Comparing location uncertainties with automatic pickers on DAS data: case studies from Canary Islands

Distributed Acoustic Sensing (DAS) offers unprecedented meter-scale spatial sampling of strain/strain-rate wavefields, enabling unaliased seismic event observations. DAS technology utilizes fiber optic cables (FOCs), extending seismological observations to extreme environments, including ocean floors. Given the logistical difficulties in deploying and maintaining traditional seismic stations in these contexts, seismological data near oceanic earthquake sources remain limited. On a positive note, telecommunication FOCs are often deployed on the ocean bottoms to connect urban areas on land, potentially bridging this observational gap.

Traditional seismological monitoring, which aims to locate earthquake sources, typically relies on phase picking and subsequent data inversion. In a standard seismometer network, automatically-retrieved arrival times can be manually validated by expert operators; however, this task becomes practically impossible with DAS due to the unprecedented data density it offers, which can easily reach tens of thousands channels, considering the current capabilities of interrogating cables up to 100 km. For the purpose of leveraging both data measurements close to the source (DAS) and the improved azimuthal coverage by land stations, DAS data flows must be automated. Potential solutions include accurately tuning automatic pickers for the specific FOC and/or employing data selection and weighing procedures. Recently, pickers based on machine learning have been tested for DAS as substitutes for standard pickers, offering promising results and efficient arrival time measurements. Despite these advancements, challenges persist in accurately estimating onsets due to spatial variability in DAS waveforms, arising from a) uniaxial signal polarization, b) sensitivity to site conditions, and c) heterogeneities in FOC coupling. These data uncertainties, in turn, affect event location accuracy.

We address this problem by conducting a preliminary comparison of two standard pickers (based on the actual amplitude-frequency content of each channel) with a machine-learning-derived picker. We focus on DAS recordings of six local earthquakes located on the ocean bottom between Fuerteventura and Gran Canaria islands during an experiment from November 2022 to April 2023. Each earthquake is provided with a reference location from the regional network of seismometers. Kurtosis and FilterPicker (standard pickers) and Phasenet-DAS (machine-learning picker) onsets are inverted for event location, with a focus on statistically comparing the solutions' uncertainty (scattering). To achieve this, we employed a Markov chain Monte Carlo method to estimate the Posterior Probability Densities (PPDs) of hypocentral parameters.

In a second stage, we test a data-weighing approach on absolute arrival times based on specific channel properties. The aim is to assess its effects on location PPDs, in comparison to the “not-weighed” inversion. We repurpose the same algorithm, previously used for the location comparison, to modify each entry of the covariance matrix in the Bayesian inversion scheme, thus enabling a differential weighting of the arrival times. These preliminary comparisons of the efficiency of automatic pickers and data weighting procedures are commonly employed for the evaluation of standard seismological networks. With DAS arrays, these approaches become even more crucial, given the reduced space for manual validation by experts.