Temporal enrichment of the seismic record of the 2018 eruption at Sierra Negra using deep neural networks

The 2018 eruption at Sierra Negra volcano, Galapagos Islands, was accompanied by 8.5 metres of caldera subsidence and intense seismicity, resulting from deflation of a shallow sill-like magma reservoir at ~2 km depth. High-precision hypocentre locations from manually picked phase arrivals show that earthquake sources are tightly constrained within a complex, multi-stranded trapdoor fault system (TDF) above the sill. However, the incompleteness of this high-precision catalogue leaves outstanding questions about the spatio-temporal evolution of seismicity through the eruption, and how it relates to deformation and magma efflux.

Here we present the results of an automated workflow to streamline the production of a ‘temporally-enriched’ seismic catalogue for 2018 eruption at Sierra Negra. We initially utilise PhaseNet, a deep-neural-network-based automatic phase picker, to identify events missing from the initial manually picked catalogue, expanding the detections from 1,618 to 9,871 events. Our catalogue identifies more events in the immediate aftermath of the M_w5.4 earthquake that initiated the eruption, and new small magnitude events (< M_L2.0) in the period more than 72 hours after the eruption onset. We then use a template matching approach to further supplement these detections. Specifically, these events fill gaps in the catalogue where tremor amplitudes make manual event detection more difficult. Hypocentre locations for newly detected events are also constrained to the TDF zone, however there is more variety in depth estimates. This has implications for how we consider the TDF with depth, and allows us to consider other potential sources of seismicity in the system.

Our workflow offers an efficient method of producing ‘temporally-enriched’ catalogues at Sierra Negra, and can be readily adapted for the sparse seismic network that remains during the current inter-eruptive phase. However, our experience at Sierra Negra suggests that applying automated earthquake detection and location methods can be challenging in volcanic settings, and requires careful parameterization and quality control.