As seismic data collection continues to grow, advanced automated processing techniques for robust phase identification and event detection are becoming increasingly important. However, the performance, benefits, and limitations of different automated detection approaches have not been fully evaluated. Our study examines how the performance of conventional techniques, including the Short-Term Average/Long-Term Average (STA/LTA) method and cross-correlation approaches, compares to that of various deep learning models. We also evaluate the added benefits that transfer learning may provide to machine learning applications. Each detection approach has been applied to three years of seismic data recorded by stations in East Antarctica. Our results emphasize that the most appropriate detection approach depends on the data attributes and the study objectives. STA/LTA is well-suited for applications that require rapid results even if there is a greater likelihood for false positive detections, and correlation-based techniques work well for identifying events with a high degree of waveform similarity. Deep learning models offer the most adaptability if dealing with a range of seismic sources and noise, and their performance can be enhanced with transfer learning, if the detection parameters are fine-tuned to ensure the accuracy and reliability of the generated catalog. Our results in East Antarctic provide new insight into polar seismicity, highlighting both cryospheric and tectonic events, and demonstrate how automated event detection approaches can be optimized to investigate seismic activity in challenging environments.

Machine learning algorithms have become a powerful tool in different areas of seismology, such as phase picking/earthquake detection, earthquake early warning and focal mechanism determination. Previously convolutional neural networks (CNN) have been applied to continuous seismic waveform recordings to perform efficient phase picking and event detection with good accuracy [Zhu et al., 2018]. However, the off-line training of current CNN requires at least a few thousands of accurately picked seismic phases, which makes it difficult to be applied to regions without sufficient picked phases. In this work, we will validate the transfer learning among different geographic regions. Our tests show that the phase picker trained on manually-labeled data acquired from Sichuan, China following the 2008 M7.9 Wenchuan earthquake [Zhu et al., 2018] works equally well on the continuous waveform acquired from Oklahoma, US [Zhu et al., 2018]. Specifically, using the CNN trained on the Wenchuan dataset, together with 895 local/regional catalog events recorded in central Oklahoma, we refine part of the networks to pick the arrival times of the local seismicity in Oklahoma. The refined CNN results are compatible with the matched filter results using the same catalog events as templates. Our next step is to extend our test to waveforms from different tectonic regions to demonstrate the generality of CNN-based phase picker. We plan to further use a New Zealand seismic dataset that includes more than 20 GeoNet stations in the North Island, where the matched-filter detected results are available to be compared with (Yao et al., 2018). Alternatively dataset include a subset of events in the waveform relocated catalog in Southern California. Updated results will be presented at the meeting.