Deep Learning Enhanced Seismic Catalog for Oklahoma

The significant increase in seismicity in the Oklahoma region due to increased wastewater injection has been well-documented and extensively investigated. Most of these investigations used seismic catalogs processed with conventional techniques that can be improved. Seismic events with small magnitude and seismograms with low signal to noise ratio are challenging to traditional techniques. Previous efforts using a convolutional neural network (CNN) have dramatically increased the number of detected seismic events in the Oklahoma region and identified time windows for these seismic events by screening through continuous seismic data. We have also successfully applied transfer learning to the PhaseNet model for a much smaller study area with a different geological setting. A transfer-learning aided double-difference tomography (TADT) workflow was designed to integrate deep learning techniques with the advanced double-difference imaging algorithm. The workflow led to better seismic event locations and increased the resolvable volume for the seismic velocity model. We will leverage the CNN detected seismic events and associated time windows and apply the TADT workflow to seismic records from Oklahoma to enhance the seismic catalog for the Oklahoma region. We will refine the PhaseNet model for seismograms from Oklahoma stations using a small dataset. The refined deep learning model will be applied to all available seismograms for the CNN detected events to measure P- and S-wave phase picks. Additional relative travel times will be measured with waveform cross-correlations. The deep-learning derived phase picks and cross-correlation derived relative travel times will then be used to constrain seismic event locations and a 3D seismic velocity model for the study area. The resulting seismic events and velocity model will be compared to existing seismic catalogs and published velocity models, respectively.