Ambient vibration characterization and monitoring of a rock slope close to collapse

Ambient vibration characterization and monitoring of a rock slope close to collapse Abstract We analyse the ambient vibration response of Alpe di Roscioro (AdR), an incipient rock slope failure located above the village Preonzo in southern Switzerland. Following a major failure in May 2012 (volume ∼210 000 m3), the remaining unstable rock mass (∼140 000 m3) remains highly fractured and disrupted, and has been the subject of intensive monitoring. We deployed a small-aperture seismic array at the site shortly after the 2012 failure. The measured seismic response exhibited strong directional amplification (factors up to 35 at 3.5 Hz), higher than previously recorded on rock slopes. The dominant direction of ground motion was found to be parallel to the predominant direction of deformation and perpendicular to open fractures, reflecting subsurface structure of the slope. We then equipped the site with two semi-permanent seismic stations to monitor the seismic response with the goal of identifying changes caused by internal damage that may precede subsequent failure. Although failure has not yet occurred, our data reveal important variations in the seismic response. Amplification factors and resonant frequencies exhibit seasonal trends related (both directly and inversely) to temperature changes and are sensitive to freezing periods (resonant frequencies increase with temperature and during freezing). We attribute these effects to thermal expansion driving microcrack closure, in addition to ice formation, which increase fracture and bulk rock stiffness. We find the site response at AdR is linear over the measured range of weak input motions spanning two orders of magnitude. Our results further develop and refine ambient vibration methods used in rock slope hazard assessment. Geomechanics, Time series analysis, Seismic noise, Site effects, Wave propagation 1 INTRODUCTION Rock slope failures represent significant hazards in populated alpine areas, and their identification, characterization and monitoring are key tasks in risk management. Although a number of well-established methods exist for these purposes (Gischig et al.2011; Jaboyedoff et al.2012; Kos et al.2016), subsurface information (e.g. the depth of the instability) is usually only extrapolated from surface observations. Boreholes are expensive in alpine environments and application of standard geophysical methods is challenging, especially in rock slopes with open fractures (Heincke et al.2005; 2006). Active seismic methods, for example, face extremely strong scattering and velocity gradients, which limit the depth resolution (Green et al.2006). Passive seismic techniques, on the other hand, present an alternative to classical geotechnical methods (see an overview by DelGaudio et al.2014). Unstable rock slopes with open, compliant fractures show a specific seismic response (e.g. Burjánek et al.2010, 2012; Lévy et al.2010; Fäh et al.2012; Bottelin et al.2013a) where ground motion is locally amplified in the direction/s of slope deformation at specific frequencies (i.e. the resonant frequencies). In particular, standing waves develop in the disrupted rock blocks (Burjánek et al.2012, henceforth BU12), and the unstable rock mass acts as a resonator. Resonant frequencies in turn carry information on the material properties and internal structure of the slope (i.e. depth, volume, fracture compliance of individual blocks), as demonstrated by numerical modelling (Moore et al.2011, 2016; Gischig et al.2015). Nevertheless, the mechanism of resonance described above is not unique. For instance, in the case of heavily fractured materials, resonance also originates from a seismic impedance contrast between the internally damaged rock mass and stiffer surrounding bedrock (Havenith et al.2002; Bottelin et al.2013a; Kleinbrod et al.2017a). Resonant frequencies of rock bodies can be retrieved relatively easily from ambient seismic data (no complex processing required), and may thus be readily used for rock slope monitoring and identification of irreversible change associated with internal damage. Lévy et al. (2010), for example, identified a decrease in the first resonance frequency of a rock column, prior to its collapse, accompanying failure of internal rock bridges. Monitoring of resonant frequencies was subsequently performed at four unstable rock slopes in the French Alps (Bottelin et al.2013b). While no change in the seismic response resulting from damage was observed at these sites, significant reversible variations caused by temperature fluctuations were detected. Similar observations were reported by Starr et al. (2015) for a short-term vibration monitoring experiment at a natural sandstone arch in Utah. Bottelin et al. (2013b) and Starr et al. (2015) proposed complementary but slightly different thermal stress mechanisms to explain these observations, the former calling on changes in fracture contact stiffness and the latter invoking material stress-stiffening. Recognition and characterization of reversible variations in resonant frequencies has thus emerged as a crucial task preceding identification of permanent changes associated with internal rock mass damage. Here we describe experiments to characterize and monitor the seismic response of a large rock slope instability at Alpe di Roscioro (AdR), Switzerland (Fig. 1). Following the last major failure in May 2012 (volume ∼210 000 m3), a large portion of the rock slope remains unstable and is pierced by a dramatic series of open tension cracks (Loew et al.2017). We deployed a small-aperture seismic array on the unstable rock mass (∼140 000 m3) shortly after the 2012 failure with the aim of characterizing the structure and seismic response. Later, we equipped the site with two semi-permanent seismic stations to monitor changes in the response over the following 4 yr. The collected time-series data allow us to distinguish and describe important variations in the observed seismic response related primarily to environmental effects. The heavily fractured slope is highly critical and will likely undergo future failure, providing a unique opportunity to study changes in the seismic response corresponding with the evolution of internal damage as a means of evaluating time-dependent progressive failure. Figure 1. View largeDownload slide Overview of the Alpe di Roscioro site: (a) Orthophoto taken before the 2012 failure. Scarps of the 2002 and 2012 major failures are indicated by orange and yellow lines, in addition to the current fracture network (in red); (b) cross-section of the site showing the location, orientation and anticipated depth of the fractures, as well as the foliation orientation; (c) photograph of the remaining unstable rock mass; d) photograph of the current fracture network and the scarp of the 2012 failure. Figure 1. View largeDownload slide Overview of the Alpe di Roscioro site: (a) Orthophoto taken before the 2012 failure. Scarps of the 2002 and 2012 major failures are indicated by orange and yellow lines, in addition to the current fracture network (in red); (b) cross-section of the site showing the location, orientation and anticipated depth of the fractures, as well as the foliation orientation; (c) photograph of the remaining unstable rock mass; d) photograph of the current fracture network and the scarp of the 2012 failure. 2 SITE DESCRIPTION The AdR instability (also known as the Preonzo rock slide due to its location near the village Preonzo) is a so-called retrogressive instability characterized by numerous catastrophic failures of different scales (Loew et al.2017). The earliest documented failure occurred on 1702 February 22 and buried part of the village Preonzo. These deposits remobilized 23 yr later in a large debris flow that again destroyed houses and caused 17 causalities. After a long period of quiescence, new tension cracks were observed in the late 1990s cutting the flat terrace behind the failure scarp (Fig. 1). In May 2002, about 150 000 m3 rock released following a period of heavy rainfall. Another smaller failure of about 25 000 m3 occurred on 2010 May 9. A monitoring system installed in 1999 (i.e. crack extensometers, reflectors measured with an automatic total station) was successful in anticipating these failures, and also successfully predicted the latest failure on 2012 May 15 involving about 210 000 m3 of rock. After the 2012 failures, a large portion of the unstable cliff remained (about 140 000m3), dissected by a network of wide fractures and exhibiting ongoing movement (Fig. 1). Local site geology consists of competent amphibolitic gneisses overlying augen gneisses that form the steep upper part of the AdR (Willenberg et al.2009). At the bottom of the steep cliff, weaker layered gneisses and schists outcrop. Foliation and lithology contacts dip roughly 25° WSW, that is, into the slope. This geological predisposition is typical for retrogressive failures that require step-wise fracturing of intact rock bridges until sufficient deformation has accrued and a sliding surface can form. These progressive deformation and internal failure processes result in gradual opening of tensile fractures along the flat bench at the back of the instability. A more detailed geological description can be found in Willenberg et al. (2009) and in Loew et al. (2017) respectively. 3 METHODS A number of studies, including our previous work on the seismic response of rock slopes (e.g. Burjánek et al.2010, henceforth BU10), have identified polarized, directional and amplified ground motion in unstable rock slopes. Since we expect a similar response at AdR, we focus on measuring similar ground motion attributes. We use the time-frequency polarization analysis (shortly polarization analysis or PA) developed by BU10 and BU12 to characterize particle motion of three-component seismic recordings as an ellipse. Three parameters of the polarization ellipse are retrieved as a function of time and frequency: (1) azimuth of the major axis (strike), (2) tilt of the major axis (dip), and (3) ratio between the length of the minor and major axes (ellipticity). In the case of ambient vibrations, we assume that observed ground motions are quasi-stationary and consider only the relative occurrence of polarization parameters for the analysed time window of 60 min. Thus, the time dependence is relaxed and the polarization parameters depend only on frequency. Details of the procedure can be found in BU10 and Burjánek et al. (2014). In addition, measured distributions of the polarization angles (strike and dip) for a single frequency can be well represented using a Wrapped Cauchy distribution (Burjánek et al.2014), while ellipticity is best represented with a Beta distribution (BU12). The Wrapped Cauchy distribution is characterized by mean and concentration parameters, while the Beta distribution by two positive shape parameters. Each of these parameters can be estimated for a given frequency using a maximum likelihood method. We evaluate site amplification using standard site-to-reference spectral ratios (SRSR). Amplitude spectra from ground motion recordings at a site with selected reference conditions (the reference site) are used to normalize spectra measured at other sites. If the interstation distance is short compared to the distance to the excitation sources, then SRSR represent frequency-dependent amplification functions with respect to the reference site. Although SRSR are commonly applied for earthquake recordings (site-to-source distance is usually well constrained), they also provide reasonable results for ambient vibration recordings (e.g. Roten et al.2006; BU12) even though the exact source distribution is usually not known. Nevertheless, noise sources must not be in close vicinity to the investigated site. For ambient vibration recordings, ground motion time-series data were split into non-overlapping windows of 100 s length, and the multitaper method was applied for each time window to estimate the Fourier amplitude spectrum (Prieto et al.2009). The time-bandwidth product, a parameter which controls spectral resolution (Prieto et al.2009), was set to 2.5 and four tapers were used. The retrieved spectra were normalized by the corresponding reference spectrum for each time window and the geometrical mean for all windows was determined. No additional smoothing of the spectra was required, as the multitaper method provides sufficiently smooth spectra while preventing spectral leakage (bias from the neighbouring frequency intervals). 4 ARRAY MEASUREMENT We performed an array measurement over 1 d on 2012 May 25, two weeks after the catastrophic failure. We used 12 three-component velocimeters with an eigenperiod of 5 s. The sensors were placed on metal trihedrons, in shallow holes when necessary, for better coupling with the ground. The precise positions of the sensors were measured with differential GPS. The geometry of the array is depicted in Fig. 2. Stations PRE001 and PRE002 were located close to a steep rock wall (to the west) within the stable area of the slope, and served as reference stations. Stations PRE003 and PRE004 were also located within the presumably stable area, but close to the potential backscarp (i.e. the rear-most tension crack). The remaining stations were located within the unstable and highly cracked area of the slope. We selected a time window of 1 hr to extract seismic data, during which all sensors of the array were correctly deployed with GPS time synchronization. Three seismic events (ML = 2.9, ML = 3.2, ML = 4.0) of the 2012 Emilia Romagna earthquake sequence were recorded by the array (Fig. S1) with good signal to noise ratio (Fig. S2) and epicentral distance ranging from 220 to 235 km. Sensor orientations were checked using the earthquake recordings. Since the epicentral distances were much larger than interstation distances, low frequency ground motion (<2 Hz considering the site effects described later) should be identical across the array. Band-passed (0.5–1.5 Hz) three-component traces of the earthquake recordings were systematically compared, which resulted in orientation corrections of up to 3°. Thus, the final error in our sensor orientations is expected to be less than 1°. Figure 2. View largeDownload slide Configuration of the seismic array (yellow dots); red squares show the positions of the semi-permanent seismic stations, while blue bars show the locations of extensometers used in this study. Green and red lines represent the boundary of the instability and the main scarp of the 2012 rockfall, respectively (after Loew et al.2017). Blue arrow indicates the mean direction of slope displacement (Loew et al.2017). The contour interval is 10 m. Figure 2. View largeDownload slide Configuration of the seismic array (yellow dots); red squares show the positions of the semi-permanent seismic stations, while blue bars show the locations of extensometers used in this study. Green and red lines represent the boundary of the instability and the main scarp of the 2012 rockfall, respectively (after Loew et al.2017). Blue arrow indicates the mean direction of slope displacement (Loew et al.2017). The contour interval is 10 m. 4.1 Polarization analysis All recordings were processed using time-frequency polarization analysis; representative results for three stations (PRE002, PRE004 and PRE012) are shown in Fig. 3. Ground motion at station PRE002, located within the stable area of the slope, showed no strong polarization characteristics below 8 Hz, with the exception of narrowband features at 1.3, 4 and 8 Hz. These features are likely caused by anthropogenic sources, which dominate the noise wavefield at these frequencies. Spectral analysis revealed nearly harmonic signals at the three frequencies (see later Fig. 4), which are unlikely to have natural origin. Since polarization attributes at these frequencies may be representative of the source and not the site response, we have excluded these frequencies from our analysis. The weak directionality between 2–4 Hz might be attributed to back-radiation from the unstable area. The ellipticity is flat in this frequency range and close to 0.4, what is typical for rock sites (Burjánek et al.2014). In contrast, the strong high-frequency directionality (>8 Hz) might be related to a very local structure, for example, to a small talus cone (see rock debris located east from PRE002 in Fig. 2). Similar observations were made for station PRE001. In contrast, the remaining stations of the array showed strongly polarized and directional ground motion between 3 and 4 Hz (Figs 3b–c, e–f, h–i). Ellipticity drops below 0.1 for these stations, and the average azimuth in this frequency range is 62°, which corresponds well with the predominant slope deformation direction of 60° (Loew et al.2017). Nevertheless, there are subtle variations in the location of the concentration peak (in terms of frequency and mean strike), which cluster within the array: (1) PRE003, PRE004 and PRE010 show the peak at 3.43 Hz and 68° ± 1°; (2) PRE006 and PRE009 show the peak at 3.43 Hz but 61° ± 1°; (3) PRE008, PRE011 and PRE012 show the peak at 3.2 Hz and 57° ± 1°; (4) PRE007 shows the peak at 3.55 Hz and 62° (which is close to the second group of the stations); and (5) PRE005 shows the peak at 3.09 Hz and 54°. While interpretation of these subtle differences is not straightforward as the stations do not form clear spatial patterns (e.g. PRE010 within the first group), we find them significant and discuss this result in Sections 4.2 and 6. The average dip ranges from −4° to 13° downward from the horizontal with median value of 4°. The larger dip values (8°–13°) cluster approximately along a profile PRE003-PRE006-PRE008-PRE010-PRE011, with the maximum value for PRE006. The concentration peaks are broader compared to the strike angle distributions. Figure 3. View largeDownload slide Polarization analysis of 1 hr recording of ambient vibrations at three stations: PRE002 (a,d,g), PRE004 (b,e,h), PRE012 (c,f,i). Colour scales represent the relative frequencies of occurrences of strike (a–c), dip (d–f) and ellipticity (g–i). Colour intensity is decreased for frequency bands affected by anthropogenic sources (1.2, 1.3, 4 and 7.9 Hz). Estimated modes of the Beta distributions representing the observed ellipticity values are connected by a black solid line. Figure 3. View largeDownload slide Polarization analysis of 1 hr recording of ambient vibrations at three stations: PRE002 (a,d,g), PRE004 (b,e,h), PRE012 (c,f,i). Colour scales represent the relative frequencies of occurrences of strike (a–c), dip (d–f) and ellipticity (g–i). Colour intensity is decreased for frequency bands affected by anthropogenic sources (1.2, 1.3, 4 and 7.9 Hz). Estimated modes of the Beta distributions representing the observed ellipticity values are connected by a black solid line. Figure 4. View largeDownload slide Mean power spectral density of ambient vibrations from stations PRE002 (black) and PRE006 (red). Geometrical mean of the two horizontal components is plotted. Figure 4. View largeDownload slide Mean power spectral density of ambient vibrations from stations PRE002 (black) and PRE006 (red). Geometrical mean of the two horizontal components is plotted. In general, most of the stations show strong peaks in the high-frequency range as well. However, these peaks are variable and not present consistently at all stations. For example, a peak at 5.4 Hz is present for PRE003 and PRE004, while points in different direction for PRE005 and is missing completely for other stations. These peaks represent likely higher-modes or local resonances, nevertheless, the interpretation is difficult. In comparison, Bottelin et al. (2013a) identified higher modes for three sites with column-like geometry, while no clear relation between spectral peaks was found at a site with similar geometrical setting as AdR (Bottelin et al.2013a; Fig. 3c). 4.2 Relative amplification Power spectral densities (PSDs) of noise and earthquake recordings were determined using the multi-taper method (Prieto et al.2009) for all stations of the array. An example of the noise PSD for two stations (PRE002 and PRE006) is shown in Fig. 4, demonstrating capabilities of the multitaper method: the PSDs are smooth but spectral resolution is preserved. Therefore, it is possible to identify narrow spectral peaks at 1.3, 4, and 8 Hz. These peaks likely originate from anthropogenic sources as discussed previously. Station PRE006 shows a broad peak centred at 3.4 Hz, which corresponds well to the polarization peak, while station PRE002 (located just less 50 m away) show just a little increase of seismic energy just below 3 Hz. Relative amplification functions (i.e. spectral amplitude ratios) with respect to reference station PRE002 were calculated for both earthquake and noise recordings. For the earthquake recordings, we analysed a single 10 s window for the ML2.9 event and two consecutive 10 s windows for the ML3.0 and ML4.0 events with the highest signal to noise ratios. Comparison of the 62° horizontal component of ground motion (i.e. crack perpendicular) is presented in Fig. 5. Spectral ratios are close to unity below 2 Hz for earthquakes and ambient vibration data, that is, ground motion is identical below 2 Hz for all stations in the array. Spectral ratios remain close to unity up to 10 Hz for station PRE001, but increase significantly above 2 Hz for other stations. Amplification functions from ambient vibration and earthquake data are in good agreement, however peak amplifications for earthquakes are slightly higher than ambient vibrations, which may result from the low number and short duration of the analysed time windows, or from a difference in the incoming wavefield (surface waves rich noise recordings versus S-waves rich earthquake recordings). Amplifications along the 152° horizontal component (parallel to the crack) and vertical directions (Fig. S3) are lower (<10). Figure 5. View largeDownload slide Site-to-reference spectral ratios for the 62° horizontal component of ground motion resulting from the ML2.9 (red), ML3.2 (green), ML4.0 (blue) earthquakes of 2012 May 25, compared to spectral ratios resolved from ambient vibrations (black). Station PRE002 was used as the reference. Figure 5. View largeDownload slide Site-to-reference spectral ratios for the 62° horizontal component of ground motion resulting from the ML2.9 (red), ML3.2 (green), ML4.0 (blue) earthquakes of 2012 May 25, compared to spectral ratios resolved from ambient vibrations (black). Station PRE002 was used as the reference. Details of spectral amplification results for ambient vibration data are shown in Fig. 6. Local maxima can be linked to the dominant polarization features (see previous section). In particular, the first local maximum for all stations (except PRE001) is coincident with the polarization peak at f0 = 3.2 Hz. The peak at 3.5 Hz consists of two not completely distinct local maxima, which match the f1 = 3.43 Hz and f2 = 3.55 Hz polarization peaks. Peak amplification of ambient ground motion reaches a factor of 35 at 3.55 Hz for station PRE006. The apparent drop in some of the amplification functions at 4 Hz is due to the anthropogenic source; assumptions required for calculating spectral ratios are likely not fulfilled for this source (e.g. it is too close). In particular, the path propagation effects (e.g. geometrical spreading), or source effects (e.g. radiation pattern) cannot be neglected across the array. The response becomes highly variable above 4 Hz. Figure 6. View largeDownload slide Mean site-to-reference spectral ratios of ambient vibrations for the 62° horizontal component of ground motion. Station PRE002 was used as the reference. Dashed black lines denote the identified resonant frequencies from polarization analysis (3.2, 3.43 and 3.55 Hz). Figure 6. View largeDownload slide Mean site-to-reference spectral ratios of ambient vibrations for the 62° horizontal component of ground motion. Station PRE002 was used as the reference. Dashed black lines denote the identified resonant frequencies from polarization analysis (3.2, 3.43 and 3.55 Hz). Amplifications along the 152° horizontal component (parallel to the crack) show a weak peak for a frequency of 4.1 Hz, and stronger peaks at 5.1 and 6 Hz respectively for most of the stations. These might represent higher perpendicular modes with respect to the fundamental mode. Moreover, a peak at 3.2 Hz is present for this component at station PRE005, which corresponds to the slight rotation of the fundamental peak observed in the polarization analysis. Amplifications in vertical directions show similar peaks as for the crack perpendicular directions, which is in agreement with the dip component identified by polarization analysis. Interpolated maps of amplification factors at f0 = 3.2 Hz and f1 = 3.45 Hz are presented in Fig. 7. The amplification values follow spatial trends; amplification is lower in stable area of the slope, increases towards the east in the unstable area for both frequencies, but is not largest at stations closest to the cliff edge (PRE009, PRE011 and PRE012). At 3.2 Hz, the area of greatest amplification trends NNW within the unstable area following a wide tension crack. The maximum factor of 20 is reached at stations PRE005, PRE006 and PRE008. At 3.45 Hz, amplification increases in the central part of the array within the unstable area. The maximum factor of 35 is reached at station PRE006. An amplification map for f2 = 3.55 Hz is not presented here as it is almost identical to the one for f1 = 3.45 Hz. Figure 7. View largeDownload slide Amplification map at frequencies of 3.2 Hz (a) and 3.45 Hz (b) for the 62° horizontal component of ground motion. Array stations are shown as black dots with red outline, while semi-permanent stations are indicated by red squares. Green and red lines represent the boundary of the instability and the main scarp of the 2012 rockfall, respectively (after Loew et al.2017). Figure 7. View largeDownload slide Amplification map at frequencies of 3.2 Hz (a) and 3.45 Hz (b) for the 62° horizontal component of ground motion. Array stations are shown as black dots with red outline, while semi-permanent stations are indicated by red squares. Green and red lines represent the boundary of the instability and the main scarp of the 2012 rockfall, respectively (after Loew et al.2017). 5 SEMI-PERMANENT STATIONS Two semi-permanent stations (PREO1 and PREO2) were deployed at AdR in summer 2012 to monitor changes in the slope's seismic response. These consist of a three-component velocimeter with eigenperiod of 1 s, recording unit, power supply, and a communication device that streams data to the Swiss Seismological Service. Station locations (Fig. 2) were selected based on various factors. The goal was to have one sensor in the unstable area and the second with the rest of the equipment in the stable area. It was also necessary to minimize the risk of losing equipment in case of a slope failure, which was not unexpected in summer 2012. Station PREO1 was thus setup in the presumably stable part, close to the position of array station PRE004, in order to explore the possibility of monitoring changes within the unstable rock mass outside the instability boundary, while Station PREO2 was placed in the southern, less active part of the instability where it could be retrieved in case of accelerating displacements. The installation was operational for only 10 d when it was destroyed by lightning. It was reinstalled later and fully operational from November 2012 to September 2013, when it was again destroyed by lightning. A new epoch of monitoring then began in December 2014. Station PREO1 was abandoned, and a new station PREO3 was setup in the stable area (20 m south of PREO1, see Fig. 2). This move was required for a new system design that provides more robust lightning protection. Station PREO2 was reinstalled at the original location. Unfortunately, the installation suffered from low signal-to-noise ratios, so further changes had to be adopted: station PREO3 was replaced by station PREO4 in the stable area (the sensor was moved a few metres away from a communication antenna that was thought to be generating interference). This configuration has been operating successfully since June 2015. 5.1 Monitoring of polarization We applied polarization analysis to all available continuous recordings from the semi-permanent stations. Continuous three-component time traces were split into non-overlapping windows of 60 min, polarization analysis was applied for each time window, and frequency-dependent distributions of the polarization attributes were estimated with the maximum likelihood method. Fig. 8 shows the concentration of strike angle for station PREO2 during the period between June 2015 and April 2016. The strong source effects visible on the PSD time history (Fig. S4) are removed by the polarization analysis. The concentration of strike angle for station PREO1 (not shown here) shows very similar patterns during the period between November 2012 and September 2013. Concentration measures the scatter of the azimuth distribution; zero concentration results in a uniform distribution (no preferential direction), while concentration of unity results in the Dirac delta function (strictly unidirectional distribution). We identified two bands (3–4 Hz and 7–9 Hz) of high strike concentration for station PREO2 (Fig. 8). We then attempted to pick concentration maxima for each time window to track potential changes in the resonant frequencies (black line, Fig. 8). The resulting curve was smoothed with moving average of 2 d. The picked values oscillate in the range of 3.3–3.8 Hz, which likely corresponds to the frequency doublet f1 = 3.43 Hz, f2 = 3.55 Hz (Fig. 6). Similarly, we picked the maximum concentration of the dip angle and the minimum of ellipticity (see Fig. 3e), which could also serve as proxies for drift of the resonant frequencies. Comparison of these values for stations PREO1, PREO2 and PREO4 is shown in Fig. 9(a) (smoothed with a 2 d moving average). Frequency picks based on ellipticity correlate well with the picks based on the concentration of strike. Frequency picks based on the concentration of dip show higher values, especially in winter/spring periods (up to 0.1 Hz), which is consistently observed on both sensors (PREO1/PREO2, resp. PREO4/PREO2). Nevertheless, the overall trends correspond well and contain meaningful signal. The time history of mean strike at the identified resonance frequency for both PREO1/PREO2 and PREO4/PREO2 couples is shown in Fig. 9(b). The mean directions correlate well for both sensors, with 4° systematic difference. This difference has decreased by 1° since January 2016. A similar difference was observed previously from the array stations (6°–7° difference for PRE004 versus PRE007 or PRE009). Figure 8. View largeDownload slide Concentration of strike angles for PREO2 (in colour). Maxima of the concentration are connected by the black line, which represents a variation of the dominant resonant frequency f1 ∼ 3.45 Hz. Figure 8. View largeDownload slide Concentration of strike angles for PREO2 (in colour). Maxima of the concentration are connected by the black line, which represents a variation of the dominant resonant frequency f1 ∼ 3.45 Hz. Figure 9. View largeDownload slide Variation of the fundamental frequency picked from (a) polarization parameters and (c) from the power spectral density (62° component), ranges of the picked resonance frequency from polarization parameters are shown in blue. (b) Mean polarization azimuth at the fundamental frequency picked from the strike angle distribution. Picking was performed in a frequency range of 2.5–3.9 Hz. Figure 9. View largeDownload slide Variation of the fundamental frequency picked from (a) polarization parameters and (c) from the power spectral density (62° component), ranges of the picked resonance frequency from polarization parameters are shown in blue. (b) Mean polarization azimuth at the fundamental frequency picked from the strike angle distribution. Picking was performed in a frequency range of 2.5–3.9 Hz. 5.2 Monitoring of PSDs PSDs of continuous recordings were calculated using the multi-taper method (Prieto et al.2009). Seismograms were split into non-overlapping windows of 60 min length and processed in the same manner as the array recordings. A shorter time window of 50 s was used for the subsequent windowing of hourly recordings to obtaining a smoother PSD. Time history of the square root of PSD for station PREO2 shows strong regular daily and weekly oscillations (Fig. S4) suggesting that the incoming noise wavefield has an origin in human activity. Nevertheless, since the dominant resonant frequencies can be easily identified from the PSD (e.g. Fig. 4), we tracked the local maximum for each time window (60 min) as a means of resolving temporal changes in the resonant frequencies, and compared these with picks from our polarization analysis (Fig. 9c). A similar approach was applied by Lévy et al. (2010) and Bottelin et al. (2013b). Frequency picks based on the PSD are in reasonable agreement with those based on polarization analysis. There is also excellent agreement between PSD frequency picks for the station couples (PREO1/PREO2, resp. PREO4/PREO2) in all other winter/spring periods. A small systematic difference in the summer/autumn periods may be attributed to the frequency doublet (f1 = 3.43 Hz, f2 = 3.55 Hz) identified earlier, where stations in the stable area (PREO1, PREO4) prefer f1, while the station in the unstable area (PREO2) prefers f2. Since the multitaper method has better spectral resolution than our polarization analysis, we further analyse only picks based on the PSD in our interpretation of frequency variations. 5.3 Monitoring of relative amplifications Relative amplification functions (SRSR) for station PREO2 with respect to PREO1 (later station PREO4) were determined for all available continuous recordings. Examples from different time windows are shown in Fig. 10: (1) SRSR for an hour of ambient vibrations recorded in November 2012; (2) SRSR for a 250 km distant ML5.3 earthquake recorded in June 2013 (strongest ground motion due to an earthquake recorded at the site); (3) SRSR for ambient vibrations preceding this earthquake; (4) SRSR for ambient vibrations recorded at the closest array stations (PRE009/PRE004) in May 2012. All spectral ratios are in close agreement. The level of excitation at the site was about 150 times stronger (127 μm peak-to-peak displacement) for the ML5.3 earthquake compared to the amplitude of ambient vibrations preceding the event (0.8 μm peak-to-peak displacement). This demonstrates that the site response at AdR is linear over the measured range of weak input motions (two orders of magnitude), and has not undergone detectable irreversible change between May 2012 and June 2013. Figure 10. View largeDownload slide Site-to-reference spectral ratios for stations PREO2/PREO1 (62° component): ambient vibrations acquired in November 2012 (red); ML5.3 earthquake recording from June 2013 (green); corresponding pre-event noise (black). Site-to-reference ratios for ambient vibration recordings of array stations PRE009/PRE004 located close to the semi-permanent stations are shown for comparison (yellow). Figure 10. View largeDownload slide Site-to-reference spectral ratios for stations PREO2/PREO1 (62° component): ambient vibrations acquired in November 2012 (red); ML5.3 earthquake recording from June 2013 (green); corresponding pre-event noise (black). Site-to-reference ratios for ambient vibration recordings of array stations PRE009/PRE004 located close to the semi-permanent stations are shown for comparison (yellow). Since PREO1 is also sensitive to the response of the unstable slope, it is difficult to interpret spectral ratios for PREO2/PREO1. The same holds true for the station combination PREO2/PREO4. Spectral ratios for PREO2/PREO1 (Fig. 10) are dominated by relative deamplification at ∼5.5 Hz, which results from the amplification of ground motion at PREO1 at this frequency (compare to PRE004 in Fig. 5). The peak close to 3.8 Hz (Fig. 10) cannot be related to any of the resonant frequencies; it is an apparent peak that results from a combination of the 5.5 Hz peak at PREO1 and higher amplification of PREO2 in the 2–4 Hz frequency band. Despite these complications, interpretation of time variations of these spectral ratios provides useful insight into seasonal changes of the dynamic behaviour. An example of the PREO2/PREO1 SRSR for November 2012 to September 2013 is presented in Fig. 11. The strong source effects visible on the PSD time history (Fig. S4) are removed by SRSR. In November 2012, the response showed almost no variation, that is, remained close to the SRSR presented in Fig. 10. However, the response changed abruptly In December 2012: below 4 Hz, SRSRs decreased close to unity (the response is almost identical for the two stations), while amplification appeared around 4.9 Hz and the overall response was shifted towards higher frequencies. Since then, the pattern has not varied notably below 6 Hz, but shows complexities at higher frequencies. At the end of the spring 2013, the response changed slowly returning to the pattern similar to that from November 2012 (Fig. 10). Similar trends have been observed during the period June 2015 to April 2016 (Fig. S5). White bands in Fig. 11 represent periods with data gaps (e.g. February 2013, May—July 2013). Moreover, there are several short periods when PREO2 recordings exhibit heavy disturbance, which are not present at PREO1 (creating the dark red bands below 2 Hz), which may be related to local weather conditions (rain, wind) at the station. The results are not biased by source effects of local earthquakes, since the region presents a low seismicity for the presented monitoring period. For example, there was just a single event with a magnitude >2 located within a 50 km range from the stations for the monitoring periods presented here. Figure 11. View largeDownload slide PREO2/PREO1 spectral ratio for the 62° horizontal component of ground motion (in colour). Variation of the fundamental frequency picked from the power spectral density is shown for PREO1 by the black line and for PREO2 by the grey line. Figure 11. View largeDownload slide PREO2/PREO1 spectral ratio for the 62° horizontal component of ground motion (in colour). Variation of the fundamental frequency picked from the power spectral density is shown for PREO1 by the black line and for PREO2 by the grey line. In a follow up experiment, we reinstalled array stations PRE002, PRE006 and PRE011 for 1 hr in June 2015 to explore potential variability in the other areas of the site. The resulting SRSRs are shown in Fig. S6, displaying very good agreement with those observed in May 2012. 5.4 Comparison with other monitoring data AdR is equipped with several crack extensometers, a rain gage, and temperature sensors all recording continuous data. Thus, it is possible to explore links between these observations and measured site response parameters. In particular, we analyse data from crack extensometers 4 and 7 (Fig. 2). Extensometer 7 measures opening of a crack that is the likely continuation of the main rear fracture, while extensometer 4 measures opening of a crack transecting the unstable mass between seismic station PREO2 and the cliff. Extensometer data following the May 2012 collapse are shown in Fig. 12. In general, extensometer 4 shows crack closure, which occurs mainly in winter/spring and appears to slow over monitoring period of 4 yr. Extensometer 7, on the other hand, shows gradual crack opening, which is also slowing. Local rainfall and temperature data were compared with those from the MeteoSwiss network station Locarno-Cimetta (NIME) located 15 km away at a similar altitude (100 m difference). Temperature and precipitation data correlate well at the two sites. As there are gaps in the data at AdR, we further analysed just meteorological data from Locarno-Cimetta. Figure 12. View largeDownload slide Crack extensometer data over the monitoring period. Time windows of the recorded and processed seismic data are shaded. Figure 12. View largeDownload slide Crack extensometer data over the monitoring period. Time windows of the recorded and processed seismic data are shaded. We compare drift of the identified resonant frequencies with other monitoring data (temperature, precipitation, crack opening) for the period November 2012 to September 2013 in Fig. 13. Broadly, we identify periods of both direct and inverse correlation between resonant frequencies and temperature change. The additive inverse of temperature (temperature multiplied by −1) is also included to highlight periods of inverse correlation. In particular, we observe inverse correlation in winter during freezing periods with a mean frequency delay of about 4 d (Fig. 14a), and direct correlation in summer–autumn without a mean frequency delay (Fig. 14b). These delays are apparent in Fig. 13, and precise values can be estimated by cross correlation (e.g. 4.8 d for the winter period presented in Fig. 14a; 0.1 d for the summer/fall period window presented in Fig. 14b). However, the exact value depends on selected time window. A shorter delay would be required to match the time-series toward the end of the winter (3.4 d), or at the beginning of the winter (2.8 d). The zero lag for the summer period is stable considering the time-series after 2012 July 15. Similar comparisons were made for the period July 2015 to April 2016 (Fig. S7), however the inverse correlation is weaker and spans a shorter time period during winter (freezing periods were shorter for the winter 2015/2016). The influence of precipitation on resonant frequencies is not clear, since the record of precipitation is more irregular than temperature. A decrease in resonant frequencies in April/May 2013 may be related to high precipitation, however, this has yet to be confirmed by further observations. Extensometers located in the vicinity of the seismic stations show closing of fractures during winter, which correlates with the onset of freezing temperatures in December 2012. The rate of crack closing decreases towards spring, followed by a period of crack opening which starts just after the end of the long freezing period (January–April 2013). The onset of the opening slightly precedes the period of higher precipitation and occurs as soon as temperatures increase above 0 °C, indicating possible correlation with snowmelt. Similar patterns with lower amplitudes are observed in the following winters (Fig. 12, Fig. S7). Figure 13. View largeDownload slide Comparison of resonant frequencies and other monitored field data for the period November 2012 to September 2013. (a) Variation of resonant frequencies (red, black), temperature (orange), and additive inverse of temperature (blue). (b) Daily precipitation. (c) Extensometer data. A 14 d moving average was applied to the resonant frequency and temperature time-series. Figure 13. View largeDownload slide Comparison of resonant frequencies and other monitored field data for the period November 2012 to September 2013. (a) Variation of resonant frequencies (red, black), temperature (orange), and additive inverse of temperature (blue). (b) Daily precipitation. (c) Extensometer data. A 14 d moving average was applied to the resonant frequency and temperature time-series. Figure 14. View largeDownload slide Relationship between temperature and resonant frequency time histories for: (a) winter period; temperature time history (light blue) was delayed by 4 d, temperature axis is oriented downwards. (b) Summer period; temperature time history (orange) was not delayed, temperature axis is oriented upwards. A 14 d moving average was applied on all presented time-series. Resonant frequencies are taken from Fig. 9(c). Figure 14. View largeDownload slide Relationship between temperature and resonant frequency time histories for: (a) winter period; temperature time history (light blue) was delayed by 4 d, temperature axis is oriented downwards. (b) Summer period; temperature time history (orange) was not delayed, temperature axis is oriented upwards. A 14 d moving average was applied on all presented time-series. Resonant frequencies are taken from Fig. 9(c). 6 DISCUSSION Through a field campaign of in-situ ambient seismic characterization and monitoring, we have resolved salient information on the site response characteristics of a large unstable rock slope close to collapse. We find that ground motion at the AdR site is strongly amplified (spectral ratios up to 35) within the unstable portion of the slope with respect to the stable areas. Such strong amplification may result in previously unanticipated large deformations during strong shaking, likely influencing the potential for earthquake triggered damage and failure. Amplification is strongest in a narrow frequency band between 3–4 Hz, where we have identified three resonant frequencies (3.2, 3.43 and 3.55 Hz) with the aid of multitaper processing (Prieto et al.2009). Spectral amplifications are strongest in the middle of the array and towards the northern portion of the slope that collapsed in 2012, where the aperture of tension cracks becomes wider as compared to the southern part of the instability. Thus, steeply dipping open fractures appear to have a stronger influence on the local site response than the nearby cliff free-face, in contrast to Burjánek et al. (2012) where the strongest amplification was observed at the cliff edge at a different unstable rock slope. We hypothesize that strong amplification factors, as well as spatial heterogeneity of polarization and amplification, are indicative of a high degree of internal fracturing at AdR and high criticality with respect to future failure, as also suggested by related recent numerical studies (Gischig et al.2016) and empirical studies (Kleinbrod et al.2017b). The high degree of internal fracturing may result in a reduction of shear wave velocities, which would contribute to the observed amplification as well. However, no seismic velocity measurements are available for this site. The resonant modes include significant vertical motion (Fig. S3b), so that horizontal-to-vertical spectral ratios (H/V curves) would have limited use in the seismic velocity characterization of AdR. In particular, H/V curves at AdR would be controlled by the observed resonance (2-D/3-D effects) and could not be interpreted using 1-D media per common procedures in earthquake engineering applications. Array stations PRE003 and PRE004, as well as semi-permanent stations PREO1 (later PREO4), also exhibit strong directional amplification despite being located on the presumably stable portion of the slope. This observation may be explained by back radiation of seismic energy from the unstable area to the stable areas in the presence of strong resonance. In contrast, however, BU10 and BU12 showed that directional amplification diminished almost completely at the border of other studied rock slope instabilities. Therefore, we cannot exclude the presence of another crack located southwest from the main scarp, which is currently not visible at the ground surface. Nonetheless, we demonstrate it is feasible in this case to detect signals of the unstable slope response using stations located in adjoining stable areas, which is useful in this and future experiments to help protect valuable seismic monitoring equipment in the case of slope failure. The three dominant resonant frequencies identified at AdR (3.2, 3.43 and 3.55 Hz) are relatively close (Fig. 6), and the corresponding mode shapes would likely be very similar (Fig. 7), that is, the amplitudes of the three peaks follow the same trends (Fig. 6). Therefore, the three peaks likely do not correspond to three different vibrational modes, rather we hypothesize that the fundamental mode is split. Normal-mode splitting is a well-known phenomenon, for example, observed normal modes of the earth are split due to earth's rotation and small-scale lateral heterogeneities (e.g. Stein and Wysession 2003). Thus, we speculate that splitting of the resonant frequencies observed at AdR may be caused by scattering produced by small-scale internal heterogeneities. The three resonant frequencies were also identified by polarization analysis. Ground motions are almost linearly polarized at these frequencies in the direction parallel to observed slope displacement (60° azimuth; Loew et al.2017). Small differences (54°–68°) exist, however, in the mean orientations at the resonant frequencies at different stations. These likely reflect local complexities of the vibrational mode shapes. Moreover, polarization analysis is not so well suited for characterization of the torsional motion, which may also be present at AdR and result in the variations of the mean orientations. The fundamental frequency of 3.2–3.5 Hz fits quite well with the physical extension of the unstable rock mass. In particular, we applied a formula for the fundamental frequency of a bending mode of a rock column proposed by Valentin et al. (2017). We assumed lateral extension of 30 m and Young's modulus of 20 GPa (Willenberg et al.2009), which resulted in a fracture depth range of 40–45 m. Although, the exact depth of the fracture network is not known, this estimate is in a rough agreement with the geological profile (Fig. 1b). We monitored the fundamental frequency of the AdR rock slope (i.e. the 3.43–3.5 Hz doublet peak) by means of polarization analysis and PSDs (Fig. 9). Frequency picks from polarization analysis show scatter and are biased by anthropogenic noise. PSDs, on the other hand, were better resolved due to the higher spectral resolution of the multi-taper method. Nevertheless, the general observed trends were the same for the both methods. Moreover, monitoring of the mean azimuth by polarization analysis provided additional information: we identified a potential irreversible change in azimuth in January 2016 (see Fig. 9b), which was not apparent in the resonant frequency or extensometer time-series, indicating that polarization angles may provide a new, sensitive measure of changes in the slope's seismic response capable of identifying changes related to damage. In general, the mean orientation of vibration at the fundamental frequency is controlled by the shape of the normal mode. Nevertheless, the physical interpretation of such change would require a detailed mechanical model and re-measurement of more points to map the extent of the change. We identified seasonal variations in amplification levels and resonant frequencies at AdR. Response of the two semi-permanent stations is similar up to 4 Hz in winter, when the doublet peak (3.43–3.55 Hz) disappears (see Fig. 13 and Fig. S7) and amplification levels become nearly equal. Thus, scattering due to small-scale internal heterogeneities is likely weaker in winter. The high frequency response (>4 Hz) reveals changes that are correlated with temperature (Figs 8 and 11). Amplification peaks are shifted toward higher frequencies with decreasing temperature. The onset of these variations is sharp and related to the onset of freezing, which lasted from December 2012 until the beginning of April 2013 (with a short break in the beginning of January; Fig. 11). Snow cover at the site, which we monitored from time-lapse photographs, was relatively low on average during that time, and we did not identify a relationship between snow cover and the seismic response. On the other hand, some observed changes may again be related to temperature. For example, a peak present at 6.7 Hz during February 2013 (Fig. 11) disappears abruptly around March 10, just after a short warm period (temperatures reached 0°C). The response does not change much in summer as compared to winter, and the transition from the ‘winter-type’ to the ‘summer-type’ response is smooth (April–May). Therefore, the mechanism is likely different compared to the rapid changes related to the onset of freezing. We did not observe any systematic changes in the slope's seismic response for summer and autumn periods during our monitoring (e.g. Fig. 10), despite an expected fluctuation in the response due to weather conditions. This was further confirmed by remeasurement of several array stations in June 2015 (Fig. S6). Observed resonant frequency drifts at AdR are linked to the temperature time histories, however, the relationship is complex. Response of the resonant frequencies depends on the season (freezing/non-freezing periods) and on the time scale of observations (days, weeks, months). Resonant frequencies are inversely correlated with air temperature during the freezing periods with a delay of about 4 d (Fig. 14a). This trend is present for observational time scales of several weeks (Figs 13 and 14), but disappears at shorter time scales. A sharp increase of the fundamental frequency of an unstable rock column with a drop in air temperature below 0 ° C was observed by Bottelin et al. (2013b), and explained by stiffening of a main rear fracture by ice formation. Similarly, at our site an increase in resonant frequencies with decreasing temperatures may be related to stiffening of the rock mass due to freezing of interstitial water in pores and small-scale fractures (as also hypothesized by Starr et al.2015) or in the infill in the open tension fractures. The time shift between temperature changes and resonant frequencies could be explained by the diffusive temperature (and freezing) front that has to penetrate a certain depth before having an effect on the overall rock mass stiffness. This would also be in agreement with the disappearing of the effect on shorter time-scales; daily or shorter-term temperature changes penetrate only few decimetres and thus do not contribute to the advancing of the freezing front a great depth. An alternative explanation may be that resonant frequencies increase due to a decrease in mass of the resonating rock slope related to gradual drainage of pore water during cold periods. However, this mechanism would result in a gradual increase of resonant frequencies while temperatures are below 0° C, and a sharp decrease as soon as melt water becomes available as temperatures increase. Instead, resonant frequencies at AdR appear to strongly correlate with temperature. In summer periods, resonant frequency variations at AdR are directly correlated with air temperature without delay. This trend is present for longer observational time scales of weeks (Figs 13 and 14), while a similar trend is also observed for shorter time scales (days) during dry periods (e.g. July/August 2015) again with almost no delay. A similar increase in the fundamental frequency of an unstable rock column with temperature at non-freezing daily periods was observed by Lévy et al. (2011) and Bottelin et al. (2013a,b). We propose that increasing thermoelastic stresses accompanying rock warming lead to closure of micro-cracks and thus to an overall stiffening of the rock mass, leading to an increase in resonant frequencies (as in Starr et al.2015). Such a thermo-elastic effect would produce faster reaction on temperature changes leading to shorter delay and also reactions observable at shorter-time scale, unlike the winter-time stiffening mechanism related to ice formation mentioned earlier. Lastly, we observed a decrease in resonant frequencies in April/May 2013 that does not correlate with air temperature. The decrease is possibly related to a rising water table due to snow melt in the broader area (not directly at the site) and higher precipitation rates (Fig. 13). Similar observations were made by Bottelin et al. (2013a). In summary, we observe two opposite and competing temperature related effects: stiffening the rock mass due to ice formation as the temperatures decrease, dominant in winter, and stiffening of the rock mass by micro-crack closure as temperatures increase, dominant in summer. The additional role of groundwater recharge may become apparent in a transitional phase in spring, when the two competing temperature-related effects are offset. We emphasize that all observed correlations with temperature are indirect. 7 CONCLUSIONS AdR is a highly disrupted unstable rock slope with a long history of catastrophic failures (Loew et al.2017). The slope will likely undergo future failures, so it provides a unique opportunity to study changes in the seismic response corresponding with the evolution of internal damage. As application of ambient vibration monitoring and characterization of unstable rock slopes becomes more prevalent, it is critical to develop refined methods and test the resolution and capabilities of these techniques as a means of detecting changes in physical properties. In this study, we measured much larger bedrock site amplifications and polarization levels than in previous studies (BU10, BU12), which have implications for co-seismic triggering and damage. Interestingly, we find that amplification at AdR is not strongest at the cliff edge, as in previous studies, but rather controlled by the internal structure of the rock mass (in particular the presence of compliant fractures). We also importantly observed that the seismic response of AdR is identical across a wide range of weak input motions, helping clarify the range of validity for modelled using linear media. Amplification levels and resonant frequencies at AdR exhibit seasonal variations related (both directly and indirectly) to temperature changes. In summer, resonant frequencies increase with increasing temperature as thermal stresses drive closure of microcracks and bulk stiffening of the rock mass. Meanwhile in winter, resonant frequencies increase with decreasing temperatures associated with the formation of pore and crack ice stiffening the rock mass. The seismic response of AdR has not changed significantly over our monitoring period, suggesting the internal structure has not sustained large irreversible changes since May 2012. This result is meaningful in the ongoing hazard evaluation. We show that ambient vibration measurements are repeatable, provide quantitative information on the internal composition of an unstable rock mass, and readily useful for monitoring slope change over time. Nevertheless, environmental conditions (such as temperature) and associated changes in seismic response parameters must also be characterized in any hazard monitoring interpretation. The ongoing long-term seismic monitoring at AdR has provided a unique data set necessary for correct interpretation of potential seismic signatures preceding future failure. Acknowledgements We thank the Ticino cantonal geologist Giorgio Valenti and colleagues Marco Franzi and Marco Andretta for providing access to temperature and extensometer data, in addition to their support for seismic measurements and monitoring without which this study would not have been possible. Meteorological data were provided by MeteoSwiss. Research presented in this study was funded by ETH Zurich project ETH-1212-2 (Characterization of unstable rock slopes through passive seismic measurements). REFERENCES Bottelin P.et al.   ( 2013a). Spectral analysis of prone-to-fall rock compartments using ambient vibrations, J. Environ. Eng. Geophys. , 18, 205– 217. Google Scholar CrossRef Search ADS   Bottelin P., Lévy C., Baillet L., Jongmans D., Guéguen P. ( 2013b). Modal and thermal analysis of Les Arches unstable rock column (Vercors massif, French Alps), Geophys. J. Int. , 194, 849– 858. Google Scholar CrossRef Search ADS   Burjánek J., Gassner-Stamm G, Poggi V., Moore J.R., Fäh D ( 2010). Ambient vibration analysis of an unstable mountain slope, Geophys. J. Int. , 180, 820– 828. Google Scholar CrossRef Search ADS   Burjánek J., Moore J.R, Yugsi-Molina F.X., Fäh D ( 2012). Instrumental evidence of normal mode rock slope vibration, Geophys. J. Int. , 188, 559– 569. Google Scholar CrossRef Search ADS   Burjánek J., Edwards B, Fäh D ( 2014). Empirical evidence of local seismic effects at sites with pronounced topography: A systematic approach, Geophys. J. Int. , 197, 608– 619. Google Scholar CrossRef Search ADS   Del Gaudio V., Muscillo S., Wasowski J. ( 2014). What we can learn about slope response to earthquakes from ambient noise analysis: An overview, Eng. Geol. , 182, 182– 200. Google Scholar CrossRef Search ADS   Fäh D.et al.   ( 2012). Coupled seismogenic geohazards in Alpine region, Boll. Geofis. Teor. Appl. , 53, 485– 508. Gischig V., Amann F., Moore J.R., Loew S., Eisenbeiss H., Stempfhuber W. ( 2011). Composite rock slope kinematics at the current Randa instability, Switzerland, based on remote sensing and numerical modeling, Eng. Geol. , 118, 37– 53. Google Scholar CrossRef Search ADS   Gischig V.S., Eberhardt E., Moore J.R., Hungr O. ( 2015). On the seismic response of deep-seated rock slope instabilities — Insights from numerical modeling, Eng. Geol. , 193, 1– 18. Google Scholar CrossRef Search ADS   Gischig V., Preisig G., Eberhardt E. ( 2016). Numerical investigation of seismically induced rock mass fatigue as a mechanism contributing to the progressive failure of deep-seated landslides, Rock Mech. Rock Eng. , 49, 2457– 2478. Google Scholar CrossRef Search ADS   Green A.G., Maurer H.R., Spillmann T., Heincke B., Willenberg H. ( 2006). High-resolution geophysical techniques for improving hazard assessments of unstable rock slopes, Leading Edge , 25, 311– 316. Google Scholar CrossRef Search ADS   Havenith H.-B., Jongmans D, Faccioli E., Abdrakhmatov K., Bard P.-Y. ( 2002). Site effect analysis around the seismically induced Ananevo rockslide, Kyrgyzstan, Bull. seism. Soc. Am. , 92, 3190– 3209. Google Scholar CrossRef Search ADS   Heincke B., Green A.G., Kruk J.V.D., Horstmeyer H. ( 2005). Acquisition and processing strategies for 3D georadar surveying a region characterized by rugged topography, Geophysics , 70, K53– K61. Google Scholar CrossRef Search ADS   Heincke B., Maurer H., Green A.G., Willenberg H., Spillmann T., Burlini L. ( 2006). Characterizing an unstable mountain slope using shallow 2D and 3D seismic tomography, Geophysics , 71, B241– B256. Google Scholar CrossRef Search ADS   Jaboyedoff M., Oppikofer T., Abellán A., Derron M.H., Loye A., Metzger R., Pedrazzini A. ( 2012). Use of LIDAR in landslide investigations: a review, Nat. Hazards , 61, 5– 28. Google Scholar CrossRef Search ADS   Kleinbrod U., Burjánek J., Hugentobler M., Amann F., Fäh D ( 2017a). A comparative study on seismic response of two unstable rock slopes within same tectonic setting but different activity level, Geophys. J. Int. , 211, 1428– 1448. Google Scholar CrossRef Search ADS   Kleinbrod U., Burjánek J., Fäh D. ( 2017b). On the seismic response of instable rock slopes based on ambient vibration recordings, Earth Planets Space , 69, 126, pp. 9, doi:10.1186/s40623-017-0712-5. Google Scholar CrossRef Search ADS   Kos A., Amann F., Strozzi T., Delaloye R., von Ruette J., Springman S. ( 2016). Contemporary glacier retreat triggers a rapid landslide response, Great Aletsch Glacier, Switzerland, Geophys. Res. Lett. , 43, 12 466–12 474. Google Scholar CrossRef Search ADS   Lévy C., Baillet L., Jongmans D., Mourot P., Hantz D., ( 2010). Dynamic response of the Chamousset rock column (Western Alps, France), J. geophys. Res. , 115, F04043, doi:10.1029/2009JF001606. Lévy C., Jongmans D., Baillet L. ( 2011). Analysis of seismic signals recorded on a prone-to-fall rock column (Vercors massif, French Alps), Geophys. J. Int. , 186, 296– 310. Google Scholar CrossRef Search ADS   Loew S., Gschwind S., Gischig V., Keller-Signer A., Valenti G. ( 2017). Monitoring and early warning of the 2012 Preonzo catastrophic rockslope failure, Landslides  14, 141– 154. Google Scholar CrossRef Search ADS   Moore J.R., Gischig V, Burjánek J., Loew S., Fäh D ( 2011). Site effects in unstable rock slopes: dynamic behavior of the Randa instability (Switzerland), Bull. seism. Soc. Am. , 101, 3110– 3116. Google Scholar CrossRef Search ADS   Moore J.R.et al.   ( 2016). Anthropogenic sources stimulate resonance of a natural rock bridge, Geophys. Res. Lett ., 43, 9669– 9676. Google Scholar CrossRef Search ADS   Prieto G.A., Parker R.L., Vernon F.L. ( 2009). A Fortran 90 library for multitaper spectrum analysis, Comput. Geosci. , 35, 1701– 1710. Google Scholar CrossRef Search ADS   Roten D., Cornou C., Fäh D., Giardini D. ( 2006). 2D resonances in Alpine valleys identified from ambient vibration wavefields, Geophys. J. Int. , 165, 889– 905. Google Scholar CrossRef Search ADS   Starr A.M., Moore J.R., Thorne M. S. ( 2015). Ambient resonance of Mesa Arch, Canyonlands National Park, Utah, Geophys. Res. Lett., 42, 6696– 6702. Stein S., Wysession M. ( 2003). An Introduction to Seismology, Earthquakes and Earth Structure , Blackwell Publishing, 498 pp. Valentin J., Capron A, Jongmans D., Baillet L., Bottelin P., Donze F., Larose E, Mangeney A. ( 2017). The dynamic response of prone-to-fall columns to ambient vibrations: comparison between measurements and numerical modelling, Geophys. J. Int. , 208( 2), 1058– 1076. Google Scholar CrossRef Search ADS   Willenberg H., Eberhardt E., Loew S., McDougall S., Hungr O. ( 2009). Hazard assessment and runout analysis for an unstable rock slope above an industrial site in the Riviera valley, Switzerland, Landslides , 6( 2), 111– 119. Google Scholar CrossRef Search ADS   SUPPORTING INFORMATION Supplementary data are available at GJI online. Figure S1. Array ML4.0 earthquake recording of 2012 May 25. The dominant wave group used for the SRSR estimation is coloured in red. The amplitude of the traces was normalized by maximum peak amplitude of the three components for each station. A peak amplitude is shown at the beginning of each trace. Figure S2. Fourier amplitude spectra of the dominant wave group (red) of the ML=4.0 earthquake (Fig. S1) used for SRSR estimation and Fourier amplitude of the pre-event noise (black). Figure S3. Mean site-to-reference spectral ratios of ambient vibrations for the 152° horizontal (a) and vertical (b) components of ground motion. Station PRE002 was used as the reference. Dashed black lines denote the identified resonant frequencies from polarization analysis (3.2, 3.43 and 3.55 Hz). Figure S4. Time history of the square root of power spectral density for station PREO2 and the 62° horizontal component of ground motion (in colour). The colour scale is saturated for values below 1st and above 99th percentile, respectively. Figure S5. PREO2/PREO4 spectral ratio for the 62° horizontal component of ground motion (in colour). Variation of the fundamental frequency identified from the power spectral density is shown for stations PREO4 (black line) and PREO2 (grey line). Figure S6. Comparison of observed SRSR in May 2012 and June 2015 for PRE006/PRE002 and PRE011/PRE002. Figure S7. Comparison of resonant frequencies and other monitored field data for the period July 2015 to April 2016. (a) Variation of resonant frequencies (red, black), temperature (orange) and additive inverse of temperature (blue). (b) Daily precipitation. (c) Extensometer data. A 14 d moving average was applied to the resonant frequency and temperature time-series. Please note: Oxford University Press is not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the paper. © The Authors 2017. Published by Oxford University Press on behalf of The Royal Astronomical Society. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Geophysical Journal International Oxford University Press

Ambient vibration characterization and monitoring of a rock slope close to collapse

Loading next page...
 
/lp/ou_press/ambient-vibration-characterization-and-monitoring-of-a-rock-slope-44xQTJkCfl
Publisher
The Royal Astronomical Society
Copyright
© The Authors 2017. Published by Oxford University Press on behalf of The Royal Astronomical Society.
ISSN
0956-540X
eISSN
1365-246X
D.O.I.
10.1093/gji/ggx424
Publisher site
See Article on Publisher Site

Abstract

Abstract We analyse the ambient vibration response of Alpe di Roscioro (AdR), an incipient rock slope failure located above the village Preonzo in southern Switzerland. Following a major failure in May 2012 (volume ∼210 000 m3), the remaining unstable rock mass (∼140 000 m3) remains highly fractured and disrupted, and has been the subject of intensive monitoring. We deployed a small-aperture seismic array at the site shortly after the 2012 failure. The measured seismic response exhibited strong directional amplification (factors up to 35 at 3.5 Hz), higher than previously recorded on rock slopes. The dominant direction of ground motion was found to be parallel to the predominant direction of deformation and perpendicular to open fractures, reflecting subsurface structure of the slope. We then equipped the site with two semi-permanent seismic stations to monitor the seismic response with the goal of identifying changes caused by internal damage that may precede subsequent failure. Although failure has not yet occurred, our data reveal important variations in the seismic response. Amplification factors and resonant frequencies exhibit seasonal trends related (both directly and inversely) to temperature changes and are sensitive to freezing periods (resonant frequencies increase with temperature and during freezing). We attribute these effects to thermal expansion driving microcrack closure, in addition to ice formation, which increase fracture and bulk rock stiffness. We find the site response at AdR is linear over the measured range of weak input motions spanning two orders of magnitude. Our results further develop and refine ambient vibration methods used in rock slope hazard assessment. Geomechanics, Time series analysis, Seismic noise, Site effects, Wave propagation 1 INTRODUCTION Rock slope failures represent significant hazards in populated alpine areas, and their identification, characterization and monitoring are key tasks in risk management. Although a number of well-established methods exist for these purposes (Gischig et al.2011; Jaboyedoff et al.2012; Kos et al.2016), subsurface information (e.g. the depth of the instability) is usually only extrapolated from surface observations. Boreholes are expensive in alpine environments and application of standard geophysical methods is challenging, especially in rock slopes with open fractures (Heincke et al.2005; 2006). Active seismic methods, for example, face extremely strong scattering and velocity gradients, which limit the depth resolution (Green et al.2006). Passive seismic techniques, on the other hand, present an alternative to classical geotechnical methods (see an overview by DelGaudio et al.2014). Unstable rock slopes with open, compliant fractures show a specific seismic response (e.g. Burjánek et al.2010, 2012; Lévy et al.2010; Fäh et al.2012; Bottelin et al.2013a) where ground motion is locally amplified in the direction/s of slope deformation at specific frequencies (i.e. the resonant frequencies). In particular, standing waves develop in the disrupted rock blocks (Burjánek et al.2012, henceforth BU12), and the unstable rock mass acts as a resonator. Resonant frequencies in turn carry information on the material properties and internal structure of the slope (i.e. depth, volume, fracture compliance of individual blocks), as demonstrated by numerical modelling (Moore et al.2011, 2016; Gischig et al.2015). Nevertheless, the mechanism of resonance described above is not unique. For instance, in the case of heavily fractured materials, resonance also originates from a seismic impedance contrast between the internally damaged rock mass and stiffer surrounding bedrock (Havenith et al.2002; Bottelin et al.2013a; Kleinbrod et al.2017a). Resonant frequencies of rock bodies can be retrieved relatively easily from ambient seismic data (no complex processing required), and may thus be readily used for rock slope monitoring and identification of irreversible change associated with internal damage. Lévy et al. (2010), for example, identified a decrease in the first resonance frequency of a rock column, prior to its collapse, accompanying failure of internal rock bridges. Monitoring of resonant frequencies was subsequently performed at four unstable rock slopes in the French Alps (Bottelin et al.2013b). While no change in the seismic response resulting from damage was observed at these sites, significant reversible variations caused by temperature fluctuations were detected. Similar observations were reported by Starr et al. (2015) for a short-term vibration monitoring experiment at a natural sandstone arch in Utah. Bottelin et al. (2013b) and Starr et al. (2015) proposed complementary but slightly different thermal stress mechanisms to explain these observations, the former calling on changes in fracture contact stiffness and the latter invoking material stress-stiffening. Recognition and characterization of reversible variations in resonant frequencies has thus emerged as a crucial task preceding identification of permanent changes associated with internal rock mass damage. Here we describe experiments to characterize and monitor the seismic response of a large rock slope instability at Alpe di Roscioro (AdR), Switzerland (Fig. 1). Following the last major failure in May 2012 (volume ∼210 000 m3), a large portion of the rock slope remains unstable and is pierced by a dramatic series of open tension cracks (Loew et al.2017). We deployed a small-aperture seismic array on the unstable rock mass (∼140 000 m3) shortly after the 2012 failure with the aim of characterizing the structure and seismic response. Later, we equipped the site with two semi-permanent seismic stations to monitor changes in the response over the following 4 yr. The collected time-series data allow us to distinguish and describe important variations in the observed seismic response related primarily to environmental effects. The heavily fractured slope is highly critical and will likely undergo future failure, providing a unique opportunity to study changes in the seismic response corresponding with the evolution of internal damage as a means of evaluating time-dependent progressive failure. Figure 1. View largeDownload slide Overview of the Alpe di Roscioro site: (a) Orthophoto taken before the 2012 failure. Scarps of the 2002 and 2012 major failures are indicated by orange and yellow lines, in addition to the current fracture network (in red); (b) cross-section of the site showing the location, orientation and anticipated depth of the fractures, as well as the foliation orientation; (c) photograph of the remaining unstable rock mass; d) photograph of the current fracture network and the scarp of the 2012 failure. Figure 1. View largeDownload slide Overview of the Alpe di Roscioro site: (a) Orthophoto taken before the 2012 failure. Scarps of the 2002 and 2012 major failures are indicated by orange and yellow lines, in addition to the current fracture network (in red); (b) cross-section of the site showing the location, orientation and anticipated depth of the fractures, as well as the foliation orientation; (c) photograph of the remaining unstable rock mass; d) photograph of the current fracture network and the scarp of the 2012 failure. 2 SITE DESCRIPTION The AdR instability (also known as the Preonzo rock slide due to its location near the village Preonzo) is a so-called retrogressive instability characterized by numerous catastrophic failures of different scales (Loew et al.2017). The earliest documented failure occurred on 1702 February 22 and buried part of the village Preonzo. These deposits remobilized 23 yr later in a large debris flow that again destroyed houses and caused 17 causalities. After a long period of quiescence, new tension cracks were observed in the late 1990s cutting the flat terrace behind the failure scarp (Fig. 1). In May 2002, about 150 000 m3 rock released following a period of heavy rainfall. Another smaller failure of about 25 000 m3 occurred on 2010 May 9. A monitoring system installed in 1999 (i.e. crack extensometers, reflectors measured with an automatic total station) was successful in anticipating these failures, and also successfully predicted the latest failure on 2012 May 15 involving about 210 000 m3 of rock. After the 2012 failures, a large portion of the unstable cliff remained (about 140 000m3), dissected by a network of wide fractures and exhibiting ongoing movement (Fig. 1). Local site geology consists of competent amphibolitic gneisses overlying augen gneisses that form the steep upper part of the AdR (Willenberg et al.2009). At the bottom of the steep cliff, weaker layered gneisses and schists outcrop. Foliation and lithology contacts dip roughly 25° WSW, that is, into the slope. This geological predisposition is typical for retrogressive failures that require step-wise fracturing of intact rock bridges until sufficient deformation has accrued and a sliding surface can form. These progressive deformation and internal failure processes result in gradual opening of tensile fractures along the flat bench at the back of the instability. A more detailed geological description can be found in Willenberg et al. (2009) and in Loew et al. (2017) respectively. 3 METHODS A number of studies, including our previous work on the seismic response of rock slopes (e.g. Burjánek et al.2010, henceforth BU10), have identified polarized, directional and amplified ground motion in unstable rock slopes. Since we expect a similar response at AdR, we focus on measuring similar ground motion attributes. We use the time-frequency polarization analysis (shortly polarization analysis or PA) developed by BU10 and BU12 to characterize particle motion of three-component seismic recordings as an ellipse. Three parameters of the polarization ellipse are retrieved as a function of time and frequency: (1) azimuth of the major axis (strike), (2) tilt of the major axis (dip), and (3) ratio between the length of the minor and major axes (ellipticity). In the case of ambient vibrations, we assume that observed ground motions are quasi-stationary and consider only the relative occurrence of polarization parameters for the analysed time window of 60 min. Thus, the time dependence is relaxed and the polarization parameters depend only on frequency. Details of the procedure can be found in BU10 and Burjánek et al. (2014). In addition, measured distributions of the polarization angles (strike and dip) for a single frequency can be well represented using a Wrapped Cauchy distribution (Burjánek et al.2014), while ellipticity is best represented with a Beta distribution (BU12). The Wrapped Cauchy distribution is characterized by mean and concentration parameters, while the Beta distribution by two positive shape parameters. Each of these parameters can be estimated for a given frequency using a maximum likelihood method. We evaluate site amplification using standard site-to-reference spectral ratios (SRSR). Amplitude spectra from ground motion recordings at a site with selected reference conditions (the reference site) are used to normalize spectra measured at other sites. If the interstation distance is short compared to the distance to the excitation sources, then SRSR represent frequency-dependent amplification functions with respect to the reference site. Although SRSR are commonly applied for earthquake recordings (site-to-source distance is usually well constrained), they also provide reasonable results for ambient vibration recordings (e.g. Roten et al.2006; BU12) even though the exact source distribution is usually not known. Nevertheless, noise sources must not be in close vicinity to the investigated site. For ambient vibration recordings, ground motion time-series data were split into non-overlapping windows of 100 s length, and the multitaper method was applied for each time window to estimate the Fourier amplitude spectrum (Prieto et al.2009). The time-bandwidth product, a parameter which controls spectral resolution (Prieto et al.2009), was set to 2.5 and four tapers were used. The retrieved spectra were normalized by the corresponding reference spectrum for each time window and the geometrical mean for all windows was determined. No additional smoothing of the spectra was required, as the multitaper method provides sufficiently smooth spectra while preventing spectral leakage (bias from the neighbouring frequency intervals). 4 ARRAY MEASUREMENT We performed an array measurement over 1 d on 2012 May 25, two weeks after the catastrophic failure. We used 12 three-component velocimeters with an eigenperiod of 5 s. The sensors were placed on metal trihedrons, in shallow holes when necessary, for better coupling with the ground. The precise positions of the sensors were measured with differential GPS. The geometry of the array is depicted in Fig. 2. Stations PRE001 and PRE002 were located close to a steep rock wall (to the west) within the stable area of the slope, and served as reference stations. Stations PRE003 and PRE004 were also located within the presumably stable area, but close to the potential backscarp (i.e. the rear-most tension crack). The remaining stations were located within the unstable and highly cracked area of the slope. We selected a time window of 1 hr to extract seismic data, during which all sensors of the array were correctly deployed with GPS time synchronization. Three seismic events (ML = 2.9, ML = 3.2, ML = 4.0) of the 2012 Emilia Romagna earthquake sequence were recorded by the array (Fig. S1) with good signal to noise ratio (Fig. S2) and epicentral distance ranging from 220 to 235 km. Sensor orientations were checked using the earthquake recordings. Since the epicentral distances were much larger than interstation distances, low frequency ground motion (<2 Hz considering the site effects described later) should be identical across the array. Band-passed (0.5–1.5 Hz) three-component traces of the earthquake recordings were systematically compared, which resulted in orientation corrections of up to 3°. Thus, the final error in our sensor orientations is expected to be less than 1°. Figure 2. View largeDownload slide Configuration of the seismic array (yellow dots); red squares show the positions of the semi-permanent seismic stations, while blue bars show the locations of extensometers used in this study. Green and red lines represent the boundary of the instability and the main scarp of the 2012 rockfall, respectively (after Loew et al.2017). Blue arrow indicates the mean direction of slope displacement (Loew et al.2017). The contour interval is 10 m. Figure 2. View largeDownload slide Configuration of the seismic array (yellow dots); red squares show the positions of the semi-permanent seismic stations, while blue bars show the locations of extensometers used in this study. Green and red lines represent the boundary of the instability and the main scarp of the 2012 rockfall, respectively (after Loew et al.2017). Blue arrow indicates the mean direction of slope displacement (Loew et al.2017). The contour interval is 10 m. 4.1 Polarization analysis All recordings were processed using time-frequency polarization analysis; representative results for three stations (PRE002, PRE004 and PRE012) are shown in Fig. 3. Ground motion at station PRE002, located within the stable area of the slope, showed no strong polarization characteristics below 8 Hz, with the exception of narrowband features at 1.3, 4 and 8 Hz. These features are likely caused by anthropogenic sources, which dominate the noise wavefield at these frequencies. Spectral analysis revealed nearly harmonic signals at the three frequencies (see later Fig. 4), which are unlikely to have natural origin. Since polarization attributes at these frequencies may be representative of the source and not the site response, we have excluded these frequencies from our analysis. The weak directionality between 2–4 Hz might be attributed to back-radiation from the unstable area. The ellipticity is flat in this frequency range and close to 0.4, what is typical for rock sites (Burjánek et al.2014). In contrast, the strong high-frequency directionality (>8 Hz) might be related to a very local structure, for example, to a small talus cone (see rock debris located east from PRE002 in Fig. 2). Similar observations were made for station PRE001. In contrast, the remaining stations of the array showed strongly polarized and directional ground motion between 3 and 4 Hz (Figs 3b–c, e–f, h–i). Ellipticity drops below 0.1 for these stations, and the average azimuth in this frequency range is 62°, which corresponds well with the predominant slope deformation direction of 60° (Loew et al.2017). Nevertheless, there are subtle variations in the location of the concentration peak (in terms of frequency and mean strike), which cluster within the array: (1) PRE003, PRE004 and PRE010 show the peak at 3.43 Hz and 68° ± 1°; (2) PRE006 and PRE009 show the peak at 3.43 Hz but 61° ± 1°; (3) PRE008, PRE011 and PRE012 show the peak at 3.2 Hz and 57° ± 1°; (4) PRE007 shows the peak at 3.55 Hz and 62° (which is close to the second group of the stations); and (5) PRE005 shows the peak at 3.09 Hz and 54°. While interpretation of these subtle differences is not straightforward as the stations do not form clear spatial patterns (e.g. PRE010 within the first group), we find them significant and discuss this result in Sections 4.2 and 6. The average dip ranges from −4° to 13° downward from the horizontal with median value of 4°. The larger dip values (8°–13°) cluster approximately along a profile PRE003-PRE006-PRE008-PRE010-PRE011, with the maximum value for PRE006. The concentration peaks are broader compared to the strike angle distributions. Figure 3. View largeDownload slide Polarization analysis of 1 hr recording of ambient vibrations at three stations: PRE002 (a,d,g), PRE004 (b,e,h), PRE012 (c,f,i). Colour scales represent the relative frequencies of occurrences of strike (a–c), dip (d–f) and ellipticity (g–i). Colour intensity is decreased for frequency bands affected by anthropogenic sources (1.2, 1.3, 4 and 7.9 Hz). Estimated modes of the Beta distributions representing the observed ellipticity values are connected by a black solid line. Figure 3. View largeDownload slide Polarization analysis of 1 hr recording of ambient vibrations at three stations: PRE002 (a,d,g), PRE004 (b,e,h), PRE012 (c,f,i). Colour scales represent the relative frequencies of occurrences of strike (a–c), dip (d–f) and ellipticity (g–i). Colour intensity is decreased for frequency bands affected by anthropogenic sources (1.2, 1.3, 4 and 7.9 Hz). Estimated modes of the Beta distributions representing the observed ellipticity values are connected by a black solid line. Figure 4. View largeDownload slide Mean power spectral density of ambient vibrations from stations PRE002 (black) and PRE006 (red). Geometrical mean of the two horizontal components is plotted. Figure 4. View largeDownload slide Mean power spectral density of ambient vibrations from stations PRE002 (black) and PRE006 (red). Geometrical mean of the two horizontal components is plotted. In general, most of the stations show strong peaks in the high-frequency range as well. However, these peaks are variable and not present consistently at all stations. For example, a peak at 5.4 Hz is present for PRE003 and PRE004, while points in different direction for PRE005 and is missing completely for other stations. These peaks represent likely higher-modes or local resonances, nevertheless, the interpretation is difficult. In comparison, Bottelin et al. (2013a) identified higher modes for three sites with column-like geometry, while no clear relation between spectral peaks was found at a site with similar geometrical setting as AdR (Bottelin et al.2013a; Fig. 3c). 4.2 Relative amplification Power spectral densities (PSDs) of noise and earthquake recordings were determined using the multi-taper method (Prieto et al.2009) for all stations of the array. An example of the noise PSD for two stations (PRE002 and PRE006) is shown in Fig. 4, demonstrating capabilities of the multitaper method: the PSDs are smooth but spectral resolution is preserved. Therefore, it is possible to identify narrow spectral peaks at 1.3, 4, and 8 Hz. These peaks likely originate from anthropogenic sources as discussed previously. Station PRE006 shows a broad peak centred at 3.4 Hz, which corresponds well to the polarization peak, while station PRE002 (located just less 50 m away) show just a little increase of seismic energy just below 3 Hz. Relative amplification functions (i.e. spectral amplitude ratios) with respect to reference station PRE002 were calculated for both earthquake and noise recordings. For the earthquake recordings, we analysed a single 10 s window for the ML2.9 event and two consecutive 10 s windows for the ML3.0 and ML4.0 events with the highest signal to noise ratios. Comparison of the 62° horizontal component of ground motion (i.e. crack perpendicular) is presented in Fig. 5. Spectral ratios are close to unity below 2 Hz for earthquakes and ambient vibration data, that is, ground motion is identical below 2 Hz for all stations in the array. Spectral ratios remain close to unity up to 10 Hz for station PRE001, but increase significantly above 2 Hz for other stations. Amplification functions from ambient vibration and earthquake data are in good agreement, however peak amplifications for earthquakes are slightly higher than ambient vibrations, which may result from the low number and short duration of the analysed time windows, or from a difference in the incoming wavefield (surface waves rich noise recordings versus S-waves rich earthquake recordings). Amplifications along the 152° horizontal component (parallel to the crack) and vertical directions (Fig. S3) are lower (<10). Figure 5. View largeDownload slide Site-to-reference spectral ratios for the 62° horizontal component of ground motion resulting from the ML2.9 (red), ML3.2 (green), ML4.0 (blue) earthquakes of 2012 May 25, compared to spectral ratios resolved from ambient vibrations (black). Station PRE002 was used as the reference. Figure 5. View largeDownload slide Site-to-reference spectral ratios for the 62° horizontal component of ground motion resulting from the ML2.9 (red), ML3.2 (green), ML4.0 (blue) earthquakes of 2012 May 25, compared to spectral ratios resolved from ambient vibrations (black). Station PRE002 was used as the reference. Details of spectral amplification results for ambient vibration data are shown in Fig. 6. Local maxima can be linked to the dominant polarization features (see previous section). In particular, the first local maximum for all stations (except PRE001) is coincident with the polarization peak at f0 = 3.2 Hz. The peak at 3.5 Hz consists of two not completely distinct local maxima, which match the f1 = 3.43 Hz and f2 = 3.55 Hz polarization peaks. Peak amplification of ambient ground motion reaches a factor of 35 at 3.55 Hz for station PRE006. The apparent drop in some of the amplification functions at 4 Hz is due to the anthropogenic source; assumptions required for calculating spectral ratios are likely not fulfilled for this source (e.g. it is too close). In particular, the path propagation effects (e.g. geometrical spreading), or source effects (e.g. radiation pattern) cannot be neglected across the array. The response becomes highly variable above 4 Hz. Figure 6. View largeDownload slide Mean site-to-reference spectral ratios of ambient vibrations for the 62° horizontal component of ground motion. Station PRE002 was used as the reference. Dashed black lines denote the identified resonant frequencies from polarization analysis (3.2, 3.43 and 3.55 Hz). Figure 6. View largeDownload slide Mean site-to-reference spectral ratios of ambient vibrations for the 62° horizontal component of ground motion. Station PRE002 was used as the reference. Dashed black lines denote the identified resonant frequencies from polarization analysis (3.2, 3.43 and 3.55 Hz). Amplifications along the 152° horizontal component (parallel to the crack) show a weak peak for a frequency of 4.1 Hz, and stronger peaks at 5.1 and 6 Hz respectively for most of the stations. These might represent higher perpendicular modes with respect to the fundamental mode. Moreover, a peak at 3.2 Hz is present for this component at station PRE005, which corresponds to the slight rotation of the fundamental peak observed in the polarization analysis. Amplifications in vertical directions show similar peaks as for the crack perpendicular directions, which is in agreement with the dip component identified by polarization analysis. Interpolated maps of amplification factors at f0 = 3.2 Hz and f1 = 3.45 Hz are presented in Fig. 7. The amplification values follow spatial trends; amplification is lower in stable area of the slope, increases towards the east in the unstable area for both frequencies, but is not largest at stations closest to the cliff edge (PRE009, PRE011 and PRE012). At 3.2 Hz, the area of greatest amplification trends NNW within the unstable area following a wide tension crack. The maximum factor of 20 is reached at stations PRE005, PRE006 and PRE008. At 3.45 Hz, amplification increases in the central part of the array within the unstable area. The maximum factor of 35 is reached at station PRE006. An amplification map for f2 = 3.55 Hz is not presented here as it is almost identical to the one for f1 = 3.45 Hz. Figure 7. View largeDownload slide Amplification map at frequencies of 3.2 Hz (a) and 3.45 Hz (b) for the 62° horizontal component of ground motion. Array stations are shown as black dots with red outline, while semi-permanent stations are indicated by red squares. Green and red lines represent the boundary of the instability and the main scarp of the 2012 rockfall, respectively (after Loew et al.2017). Figure 7. View largeDownload slide Amplification map at frequencies of 3.2 Hz (a) and 3.45 Hz (b) for the 62° horizontal component of ground motion. Array stations are shown as black dots with red outline, while semi-permanent stations are indicated by red squares. Green and red lines represent the boundary of the instability and the main scarp of the 2012 rockfall, respectively (after Loew et al.2017). 5 SEMI-PERMANENT STATIONS Two semi-permanent stations (PREO1 and PREO2) were deployed at AdR in summer 2012 to monitor changes in the slope's seismic response. These consist of a three-component velocimeter with eigenperiod of 1 s, recording unit, power supply, and a communication device that streams data to the Swiss Seismological Service. Station locations (Fig. 2) were selected based on various factors. The goal was to have one sensor in the unstable area and the second with the rest of the equipment in the stable area. It was also necessary to minimize the risk of losing equipment in case of a slope failure, which was not unexpected in summer 2012. Station PREO1 was thus setup in the presumably stable part, close to the position of array station PRE004, in order to explore the possibility of monitoring changes within the unstable rock mass outside the instability boundary, while Station PREO2 was placed in the southern, less active part of the instability where it could be retrieved in case of accelerating displacements. The installation was operational for only 10 d when it was destroyed by lightning. It was reinstalled later and fully operational from November 2012 to September 2013, when it was again destroyed by lightning. A new epoch of monitoring then began in December 2014. Station PREO1 was abandoned, and a new station PREO3 was setup in the stable area (20 m south of PREO1, see Fig. 2). This move was required for a new system design that provides more robust lightning protection. Station PREO2 was reinstalled at the original location. Unfortunately, the installation suffered from low signal-to-noise ratios, so further changes had to be adopted: station PREO3 was replaced by station PREO4 in the stable area (the sensor was moved a few metres away from a communication antenna that was thought to be generating interference). This configuration has been operating successfully since June 2015. 5.1 Monitoring of polarization We applied polarization analysis to all available continuous recordings from the semi-permanent stations. Continuous three-component time traces were split into non-overlapping windows of 60 min, polarization analysis was applied for each time window, and frequency-dependent distributions of the polarization attributes were estimated with the maximum likelihood method. Fig. 8 shows the concentration of strike angle for station PREO2 during the period between June 2015 and April 2016. The strong source effects visible on the PSD time history (Fig. S4) are removed by the polarization analysis. The concentration of strike angle for station PREO1 (not shown here) shows very similar patterns during the period between November 2012 and September 2013. Concentration measures the scatter of the azimuth distribution; zero concentration results in a uniform distribution (no preferential direction), while concentration of unity results in the Dirac delta function (strictly unidirectional distribution). We identified two bands (3–4 Hz and 7–9 Hz) of high strike concentration for station PREO2 (Fig. 8). We then attempted to pick concentration maxima for each time window to track potential changes in the resonant frequencies (black line, Fig. 8). The resulting curve was smoothed with moving average of 2 d. The picked values oscillate in the range of 3.3–3.8 Hz, which likely corresponds to the frequency doublet f1 = 3.43 Hz, f2 = 3.55 Hz (Fig. 6). Similarly, we picked the maximum concentration of the dip angle and the minimum of ellipticity (see Fig. 3e), which could also serve as proxies for drift of the resonant frequencies. Comparison of these values for stations PREO1, PREO2 and PREO4 is shown in Fig. 9(a) (smoothed with a 2 d moving average). Frequency picks based on ellipticity correlate well with the picks based on the concentration of strike. Frequency picks based on the concentration of dip show higher values, especially in winter/spring periods (up to 0.1 Hz), which is consistently observed on both sensors (PREO1/PREO2, resp. PREO4/PREO2). Nevertheless, the overall trends correspond well and contain meaningful signal. The time history of mean strike at the identified resonance frequency for both PREO1/PREO2 and PREO4/PREO2 couples is shown in Fig. 9(b). The mean directions correlate well for both sensors, with 4° systematic difference. This difference has decreased by 1° since January 2016. A similar difference was observed previously from the array stations (6°–7° difference for PRE004 versus PRE007 or PRE009). Figure 8. View largeDownload slide Concentration of strike angles for PREO2 (in colour). Maxima of the concentration are connected by the black line, which represents a variation of the dominant resonant frequency f1 ∼ 3.45 Hz. Figure 8. View largeDownload slide Concentration of strike angles for PREO2 (in colour). Maxima of the concentration are connected by the black line, which represents a variation of the dominant resonant frequency f1 ∼ 3.45 Hz. Figure 9. View largeDownload slide Variation of the fundamental frequency picked from (a) polarization parameters and (c) from the power spectral density (62° component), ranges of the picked resonance frequency from polarization parameters are shown in blue. (b) Mean polarization azimuth at the fundamental frequency picked from the strike angle distribution. Picking was performed in a frequency range of 2.5–3.9 Hz. Figure 9. View largeDownload slide Variation of the fundamental frequency picked from (a) polarization parameters and (c) from the power spectral density (62° component), ranges of the picked resonance frequency from polarization parameters are shown in blue. (b) Mean polarization azimuth at the fundamental frequency picked from the strike angle distribution. Picking was performed in a frequency range of 2.5–3.9 Hz. 5.2 Monitoring of PSDs PSDs of continuous recordings were calculated using the multi-taper method (Prieto et al.2009). Seismograms were split into non-overlapping windows of 60 min length and processed in the same manner as the array recordings. A shorter time window of 50 s was used for the subsequent windowing of hourly recordings to obtaining a smoother PSD. Time history of the square root of PSD for station PREO2 shows strong regular daily and weekly oscillations (Fig. S4) suggesting that the incoming noise wavefield has an origin in human activity. Nevertheless, since the dominant resonant frequencies can be easily identified from the PSD (e.g. Fig. 4), we tracked the local maximum for each time window (60 min) as a means of resolving temporal changes in the resonant frequencies, and compared these with picks from our polarization analysis (Fig. 9c). A similar approach was applied by Lévy et al. (2010) and Bottelin et al. (2013b). Frequency picks based on the PSD are in reasonable agreement with those based on polarization analysis. There is also excellent agreement between PSD frequency picks for the station couples (PREO1/PREO2, resp. PREO4/PREO2) in all other winter/spring periods. A small systematic difference in the summer/autumn periods may be attributed to the frequency doublet (f1 = 3.43 Hz, f2 = 3.55 Hz) identified earlier, where stations in the stable area (PREO1, PREO4) prefer f1, while the station in the unstable area (PREO2) prefers f2. Since the multitaper method has better spectral resolution than our polarization analysis, we further analyse only picks based on the PSD in our interpretation of frequency variations. 5.3 Monitoring of relative amplifications Relative amplification functions (SRSR) for station PREO2 with respect to PREO1 (later station PREO4) were determined for all available continuous recordings. Examples from different time windows are shown in Fig. 10: (1) SRSR for an hour of ambient vibrations recorded in November 2012; (2) SRSR for a 250 km distant ML5.3 earthquake recorded in June 2013 (strongest ground motion due to an earthquake recorded at the site); (3) SRSR for ambient vibrations preceding this earthquake; (4) SRSR for ambient vibrations recorded at the closest array stations (PRE009/PRE004) in May 2012. All spectral ratios are in close agreement. The level of excitation at the site was about 150 times stronger (127 μm peak-to-peak displacement) for the ML5.3 earthquake compared to the amplitude of ambient vibrations preceding the event (0.8 μm peak-to-peak displacement). This demonstrates that the site response at AdR is linear over the measured range of weak input motions (two orders of magnitude), and has not undergone detectable irreversible change between May 2012 and June 2013. Figure 10. View largeDownload slide Site-to-reference spectral ratios for stations PREO2/PREO1 (62° component): ambient vibrations acquired in November 2012 (red); ML5.3 earthquake recording from June 2013 (green); corresponding pre-event noise (black). Site-to-reference ratios for ambient vibration recordings of array stations PRE009/PRE004 located close to the semi-permanent stations are shown for comparison (yellow). Figure 10. View largeDownload slide Site-to-reference spectral ratios for stations PREO2/PREO1 (62° component): ambient vibrations acquired in November 2012 (red); ML5.3 earthquake recording from June 2013 (green); corresponding pre-event noise (black). Site-to-reference ratios for ambient vibration recordings of array stations PRE009/PRE004 located close to the semi-permanent stations are shown for comparison (yellow). Since PREO1 is also sensitive to the response of the unstable slope, it is difficult to interpret spectral ratios for PREO2/PREO1. The same holds true for the station combination PREO2/PREO4. Spectral ratios for PREO2/PREO1 (Fig. 10) are dominated by relative deamplification at ∼5.5 Hz, which results from the amplification of ground motion at PREO1 at this frequency (compare to PRE004 in Fig. 5). The peak close to 3.8 Hz (Fig. 10) cannot be related to any of the resonant frequencies; it is an apparent peak that results from a combination of the 5.5 Hz peak at PREO1 and higher amplification of PREO2 in the 2–4 Hz frequency band. Despite these complications, interpretation of time variations of these spectral ratios provides useful insight into seasonal changes of the dynamic behaviour. An example of the PREO2/PREO1 SRSR for November 2012 to September 2013 is presented in Fig. 11. The strong source effects visible on the PSD time history (Fig. S4) are removed by SRSR. In November 2012, the response showed almost no variation, that is, remained close to the SRSR presented in Fig. 10. However, the response changed abruptly In December 2012: below 4 Hz, SRSRs decreased close to unity (the response is almost identical for the two stations), while amplification appeared around 4.9 Hz and the overall response was shifted towards higher frequencies. Since then, the pattern has not varied notably below 6 Hz, but shows complexities at higher frequencies. At the end of the spring 2013, the response changed slowly returning to the pattern similar to that from November 2012 (Fig. 10). Similar trends have been observed during the period June 2015 to April 2016 (Fig. S5). White bands in Fig. 11 represent periods with data gaps (e.g. February 2013, May—July 2013). Moreover, there are several short periods when PREO2 recordings exhibit heavy disturbance, which are not present at PREO1 (creating the dark red bands below 2 Hz), which may be related to local weather conditions (rain, wind) at the station. The results are not biased by source effects of local earthquakes, since the region presents a low seismicity for the presented monitoring period. For example, there was just a single event with a magnitude >2 located within a 50 km range from the stations for the monitoring periods presented here. Figure 11. View largeDownload slide PREO2/PREO1 spectral ratio for the 62° horizontal component of ground motion (in colour). Variation of the fundamental frequency picked from the power spectral density is shown for PREO1 by the black line and for PREO2 by the grey line. Figure 11. View largeDownload slide PREO2/PREO1 spectral ratio for the 62° horizontal component of ground motion (in colour). Variation of the fundamental frequency picked from the power spectral density is shown for PREO1 by the black line and for PREO2 by the grey line. In a follow up experiment, we reinstalled array stations PRE002, PRE006 and PRE011 for 1 hr in June 2015 to explore potential variability in the other areas of the site. The resulting SRSRs are shown in Fig. S6, displaying very good agreement with those observed in May 2012. 5.4 Comparison with other monitoring data AdR is equipped with several crack extensometers, a rain gage, and temperature sensors all recording continuous data. Thus, it is possible to explore links between these observations and measured site response parameters. In particular, we analyse data from crack extensometers 4 and 7 (Fig. 2). Extensometer 7 measures opening of a crack that is the likely continuation of the main rear fracture, while extensometer 4 measures opening of a crack transecting the unstable mass between seismic station PREO2 and the cliff. Extensometer data following the May 2012 collapse are shown in Fig. 12. In general, extensometer 4 shows crack closure, which occurs mainly in winter/spring and appears to slow over monitoring period of 4 yr. Extensometer 7, on the other hand, shows gradual crack opening, which is also slowing. Local rainfall and temperature data were compared with those from the MeteoSwiss network station Locarno-Cimetta (NIME) located 15 km away at a similar altitude (100 m difference). Temperature and precipitation data correlate well at the two sites. As there are gaps in the data at AdR, we further analysed just meteorological data from Locarno-Cimetta. Figure 12. View largeDownload slide Crack extensometer data over the monitoring period. Time windows of the recorded and processed seismic data are shaded. Figure 12. View largeDownload slide Crack extensometer data over the monitoring period. Time windows of the recorded and processed seismic data are shaded. We compare drift of the identified resonant frequencies with other monitoring data (temperature, precipitation, crack opening) for the period November 2012 to September 2013 in Fig. 13. Broadly, we identify periods of both direct and inverse correlation between resonant frequencies and temperature change. The additive inverse of temperature (temperature multiplied by −1) is also included to highlight periods of inverse correlation. In particular, we observe inverse correlation in winter during freezing periods with a mean frequency delay of about 4 d (Fig. 14a), and direct correlation in summer–autumn without a mean frequency delay (Fig. 14b). These delays are apparent in Fig. 13, and precise values can be estimated by cross correlation (e.g. 4.8 d for the winter period presented in Fig. 14a; 0.1 d for the summer/fall period window presented in Fig. 14b). However, the exact value depends on selected time window. A shorter delay would be required to match the time-series toward the end of the winter (3.4 d), or at the beginning of the winter (2.8 d). The zero lag for the summer period is stable considering the time-series after 2012 July 15. Similar comparisons were made for the period July 2015 to April 2016 (Fig. S7), however the inverse correlation is weaker and spans a shorter time period during winter (freezing periods were shorter for the winter 2015/2016). The influence of precipitation on resonant frequencies is not clear, since the record of precipitation is more irregular than temperature. A decrease in resonant frequencies in April/May 2013 may be related to high precipitation, however, this has yet to be confirmed by further observations. Extensometers located in the vicinity of the seismic stations show closing of fractures during winter, which correlates with the onset of freezing temperatures in December 2012. The rate of crack closing decreases towards spring, followed by a period of crack opening which starts just after the end of the long freezing period (January–April 2013). The onset of the opening slightly precedes the period of higher precipitation and occurs as soon as temperatures increase above 0 °C, indicating possible correlation with snowmelt. Similar patterns with lower amplitudes are observed in the following winters (Fig. 12, Fig. S7). Figure 13. View largeDownload slide Comparison of resonant frequencies and other monitored field data for the period November 2012 to September 2013. (a) Variation of resonant frequencies (red, black), temperature (orange), and additive inverse of temperature (blue). (b) Daily precipitation. (c) Extensometer data. A 14 d moving average was applied to the resonant frequency and temperature time-series. Figure 13. View largeDownload slide Comparison of resonant frequencies and other monitored field data for the period November 2012 to September 2013. (a) Variation of resonant frequencies (red, black), temperature (orange), and additive inverse of temperature (blue). (b) Daily precipitation. (c) Extensometer data. A 14 d moving average was applied to the resonant frequency and temperature time-series. Figure 14. View largeDownload slide Relationship between temperature and resonant frequency time histories for: (a) winter period; temperature time history (light blue) was delayed by 4 d, temperature axis is oriented downwards. (b) Summer period; temperature time history (orange) was not delayed, temperature axis is oriented upwards. A 14 d moving average was applied on all presented time-series. Resonant frequencies are taken from Fig. 9(c). Figure 14. View largeDownload slide Relationship between temperature and resonant frequency time histories for: (a) winter period; temperature time history (light blue) was delayed by 4 d, temperature axis is oriented downwards. (b) Summer period; temperature time history (orange) was not delayed, temperature axis is oriented upwards. A 14 d moving average was applied on all presented time-series. Resonant frequencies are taken from Fig. 9(c). 6 DISCUSSION Through a field campaign of in-situ ambient seismic characterization and monitoring, we have resolved salient information on the site response characteristics of a large unstable rock slope close to collapse. We find that ground motion at the AdR site is strongly amplified (spectral ratios up to 35) within the unstable portion of the slope with respect to the stable areas. Such strong amplification may result in previously unanticipated large deformations during strong shaking, likely influencing the potential for earthquake triggered damage and failure. Amplification is strongest in a narrow frequency band between 3–4 Hz, where we have identified three resonant frequencies (3.2, 3.43 and 3.55 Hz) with the aid of multitaper processing (Prieto et al.2009). Spectral amplifications are strongest in the middle of the array and towards the northern portion of the slope that collapsed in 2012, where the aperture of tension cracks becomes wider as compared to the southern part of the instability. Thus, steeply dipping open fractures appear to have a stronger influence on the local site response than the nearby cliff free-face, in contrast to Burjánek et al. (2012) where the strongest amplification was observed at the cliff edge at a different unstable rock slope. We hypothesize that strong amplification factors, as well as spatial heterogeneity of polarization and amplification, are indicative of a high degree of internal fracturing at AdR and high criticality with respect to future failure, as also suggested by related recent numerical studies (Gischig et al.2016) and empirical studies (Kleinbrod et al.2017b). The high degree of internal fracturing may result in a reduction of shear wave velocities, which would contribute to the observed amplification as well. However, no seismic velocity measurements are available for this site. The resonant modes include significant vertical motion (Fig. S3b), so that horizontal-to-vertical spectral ratios (H/V curves) would have limited use in the seismic velocity characterization of AdR. In particular, H/V curves at AdR would be controlled by the observed resonance (2-D/3-D effects) and could not be interpreted using 1-D media per common procedures in earthquake engineering applications. Array stations PRE003 and PRE004, as well as semi-permanent stations PREO1 (later PREO4), also exhibit strong directional amplification despite being located on the presumably stable portion of the slope. This observation may be explained by back radiation of seismic energy from the unstable area to the stable areas in the presence of strong resonance. In contrast, however, BU10 and BU12 showed that directional amplification diminished almost completely at the border of other studied rock slope instabilities. Therefore, we cannot exclude the presence of another crack located southwest from the main scarp, which is currently not visible at the ground surface. Nonetheless, we demonstrate it is feasible in this case to detect signals of the unstable slope response using stations located in adjoining stable areas, which is useful in this and future experiments to help protect valuable seismic monitoring equipment in the case of slope failure. The three dominant resonant frequencies identified at AdR (3.2, 3.43 and 3.55 Hz) are relatively close (Fig. 6), and the corresponding mode shapes would likely be very similar (Fig. 7), that is, the amplitudes of the three peaks follow the same trends (Fig. 6). Therefore, the three peaks likely do not correspond to three different vibrational modes, rather we hypothesize that the fundamental mode is split. Normal-mode splitting is a well-known phenomenon, for example, observed normal modes of the earth are split due to earth's rotation and small-scale lateral heterogeneities (e.g. Stein and Wysession 2003). Thus, we speculate that splitting of the resonant frequencies observed at AdR may be caused by scattering produced by small-scale internal heterogeneities. The three resonant frequencies were also identified by polarization analysis. Ground motions are almost linearly polarized at these frequencies in the direction parallel to observed slope displacement (60° azimuth; Loew et al.2017). Small differences (54°–68°) exist, however, in the mean orientations at the resonant frequencies at different stations. These likely reflect local complexities of the vibrational mode shapes. Moreover, polarization analysis is not so well suited for characterization of the torsional motion, which may also be present at AdR and result in the variations of the mean orientations. The fundamental frequency of 3.2–3.5 Hz fits quite well with the physical extension of the unstable rock mass. In particular, we applied a formula for the fundamental frequency of a bending mode of a rock column proposed by Valentin et al. (2017). We assumed lateral extension of 30 m and Young's modulus of 20 GPa (Willenberg et al.2009), which resulted in a fracture depth range of 40–45 m. Although, the exact depth of the fracture network is not known, this estimate is in a rough agreement with the geological profile (Fig. 1b). We monitored the fundamental frequency of the AdR rock slope (i.e. the 3.43–3.5 Hz doublet peak) by means of polarization analysis and PSDs (Fig. 9). Frequency picks from polarization analysis show scatter and are biased by anthropogenic noise. PSDs, on the other hand, were better resolved due to the higher spectral resolution of the multi-taper method. Nevertheless, the general observed trends were the same for the both methods. Moreover, monitoring of the mean azimuth by polarization analysis provided additional information: we identified a potential irreversible change in azimuth in January 2016 (see Fig. 9b), which was not apparent in the resonant frequency or extensometer time-series, indicating that polarization angles may provide a new, sensitive measure of changes in the slope's seismic response capable of identifying changes related to damage. In general, the mean orientation of vibration at the fundamental frequency is controlled by the shape of the normal mode. Nevertheless, the physical interpretation of such change would require a detailed mechanical model and re-measurement of more points to map the extent of the change. We identified seasonal variations in amplification levels and resonant frequencies at AdR. Response of the two semi-permanent stations is similar up to 4 Hz in winter, when the doublet peak (3.43–3.55 Hz) disappears (see Fig. 13 and Fig. S7) and amplification levels become nearly equal. Thus, scattering due to small-scale internal heterogeneities is likely weaker in winter. The high frequency response (>4 Hz) reveals changes that are correlated with temperature (Figs 8 and 11). Amplification peaks are shifted toward higher frequencies with decreasing temperature. The onset of these variations is sharp and related to the onset of freezing, which lasted from December 2012 until the beginning of April 2013 (with a short break in the beginning of January; Fig. 11). Snow cover at the site, which we monitored from time-lapse photographs, was relatively low on average during that time, and we did not identify a relationship between snow cover and the seismic response. On the other hand, some observed changes may again be related to temperature. For example, a peak present at 6.7 Hz during February 2013 (Fig. 11) disappears abruptly around March 10, just after a short warm period (temperatures reached 0°C). The response does not change much in summer as compared to winter, and the transition from the ‘winter-type’ to the ‘summer-type’ response is smooth (April–May). Therefore, the mechanism is likely different compared to the rapid changes related to the onset of freezing. We did not observe any systematic changes in the slope's seismic response for summer and autumn periods during our monitoring (e.g. Fig. 10), despite an expected fluctuation in the response due to weather conditions. This was further confirmed by remeasurement of several array stations in June 2015 (Fig. S6). Observed resonant frequency drifts at AdR are linked to the temperature time histories, however, the relationship is complex. Response of the resonant frequencies depends on the season (freezing/non-freezing periods) and on the time scale of observations (days, weeks, months). Resonant frequencies are inversely correlated with air temperature during the freezing periods with a delay of about 4 d (Fig. 14a). This trend is present for observational time scales of several weeks (Figs 13 and 14), but disappears at shorter time scales. A sharp increase of the fundamental frequency of an unstable rock column with a drop in air temperature below 0 ° C was observed by Bottelin et al. (2013b), and explained by stiffening of a main rear fracture by ice formation. Similarly, at our site an increase in resonant frequencies with decreasing temperatures may be related to stiffening of the rock mass due to freezing of interstitial water in pores and small-scale fractures (as also hypothesized by Starr et al.2015) or in the infill in the open tension fractures. The time shift between temperature changes and resonant frequencies could be explained by the diffusive temperature (and freezing) front that has to penetrate a certain depth before having an effect on the overall rock mass stiffness. This would also be in agreement with the disappearing of the effect on shorter time-scales; daily or shorter-term temperature changes penetrate only few decimetres and thus do not contribute to the advancing of the freezing front a great depth. An alternative explanation may be that resonant frequencies increase due to a decrease in mass of the resonating rock slope related to gradual drainage of pore water during cold periods. However, this mechanism would result in a gradual increase of resonant frequencies while temperatures are below 0° C, and a sharp decrease as soon as melt water becomes available as temperatures increase. Instead, resonant frequencies at AdR appear to strongly correlate with temperature. In summer periods, resonant frequency variations at AdR are directly correlated with air temperature without delay. This trend is present for longer observational time scales of weeks (Figs 13 and 14), while a similar trend is also observed for shorter time scales (days) during dry periods (e.g. July/August 2015) again with almost no delay. A similar increase in the fundamental frequency of an unstable rock column with temperature at non-freezing daily periods was observed by Lévy et al. (2011) and Bottelin et al. (2013a,b). We propose that increasing thermoelastic stresses accompanying rock warming lead to closure of micro-cracks and thus to an overall stiffening of the rock mass, leading to an increase in resonant frequencies (as in Starr et al.2015). Such a thermo-elastic effect would produce faster reaction on temperature changes leading to shorter delay and also reactions observable at shorter-time scale, unlike the winter-time stiffening mechanism related to ice formation mentioned earlier. Lastly, we observed a decrease in resonant frequencies in April/May 2013 that does not correlate with air temperature. The decrease is possibly related to a rising water table due to snow melt in the broader area (not directly at the site) and higher precipitation rates (Fig. 13). Similar observations were made by Bottelin et al. (2013a). In summary, we observe two opposite and competing temperature related effects: stiffening the rock mass due to ice formation as the temperatures decrease, dominant in winter, and stiffening of the rock mass by micro-crack closure as temperatures increase, dominant in summer. The additional role of groundwater recharge may become apparent in a transitional phase in spring, when the two competing temperature-related effects are offset. We emphasize that all observed correlations with temperature are indirect. 7 CONCLUSIONS AdR is a highly disrupted unstable rock slope with a long history of catastrophic failures (Loew et al.2017). The slope will likely undergo future failures, so it provides a unique opportunity to study changes in the seismic response corresponding with the evolution of internal damage. As application of ambient vibration monitoring and characterization of unstable rock slopes becomes more prevalent, it is critical to develop refined methods and test the resolution and capabilities of these techniques as a means of detecting changes in physical properties. In this study, we measured much larger bedrock site amplifications and polarization levels than in previous studies (BU10, BU12), which have implications for co-seismic triggering and damage. Interestingly, we find that amplification at AdR is not strongest at the cliff edge, as in previous studies, but rather controlled by the internal structure of the rock mass (in particular the presence of compliant fractures). We also importantly observed that the seismic response of AdR is identical across a wide range of weak input motions, helping clarify the range of validity for modelled using linear media. Amplification levels and resonant frequencies at AdR exhibit seasonal variations related (both directly and indirectly) to temperature changes. In summer, resonant frequencies increase with increasing temperature as thermal stresses drive closure of microcracks and bulk stiffening of the rock mass. Meanwhile in winter, resonant frequencies increase with decreasing temperatures associated with the formation of pore and crack ice stiffening the rock mass. The seismic response of AdR has not changed significantly over our monitoring period, suggesting the internal structure has not sustained large irreversible changes since May 2012. This result is meaningful in the ongoing hazard evaluation. We show that ambient vibration measurements are repeatable, provide quantitative information on the internal composition of an unstable rock mass, and readily useful for monitoring slope change over time. Nevertheless, environmental conditions (such as temperature) and associated changes in seismic response parameters must also be characterized in any hazard monitoring interpretation. The ongoing long-term seismic monitoring at AdR has provided a unique data set necessary for correct interpretation of potential seismic signatures preceding future failure. Acknowledgements We thank the Ticino cantonal geologist Giorgio Valenti and colleagues Marco Franzi and Marco Andretta for providing access to temperature and extensometer data, in addition to their support for seismic measurements and monitoring without which this study would not have been possible. Meteorological data were provided by MeteoSwiss. Research presented in this study was funded by ETH Zurich project ETH-1212-2 (Characterization of unstable rock slopes through passive seismic measurements). REFERENCES Bottelin P.et al.   ( 2013a). Spectral analysis of prone-to-fall rock compartments using ambient vibrations, J. Environ. Eng. Geophys. , 18, 205– 217. Google Scholar CrossRef Search ADS   Bottelin P., Lévy C., Baillet L., Jongmans D., Guéguen P. ( 2013b). Modal and thermal analysis of Les Arches unstable rock column (Vercors massif, French Alps), Geophys. J. Int. , 194, 849– 858. Google Scholar CrossRef Search ADS   Burjánek J., Gassner-Stamm G, Poggi V., Moore J.R., Fäh D ( 2010). Ambient vibration analysis of an unstable mountain slope, Geophys. J. Int. , 180, 820– 828. Google Scholar CrossRef Search ADS   Burjánek J., Moore J.R, Yugsi-Molina F.X., Fäh D ( 2012). Instrumental evidence of normal mode rock slope vibration, Geophys. J. Int. , 188, 559– 569. Google Scholar CrossRef Search ADS   Burjánek J., Edwards B, Fäh D ( 2014). Empirical evidence of local seismic effects at sites with pronounced topography: A systematic approach, Geophys. J. Int. , 197, 608– 619. Google Scholar CrossRef Search ADS   Del Gaudio V., Muscillo S., Wasowski J. ( 2014). What we can learn about slope response to earthquakes from ambient noise analysis: An overview, Eng. Geol. , 182, 182– 200. Google Scholar CrossRef Search ADS   Fäh D.et al.   ( 2012). Coupled seismogenic geohazards in Alpine region, Boll. Geofis. Teor. Appl. , 53, 485– 508. Gischig V., Amann F., Moore J.R., Loew S., Eisenbeiss H., Stempfhuber W. ( 2011). Composite rock slope kinematics at the current Randa instability, Switzerland, based on remote sensing and numerical modeling, Eng. Geol. , 118, 37– 53. Google Scholar CrossRef Search ADS   Gischig V.S., Eberhardt E., Moore J.R., Hungr O. ( 2015). On the seismic response of deep-seated rock slope instabilities — Insights from numerical modeling, Eng. Geol. , 193, 1– 18. Google Scholar CrossRef Search ADS   Gischig V., Preisig G., Eberhardt E. ( 2016). Numerical investigation of seismically induced rock mass fatigue as a mechanism contributing to the progressive failure of deep-seated landslides, Rock Mech. Rock Eng. , 49, 2457– 2478. Google Scholar CrossRef Search ADS   Green A.G., Maurer H.R., Spillmann T., Heincke B., Willenberg H. ( 2006). High-resolution geophysical techniques for improving hazard assessments of unstable rock slopes, Leading Edge , 25, 311– 316. Google Scholar CrossRef Search ADS   Havenith H.-B., Jongmans D, Faccioli E., Abdrakhmatov K., Bard P.-Y. ( 2002). Site effect analysis around the seismically induced Ananevo rockslide, Kyrgyzstan, Bull. seism. Soc. Am. , 92, 3190– 3209. Google Scholar CrossRef Search ADS   Heincke B., Green A.G., Kruk J.V.D., Horstmeyer H. ( 2005). Acquisition and processing strategies for 3D georadar surveying a region characterized by rugged topography, Geophysics , 70, K53– K61. Google Scholar CrossRef Search ADS   Heincke B., Maurer H., Green A.G., Willenberg H., Spillmann T., Burlini L. ( 2006). Characterizing an unstable mountain slope using shallow 2D and 3D seismic tomography, Geophysics , 71, B241– B256. Google Scholar CrossRef Search ADS   Jaboyedoff M., Oppikofer T., Abellán A., Derron M.H., Loye A., Metzger R., Pedrazzini A. ( 2012). Use of LIDAR in landslide investigations: a review, Nat. Hazards , 61, 5– 28. Google Scholar CrossRef Search ADS   Kleinbrod U., Burjánek J., Hugentobler M., Amann F., Fäh D ( 2017a). A comparative study on seismic response of two unstable rock slopes within same tectonic setting but different activity level, Geophys. J. Int. , 211, 1428– 1448. Google Scholar CrossRef Search ADS   Kleinbrod U., Burjánek J., Fäh D. ( 2017b). On the seismic response of instable rock slopes based on ambient vibration recordings, Earth Planets Space , 69, 126, pp. 9, doi:10.1186/s40623-017-0712-5. Google Scholar CrossRef Search ADS   Kos A., Amann F., Strozzi T., Delaloye R., von Ruette J., Springman S. ( 2016). Contemporary glacier retreat triggers a rapid landslide response, Great Aletsch Glacier, Switzerland, Geophys. Res. Lett. , 43, 12 466–12 474. Google Scholar CrossRef Search ADS   Lévy C., Baillet L., Jongmans D., Mourot P., Hantz D., ( 2010). Dynamic response of the Chamousset rock column (Western Alps, France), J. geophys. Res. , 115, F04043, doi:10.1029/2009JF001606. Lévy C., Jongmans D., Baillet L. ( 2011). Analysis of seismic signals recorded on a prone-to-fall rock column (Vercors massif, French Alps), Geophys. J. Int. , 186, 296– 310. Google Scholar CrossRef Search ADS   Loew S., Gschwind S., Gischig V., Keller-Signer A., Valenti G. ( 2017). Monitoring and early warning of the 2012 Preonzo catastrophic rockslope failure, Landslides  14, 141– 154. Google Scholar CrossRef Search ADS   Moore J.R., Gischig V, Burjánek J., Loew S., Fäh D ( 2011). Site effects in unstable rock slopes: dynamic behavior of the Randa instability (Switzerland), Bull. seism. Soc. Am. , 101, 3110– 3116. Google Scholar CrossRef Search ADS   Moore J.R.et al.   ( 2016). Anthropogenic sources stimulate resonance of a natural rock bridge, Geophys. Res. Lett ., 43, 9669– 9676. Google Scholar CrossRef Search ADS   Prieto G.A., Parker R.L., Vernon F.L. ( 2009). A Fortran 90 library for multitaper spectrum analysis, Comput. Geosci. , 35, 1701– 1710. Google Scholar CrossRef Search ADS   Roten D., Cornou C., Fäh D., Giardini D. ( 2006). 2D resonances in Alpine valleys identified from ambient vibration wavefields, Geophys. J. Int. , 165, 889– 905. Google Scholar CrossRef Search ADS   Starr A.M., Moore J.R., Thorne M. S. ( 2015). Ambient resonance of Mesa Arch, Canyonlands National Park, Utah, Geophys. Res. Lett., 42, 6696– 6702. Stein S., Wysession M. ( 2003). An Introduction to Seismology, Earthquakes and Earth Structure , Blackwell Publishing, 498 pp. Valentin J., Capron A, Jongmans D., Baillet L., Bottelin P., Donze F., Larose E, Mangeney A. ( 2017). The dynamic response of prone-to-fall columns to ambient vibrations: comparison between measurements and numerical modelling, Geophys. J. Int. , 208( 2), 1058– 1076. Google Scholar CrossRef Search ADS   Willenberg H., Eberhardt E., Loew S., McDougall S., Hungr O. ( 2009). Hazard assessment and runout analysis for an unstable rock slope above an industrial site in the Riviera valley, Switzerland, Landslides , 6( 2), 111– 119. Google Scholar CrossRef Search ADS   SUPPORTING INFORMATION Supplementary data are available at GJI online. Figure S1. Array ML4.0 earthquake recording of 2012 May 25. The dominant wave group used for the SRSR estimation is coloured in red. The amplitude of the traces was normalized by maximum peak amplitude of the three components for each station. A peak amplitude is shown at the beginning of each trace. Figure S2. Fourier amplitude spectra of the dominant wave group (red) of the ML=4.0 earthquake (Fig. S1) used for SRSR estimation and Fourier amplitude of the pre-event noise (black). Figure S3. Mean site-to-reference spectral ratios of ambient vibrations for the 152° horizontal (a) and vertical (b) components of ground motion. Station PRE002 was used as the reference. Dashed black lines denote the identified resonant frequencies from polarization analysis (3.2, 3.43 and 3.55 Hz). Figure S4. Time history of the square root of power spectral density for station PREO2 and the 62° horizontal component of ground motion (in colour). The colour scale is saturated for values below 1st and above 99th percentile, respectively. Figure S5. PREO2/PREO4 spectral ratio for the 62° horizontal component of ground motion (in colour). Variation of the fundamental frequency identified from the power spectral density is shown for stations PREO4 (black line) and PREO2 (grey line). Figure S6. Comparison of observed SRSR in May 2012 and June 2015 for PRE006/PRE002 and PRE011/PRE002. Figure S7. Comparison of resonant frequencies and other monitored field data for the period July 2015 to April 2016. (a) Variation of resonant frequencies (red, black), temperature (orange) and additive inverse of temperature (blue). (b) Daily precipitation. (c) Extensometer data. A 14 d moving average was applied to the resonant frequency and temperature time-series. Please note: Oxford University Press is not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the paper. © The Authors 2017. Published by Oxford University Press on behalf of The Royal Astronomical Society.

Journal

Geophysical Journal InternationalOxford University Press

Published: Jan 1, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 12 million articles from more than
10,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Unlimited reading

Read as many articles as you need. Full articles with original layout, charts and figures. Read online, from anywhere.

Stay up to date

Keep up with your field with Personalized Recommendations and Follow Journals to get automatic updates.

Organize your research

It’s easy to organize your research with our built-in tools.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve Freelancer

DeepDyve Pro

Price
FREE
$49/month

$360/year
Save searches from
Google Scholar,
PubMed
Create lists to
organize your research
Export lists, citations
Read DeepDyve articles
Abstract access only
Unlimited access to over
18 million full-text articles
Print
20 pages/month
PDF Discount
20% off