3.2 Receiver Function Processing
Previous work shows that the autonomous three-component 5-Hz geophones used in this study can yield high quality RFs comparable with co-located broadband seismometers (Ward et al., 2018; Ward & Lin, 2017). Like those earlier studies, our short deployment period limited the number of teleseismic events for RF calculation. Out of ~86 teleseismic events >Mw 5.0 occurring within the 30°- 90° search radius, we retained 7 events (Table S1; Fig. S1(a) and S1(b)) that met the selection criteria: (1) a magnitude >5.5, (2) a 30° – 90° epicentral distance from the center of the array, and (3) a signal-to-noise ratio (SNR)>3 and an identifiable incident P wave across the array (Figure S1(c)).
Prior to calculating RFs, we windowed the seismograms from 15 s before to 75 s after the theoretical P arrival. Next, we decimated the waveforms to 50 samples per second using an finite impulse response filter to prevent aliasing. We then removed the mean and the trend and applied a Hanning taper. Finally, we removed the instrument response from the nodal geophones (5 Hz corner frequency). We followed the above steps as outlined by Ward et al. (2018). We then filtered the resulting time series using a bandpass of 0.2 – 2.0 Hz. To groundtruth our waveform processing workflow, we retrieved waveforms for the selected 7 events recorded by AACSE broadband stations deployed within the node array footprint (Z. Li et al., 2020), performed the same pre-processing procedure, and compared the resultant broadband waveforms with the pre-processed nodal time series (Fig. S2).
After preprocessing, we culled additional noisy signals by applying a SNR-based noise reduction procedure which eliminated traces with SNR< 2.0 on the vertical component or SNR < 1.25 on the north component. Then we rotated from the station ZNE (vertical, north, east) coordinate system to the earthquake ZRT (vertical, radial, transverse) system. To compute the RFs for each event, we deconvolved the radial component seismograms with vertical component seismograms at each station using the time-domain iterative deconvolution method (Ligorria & Ammon, 1999) with a Gaussian filter parameter of 2.5 (~1.2 Hz) and 5.0 (~2.4 Hz). All analysis was performed via Python using the open-source rf software package (Eulenfeld, 2020).
Before stacking the RFs, we applied a Ps phase moveout correction using the iasp91 (Kennett & Engdahl, 1991) model and calculated piercing points. We set the piercing point depth at 20 km based on estimates of slab depth (20 – 27 km) beneath the study area from the Slab2.0 model (Hayes et al., 2018), created equal profile boxes along the array (Fig. S3), and then stacked the receiver functions by common conversion points (Fig. 2). Both the stacked 1.2 and 2.4 Hz RFs were converted to depth (Fig. 2b and 2c) using the rf software and the iasp91 velocity model.