Application of a Deep Learning Phase Picker to Improve the Performance of Deep Borehole Seismic Data Analysis

Deep learning has demonstrated significant advancements in various fields that previously required human intervention.In particular, artificial intelligence (AI) has shown superior efficiency in areas where it is difficult to define clear rules or theories, outperforming traditional methods.However, these technologies have the drawback of requiring huge amounts of training data.The field of seismology is particularly well-suited for the development and application of deep neural networks due to the vast amount of digital seismic data accumulated alongside the analysis of skilled professionals.In this study, a deep learning method was applied to seismograms recorded at deep borehole seismometers installed for microseismic monitoring, and potential future applications of AI technology were explored.The deep borehole seismometers are installed at depths of 500 m and 1000 m, recording ground motions at a sampling rate of 1,000 sps.Since a deep neural network-based seismic phase picker has been trained on seismic data sampled at 100 sps, the deep borehole seismic data were resampled for phase picking.While 94 % of P-waves and 93 % of S-waves picked by visual inspection could be detected, the technique also detected many more phases that were not identified in the process of visual inspection, and approximately 131 % more earthquakes were detected compared to manual analysis.This indicates that while the existing deep learning-based seismic phase picker performs very well, there is still room for improvement.Furthermore, if deep neural networks trained using deep borehole seismic data to eliminate the preprocessing steps required for resampling, it is expected that earthquakes could be analyzed more rapidly and precisely.