Optimizing automated earthquake detection methods in Delaware Basin, Southeastern New Mexico

Recent increases in seismicity due to anthropogenic activities in southeast New Mexico, USA, demonstrate the need for timely earthquake detection and updated seismicity catalogs. The New Mexico Tech Seismological Observatory (NMTSO) has for many years performed earthquake detection and location in New Mexico using only manual analyst review, but as the number of earthquakes has increased, the implementation of automated methods in the earthquake detection workflow can improve efficiency, accuracy, and completeness of earthquake catalogs. In this study, we applied two machine-learning-based seismic phase auto-pickers to continuous waveform data from multiple seismic networks in southeast New Mexico and west Texas for 2021 to assess the efficiency of these tools for earthquake detection in this region. Results from deep learning tools (the EQTransformer and the PhaseNet auto-pickers) for events located within the New Mexico region were benchmarked against the manually picked NMTSO earthquake catalog derived for the study period. These catalog comparisons, specifically the numbers of missed and false detections, provide a baseline for evaluating machine learning tools in the southeast New Mexico region. We find that both machine learning methods allow detection of significantly more earthquakes than our manual detection, but due to significant numbers of false positives and missed events, they are best employed in conjunction with analyst review. We now employ EQTransformer in our earthquake detection process, thus advancing our ability to provide timely earthquake detection and location in this seismically active area. Article highlights. This study evaluated two machine-learning-based phase pickers for application within the New Mexico Seismic Network. The results demonstrate the importance of parameter optimization to minimize false positive (detections identified as noise events based on manual review) and false negative (earthquakes present in manually reviewed catalog but not present in the machine learning catalog) detections and improve event identification accuracy. By comparing the performance of the different tools, we implemented the most effective workflow for network earthquake monitoring, thereby improving the catalog completeness.