Distributed acoustic sensing as a tool for exploration and monitoring: a proof-of-concept

We use PoroTOMO experimental data to compare the performance of Distributed Acoustic Sensing (DAS) and geophone data in executing standard exploration and monitoring activities. The PoroTOMO experiment consists of two "seismic systems": (a) a 8.6 km long optical fibre cable deployed across the Brady geothermal field and covering an area of 1.5 x 0.5 km with 100m long segments, and (b) an array of 238 co-located geophones with an average spacing of 60m. The PoroTOMO experiment recorded continuous seismic data between March 10th and March 25th 2016. During such period, a ML 4.3 regional 5 event occurred in the southwest, about 150 km away from the geothermal field, together with several microseismic local events related to the geothermal activity. The seismic waves generated from such seismic events have been used as input data in this study. For the exploration tasks, we compare the propagation of the ML 4.3 event across the geothermal field in both seismic systems in term of relative time-delay, for a number of configurations and segments. Defined the propagation, we analyse and compare the amplitude and the signal-to-noise ratio (SNR) of the P-wave in the two systems at high resolution. For testing the 10 potential in monitoring local seismicity, we first perform an analysis of the geophone data for locating a microseismic event, based on expert opinion. Then, we a adopt different workflow for the automatic location of the same microseismic event using DAS data. For testing the potential in monitoring distant event, data from the regional earthquake are used for retrieving both the propagation direction and apparent velocity of the wavefield, using a standard plane-wave-fitting approach. Our results indicate that: (1) at a local scale, the seismic P-waves propagation and their characteristics (i.e. SNR and ampli15 tude) along a single cable segment are robustly consistent with recordings from co-located geophones (delay-times δt∼ 0.3 over 400 m for both seismic systems) ; (2) the interpretation of seismic wave propagation across multiple separated segments is less clear, due to the heavy contamination of scattering sources and local velocity heterogeneities; nonetheless, results from the plane-wave fitting still indicate the possibility for a consistent detection and location of the event; (3) at high-resolution (10m), large amplitude variations along the fibre cable seem to robustly correlate with near surface geology; (4) automatic monitoring 20 of microseismicity can be performed with DAS recordings with results comparable to manual analysis of geophone recordings (i.e. maximum horizontal error on event location around 70 m for both geophones and DAS data) ; and (5) DAS data preconditioning (e.g., temporal sub-sampling and channel-stacking) and dedicated processing techniques are strictly necessary for making any real-time monitoring procedure feasible and trustable. 1 https://doi.org/10.5194/se-2021-37 Preprint. Discussion started: 15 April 2021 c © Author(s) 2021. CC BY 4.0 License.

1 Introduction ments; on average, the length of individual segments was on the order of 150 m. Relevant segments used in the next sections are numbered in Figure 1. Acquisition gauge length is fixed to 10 m, with channel sampling every 1 m, (Wang et al., 2018).
For the exploration tasks, described in section 3.1, we use one channel every gauge length (10 m) and discard the channels too close (10 m) to the bending point of the cable (as done in Wang et al., 2018). Given the segment length and the gauge length, we obtain about 10-15 independent acoustic records for each segment. The acoustic signal are recorded at 1000 sps. DAS data are organised in 30-seconds long HDF5 files, about one 1Gb for each file. In parallel, we also downloaded and analysed data 95 recorded by a co-located array of 238 Nodal geophones (Fairfield Nodal ZLand 3-component short period seismometers, with a peak frequency of 4.5Hz, called "nodes" hereinfater). Nodes had been deployed as close as possible to the fibre, allowing a detailed comparison of the signal recorded by the DAS and nodes systems.
We make use of the waveforms relative to the M L 4.3 Hawthorne event. For the exploration tasks, the analysis is limited to P-wave arrival. P-wave can be clearly seen in both DAS ad Nodes recording ( Figure 2). Here, we filter the waveforms between 100 0.5 and 2 Hz, enhancing the earthquake signal. DAS recordings are downsampled to 500 sps, to be consistent with nodes recordings. To compare nodes to DAS recordings, horizontal components of the nodes are rotated to match the azimuth of the closest segment of fibre cable ( Figure 1). All computations on nodal recording are operated on the rotated seismograms, if not specified.
The analysis of the M L 4.3 event consists in the automatic determination of the relative P-wave arrival times for each 105 segment of fibre cable and for all co-located nodes. The P-wave time delays are computed following the approach described in VanDecar and Crosson (1990) and in Piana Agostinetti and Martini (2019). For each of the three experiments (see Section 3.1), we compute the P-wave time delays for the selected DAS channels (one every 10 m). Following VanDecar and Crosson (1990), we first compute (1) the time-delay between all couples of selected channels, cross-correlating a 2 s long time window; and then (2) the absolute time-delay δt i for the i-th selected channel under the condition of i δt i = 0. The same procedure is 110 performed for the oriented horizontal recordings of the co-located nodes. After the definition of the time-delays, we measure the SNR for each channel and each node, considering a 5 s long time-window before and after the P-wave arrival ( Figure 2). On the same time-window that contains the P-wave, we also measure the maximum amplitude.
For monitoring tasks, we consider two test cases of earthquakes recorded at regional and very local ranges. For the former case, we use the aforementioned Hawthorne earthquake. For the latter one, we review the catalogue of microseismicity found 115 in Li and Zhan (2018) spanning the entire Brady geothermal field, and select one event that occurred on 2016 March 14 -10:41:576UTC, in close proximity of the node and DAS deployments.
The local earthquake as recorded by the nodes is processed to obtain a reference location. Precise P-and S-wave arrival times are derived from manual picking at 75 vertical and 48 horizontal channels. Theoretical travel times are calculated in a homogeneous half-space, with compressional (V P ) wave velocity of 3 km/s (Zeng et al., 2017;Parker et al., 2018) and a ratio 120 between Vp and S-wave velocity (V S ) V P /V S = 2.8, as indicated by the modified Wadati diagram. The likelihood function for source location has been explored using the Octet-tree sampling method (Lomax et al., 2009). Analysis of the same event at the DAS deployment begins by band-pass filtering over the 15-40Hz frequency band, followed by subsampling at 100Hz. A spatial sub-sampling is then performed by stacking the recordings from 11 adjacent channels within each profile, with a 20-channel 274 channels, thus similar to the nodal array. For each stacked trace, we then calculate a characteristic function given by the Kurtosis (Langet et al., 2014), and use an AIC autopicker (Sleeman and van Eck, 1999) to automatically identify the onset time of these functions. Finally, we use the DBSCAN algorithm (Ester et al., 1996) to discard the most obvious outliers, thus obtaining a set of 243 estimates of arrival times. Using the same simplified velocity structure described above, these data are finally inverted for a source location using both a conventional travel-time inversion as for the nodal array data, and a likelihood 130 function based on the Equal Differential Time [EDT] formulation (Font et al., 2004;Lomax, 2005). In this latter method, the likelihood of a given model is expressed as the sum of a set of probability density functions which, for any independent channel pair, incorporate the squared difference between the observed and predicted differential times at that pair. The EDT likelihood function is thus particularly appealing given its robustness in the presence of outliers, i.e. observations whose residuals are greater than the nominal uncertainty (Lomax, 2005).

135
For the Hawthorne earthquake, 2-minute-long DAS recordings are first band-pass filtered over the 0.5-2Hz frequency band, and then sub-sampled at 100Hz. Spatial sub-sampling is performed in a similar manner to what described above, but stacking over 41 adjacent channels with a 40-channel step, thus obtaining a virtual array of 57 elements. Delay times ∆T ij , i = j between all the independent array channels are derived from the maxima CC max of the corresponding cross-correlation function; these differential times are used to derive the horizontal slowness vector according to the linear relationship (Del Pezzo and where the matrix ∆x contains the differences between the x (EW) and y (NS) components of the array channel's coordinates, and s is the horizontal slowness vector, from which we derive the propagation azimuth (measured clockwise from the N direction) and apparent velocity.

145
Equation 1 is solved using a weighted least-squares approach, with weights defined as: (2) so to emphasize the contribution of the most correlated channels. For the inversion, we considered only those channel pairs exhibiting a CC max larger than an arbitrary threshold, here set equal to 0.85. The procedure is iterated over 4-s-long time windows, sliding along the DAS recordings with 80% overlap.

Exploration
Direct comparison of DAS and Nodal recordings is not possible, due to the sensitivity of the two seismic systems to different geo-observables. Many different procedures have been developed to convert DAS recordings to Nodal recordings and viceversa (Wang et al., 2018, and references therein). Here we do not aim to use the two systems together in the same analysis, but to  5b). Those could be apparent anomalies given by local surface waves generated from interaction of the P-wave with the local topography and erroneously cross-correlated with the correct P-wave. Nevertheless, this result supports our workflow where DAS data need to be strictly compared to co-located, re-oriented Nodal data. 180 We present a first analysis on the P-wave recorded by the DAS system, considering DAS segment 48. This segment is the longest one (about 350 m), with 7 geophones almost co-located along the cable (within 30m), and it gives us the possibility of following the P-wave over a long distance (Figure 7). P-waves recorded along the cable are generally similar, but not as much as on the vertical components of the Nodal system, giving smaller cross-correlation coefficients ( about 0.95 on the average, To test the possibility of appreciating 3D features, we apply our workflow to two parallel segments, 3 and 5 that run along the longer side of a rectangular area of about 150 x 70 m. Eight geophones are deployed along the two segments (4 geophones/segments, Figure 9). The results indicate that both time-delays and maximum amplitude display the same spatial variation. In particular, time-delay ranges between -0.12 s and 0.12 s for both systems and indicate an un-expected East-to-West