Towards a standard typology of endogenous landslide seismic sources

. The objective of this work is to propose a standard classiﬁcation of seismic signals generated by gravitational processes and detected at close distances ( < 1 km). We review the studies where seismic instruments have been installed on unstable slopes and discuss the choice of the seismic instruments and the network geometries. Seismic observations acquired at 13 unstable slopes are analyzed in order to construct the proposed typology. The selected slopes are affected by various landslide types 5 (slide, fall, topple, and ﬂow) triggered in various material (from unconsolidated soils to consolidated rocks). We investigate high frequency


Introduction
Seismology can be used to record (remotely and in a non-invasive way) ground deformation processes and to measure stress/strain conditions through the hydro-mechanical interactions occurring in the media.Seismology is widely used to understand the physical processes taking place on tectonic faults or volcanoes, to investigate fluid reservoir circulation, and more recently to analyze the dynamics of Earth surface processes such as glaciers (Podolskiy and Walter, 2016), snow avalanches (Leprettre et al., 1996;Sabot et al., 1998;Surin et al., 2000;Lacroix et al., 2011;Pérez-Guillén et al., 2016) and landslides (Deparis et al., 2008;Ekström and Stark, 2013;Gomberg et al., 1995;Rouse et al., 1991).In this manuscript, the term landslide describes a wide variety of processes resulting from the downslope movement of slope-forming materials by falling, toppling, sliding or flowing mechanisms (Hungr et al., 2014).Thus, landslides cover a large range of deformation processes, that can be differentiated in terms of sizes and volumes (smaller than 1 m 3 up to more than 10 7 m 3 ), in terms of displacement rates (mm.yr− 1 to m.s − 1), and in terms of mobilized material (hard/soft rocks, debris, poorly consolidated soils, and artificial fills).
With the increasing number of seismic sensors deployed worldwide and to the development of automatic seismological processing chains, the construction of landslide catalogs using seismology is now possible, especially at the regional scale (e.g.Switzerland, Hammer et al. (2013); Dammeier et al. (2016); France, Deparis et al. (2008)).However, the forecast of a particular landslide rupture or acceleration is still challenging at the slope scale, which is the focus of this work.In the 1960s, Cadman and Goodman (1967) observed an increase of Acoustic Emissions (AE) generated by slopes tilted towards failure at both laboratory and field scales.AEs are high frequency (10-1000 kHz) body waves generated by the release of strain energy through grain rearrangement (Michlmayr et al., 2012).Further studies confirmed these results for several slopes (Rouse et al., 1991;Smith et al., 2014;Dixon et al., 2015Dixon et al., , 2018) ) where correlations between AE, surface displacement and heavy rainfall were documented.AEs record deep deformation processes before signs of displacement are identifiable at the surface.However, AEs are rapidly attenuated with the distance to the sources.The location of the sensors and the type of waveguide are also critical to capture the slope behavior.Recent developments of Fiber Optic Distributed Acoustic Systems (FO-DAS) offer the opportunity to overcome attenuation limitations and deploy measures over long distances (Michlmayr et al., 2017).More recently, several studies focused on the analysis of the micro-seismicity (MS) observed on unstable slopes.MS studies analyze the seismic waves generated by the release of strain energy in the ground at larger scale than the grain to grain interactions in the frequency range of 1 to 500 Hz.The method offers the opportunity to remotely record the spatial distribution of the deformation through time (McCann and Forster, 1990;BRGM, 1995) and is less sensitive to attenuation than AE methods.Gomberg et al. (1995) installed seismometers on the Slumgullion slow-moving landslide (Colorado, USA) in order to understand the mechanical processes taking place during landslide deformation.Further studies used the same method for several slope configurations (hard/soft rocks, soils, very slow to rapid movements) but also investigated the possible links between the displacement rate and the seismic energy release (Spillmann et al., 2007;Helmstetter and Garambois, 2010;Walter et al., 2012Walter et al., , 2013b;;Tonnellier et al., 2013).Helmstetter and Garambois (2010) correlated the seismic response of the Séchilienne rockslide with the surface displacement rate and the rainfall amount.The analysis of the seismic waves generated by landslides allows monitoring spatiotemporal changes of the stress-strain field in the material from the scale of microscopic internal damage (Dixon et al., 2003;Michlmayr et al., 2012;Smith et al., 2017) to the initiation (e.g.pre-failure) of large ruptures (Amitrano et al., 2005;Yamada et al., 2016b;Poli, 2017;Schöpa et al., 2018).Both the failure and surface processes (e.g.rockfall, debris flow) generate seismic waves.Physical properties (mass, bulk momentum, velocity, trajectory) of the landslide can be inferred from the analysis of the seismic signals (Kanamori et al., 1984;Brodsky et al., 2003;Lacroix and Helmstetter, 2011;Ekström and Stark, 2013;Tang et al., 2015;Hibert et al., 2014a;Levy et al., 2015).On clayey landslides, drops of shear wave velocity have been observed before acceleration episodes.This shear wave variation through time has been documented using noise correlation techniques for laboratory experiments (Mainsant et al., 2012b), and for a few cases in the field at Pont-Bourquin landslide (Switzerland, Mainsant et al. (2012a)), at Harmaliére landslide (France, Bièvre et al. (2017)) and at Just-Tegoborze landslide (Poland, Harba and Pilecki (2017)).Precursory seismic signals are also expected and documented before large failures.Precursory increase in micro-seismic activity (in terms of event rates and/or average amplitudes) has been observed first before the fall of a coastal cliff (Mesnil-Val, France, Amitrano et al. (2005)) and was interpreted as the propagation of a fracture.More recently, repeating events have been detected before the Rausu landslide (Japan, Yamada et al. (2016b)) and the Nuugaatsiaq landslide (Greenland, Poli (2017)).These events are likely associated with the repeated failure of asperities surrounded by aseismic slip, driven by the acceleration of the slope displacement during the nucleation phase of the landslide rupture.Schöpa et al. (2018) recorded harmonic tremors that started 30 min before the failure of the Askja caldera landslide (Iceland) with temporal fluctuations of resonance frequency around 2.5 Hz.This complex tremor signal was interpreted as repeating stick-slip events with very short recurrence times (less than 1 s) producing a continuous signal.However, the characterization of the size of the asperity and the velocity of the ruptures associated to these precursory signals are difficult to invert mostly because of the lack of dense seismic network at close proximity of the slope instability (Schöpa et al., 2018).Therefore, the monitoring of endogenous MS may represent a promising approach especially, with the advent of robust, cheaper and portable seismic sensors and digitizers.It is now possible to install dense sensor networks close to the unstable slopes and record low amplitude signals in broad frequency bands.A wide variety of unstable slopes are currently monitored (i.e. through permanent or campaign installation) with seismic networks of different sizes and instruments (Table 1).
Understanding the possible mechanisms generating these seismic signals needs to be achieved.The discrimination of the endogenous landslide seismic signals is difficult and need to be established.The objective of this paper is thus to propose a typology of the landslide micro-seismic signals recorded in the field.The proposed typology is based on the analysis of observations from 13 monitored sites.The typology includes all the seismic sources recorded at near distances (< 1 km) and in the frequency range of MS studies (1-500 Hz), and generated by landslides 1) developed in hard/soft rocks and soils, 2) characterized by fragile (i.e.rupture) and ductile (i.e.viscous) deformation mechanisms.
In our work, we first discuss all the physical processes that occur on landslides and may generate seismic signals.We further present the available seismic sensors, the most commonly used network geometry and the instrumented sites.Then we establish a classification scheme of the landslide seismic signals from relevant signal features based on the analysis of the datasets of 13 sites.We further discuss the perspectives and remaining challenges of monitoring landslide deformation with MS approaches.
The seismic signals associated with very large rock/debris avalanches and slides observed at regional distances are out of the scope of this work.II: sliding and III: tearing) depending on the movement of the medium on the sides of the fracture plane.They result from either brittle failure of the media or from dessication effects forming polygonal failures during soil drying.On landslides, most of the fractures occur in a tensile mode because of the low tensile toughness of the landslide material and the shallow depth (Stumpf et al., 2013).The formation of fractures can also be generated in depth by progressive degradation of the rock through ground shaking and/or through weathering and long-term damage due to gravitational load.At the base and on the edges of the landslide, the movement is assumed to develop fractures in shear mode, creating sliding surfaces.Shearing on the fracture plane and tensile fracture opening/closing generate seismic signals.Shearing takes place at different scales from earthquakes on tectonic plates to grain friction and generates a variety of seismic signals (Zigone et al., 2011).Unstable regime leads to stick-slip behavior where the stress is regularly suddenly released generating impulsive seismic events.Tremor like signals or isolated impulsive or emergent events are also generated during plate motions.A variety of signals are observed during glacier motion.Deep icequakes are usually associated to basal motion (Winberry et al., 2011;Pratt et al., 2014;Helmstetter et al., 2015a, b;Roeoesli et al., 2016a;Podolskiy and Walter, 2016).Tremor like signals are also recorded during glacier motion (Lipovsky and Dunham, 2016).They are characterized by long duration signals of low amplitudes with no clear phase onsets.
They are associated with repetitive stick-slip events on the fracture plane.Tensile fracture opening/closing generate similar signals on glacier at the surface and at depth (Walter et al., 2013a;Helmstetter et al., 2015b;Podolskiy and Walter, 2016).
Focal mechanism and location of the source allow to differentiate between tensile and shear mechanism.

Topple and fall related sources
On vertical to sub-vertical slopes, mass movement occurs as the topple of rock columns or as the free-fall (and possibly bouncing and rolling) of rocky blocks (Hungr et al., 2014).In the case of toppling, the movement starts with a slow rotation of the rock blocks under the effects of water infiltration or ground shaking and ends with the free fall of larger blocks.Rockfalls, during the propagation phase, impact the ground at some location along their trajectory.These impacts generate seismic waves that can be recorded remotely by seismometers.The range of rockfall volumes can be very large, varying from less than one cubic meter to thousands cubic of meters.

Mass flow related sources
Mass flows gather different run-out processes of debris or of a mixture of water and debris.They cover a large range of volumes from large rock avalanches of several millions cubic meters to small (hundreds cubic of meters) debris falls and flows (Hungr et al., 2001).They can occur in wet or dry conditions.The contacts of the rock/debris fragments with the bedrock and in the mass flow generate seismic radiations (Suriñach et al., 2001;Burtin et al., 2009;Schneider et al., 2010;Hibert et al., 2011;Abancó et al., 2012;Burtin et al., 2013;Levy et al., 2015;Kean et al.;Vázquez et al., 2016;Hibert et al., 2017b).The seismic signal is hence a combination of grain contacts within the granular flow and of grain to ground surface contacts and hence generate a complex seismic signal.

Fluid related sources
Hydrological forcing (e.g.precipitation, snow-melt) is one of the most common landslide triggers.The presence of fracture networks, water pipes and the heterogeneity of the rock/soil media result in the development of preferential water flow paths (Richards and Reddy, 2007;Hencher, 2010).These preferential flows induced local saturated area where the increase of pore water pressure may destabilize shallow or deep shear surfaces.In soils, the dissolution of material into finer granular debris creates weak zones prone to collapse either by suffusion (i.e.non cohesive material wash out under mechanical action) or by dispersion (i.e. chemical dissolution of fractured clay soils; Richards, Jones, 1981).In rocks, pipes may develop by erosion.

Sensors used in landslide monitoring
Body and surface mechanical waves may be generated by the sources described in Section 2. Body waves (Primary -P-, Secondary -S-) radiate inside the media.P-waves shake the ground in the same direction they propagate while S-waves shake the ground perpendicularly to their propagation direction.Surface waves only travel along the surface of the ground and their velocity, frequency content and intensity change with the depth of propagation.Acoustic waves can be generated by the conversion of body waves at the surface.These waves travel in the air at a velocity of about 340 m.s −1 , slightly varying with temperature and air pressure.Acoustic waves are often generated by anthropogenic or atmospheric sources (e.g., gun shots, explosions, storms...), but can also be generated by rockfalls, debris flows or shallow fracture events.All these mechanical waves are subject to attenuation with the travel distance; the high frequency waves are attenuated faster than the low frequency waves.The relatively low energy released by the landslide related sources makes the choice of the seismic instruments to deploy very important.Four types of instruments are used to record ground motion for different frequency ranges and sensitivities.
For landslide monitoring, Short-Period (SP) seismometers and geophones, Broad-Band (BB) seismometers, accelerometers, and AE sensors are commonly installed in the field.
-Broad-Band seismometers are force-balanced sensors with very low corner frequency (< 0.01 Hz) that can record the ground motion with a flat response in a large frequency range [0.01-25] Hz.They require a careful mass calibration during their installation and are sensitive to temperature and pressure variations.They are mostly used to record very weak ground motion and ambient noise; -SP-seismometers are passive or force-balanced instruments with high corner frequency (> 1Hz).They measure the velocity of the ground with high sensitivity and a flat response in the [1-100] Hz frequency band.They are recommended for volcanic and glacier monitoring among other applications.They are less sensitive to air temperature and pressure variations and do not require mass calibration.They are hence particularly suitable for landslide monitoring.Geophones are similar to SP-seismometers but usually cover higher frequencies [1-600]Hz with lower sensitivity.They are mainly used for active seismic campaigns but may also be installed for the same purposes as SP-seismometers; -Accelerometers are strong motion sensors able to record high amplitudes and high frequencies seismic waves.They can resolve accelerations in the frequency bands from 0.1 to 10 kHz.The response of the sensor is proportional to ground acceleration for all frequencies (there is no corner frequency).But the noise level is important for low frequencies and the sensitivity is not as good as for velocimeters.They are used to record strong ground motion in particular when installed close to epicenters (< 100 km) of large earthquakes where seismometers usually saturate.For landslide, they are usually used as inclinometers; -AE (Acoustic Emission) sensors can record ground vibrations at very high frequencies (10 kHz-10 MHz) and low amplitudes.There are two types of AE sensors: the first type is very sensitive to a narrow frequency band only while the second type is sensitive to a broader frequency band (Michlmayr et al., 2012).In the field, a waveguide is often installed together with AE sensors in order to counteract the attenuation of the signal.They are used in combination with accelerometers for structural monitoring and for laboratory experiments (e.g.loading, shear, flume tests) and can be used on landslide to monitor very low magnitude sources at the grain-to-grain interactions (Dixon et al., 2003;Michlmayr et al., 2012;Smith et al., 2017); in addition, microphones or infrasound sensors can be useful to detect, locate and classify landslides seismic signals (Kogelnig et al., 2014;Schimmel and Hübl, 2016;Helmstetter and Janex, 2017).The detection of acoustic waves and body waves at one point, because they propagate at different velocities, can be used to estimate the distance from the source.The relative amplitude of seismic and acoustic waves can also provide information on the depth of the source, because shallow sources generate more acoustic waves than deeper ones.
It must be noted that AE sensors only record acoustic emissions generated at very high frequencies (> 10kHz) and consequently are very sensitive to attenuation.Indeed, attenuation factor Q is estimated to range between 10 −2 and 10 1 dB.cm −1 (Michlmayr et al., 2012).Even with a waveguide, they must be collocated with the cracks or the sliding surfaces observed on the slope (Dixon et al., 2015).BB, SP seismometers and geophones record seismic signals in the common band of 10 0 -10 2 Hz and hence offer a solution to monitor more distant sources.The detection of a seismic sources by MS sensors depends on the seismic energy released by the source, the sensor to the source distance and the attenuation of the media.Installation of MS sensors at the proximity of the geomorphological features of interest (e.g.scarp, faults, sliding surfaces, superficial crack networks, etc.) optimize the detection of the seismic signals generated by those processes but distant sources (> 1 m) can also be recorded by MS sensors.The latter do not need to be co-located with the geomorphological features of interest.After correcting the sensor response, the signals generated by these sensors can be analyzed and compared in their common frequency range.Installation of BB seismometers can complete SP network and enable to investigate the low-frequency signals generated by the slope while geophones are more adapted to explore very high frequency content (> 100 Hz).Dense networks of the latter instruments are recommended to investigate the seismicity induced by landslide deformation while the installation of one unique BB seismometer is enough to investigate the low-frequency radiations of the landslide.

Network geometry
Several network configurations have been tested in different studies.It must be noted that the network geometry in the case of landslides is constrained by the site configuration.Indeed, the maintenance of seismic sensors may be very challenging when installed on the moving parts of the landslide; therefore, an installation on the most stable parts of the landslide or at its vicinity is often preferred for permanent monitoring (Spillmann et al., 2007;Helmstetter and Garambois, 2010;Walter et al., 2017).
During field campaigns, maintenance of sensors installed on the unstable slopes is possible and often realized (Gomberg et al., 2011;Walter et al., 2012;Tonnellier et al., 2013).Therefore, the main challenges for seismic sensor installation at this scale is 1) to locate the sensor at close distance to the sources, 2) to maximize the number of stations and to locate the sensor close to each other to record the same event at different seismic station and 3) minimize the azimuthal gap between the sensors.
The number of deployed sensors plays an important in the magnitude of completeness (M c ) of the seismic network.While the geometry of the network (i.e.inter-sensor distances, azimuthal gap) mostly control the accuracy of source locations.
Seismic sensors can be deployed in network of single sensors or network of sensor arrays.The difference between seismic network and seismic arrays is related to the distance at which the signals recorded by two sensors can be correlated.In the case of seismic arrays, the distance between the sensors is reduced to maximize the correlation of the signals recorded by each sensor.Otherwise the installation is called a seismic network (Podolskiy and Walter, 2016).Although the inter-sensor distance is often small (< 1 km) in the case of landslide monitoring, decorrelation of the signals is often observed even at small distances due to the complexity of the underground structure especially at high frequencies.The use of the "seismic array" approach in landslide monitoring often refers to specific geometries of collocated sensors (inter-sensors distances < 50 m) organized with a central sensor (often a three-component seismometer) and several satellite sensors (often vertical sensors).This kind of installation presents many advantages such as enhancing the Signal-to-Noise (SNR) ratio and allowing the computation of the back-azimuth of the source with beam-forming methods.
For the majority of the instrumented landslides, seismic networks are organized with single sensors located on or at close distance of the unstable slopes.The inter-sensors distance and the azimuthal gap are often controlled by the location of easily accessible or stable portions of the slopes.However, specific geometry can be adopted such as (almost) linear geometry.This is particularly the case for the monitoring the propagation of debris flows in stream channels.Dense networks (number of sensors > 50) can also be deployed.In this case the sensors are installed using a grid geometry with regular inter-sensor distances.
This kind of installation is probably the most optimal but is currently mostly realized during short acquisition campaigns due to the difficulty to maintain a large number of sensors over long periods (battery, data storage, possible movement of the sensor), especially when installed directly on the unstable zones of landslides.Finally, the installation of sensors at depth (> 1 m) is challenging for landslide and it has currently only been realized on hard-rock slopes (e.g.Randa, Spillmann et al. (2007) or Séchilienne, RESIF/OMIV ( 2015)).This kind of installation are however very valuable to constrain the depth of the sources.

MS processing chains
One of the current challenge for landslide MS analysis is the development of dedicated processing chains able to analyze the unconventional seismic signals observed on landslides.The three steps of MS processing are successively: the detection, the classification and the location of the endogenous seismic events.The development of robust and versatile processing chains for analyzing landslide micro-seismicity is challenging because of 1) the low magnitude of the events and the attenuation of the media that results in emergent and low Signal-to-Noise Ratio (SNR) records, 2) the seismic source radiation patterns that may be single centroid source, double couple source or volumetric source, and, 3) the heterogeneity and variation in time (i.e.topography, water table levels, fissures) of the underground structure preventing the construction of precise velocity models and hence, accurate source locations.
First, for detecting automatically or manually the seismic events, the use of spectrograms is common.Spectrograms represent the evolution of the frequency content in time by computing the Fourier Transform on small moving time windows (e.g.< 1 s).Automatic detection is usually carried out with the STA/LTA (Short-Term Average/Long-Term Average) detector (Allen, 1982) applied on the summed energy of the spectrogram (Spillmann et al., 2007;Helmstetter and Garambois, 2010;Tonnellier et al., 2013).
Second, classifying the detected signals can be carried out automatically by discarding exogenous events with simple criteria (i.e.threshold on the signal duration, inter-trace correlation, apparent velocity) but the determination of the threshold to differentiate the class of signals may be difficult.Machine learning algorithms offer nowadays the possibility to automatize and improve this step.Dammeier et al. (2016) developed a Hidden Markov Model (HMM) that can detect automatically in the time series the occurrence of one particular type of events.The success rate of HMM is reasonable and this technique has the advantage of requiring only one single example to scan the time series.The Random Forest algorithm has proven its efficiency for volcanic and landslide signals classification with higher success rate and versatility (Provost et al., 2017a;Hibert et al., 2017c).New signals are successfully classified in multiple pre-defined classes and changes in the source properties may be detected by change on the uncertainties (Hibert et al., 2017c).It must be noticed that this approach requires a training set with sufficient elements to build the model.Good success rates (i.e.> 85 %) are rapidly reached with 100 elements or more per class.Template-matching filters have also been used in many studies of landslide collapse and glaciers (Allstadt and Malone, 2014;Yamada et al., 2016a;Poli, 2017;Helmstetter et al., 2015a, b;Bièvre et al., 2017;Helmstetter et al., 2017a)  Finally, the location of the sources is the most challenging step.Common location methods (such as NonLinLoc; Lomax et al. (2000Lomax et al. ( , 2009))) were used in combination to 3D-velocity models for locating impulsive micro-earthquakes occurring at the Randa rockslide (Spillmann et al., 2007).However, a certain number of recorded signals do not exhibit impulsive first arrivals and clear P-and S-waves onsets.For this kind of signal, location methods based on the inter-trace correlation of the surface waves waveform (Lacroix and Helmstetter, 2011) or on the amplitude (Burtin et al., 2016;Walter et al., 2017) are more suitable and easier to automatize.Other methods such as HypoLine (Joswig, 2008) aim at integrating different strategies (i.e.first arrival picking, inter-trace correlation and beam-forming) to locate accurately the epicenter under the control of an operator while (Provost et al., 2018) developed a method combining Amplitude Source Location (ASL) and inter-trace correlation of the first arrivals in an automatic scheme.In most of the studies, the media attenuation field and/or the ground velocity is approximated to an 1D model, and/or do not take into account the topography.Both the complexity of the landslide underground structure and of the recorded seismic signals lead to mis-location of the events that prevents for accurate interpretation of certain sources and leads to false alarms (Walter et al., 2017).

Instrumented sites
In the last two decades, seismic networks have been installed on several unstable slopes worldwide.Table 1 synthesizes the unstable slopes or debris flow prone catchments instrumented with seismic sensors worldwide.The sites are classified in terms of landslide types (i.e.slide, fall and flow) according to the geomorphological typology of (Cruden and Varnes, 1996).Studies on snow avalanches (Lawrence and Williams, 1976;Kishimura and Izumi, 1997;Sabot et al., 1998;Suriñach et al., 2001;Biescas et al., 2003) are not integrated.Most of the instrumented sites are located in the European Alps (France, Italy and Switzerland).Short-Period (SP) seismometers and Geophones (G) are the most common type of instruments.Their installation and maintenance is easy as they do not require mass calibration in comparison to Broad-band (BB) or long-period (LP) seismometers.

Data
Seismic observations from 13 sites are used to propose the typology.The sites are representative of various types of slope movements and lithology (Table 1) with four slides occurring in hard rocks, four slides occurring in soft rocks, three rockfallprone cliffs occurring in hard and soft rocks and one catchment prone to debris flows.The seismic instruments installed on these sites are recording the seismicity generated by the slope deformation and are installed either permanently or were acquired during short campaigns (Table 1).The Riou-Bourdoux catchment is the only site where the seismic signals were manually triggered as rock blocks were thrown down the cliff and monitored with cameras, LiDAR and seismic sensors (Hibert et al. (2017a)).
The dimension of the unstable slopes range from 60 m × 30 m for the Chamousset cliff to 7 km × 300 m for the St.-Eynard cliff (Table 2).The seismic networks are deployed with various geometry depending on the configuration of the slope, its activity and the duration of the installation.For most of the sites, at least one seismic sensor is deployed on the active zone or very close to (Table 2).The maximal distance to the slope instabilities is 500 m for the St.-Eynard cliff being the largest investigated site of our study.
The seismic network geometry of the majority of sites are distributed seismic network where sensors location are regularly installed over the active zone or at its vicinity.In the case of the Rebaixader catchment, the seismic network is installed at the border of the stream channel almost linearly.At the Slumgullion landslide, a dense network has been installed with regular spacing of the seismic sensors.Seismic arrays are installed at the other sites.The geometry of the seismic arrays are triangular shape with the exception of the Séchilienne landslide where an hexagonal shape is used.
The instruments are mostly SP seismometers with natural frequencies of 1 Hz to 5 Hz.Fewer geophones and BB seismometers are installed at the sites.The instrument response is corrected for all the dataset.To be consistent with the sensitivity of all the sensors, we do not investigate the data below 1 Hz for BB seismometers and above 100 Hz for SP seismometers and geophones.
The dataset being analyzed is composed of either published seismic events or published catalogs.The comparison of these events and catalogs enable to compare the signals and to compose the classes of the typology.In the case that no published events or catalogs are available, we analyzed manually the dataset to complete the number of examples for each proposed class (see Section 5 for detailed information).

Methodology
The seismic signals recorded at different sites are compared in order to identify common features.Seismic signals result from the convolution of both the wave propagation and of the seismic source mechanism.Consequently, the observation of common signal features in signals recorded at different sites can only be explained by similar source mechanisms.The proposed typology is hence based on the analysis of these common features.We then selected nine signal features in order to quantify the differences and similarities between the different classes.The nine parameters are chosen because they correspond to the criteria used by experts to analyze and classify a seismic signal and also because they can be used in automatic classification algorithms (Fäh and Koch, 2002;Langer et al., 2006;Curilem et al., 2009;Hammer et al., 2012Hammer et al., , 2013;;Hibert et al., 2014a;Ruano et al., 2014;Maggi et al., 2017;Provost et al., 2017a;Hibert et al., 2017c).They can be computed for any signal types and present a robust framework for future comparison.The selected signal features are: the duration of the signal T (expressed in second), computed on the stacked spectrogram of the traces (Helmstetter and Garambois, 2010).
the dissymetry coefficient of the signal (expressed in percent), computed as: with t 1 , t 2 and t m the time of the signal onset, ending and maximum respectively.
the number of peaks of the signal envelop N peaks , computed as the number of local maximum above 50% of maximal value of the signal envelop.The envelop of the signal is computed as the absolute value of the Hilbert transform of the signal.The envelop is smoothed by a computing the average on a moving window of length: δt = 100 fsT .
the duration of the signal auto-correlation, defined as: with, with C equal to the signal auto-correlation.A max is expressed in percent (%) and represents the duration of the signal correlating with itself.As an example, a signal with a rapid and abrupt change in frequency content will rapidly be uncorrelated (low A max ) while a signal with a constant frequency content will have a long auto-correlation (high A max ).
the mean frequency (expressed in Hertz), computed as: with the Power Spectral Density (PSD) defined as: with f s and N being the sampling frequency of the signal and the number of samples respectively.The mean frequency is chosen here as it more representative of the signal spectrum energy and less sensitive to noise than the frequency of maximum energy.(Farin et al., 2014).
the frequency corresponding to the maximal energy of the spectrum F max (expressed in Hertz).
the frequency bandwidth F w defined as: the minimal frequency of the signal spectrum, computed as: the maximal frequency of the signal spectrum, computed as: the maximal frequency of the signal spectrum f max (not to be confused with parameter F max defined above).(Hibert et al., 2014a;Farin et al., 2014).Scaling laws have also been established between seismic energy, momentum, block mass and velocity before impacts (Hibert et al., 2017a).The frequency content is mainly controlled by the block mass.
The frequency of the spectral maximum energy decreases when the block mass increases (Farin et al., 2014;Huang et al., 2007;Burtin et al., 2016).If the rockfalls are well isolated, each impact generates impulsive waves.In the case of multiple rockfalls or short distances between the seismic sources and the sensors, the first arrivals may be emergent due to simultaneous arrivals of waves generated by impactors of different sizes impacting the ground at closely spaced time intervals (Levy et al., 2015;Hibert et al., 2014a).distinguishable and the signal is dominated by surface waves (Deparis et al., 2008;Dammeier et al., 2011;Helmstetter and Garambois, 2010;Hibert et al., 2014a;Levy et al., 2015).The dissymetry coefficient of the signal varies between 30% and 75% and depends on the acceleration and the volume of mass involved in the flow through time (Suriñach et al., 2001;Suriñach et al., 2005;Schneider et al., 2010;Levy et al., 2015;Hibert et al., 2017b).The frequency ranges from (> 500 m).Like dry granular flow, the duration auto-correlation is very weak (T corr =0%) and no seismic phase can be distinguished.The seismic sensors are often installed at very close distance to the flow path so high frequencies up to 100 Hz may be recorded (Abancó et al., 2014;Burtin et al., 2016;Walter et al., 2017).Little energy is present in the lowfrequencies (< 10 Hz) depending on the amount of water and the size of the rocky blocks integrated in the flow (Burtin et al., 2016).The signal is emergent and the amplitude variation depends on the mass involved in the flow passing in the vicinity of the sensor.Debris flows are very often divided in a front with the largest boulders and the highest velocity followed by a body and a tail where the sediment concentration and the velocity decreases (Pierson, 1995).The seismic signal amplitude hence increases progressively as the front is passing at the vicinity of the sensor (Abancó et al., 2012;Hürlimann et al., 2014;Burtin et al., 2016;Walter et al., 2017) and decreases progressively, as the front is moving away from the sensor (dissymetry coefficient > 50%).Large spikes and low-frequencies may be observed in the seismic signal corresponding to the front of the debris flow generated by large boulders impacts.The frequency content also changes and, progressively, energy in the lower frequencies decreases (Fig 5 .a).

Slopequake (SQ)
The "Slopequake" class gathers all the seismic signals generated by sources located within the slope at the sub-surface or at depth such as fracture related sources or fluid migration (cf.section 2).Different names have already been proposed for this kind of signals: "slidequakes" (Gomberg et al., 2011), "micro-earthquake" (Helmstetter and Garambois, 2010;Lacroix and Helmstetter, 2011), "quakes" (Tonnellier et al., 2013;Vouillamoz et al., 2017) or "Landslide Micro-Quake (LMQ)" (Brückl, 2017).We here proposed the term "Slopequake" as a general name for these events.They are characterized by short duration (< 10 s) and are sub-divided in two classes "Simple" and "Complex".

Simple Slopequake
"Simple Slopequake" signals are of short (< 2 s) to very short duration (< 1 s) signals.Their main feature is the triangularshape of the spectrogram with largest amplitudes being recorded in the first part of the signal (dissymetry coefficient < 50%).
The first arrivals contain the highest frequencies of the signal and are followed by a decrease of the frequencies.Depending on the frequency content, these signals can be sub-divided into three classes: -Low-Frequency Slopequake (LF-SQ) ( -High-Frequency Slopequake (HF-SQ) (Fig 7)): The signal lasts between 1 and 5 s.The maximal amplitude of the signal waveform occurs close to the beginning of the signal (dissymetry coefficient < 30%).The waveform presents only one peak and the first arrivals are mainly impulsive.Different phases may be observed (Spillmann et al., 2007;Lévy et al., 2010): P-arrivals are detected at the beginning of the signal and correspond to the high frequency waves, surface waves are then observed at the time the frequency decreases.However, in general the short sensor to source distance makes difficult the differentiation between the different seismic phases.The auto-correlation these signals is hence lower than for LF-SQ (< 10 %).In most of the cases, the picking of the different waves onset is made difficult because of the sensor-to-source distances and the low frequency sampling.The largest PSD values are observed between 3 and 45 Hz with a mean frequency ranging between 20 and 30 Hz.
-Hybrid Slopequake (Hybrid-SQ) (Fig 8)): The signal lasts between 1 and 2 s.It presents the characteristics of the two precedent signals.The brief first arrivals are very impulsive and last less than one second.They are followed by a low-frequency coda similar to the LF-SQ.The maximal amplitude of the signal waveform occurs close to the beginning of the signal (dissymetry coefficient < 40%).The waveform presents only one peak and the first arrivals are impulsive.
These signals are suspected to be associated to boundary or basal sliding (Helmstetter and Garambois, 2010;Gomberg et al., 2011;Walter et al., 2013b;Tonnellier et al., 2013) or fracturing of the slope (Helmstetter and Garambois, 2010;Colombero et al., 2018).Currently, only few studies have proposed inversion of the source tensor (Lévy et al., 2010).To the best of our knowledge, for soft-rock landslides, no source mechanism was modeled.Therefore, it remains difficult to set if the observation of LF-and HF-slopequakes is due to attenuation of the high frequencies with the distance or to the source mechanism.Indeed, the rupture velocity may explain the difference of frequency content and low-frequency earthquakes are observed on tectonic faults (Shelly et al., 2006;Brown et al., 2009;Thomas et al., 2016).They are characterized by low magnitude (M w < 2) and short duration (< 1 s) and constitute at least part of the seismic tremor signal.Therefore, the main assumption for the source of these events are slow rupture (Thomas et al., 2016).Another interpretation for the low frequency quakes dominated by surface waves is crevasse opening (at the surface) as observed in glacier (Deichmann et al., 2000;Mikesell et al., 2012).Colombero et al. (2018) analyzed AE at laboratory scales generated during thermal fracturing.During this experiment, high-frequency AEs are recorded during the heating stage up to the failure of the rock sample and are interpreted as thermal cracking events (Colombero et al., 2018).Low-frequency AEs are recorded during cooling stage (after failure) and are associated to stick-slip events (Colombero et al., 2018).
Hybrid slopequakes are very similar to the events recorded on volcanoes and glaciers with the presence of fluids in conduits or crevasses (Chouet, 1988;Helmstetter et al., 2015b).The sources of these events are assumed to be related to hydro-fracturing.
The first high-frequency events corresponding to a brittle failure is followed by water flow into the newly opened cracks (Chouet, 1988;Benson et al., 2008).
The frequency content depends on the sensor to source distance and on the source mechanism.Observation of LF-and HF-SQ may be the signature of on-going processes taking place within the slope instabilities justifying the three proposed classes for simple slopequakes.

Complex Slopequake
The second class of short duration signals has the same general properties as the simple slopequakes but exhibits particular frequency content or precursory events.These additional characteristics change the possible interpretation of the sources.
Consequently, these signals are gathered in the class "Complex Slopequake".Three different sub-classes are proposed: -Slopequake with precursors ( ranges from 5 to 100 Hz but varies significantly at each site.At all sites, the amplitude of the signal is significantly higher for one of the sensor (3 to 50 times higher) when considering vertical traces.The precursory signal is buried in the noise at the sensors with lowest amplitudes and the signal is similar to a LF-slopequake.Such events have never been documented to our knowledge.They are likely to be generated by a strong and local source located at the very close vicinity of one of the sensor (< 10 m) due to the maximal amplitude (> 10 5 nm.s −1 ) and the rapid decrease of the amplitude recorded by the other sensors.Although the signal is similar to certain earthquakes (the precursory signals interpreted as P-waves arrivals and the strong arrivals as surface waves), no earthquake location can explain the signal recorded at the time these events are recorded.Their occurrence in the night time makes a human activity unlikely to be the source.The most probable source would then be the detachment of a single block and its fall in the vicinity to one of the sensor.This kind of precursory signals are observed for some rockfalls ( events may also be a very rapid succession (< 1 s) of shear events along the basal or the side bounding strike-slip faults (Hawthorne and Ampuero, 2017).Further investigations are needed to analyze their occurrences over time and their location to confirm one or the other assumptions.

Discussion
The proposed typology is summarized in In the case of the Super-Sauze datasets, rockfall are characterized by a lack of energy in high frequencies due in this case to the distance between the seismic network and the scarp.Installation of additional sensors could be the easiest way to get rid of this variability.It must also be noted that, differentiating flow and fall signals may be challenging.Indeed, some of the events are very likely a mix of these two sources.Rockfalls of various blocks may generate granular flows with metric block impacts, both overlapping in the recorded seismic signals.Presence of metric rocks is also observed in debris flow prone torrents; for this type of events, the block impacts within the mass flows are recorded in the seismic signals (Burtin et al., 2016).
Our analysis does not allow at this stage, to conclude whether the frequency content of simple slopequake is associated to source mechanism because complete catalogs differentiating these two classes are not yet available.Colombero et al. (2018) suggested that HF-slopequake are the dominant class of slopequake at the Madonna del Sasso cliff (hard-rock) and were generated by thermal cracking while LF-slopequake associated to frictional sliding are less frequent.Although we did not investigate the whole datasets, no LF-slopequakes were provided at two hard-rock cliffs: Aaknes and Chamousset (Fig 12) while LF-slopequake are recorded at hard-rock slides: La Clapiè re and Séchilienne (Fig 12).This observation seems to confirm the results of Colombero et al. (2018).However, further comparison of the occurrence of the different slopequakes at specific sites in space and time must be done to improve the comprehension of these sources and confirm this statement.
Harmonic signals have also been documented at the Pechgraben and Super-Sauze landslides (Vouillamoz et al., 2017).
These signals last from 1 to 5 s and may repeat during minute-lasting sequences.The proposed interpretation includes hydrofracturing or repetitive swarms of micro-earthquakes (Vouillamoz et al., 2017) while Gomberg et al. (2011)  No long-lasting tremors are presented in this study.Schöpa et al. (2018) recorded a tremor with gliding before the occurrence of the Askja caldera landslide.Similar tremors have been found on the Whillans ice stream in Antarctica during slow slip events (Paul Winberry et al., 2013;Lipovsky and Dunham, 2016), which repeat twice a day with a slip of about 10 cm lasting for about 20 minutes.Therefore, such signals may also occur during the nucleation phase of landslide failure.The question remains if they are not observed because landslide acceleration is aseismic due to high pore fluid pressure (Scholz, 1998) or low normal stress at the sub-surface of the slope.
Difficulties still arise in providing an exhaustive description and interpretation of all the sources from the simple analysis of the proposed signal features, particularly those generating short-duration signals, in particular for the classes discriminated by the frequency content such as LF-SQ and HF-SQ.The ambiguity between propagation effect and source mechanisms prevents further interpretation due to several limitations.Firstly, the location of the sources remain difficult to establish due to the complexity of some of the signals (Gomberg et al., 2011;Lacroix and Helmstetter, 2011;Tonnellier et al., 2013;Provost et al., 2018), the size of the instrumented sites and the complexity of the underground structure that influences the polarisation of the waves (Neuberg et al., 2000) and the sensors (ie.number, location and type: 1C/3C sensor) installed close to the unstable slopes (Godano et al., 2009).The location of the epicenter of most of the events seems coherent with the instabilities deformation field at the surface (Helmstetter and Garambois, 2010;Levy et al., 2011;Walter et al., 2013b;Provost et al., 2018) although resolving dispersion and 3-D heterogeneities of the velocity fields currently prevent to infer the depth of the events and their focal mechanisms.Secondly, a complementary approach to explain the origin of the sources is the analysis of their occurrence with respect to surface or basal displacement and monitoring of the water content and pore fluid pressures.It requires both exhaustive catalogs of landslide seismicity over long time periods and continuous and distributed datasets of displacements and pore fluid pressures which remain challenging to acquire.Finally, on addition to the characteristics of seismic signals, further information on the sources processes can be obtained from the distribution of the events in time, space and size.Events that occur regularly in time with similar amplitudes are likely associated with the repeated failure of an asperity surrounded by aseismic slip, for instance, at the base of a glacier (Helmstetter et al., 2015a) or of a landslide (Yamada et al., 2016a;Poli, 2017).
Signal amplitudes and recurrence times often display progressive variations in time.In contrast, events that are clustered in time and space, with a broad distribution of energies, are more likely associated with the propagation of a fracture (Helmstetter et al., 2015b).The daily distribution of events time can also be helpful to identify anthropogenic sources, that occur mostly during the day.In contrast, natural events are more frequently detected at night, when the noise level is smaller.
Simulations and models are also required to explain the current observations.Indeed, experimental results suggest an increase of acoustic emissions correlated with the increase of the slope velocity (Smith et al., 2017) or an increase of acoustic emission due to the creation of the rupture area (Lockner et al., 1991).Acceleration of pre-existing rupture surface(s) seems to be the mechanism responsible for the seismicity recorded before large rockslide collapse.Yamada et al. (2016a); Poli (2017) argued that the high correlation between the repetitive events could only be explained by stick-slip movement of the locked section(s), while a cracking process would imply a migration of the location of the events and a change in the events waveforms.Schöpa et al. (2018) argued that the presence of gliding frequencies could only be produced by similar sources and hence close location.On the contrary, in the case of the Mesnil-Val column, Senfaute et al. (2009) interpreted the evolution from high frequency to low frequency events as the progressive formation of the rupture surface followed by the final rupture process immediately before the column collapse where both tensile cracks and shearing motion on the created rupture are generated.

Conclusions
Over the last decades, numerous studies have recorded seismic signals generated by various types of landslides (i.e.slide, topple, fall and flow), for different kinematic regimes and rock/soil media.These studies demonstrated the added-value of analyzing landslide-induced micro-seismicity to improve our understanding of the mechanisms and to progress in the forecast of landslide evolution.
In this work we propose a review of the endogenous seismic sources generated by the deformation of unstable slopes.A dataset of fourteen slopes is gathered and analyzed.Each of the source is described by nine quantitative features of the recorded seismic signals.Those features provide distinct characteristics for each type of source.A library of relevant signals recorded at relevant site is shared as supplementary material.We propose three main classes "slopequake", "rockfall" and "granular flow" to describe the main type of deformation observed on the slopes.Slopequakes are related to shearing or fracturing processes.
This family exhibits the most variability due to the complexity of the sources.These variations are likely to be generated by different source mechanisms."Rockfall" and "granular flow" classes are associated to mass propagation on the slope surface.
They are distinguishable by the number of peaks clearly identified in the seismic signals.
Presently, several descriptions of the seismic sources are proposed for each study case.We believe that a standard typology will allow to discuss and compare seismic signals recorded at many unstable slopes.We encourage future studies to use and possibly enrich the proposed typology.This also requires publication of the datasets and/or catalogs to progress towards a common interpretation.Recently, organizations such as the United States Geological Survey (USGS) or the French Landslide Observatory (OMIV) have started this work (RESIF/OMIV, 2015).
Recent arrival on the market of relatively cheap and autonomous seismometers (eg.ZLand node systems, Raspberry-Shake systems) will allow the deployment of denser seismic networks of 3C sensors.The latter will certainly improve the location accuracy and enable inversion of the focal mechanism of the sources.Moreover, the recent operational applications of Ground-Based SAR (Synthetic Aperture Radar) and terrestrial LiDAR technologies for monitoring purposes shows their relevance to monitor distributed surface displacements.On-going monitoring on several landslides combining those innovative approaches will certainly help to associate SQ events to deformation processes (Dietze et al., 2017b;RESIF/OMIV, 2015;Provost et al., 2017b).
The proposed typology will help to constrain the design of new models to confirm the assumptions on the nature and the properties of the seismic sources.This will be particularly important for 1) explaining the variability of the SQ sources observed at the sites, 2) progressing in the physical understanding of the SQ sources, and 3) ascertaining the spatio-temporal variations of the seismic activity observed at some unstable slopes in relation with their deformation as well as, with external forcings such as intense rainfalls and earthquakes.
acquisition at, respectively, the Pechgraben site and the Aiguilles-Pas de l'Ours site.The authors gratefully acknowledge Dr. Velio Coviello and Wei-An Chao for their carefull reviews and thorough comments which helped improving the manuscript.(Abancó et al., 2012;Hürlimann et al., 2014;Arattano et al., 2016).See in order to detect and classify seismic signals.This method consist in scanning continuous data to search for signals with waveforms similar to template signals.It can detect seismic signals of very small amplitude, smaller than the noise level.Seismic signals are grouped in clusters of similar waveforms, implying similar source locations and focal mechanism.
Fig 2 displays the seismic waves recorded for a single block fall at the Riou-Bourdoux catchment (French Alps).The block was manually launched in the catchment and recorded with seismic sensors and cameras(Hibert et al., 2017a).The signal is characterized by successive impacts visible both on the waveform and on the spectrograms and lasts around 20 s.The spectral content contains mostly frequencies above 10 Hz but energy below 10 Hz is present for certain impacts (Fig2a).At closer distance, very high frequencies can be recorded up to 100 Hz (Fig2a).The auto-correlation remains large over time due to the similitude of the individual impacts signals (T corr > 10%).P-and S-waves are hardly distinguishable on the record and the signals recorded at the seismic sensors are dominated by surface waves(Dammeier et al., 2011;Helmstetter and Garambois, 2010;Hibert et al., 2014a;Levy et al., 2015).Seismic signals of natural masses detaching from cliffs are presented in Fig 3.They present similar characteristics to the artificially triggered rockfall.Depending on the height of the cliff, the signal lasts between 5 and tens of seconds.The symmetry of the signal ranges from 0 to 80 % depending on the cliff configuration.In general, the most energetic impacts are recorded Granular flows are characterized by cigar-shape signals lasting between tens to thousands of seconds.They are subdivided in two classes: -Dry granular flow (Fig 4): These signals are characterized by cigar-shape waveforms of long duration (< 500 s).Due to the absence of water, the source generally propagates over small distances.The duration of auto-correlation is very weak (T corr ≈0%) and no seismic phase can be distinguished.No distinguishable impacts can be observed in the waveform nor in the spectrogram at the opposite of rockfall signals.The signal onsets is emergent and P-and S-waves are hardly

1
to 35 Hz.The maximal frequency of the PSD varies between 5 and 10 Hz and can be larger (up to 20 Hz) when the seismic sensors are located close to the propagation path.The PSD values are significantly low below 3 Hz and incrase rapidly between 3 and 20 Hz. -Wet granular flow (Fig 5): These signals last several thousands of seconds to several hours and correspond to debris flows.They occur during rainfall episodes when fine material and boulders propagates downstream over long distances Fig 6): The signal lasts between 1 and 5 s.The maximal amplitude of the signal waveform occurs at the beginning or at the center of the signal (15% < dissymetry coefficient < 50%).The waveform presents only one peak and most of the first arrivals are emergent.Phase onsets are difficult to identify.The signals are mostly dominated by surface waves.Consequently, the duration auto-correlation of the signals is large (> 10%).The largest PSD values are observed between 5 and 25 Hz with a mean frequency ranging between 10 and 15 Hz.
Fig 9): The third class of short duration signals are similar to the slopequake signals but are preceded by a precursory signal of smaller amplitude (Fig 9).The content of the precursory signal ranges from 5 to 100 Hz depending on the site and is slightly lower than the highest frequency generated by slopequake-like event.The precursory arrival last up to 1.2 s in the presented examples and no clear phases are detected.The frequency content Fig 3.a) and at a the Saint-Martin-le-Vinoux quarry (France; Helmstetter et al. (2011)).At the Saint-Martin-le-Vinoux underground quarry, the duration between the detachment and the signal impact is well correlated to the room height.This interpretation is coherent with the drop of amplitude before the more energetic event at the Chamousset rock column (Fig 9.c) where a progressive decrease of the precursory signal is observed.However, on the other sites (Fig 9.a,.b)such decrease is not present.The one second lasting precursory signal has a constant amplitude and frequency content.Another interpretation could be that these precursory signals are a succession of overlaping slip or fracture events.The interpretation of these signals cannot be established with certainty and further analysis (i.e.location, time of occurrence) and other examples are needed to discriminate the mechanism at work.-Tremor-like slopequake (Fig 10): The last class of short duration signals often last between 1 and 5 seconds (Fig 10).They present a symmetrical waveform (S=50%) with emergent arrivals and slow decrease of the amplitude to the noise level.The frequency ranges from 5 Hz to 25 Hz.High-frequencies may be briefly recorded in certain events(Fig 10.c)    .The maximal energy of the PSD corresponds to a frequency of 8 to 13 Hz while the mean energy corresponds to a frequency of 13 to 17 Hz.No seismic phases are identified.The signal is not recorded by all the sensors even when the sensors are organized in small arrays with short inter-sensor distances (< 50 m).Their waveforms and frequency content are similar to the one of the granular flows(Fig 4).Small debris flows have been observed at La Clapière and Super-Sauze landslides and are likely to generate seismic waves; however, small debris flows are not observed at the Pas de l'Ours landslide when these kinds of seismic signals are recorded.Another possible source mechanisms for such Fig 11.The approach consisted of comparing the datasets of different sites in order to identify the common features of the recorded seismic signals.Three main classes can be differentiated mainly from the length of the signals, the number of peaks and the duration of the auto-correlation.Figure 12 shows more examples of the signal variability for the sites where long seismic catalogs have been recorded (e.g.Aaknes, Chamousset, Séchilienne, Super-Sauze and La Clapière).Only the signals classified as Rockfall, LF-and HF-slopequake are presented because fewer events of the other classes are present in the investigated datasets.The signal features are in good agreement with the defined classes proposed in the present classification (Fig 11).In general, narrow variability is observed on the feature values among the different sites and consequently, the observed features are likely associated to the source mechanism.However, some variability exist for rockfall events.Indeed, the volume of the blocks and possible breaks control the frequency content and the auto-correlation duration while the height of the scarp will play a significant role in the duration of the event.Depending on the site, rock fall signals can have similar features to each other (e.g.Séchilienne, Fig 12) suggesting a constant source mechanism or very different features suggesting the presence of different rockfall mechanisms and/or trajectories (e.g.Super-Sauze, Fig 12).
hypothesized that these kinds of signals were caused by trapped waves along the side-bounding strike-slip fault generated by shear events.In the investigated datasets, similar signals are recorded at the La Clapière and the Aiguilles landslides with a fundamental frequency of 8 ± 1 Hz(Fig 13.b,c).At the Séchilienne landslide, harmonic signals are also detected(Fig 13.d), mostly during the day, with different resonant frequencies between 2 and 12 Hz.Similar signals are observed at the Slumgullion and Super-Sauze sites but without clear harmonics in the PSD(Fig 13.e,f).The presence of pipes and drains on or in the vicinity of these sites could also explain the origin of these signals justifying that these signals are not included in the Slopequake class as they may not be generated by a slope deformation process.The location of the source, the distribution of the amplitude, the stability of the fundamental frequency and the daily temporal occurrence of the source supports this assumption.Systematic location of these events is needed to determine if they must be integrated or not in the general typology in the case that they are generated by fluid resonance in fractures.For certain signals, the coda is dominated by resonance frequencies(Fig 3d, Fig 9c)  at high frequencies (i.e.20 and 43Hz), well observed in the spectrogram of the signal.The resonance is not present before the beginning of the signal and hence can not be due to anthropogenic noise (i.e.motors).In the case of Chamousset cliff,Levy et al. (2011) explained the presence of this monochromatic coda by the resonance of the rock column after the occurrence of the rock bridge breakage.At the Super-Sauze, similar resonant coda are observed at the end of certain rockfalls(Figure 4.d).Considering the distance between the main scarp and the seismic arrays (> 300 m) and the absence of large fracture on the scarp, the occurrence of this kind of resonance is very surprising in this case.This signals feature could also result from the wave propagation (i.e.trapped waves).

Figure 1 .
Figure 1.Conceptual scheme of the landslide endogenous seismic sources with a) wet granular flow, b) dry granular flow, c) rockfall, d) tensile fracture opening, e) tensile cracks opening, f) shearing and h) fluid migration in fracture.

GFigure 2 .
Figure 2. Example of one controlled rockfall (mass= 430kg) at the Riou-Bourdoux catchment (Hibert et al., 2017a) recorded by SP seismometer located at 50 m of the rock departure (left) and recorded by BB seismometer near the rock arrival (right).The waveforms of the vertical traces are plotted on the upper part of the figure.The amplitude are normalized on the trace with the maximal amplitude (black), the signal recorded by the other sensors (when available) are represented in color below.The maximal amplitudes (Amax) of all the traces are plotted on the sub-plot in nm.s −1 .The spectrogram is plotted on the middle part of the figure and normalized to the maximal energy.The lower part of the figure represents the PSD of the most energetic trace and the frequency corresponding to the maximum and the mean of the PSD are plotted in red and gray respectively.

Figure 4 .
Figure 4. Dry granular flow events recorded at a) Séchilienne and b) the Piton de la Fournaise Caldera.See Fig. 2 for description of the figure.

Fig. 2
Fig. 2 for description of the figure.

Figure 6 .
Figure 6.Low-Frequency Slopequakes recorded at the a) Slumgullion (Gomberg et al., 2011), b) Pont-Bourquin, c) La Clapière and d) Aiguilles-Pas de l'Ours slopes.See Fig. 2 for description of the figure.Note that for the Slumgullion signals (a), the amplitudes are expressed in counts.

Figure 8 .
Figure 8. Hybrid-Slopequake recorded at the a) Pechgraben and b) Super-Sauze landslide.See Fig. 2 for description of the figure.

Figure 9 .
Figure 9. Examples of Slopequakes with presursory event recorded at the a) Super-Sauze, b) Séchilienne and c) Chamousset slopes.See Fig. 2 for description of the figure.

Figure 10 .
Figure 10.Examples of repetitive Slopequakes recorded at the a),c) Super-Sauze, b) La Clapière and d) Aiguilles-Pas de l'Ours slopes.See Fig. 2 for description of the figure.

Figure 11 Figure 12 .
Figure 11.a) Summary of the proposed classification with plot of the attributes for the examples presented in the precedent figures and an example of waveform for each class.The convention for the attribute plot is presented in b); n being the number of seismic signals examples used to plot the feature diagram.

Figure 13 .
Figure 13.Examples of pure harmonic signals recorded at the a) Pechgraben, b) La Clapière and c) Aiguilles-Pas de l'Ours, d) Séchilienne, e) Slumgullion (Gomberg et al., 2011) and f) Super-Sauze slopes.See Fig. 2 for description of the figure.Note that for the Slumgullion signals (a), the amplitudes are expressed in counts.

Table 1 :
Table of the instrumented sites.The bolded names correspond to the sites investigated in the present paper to establish the typology.