Real-time multi-station microseismic event detection via transfer learning of computer vision deep learning model

Summary With the shale gas exploration and development into deep strata (> 3500 m in depth), weak signal detection becomes more challenging and vital in surface-based microseismic monitoring. Most existing deep learning approaches are trained on synthetic or single-station data, which limits their performance in detecting weak events in field seismic recordings. Here, we propose a deep learning approach based on the computer vision object detection model YOLOv5 via transfer learning and train it on a dataset from field observation to better detect microseismic events in multi-channel seismic data. First, we convert continuous seismic data into waveform images in grayscale with a record length of 20 s, and manually annotate the microseismic events containing the first arrival signals using rectangular boxes. To train the model, the transfer learning strategy is used by using parameters of YOLOs in computer vision to initiate the model. After training and validation, we conducted comprehensive tests to examine the performance of the model. Compared with the single-station short-term average/long-term average (STA/LTA) detector, our method reduces false positives markedly, yet it also delivers higher recall rate than the available multi-station detection scheme. When applied to continuous dataset from a different area, the model achieved good performance with a precision of 0.9832 and a recall of 0.9702, indicating it learns representative features of microseismic signals. All these tests illustrate that the proposed method achieves high detection accuracy, low false positive rates, and robustness. Meanwhile, the method possesses the superiority for real-time monitoring and can be easily updated by adding a few new samples to the training dataset.