious overfitting even for a small training set. Statistics further show that the model has no special requirements for the locations of time windows and can handle seismic records with large epicentral distances. By comparison with other models (model with short time windows or the one of simultaneous picking of P- and S-phases), we find that APP can successfully process event-crowded waveform data and effectively reduce the influence of noise. Finally, we apply the trained model to a seismic network in the Inner Mongolia. Results show that the number of earthquakes detected by the APP is 30 times the manually-detected one, and that 98.1% of the events in the catalog are detected by the APP. The successful application to a seismic network different from the training set suggests that the APP is robust and has strong generalization ability.

3-D Seismic First Break Picking Based on Two-Channel Mask Strategy

Seismic First Break Picking in a Higher Dimension Using Deep Graph Learning

DSU-Net: Dynamic Snake U-Net for 2-D Seismic First Break Picking

Active Source Seismic Identification and Automatic Picking of the P ̄wave First Arrival Using a Convolutional Neural Network 1

MSSPN: Automatic First Arrival Picking using Multi-Stage Segmentation Picking Network

UPNet: Uncertainty-based Picking Deep Learning Network for Robust First Break Picking

A fine-tuning workflow for automatic first-break picking with deep learning

Microseismic First-Arrival Picking Using Fine-Tuning Feature Pyramid Networks

An Array-Assisted Deep Learning Approach to Seismic Phase-Picking

Improved U-Net3+ Network for First Arrival Picking of Noisy Earthquake Recordings

Variable seismic waveforms representation: Weak-supervised learning based seismic horizon picking

Deep-Learning-Based Seismic-Signal P-Wave First-Arrival Picking Detection Using Spectrogram Images

3-D Seismic Fault Detection Using Recurrent Convolutional Neural Networks With Compound Loss

Ground-truth-free Deep Learning for 3-D Seismic Denoising and Reconstruction with Channel Attention Mechanism

Improving the accuracy of predicted first-arrival times using deep learning

MFFnet: A Seismic Phase Picking Network Based on Multiple Feature Fusion

Automatic Velocity Picking Using a Multi-Information Fusion Deep Semantic Segmentation Network

Phase arrival picking for bridging multi-source downhole microseismic data using deep transfer learning

Identification of Seismic Wave First Arrivals from Earthquake Records via Deep Learning.

MSNet: A Seismic Phase Picking Network Applicable to Microseismic Monitoring

CNN-Based Ringing Effect Attenuation of Vibroseis Data for First-Break Picking