Automated picking of seismic first arrivals using a single- to multidomain self-trained network

Abstract

Over the past few decades, seismic data volume has increased owing to the evolution of survey acquisition designs that comprise numerous receivers and shotpoints. Consequently, the demand for efficient algorithms to detect seismic first arrivals has been increasing. Machine-learning methods, particularly convolutional neural networks, have shown great potential in achieving this. However, most existing methods lack generalization abilities or are too sensitive to noise. We develop a new picking strategy based on a single- to multidomain self-trained network to achieve automated picking of first arrivals. Our self-trained network learns from single to multiple domains while rejecting the incorrectly predicted picks and assigning pseudo-first-arrival times to the unlabeled data set. The model inherits the advantages of single- and multidomain picking strategies and has high generalizability and noise robustness. In experiments on an open hard-rock seismic multisurvey data set, our picking method outperforms the benchmark for most evaluation metrics. Remarkably, our approach achieves an accuracy improvement of approximately 10% against the data from one mine site (Lalor), with a sampling rate that differed from the training data. In addition, our method indicates robust results considering the limited amount of labeled data. The experimental results demonstrate the high generalizability and robustness of our method for unseen data sets and reveals the effectiveness of single- to multidomain self-trained networks to improve the accuracy and stability of automated first-arrival picking. Our method contributes to reducing human labor and the total seismic processing time.