Because this experiment generated many tens of terabytes of data, we connected the DAS interrogator unit to a network-attached storage (NAS) server. The server was connected to an internet network, providing us with remote data access in real time. The interrogator and NAS were hosted in a Penn State IT building connected with an uninterruptible power supply (UPS) for backup power in case of station power interruption. Figure  (top left panel) shows the setup of the DAS system of the FORESEE pilot experiment in a standard computer rack. A GPS clock and antenna were connected for precise timing. We installed the GPS antenna outside of the building, roughly 2.0 m above the ground. Raw DAS recordings are saved in TDMS format (Silixa iDAS default data format, for which Silixa provides MATLAB and Python reader scripts). The selected DAS recording settings yielded over 200 GBd-1 and ultimately about 76 TB of raw data per year.

Before any seismic array processing, we needed to determine the precise location of each sensor along the fiber. Because the fiber sits in underground conduits, it is invisible from the surface, so traditional methods of obtaining GPS coordinates at individual surface seismic sensors were not feasible. Penn State's Enterprise Networking and Communication Services department manages the cable, so they provided a map of the fiber cable path and telecommunication manholes visible at the surface. Following this map, we ran tap tests to determine a set of representative channel indices at particular locations. For each location, we used a hammer drop source to generate six distinguished shots on the ground. Ideally the channel centered around a shot will respond most strongly to the shot and correspond to the GPS position of the hammer source. At many locations, strong noise inhibited our ability to identify hammer shots in raw DAS data. Instead, we computed the spectral energy, applying the Fourier transform to windowed trace by a short sliding window. . As seen in Fig. b, the hammer shots are easily identified and the center channel is relatively unambiguous. We refer readers to our previous work for examples . We ran more than 101 tap tests to obtain the positions of 101 channels. Finally we used interpolation to estimate locations of all 2137 channels with 2 m spacing. Note that some channels were removed from analysis after determining locations where fiber was looped back on itself.

To verify that this DAS system (interrogator, fiber and cable in conduit) records a useful metric of ground motion, we convert our DAS recordings to particle velocity and compare to the nearest broadband seismometer. This calibration procedure is based on several previous studies . To ensure this paper is self-contained, we briefly review this procedure.

The interrogator that we used records axial strain rate in the direction of the cable. Specifically, the actual output of the DAS system is the phase changes between consecutive pulses during the laser repetition rate (Fig. ). The phase of light traveling a distance x in a cable with refractive index n can be expressed as Φ=4πnxλ. The wavelength of Rayleigh backscattered light equals the incident wavelength (1500 nm). We assume refractive index changes linearly with strain rate (Δnn=(ζ-1)Δxx). The scalar multiplicative factor (ζ) is determined by the material properties: ζ=0.735 for single-mode fiber glass with light propagating inside. Hence the phase changes at the same fiber section separated by the gauge length LG (e.g., LG=10m) can be expanded as ΔΦ=∂Φ∂xx+∂Φ∂nn=4πnxλxΔx+4πnxλnΔn1=4πnLGλΔxx+Δnn=4πnζLGΔxλx. After rearranging the equation above, we can link raw DAS output (strain rate in iDAS) to strain rate ϵ˙xx by applying a constant scaling factor as follows: ϵ˙xx=λ4πnζLGdtΔΦ=1500×10-9(m)4π×1.445×10(m)×0.735dtΔΦ2=11.6×10-9fsΔΦ, where fs=1/dt is the sampling frequency and dt is the time sampling rate. The strain rate ϵ˙xx is expressed in terms of strain per second. Note that the raw DAS output ΔΦ can be different in terms of optical unit (either phase change or phase change rate), depending on the instrumental setup .

To illustrate the calibration procedure, we take the Peru M 8.0 earthquake on 26 May 2019 as an example to convert DAS recordings to particle velocity. We chose over 80 channels of an L-shape subarray with 40 channels in each direction, shown in Fig. . Traces from the DAS cable segment of the same orientation are compared with one horizontal component rotated to the cable direction of a nearby seismic station (SSPA) about 17.7 km away. This distance would be too far for a comparison in response to a local seismic source, but this teleseismic event is expected to yield similar responses by seismometers at this distance. Figure  shows individual steps to convert strain rate to particle velocity. First we multiply the raw data ΔΦ by a constant scaling factor (11.6fs) to obtain the strain rate (nanostrain per second). The strain rate is integrated along the time axis to obtain the strain. Low-frequency artifacts resulting from integration are removed by applying a band-pass filter between 0.02 and 0.5 Hz. Then DAS array strain values are converted to particle velocity by scaling the apparent velocity c as follows: 3v=∂u∂t=-c∂u∂x=-ωk∂u∂x=-ωkϵxx, where c is the apparent phase velocity along the cable axial direction. In the test, we apply the f–k transform of the seismograms and scale Fourier coefficients by V(ω,k)=-(ω+σ)E(ω,k)(k+σ), where σ is the threshold to avoid instabilities in division by zero frequency and wavenumber values . Finally we apply inverse Fourier transform to V(ω,k) to get particle velocity in the time domain.

Here we use 40 traces (with coherent waveforms) for the f–k transform. Then we select one best σ=1×10-5 by trial and error for matching particle velocity waveform. We note that the waveform matching is somehow dependent on the threshold number σ , which may arise from an ambiguous velocity factor determined by the f–k transform with limited offset in the case of plane wave. We also include a longer offset (800 m) (maximum in the fiber direction avoiding the turning point) with σ=1×10-3. Although the waveforms are highly varying between channels due to possible coupling effects, the resulted particle velocity is very close (see red trace in the bottom panel in Fig. ). We refer readers to and for details of the f–k conversion method.

Figure  shows the comparisons of DAS recordings in two orthogonal directions (Fig. a) to reference seismograms (BH1 and BH2) (particle velocity in µms-1) at the nearest seismic station (SSPA) after applying a band-pass filter (0.02–0.5 Hz). It is clear that the DAS data show an agreement with the reference seismogram (particularly S wave in BH2). Small inconsistencies in coda wave amplitudes are possibly due to the DAS data representing the average signals over a gauge length rather than an isolated point sensor , although many other factors may play a role in waveform differences, such as different locations of the seismic station and DAS array, directional sensitivity, unknown DAS instrument response, and the host environment. Overall, this calibration process suggests that this DAS array is an acceptable system for recording single-component low-frequency seismic data.

To calibrate higher frequencies, we performed a similar comparison using an active source, shown in Fig. . We co-located a 24-channel geophone array (4 m spacing) on the ground just above the fiber cable. After synchronizing the GPS timing, we searched for DAS shot gather recordings. We applied a band-pass filter (10–60 Hz) and auto-gain control (AGC) to geophone and DAS data. In the geophone data, we identified a direct wave (∼800 ms-1) and refraction (∼4000 ms-1). From this we inferred that the depth of the first layer is approximately 8 m. In DAS, there is more coherence in the moveout of the waveform, and it appears that the direct wave is about 2250 ms-1. We hypothesize that this seems to be the average of the direct arrival and refraction on the geophone gather. We identified four possible reflections (red arrows) that are kinematically consistent in both the geophone and DAS data in Fig. . Dynamically, these phases in DAS are more continuous and coherent, possibly owing to the continuous nature of the overlapping DAS channels, compared to discrete geophones.

Ground roll (surface waves ∼370 ms-1) is clear in the geophone gather but not manifested in DAS. Results on this have been mixed in different experiments. conducted an active-source seismic experiment with fiber in underground telecommunication conduits and did not report strong ground roll that was clear on 3C nodes. However, conducted an active-source seismic experiment to compare geophone and DAS data where the fiber was left on the surface of the ground. Their data showed clear surface wave moveout in DAS. There are several possibly hypotheses behind the lack of surface waves in Fig. b, for example fiber buried at shallow depths (∼1 m) in the conduit and loose contact of fiber to the conduit.

Another unique and interesting observation on DAS is that the energy around the hammer shot shifted down (highlighted yellow zone), which is likely the trapped waves propagated vertically inside a low-velocity zone (e.g., sinkhole in this area). This observation echoes the study of fault-zone-trapped waves using dense seismic geophones and DAS by . The ability to detect such signals is a benefit of using dense sensors like DAS. We suggest that further seismic wave modeling should be conducted to verify the presence of the low-velocity zone.

Figure  shows DAS recordings of four earthquakes (Peru M 8.0 earthquake on 26 May 2019, Ridgecrest M 7.1 earthquake on 6 July 2019, Tennessee (TE) M 3.8 earthquake on 20 January 2020 and PA M 1.1 earthquake on 27 August 2019). We applied the band-pass filtering to both PSRS and DAS seismograms. All top panels show seismograms (particle velocity) from nearby seismic station PSRS (Nyquist frequency: 50 Hz) as references and one trace (100th sensor) from DAS, while the bottom in Fig.  shows full DAS recordings of 2137 channels. DAS of the Peru M 8.0 earthquake shows strong P waves, S waves and surface waves. It serves well to calibrate kinematics and dynamics of seismograms in the above section. While, for the Ridgecrest M 7.1 earthquake, P waves are relative invisible in DAS, S waves are clearly identified at 03:27:30 UTC. Surprisingly, DAS has strong S waves, and seismometer surface wave energy is much stronger than S-wave energy. The Tennessee M 3.8 earthquake appears weak in energy in DAS recordings in Fig. c. Due to rush hour (15:10 LT), this DAS record is very noisy and contaminated with traffic noise (linear events), but surface waves are still identifiable. The last event, shown in d, was the PA M 1.1 earthquake near State College (10 km away from the array) in Pennsylvania. Two body wave phases are clearly visible, and surface waves likely follow S waves.

While there have been other examples of DAS arrays using existing telecommunication infrastructure on land, these have been concentrated in the western USA, which rarely experiences thunderstorm lightning . Thus, in deploying an array in the eastern USA, we had a unique opportunity to study how thunder and lightning couple to the ground to induce seismic waves. Prior studies with a single seismometer or a small handful of seismometers have observed that thunder can induce ground motion , but to our knowledge this has never been studied with a large dense array . On 15 April 2019, a severe thunderstorm crossed over the array, verified by the National Lightning Detection Network. We were able to detect clear thunder events across the array, one of which that occurred on 15 April 2019 is pictured in Fig.  . Figure a shows 2 min raw recordings of six (e4-e10) events between03:33 and 03:35 UTC. Spectrograms of two channel traces (black traces overlain in Fig. a) show the spatial variation of recordings between two channels. We hypothesize that this recorded seismic energy is induced by thunder and/or lightning electromagnetic waves coupling to the ground to induce surface waves propagating in the shallow subsurface. The spatial variation could suggest the spatial attenuation of the shallow subsurface layer. The travel time moveout selected enables us to characterize the thunderquake events and wave propagation across the FORESEE array. Even State College also be reconstructed from DAS data. Up to now, we have manually identified more than 120 thunderquakes from other four severe thunderstorms from April to August 2019.

Figure  shows the average power spectral density (PSD) of raw DAS recordings of the earthquake and thunderquake events in Figs.  and , which is computed by averaging the power spectrum of each channel for all channels. The Nyquist frequency of DAS recordings is 250 vs. a 50 Hz Nyquist frequency of seismometers. Distant earthquakes were downsampled in preprocessing. We observed low-frequency local and regional earthquakes (PA M 1.1 and TE M 3.3) (0.05–20 Hz) and a long-period global earthquake (0.01–1 Hz). The peak of thunderquakes lies in the range of high frequency (10–130 Hz). We can conclude that the FORESEE DAS fiber is able to record the broadband quakes (0.01–250 Hz).

It is surprising that, despite relying on just friction and gravity for coupling existing fiber optics to conduits, we were able to detect walking individuals in Fig. . After band-pass filtering (1–5 Hz), clear footstep signals are recorded by a subset of the array beneath a straight path without moving cars. With offsets in this dense array in the bottom panel of Fig. , we are able to identify the walking direction and whether vibrations are caused by individuals or groups (see examples of individuals and groups in Fig. ); the second author (dashed black line) following a group of people (blue line) who departed from the bus station (around channel 1270) and headed south; and later a walker (red line) who moved in the opposite direction towards the bus station, corresponding well with the timings of in-person field observations made on 15 July 2019 (Fig. ). We can also estimate the walking speed as about 1.2 ms-1 from the slope of dashed lines. To be general, we analyze the power spectral density of 1 min data at noon (across channels 800–2137, above central campus) from three very different days (16 April 2019, 16 February 2020 and 16 April 2020). Figure  shows significant peaks at around 2.0 and 4.0 Hz (harmonic), while they disappear at midnight. This 2.0 Hz signal corresponds to walking steps on campus, equivalent to 120 steps per minute, which is similar to the people walking on shopping floors with an average frequency of 2.0 Hz and a velocity of 1.4 ms-1 , slower than marathon runners with 2.8 Hz . Not surprisingly, these footstep signals do not show up on 16 April 2020 during the COVID-19 crisis after the stay-at-home order in PA on 1 April 2020. And the noise power spectra level is close to that at midnight. Spatially, we also calculated the power spectra of data at channels 1–600 (on the edge of campus), and the peaks at 2 and 4 Hz are not found (not shown here). We can see that DAS recordings provide the spatial and temporal distribution of the footsteps that may be useful for designing and analysing campus traffic.

For the goals of the FORESEE array, these footsteps are likely to hinder efforts towards subsurface imaging with ambient-noise interferometry, so we are interested in removing these signals prior to imaging. Further, as DAS arrays are deployed in a wider variety of locations, there may be areas where removing the footsteps is needed to ensure privacy of people in the area. We recently developed a convolutional neural network to automatically identify pedestrian footsteps, the first step towards removing these signals .

Several mining sites around State College (Fig. ) provide well-repeatable blast sources for further calibration and near-surface monitoring. The distances of four mining sites (site 1, site 2, site 3 and site 4) to the campus are about 31, 41, 84 and 15 km, respectively. The reference seismic station indicated by the blue triangle is about 10 km southwest from State College. Figure  shows three quarry blast events, one each from sites 1, 2 and 3. With proper band-pass filtering (1–5 Hz), these events are clearly visible in the noisy records since they were occurring during traffic hours. Their recordings are kinematically consistent with reference seismograms by 1–2.5 Hz band-pass filtering. Strong surface waves are identified despite these explosive sources. In 2019 there are several hundred repeatable blast events cataloged from these sites, which could allow yearly near-surface monitoring of geotechnical engineering activities and/or hydrological systems owing to a significant ground water level variation. One strong event, shown in Fig. , was recorded on 14 May 2019 from site 4 (Pleasant Gap, PA). Since this event has equal-distance seismic station PSRS and DAS (see Fig. ), there is a 2 s time delay between two data. Figure a shows high-frequency P and S waves (10–20 Hz). In Fig. b, we can see strong low-frequency transverse motions. As in previous observations, this low-frequency transverse wave exhibits flipped polarity at the orthogonal fiber locations (e.g., indicated by arrows in channels 170 and 600), which is either a SH or Love wave and was also observed in previous DAS recordings from the Stanford DAS array .

In general, seismic sources that only couple to the ground through sound often have weaker coupling, and prior studies of active-seismic-source experiments detected by buried fiber optics did not show air waves . However, our DAS data recorded distinctive signals corresponding to live music during the 26 April 2019 Penn State Movin' On music festival . The live music stage was directly above channel 120–150. Figure  shows an example of 20 min DAS recordings of four songs – “Welcome to Your Life”, “Cannonball”, “Good Morning”, “Ways to Go” – based on the timings of the concert playlist. Figure b shows large variations of recorded seismic amplitudes, even within the different parts of a single song. The break between songs is easily identified as a gap between waveforms and spectrograms in Fig. c. The spectrogram of trace 130 shows that different songs result in characteristic spectra composed of narrow and evenly spaced energy (zoomed details in Fig. d). We can detect sustained notes in the bass range (40–140 Hz). Played back and visualized with the IRIS SeisSound tool, these signals are clearly fat bass riffs with much of the energy below 100 Hz (refer to audio of the song “Good Morning” in the supporting materials of ). This is not the first time seismometers have detected the bass line of concerts; showed similar results from a single broadband seismic station during a Bruce Springsteen concert in Barcelona. The difference here is the densely sampled spatial data, which enable us to see that these signals were clearly sensed more than half a kilometer away in Fig. . reported similar seismic recordings of parade floats and bands by the Pasadena distributed acoustic sensing array.