ALBERT

Description: Automated procedures for seismic arrival-time picking on large and real-time seismological waveform datasets are critical for many seismological tasks. Recent, high-performance, automated arrival-time pickers mainly use deep-neural-networks applied to nearly raw, seismogram waveforms as input data. However, there is a long history in earthquake seismology of rule-based, automated arrival detection and picking algorithms that efficiently exploit variations in amplitude, frequency and polarization of seismogram waveforms.Here we use this classical, seismological domain-knowledge to transform raw seismogram waveforms into input features for a deep-learning picker. We preprocess 3-component, broadband seismograms into 3-component characteristic functions of a multi-band picker (FilterPicker), plus the instantaneous modulus and inclination of the waveforms. We use these five time-series as input instead of the 3-component, raw seismograms to extend the deep-neural-network picker PhaseNet within the SeisBench platform. We compare the original, purely data-driven PhaseNet and our extended, domain-knowledge PhaseNet (DKPN), using identical training and validation datasets, with application to in- and cross-domain testing datasets.We find that the explicit information targeting arrival-time detection and picking introduced by the domain-knowledge processing enables DKPN to be trained with smaller datasets than PhaseNet. Relative to PhaseNet, DKPN shows improved performance and stability for P picking and slightly improved S picking, especially for cross-domain application. With increasing training dataset size PhaseNet performance generally improves and converges to that of DKPN, except for cross-domain P picking. The results suggest that DKPN primarily needs to learn pick characterization, while PhaseNet additionally requires learning the more difficult task of arrival detection.