Comparison of Single-Trace and Multiple-Trace Polarity Determination for Surface Microseismic Data Using Deep Learning

For surface microseismic monitoring, determination of the P-wave first-motion polarity is important because (1) it has been widely used to determine focal mechanisms and (2) the location accuracy of the diffraction-stack-based method is improved greatly using polarization correction. The convolutional neural network (CNN) is a form of deep learning algorithm that can be applied to predict the polarity of a seismogram automatically. However, the existing network designed for polarity detection utilizes only individual trace information. In this study, we design a multitrace-based CNN (MT-CNN) architecture using several neighbor traces combined as training samples, which could utilize the polarity information of neighbor sensors in the surface microseismic array. We use 17,227 field seismograms with labeled polarities to train two different neural networks that predict the polarities by a single trace or by multiple traces. The performance of the test set and field example of two CNN architectures shows that the MTCNN significantly produces fewer polarity prediction errors and leads to more accurate focal mechanism solutions for microseismic events.