Local Distance Seismic Source Discrimination in the United Kingdom Using P/S Amplitude Ratios and Machine Learning

ABSTRACT Seismic source discriminants comparing earthquakes and explosions have been validated for large-magnitude events (local magnitude, ML>3.5) detected at regional and greater distances (>350 km). However, there is less evidence that existing seismic source discriminants apply to small-magnitude seismic events (ML≤3.5), which are predominantly detected at local distances (≤350 km). In this study, we investigate the performance of a P/S amplitude ratio discriminant for 210 small-magnitude seismic events in the United Kingdom (UK). High-frequency (> 8 Hz) network-averaged P/S ratios can discriminate between earthquakes and explosions, whereas low frequencies (2–4 Hz) can separate offshore explosions and quarry blasts. We derive a path and site amplitude correction model that reduces P/S ratio distribution variability, but using discrimination metrics we show that simple across-network averaging is the most important step to allow effective discrimination. We also applied a machine learning classification model, previously trained to discriminate earthquakes and explosions in France, to our dataset. The model demonstrates high classification accuracy, likely due to similarities between geology and seismic sources in France and the UK. When the predictions from the convolutional neural network are combined with high-frequency P/S ratios, we are able to distinguish other seismic sources, such as longwall mining-induced earthquakes. These results suggest that P/S ratios can be considered a useful seismic source discriminant at local distances across the UK, and network-averaging sufficiently accounts for non-source factors.