Enhanced Microseismic Event Detection Using Deep Neural Networks

Summary A reliable automatic event detection procedure is a vital tool in any passive seismic monitoring scenario, where recordings often span long periods of time and events are typically below the noise level. A commonly-used automatic detection procedure involves taking the ratio of a Short-Time Average to Long-Time Average (STALTA). This approach has two pitfalls: its high susceptibility to noise, and, its highly sensitive parameters which need careful tuning. Whilst a number of alternatives have been proposed they are often too computationally heavy or time consuming to provide an alternative for realtime detection. This work adapts a common procedure from the field of computer vision to develop and train a deep learning model to automatically detect seismic events in a passive monitoring scenario. Trained on synthetic data, the proposed approach requires no manual annotation of data - a time-consuming task. Benchmarked against the STALTA approach on a recorded event observed in the North Sea, the deep learning model was shown to accurately detect the event even on the noisy traces where the STALTA failed. With a compute time faster than the data flow, the proposed approach is a strong contender for realtime monitoring applications.