Earthquake Detection in Slowly Deforming Iberia using a Convolutional Neural Network Phase Picker

Abstract:

Iberia, in Southwest Europe, is a region with tectonic deformation rates ≲1 mm/yr, defined as a slowly deforming region (SDR), with the exception of the Betics in the South near the plate boundary between Nubia and Eurasia. Understanding the seismic behavior of SDRs remains a challenge to seismologists due to the low tectonic loading rates, complex systems of non-linear faults and episodic and migrating seismic activity that complicate the imaging of the seismic cycle and hazard assessment. Accordingly, seismic activity in Iberia is characterized by frequent low-magnitude earthquakes at shallow depths (<30 km) and infrequent moderate to high magnitude earthquakes, which are rare but destructive. These earthquakes are documented mostly in the historical and geological records. The seismicity of Iberia appears to be spatially diffuse based on the instrumental recordings. However, a question that remains open is whether instrumental seismicity is in fact diffuse or just poorly located. From 2007 to 2014 dense temporary seismic networks, such as the IberArray and WILAS networks, were deployed in this region. The continuous seismic data collected during these deployments hold untapped information that can be used to improve the seismic catalog and to gain insight on the tectonic processes of SDRs.

We trained a convolutional neural network-based phase-identification classifier (Zhu et al., 2019) using data from permanent seismic networks in the region and their respective phase catalogs. We built a training dataset of 20 s waveform windows of 80709 P-phases, 36568 S-phases and added noise windows right before and after the P and S-windows, respectively. Due to the regional distances that comprise our study area, we preprocessed data with a 0.5 Hz high-pass filter and a soft-clipping function. We achieved a model with an overall 95.5% training accuracy. The model was applied for phase picking at the permanent and the temporary networks. Preliminary results show a recall of 79% of catalogued phases in 2007. The detected phases will be linked by phase association and located to compile a new earthquake catalog for the region. The new catalog will be compared with local catalogs compiled using matched filter detection, a well-established method that uses known events to search for similar waveforms through cross-correlation.