Brief communication: Seismological analysis of ﬂood dynamics and hydrologically-triggered earthquake swarms associated with storm Alex

. On October 2, 2020, the Maritime Alps in southern France were struck by the devastating storm Alex that caused locally more than 600 mm of rain in less than 24 hours. The extreme rainfall and ﬂooding destroyed regional rain and stream gauges. That hinders our understanding of the spatial and temporal dynamics of rainfall-runoff processes during the storm. Here, we show that seismological observations from permanent seismic stations constrain these processes at a catchment scale. The 5 analysis of seismic power, peak frequency, and backazimuth provide us with the timing and velocity of the propagation of ﬂash-ﬂood waves associated with bedload-dominated phases of the ﬂood on the Vésubie river. Moreover, the combined short-term average to long-term average ratio and template matching earthquake detection reveal that 114 local earthquakes between local magnitude ML=-0.5 and ML=2 were triggered by the hydrological loading and/or the resulting in-situ underground pore pressure increase. This study shows the impact of storm Alex on the Earth’s surface and deep layer processes and paves the 10 way to future works that can reveal further details of these processes. Spectral Densities (PSD,Solomon (1991b)) during the storm. Then, we perform additional analysis on station SPIF by assessing temporal changes in: (1) peak frequency, (2) dominant backazimuthal orientation of seismic noise, and (3) relation between high frequency (10-45 Hz) and low frequency (1-10 Hz) seismic noise. We contextualize these seismological observations by comparing them with runoff temporal series. A simple KLEM rainfall-runoff model (Kinematic Local Excess Model, Borga et al. is used for runoff simulation. A full 75 description Code and data availability. Obspy Python routines (www.obspy.org) were used to download waveforms and pre-process seismic data.The seismic data is collected under the network code FR (10.15778/RESIF.FR, SPIF, TURF, MVIF stations) and RA (10.15778/RESIF.RA, BELV station) and all seismic data are openly available in the archives of French seismological and geodetic network Résif (https://seismology.resif.fr/). The code used for backazimuth analysis can be found in the online supplement of Goodling et al. (2018) paper. Rainfall data (ANTILOPE 260 and COMEPHORE) were provided by Météo-France and are available on request. To gain access please contact Pierre Brigode.

Seismic methods have the potential to monitor surface and subsurface processes associated with extreme weather events.
In particular, both turbulent flow and sediment transport during floods generate ground motion in different frequency bands 20 (Schmandt et al., 2013;Gimbert et al., 2014) that can be used to track the flood dynamics (e.g., Cook et al., 2018). Surface seismic waves are generated by impact forces exerted by mobile particles on the river bed (e.g., Tsai et al., 2012;Gimbert et al., 2019) and ambient seismic measurements have recently been used to monitor fluxes associated with transported bed material Bakker et al. (2020); Lagarde et al. (2021). In the past decade, near-river seismic monitoring has been conducted during moderate-magnitude floods (e.g., Burtin et al., 2016;Roth et al., 2016) and controlled small-magnitude flow events 25 (Schmandt et al., 2013(Schmandt et al., , 2017. To date, extensive seismic investigations of large-magnitude flood events are rare and mostly associated with glacier lake outburst floods (Cook et al., 2018;Maurer et al., 2020) and natural hazard cascade (Cook et al., 2021). Yet, improved understanding of flood dynamics is crucial for early warning, risk mitigation, and modeling landscape evolution (Raynaud et al., 2015;Borga et al., 2019).
Exceptionally intense rainfall can reactivate existing faults through changing crustal stress conditions due to additional fluid 30 mass load or in situ stress changes, resulting in hydrologically trigerred earthquakes (e.g., Hainzl et al., 2006). Over the past two decades, a growing number of studies has shown a correlation between meteorological events and earthquake activity in various geological contexts (Costain and Bollinger, 2010). Several sites show the seasonal modulation of the seismicity due to rainfall or snowmelt periods in Japan (Ueda and Kato, 2019); Nepal (Kundu et al., 2017); Taiwan (Hsu et al., 2021); Oregon, USA (Saar and Manga, 2003); California, USA (Johnson et al., 2017;Montgomery-Brown et al., 2019); and Italy (D'Agostino 35 et al., 2018). Other observations display a punctual increase of seismic activity following an exceptional rainfall episode, for example in the Swiss Alps (Roth et al., 1992), German Alps (Kraft et al., 2006), and southern France (Rigo et al., 2008).
Here, we present a set of seismological observations from 11 stations from the permanent French RESIF Network that captured the October 2020 extreme rainfall and flash flood caused by storm Alex (Carrega and Michelot, 2021) in the southwestern Alps (the Maritime Alps), South-East France. This unique dataset not only allows us to study surface flash-flood related hazard, 40 but also the seismogenic subsurface response to an unusually intense rainfall which locally exceeded 600 mm in less than day.
Three rivers were strongly impacted by the flash floods: the Vésubie, the Roya, and the Tinée rivers (Figure1A). We first gain insights onto the Vésubie river dynamics during the flash flood by analyzing seismic power, peak frequency, and dominant backazimuthal orientation of seismic noise. These observation are compared with simple rainfall-runoff modeling (Brigode et al., 2021a). Then, by using template matching we detect a series of impulsive signals that correspond to small earthquakes 45 [down to local magnitude (ML) of -0.5] in the area where rainfall rate in the Tinée river catchment area was maximum.
These preliminary analyses demonstrate that the seismological observations reported herein provide a better understanding and quantification of the hydro-geological impact of extreme weather phenomena on the mountainous terrain and the related fluvial hazards. The latter is important for catchment areas with few "classical" hydrological observations, as the Vésubie river catchment presented here. 50 On October 2, 2020, the Maritime Alps were struck by a violent meteorological event called a "Mediterranean episode" caused by storm Alex (Carrega and Michelot, 2021). Although heavy rainfalls occur regularly in autumn in the Mediterranean region, the storm Alex maximum daily rainfall was the highest measured since 1997 ( Figure 1C). The rainfall started at 06:00 UTC on October 2, 2020, lasted for less than 24 hours and generated a cumulative intensity that locally exceeded the typical yearly 55 average (>600 mm/day, Figure 1A). These estimates have been obtained hourly with the ANTILOPE model (Laurantin, 2008) with 1 km 2 spatial resolution. The ANTILOPE model was produced by Météo-France and constrained by radar data and 40 rain gauges located in the region ( Figure 1A), although the estimation of rainfall maps is highly uncertain in this context as a result of few rain gauges available, rainfall measurement uncertainties due to observed intensities, limits of the radar observations, and spatial interpolation.

60
The torrential rains triggered hazardous flash floods and landslides of an intensity and spatial extent that have never been observed yet in this area, causing several casualties as well as large infrastructure and economic damage [ Figure 1A,B1, Brigode et al. (2021b)]. To date, the spatial and temporal evolution of the flood remain poorly understood. This is partially due to a limited number of observations caused by the instrument destruction during the flood. Stream gauge measurements during the episode are incomplete and highly uncertain due to scale saturation, destruction of measuring devices, and changes in the 65 river bed level.
We focus our flood-dynamics analysis on the Vésubie river because (1) the Vésubie catchment has been one of the most strongly affected in the region (Figure 1, B1, B2) and (2) the seismic station coverage is particularly adequate in this catchment with three seismic stations being located in the proximity: SPIF (3-component velocimeter), BELV (3-component accelerometer) and TURF (3-component velocimeter) at respectively about 1,570, 630 and 5,970 m from the Vésubie river. We investigate 70 the level of seismic power recorded by these stations by calculating the Power Spectral Densities (PSD, Solomon (1991b)) during the storm. Then, we perform additional analysis on station SPIF by assessing temporal changes in: (1) peak frequency, (2) dominant backazimuthal orientation of seismic noise, and (3) relation between high frequency (10-45 Hz) and low frequency (1-10 Hz) seismic noise. We contextualize these seismological observations by comparing them with runoff temporal series. A simple KLEM rainfall-runoff model (Kinematic Local Excess Model, Borga et al. (2007)) is used for runoff simulation. A full 75 description of the methods used in this paper is provided in Appendix A.  Location in France of the study area is marked by a blue dot in the inset map. B. The village of Saint-Martin Vésubie before (summer, 2020) and after the storm episode (source: IGN 2020). C. Annual maximum daily rainfall rate from the COMEPHORE database (hourly rainfall estimated on pixels of 1 km 2 , available since 1997, Tabary et al. (2012)) calculated over the 25 km 2 rectangle shown in Figure B2A. .We contextualize our observation using a rainfall estimation obtained with the ANTILOPE model (Laurantin, 2008). The ANTILOPE model was produced by Météo-France based on radar data and 40 rain gauges located in the region ( Figure 1A).
The estimation of rainfall maps is highly uncertain in this context (few rain gauges available, rainfall measurement uncertainties due to observed intensities, limits of the radar observations, and spatial interpolation). 1-20 Hz. The seismic power during storm Alex is 20 dB higher than the pre-flood "background" ambient seismic noise power levels. Since the decibel scale is a base-10 logarithmic scale, the 20 dB observed difference means 100 times higher seismic power. For the TURF station, the seismic power is increased by about 10 times relative to the pre-flood conditions, mostly at frequencies lower than 5 Hz. The seismic power averaged in the 1-20 Hz frequency band for SPIF and BELV stations ( Figure   2A-B) show a rapid increase in recorded seismic power from 10:00 and 11:00 UTC, respectively. Three local seismic power 90 maxima are visible on SPIF and BELV stations. Their arrival times are marked in color in Figure 2 and the seismic power thresholds used to define the maxima are shown in Figure B3.
The first two seismic power maxima have pronounced high-amplitude peaks and arrive at 12:05 and 13:15 (SPIF), and 12:30 and 13:35 UTC (BELV), respectively. The third maximum has a distributed amplitude in time and occurs between ∼16:00 and ∼20:00 UTC at SPIF and ∼16:00 and ∼22:00 UTC at BELV. The seismic power recorded at the TURF station shows a 95 progressive increase with a single broad peak between ∼17:30 and ∼22:00 UTC. The peak associated with the first maximum has the highest magnitude at the SPIF station, while all three maxima have similar magnitudes on the BELV station. The peak associated with the first maximum lasts for ∼ 30 min and that associated with the second maximum lasts for ∼ 90 min. The peak associated with the third maximum is the broadest, lasting for 4 and 6 hours on the SPIF and BELV stations, respectively.
For the sake of comparison with the runoff modelling, we use a linear scale for the seismic power representation in Figure 2A, to the river points with the shortest distance between the seismic stations and the Vésubie river and are shown in Figure B2A.
Modelling predictions indicate that the runoff maxima occur at 14:00, 14:25, and 15:00 UTC (the first runoff maximum), and 18:00, 18:25, 19:00 UTC (the second runoff maximum), from upstream to downstream. The available stream gauge measure-105 ments at Utelle ( Figure B2A) show a similar rapid increase in runoff as the seismic power and the rainfall-runoff model ( Figure   B4). However, no maximum runoff measurements are available since the stream gauge (marked as a dark-blue diamond in Figure 1A and B2A) was destroyed during the flood. To investigate potential changes in seismic noise sources, we calculate the peak frequency and the backazimuth ( Figure   2E-F). In figure 2E peak frequency values are time color-coded meaning that each color corresponds to the consecutive 200 s 110 long time windows shown in Figure 2A). The peak frequency corresponds to the frequency that has the maximum seismic power value in the analyzed time window ( Figure B5D). Peak frequency and backazimuth (θ, averaged in the 3-8 Hz frequency band, Figure 2F) show distinct value shift at the SPIF station before and during the flood. Starting from 08:30 UTC multiple lightning strikes occurred at the distance of 15 km from the SPIF station (https://www.blitzortung.org/en/, Figure B5). At this time there are higher-amplitude arrivals visible at the SPIF station causing jumps in the peak frequency from 2 Hz to 115 higher values up to 40-50 Hz at 09:30 UTC ( Figure B5). These arrivals can be associated with the lightning/thunder, rain, or anthropogenic activity. However, at 11:00 UTC the peak frequency stabilizes at 6 Hz. Then, the peak frequency drops to 4 Hz at ∼13:20 and comes back to 6 Hz at ∼15:00 UTC. This drop in the peak frequency coincides in time with the second seismic power maximum visible at the SPIF station. The backazimuth starts pointing along a 100°-120°axis at 10:00 UTC ( Figure 2F) although the degree of polarization is relatively weak (β 2 ∼0.5, Figure B6). The dominant degree of polarization (β 2 in the 120 range 0-1), based on Koper and Hawley (2010), provides a measure for the confidence with which the horizontal azimuth can be interpreted, where β 2 >0.5 is recommended by Goodling et al. (2018). Therefore the backazimuth observations should be taken with caution.
The relative contributions of low-(2-10 Hz) and high-(10-45 Hz) frequency seismic power are shown in Figure 2G. Different time periods characterized by a varying relationship between low frequency and high frequency seismic power can be identified: 125 between 08:30 and 10:00 UTC the seismic power increases similarly in the two-frequency range (slope ∼1), between 10:00 and 16:00 UTC the high frequency seismic power increases more strongly (slope >1), and finally between 16:00 UTC October 2 and 07:00 UTC October 3 the seismic power decreases similarly. The equivalent of Figure 2G in the linear amplitude scale [(m/s) 2 Hz -1 ] is presented in Figure B7. We discuss the significance of slope changes in the discussion section.

Earthquake swarm detection
Since 2014, the seismic activity of the studied area is permanently monitored by the Seiscomp3 ( ( Figure 3D). The earthquakes form three distinct swarms in space and time that were mostly successively activated from South to North ( Figure 3B-C). The location error is estimated to be about ±2km. We detected 91 additional earthquakes by applying 140 the template matching detection method (Gibbons and Ringdal, 2006) to the continuous data recorded by the MVIF station ( Figure 3D). The template matching increases the number of detected earthquakes by about 400 % and decreases the minimum magnitude by one unit compared to the Seiscomp3 detections based on STA/LTA. Most of the newly identified events occur on October 8 and may be related to the middle swarm since they best correlate with one of the templates constituting this cluster. the SPIF station points towards 110°direction ( Figure B2, black arrow), which does not point towards the closest river section (located at a backazimuth of 66°). The backazimuth of ∼110°may be associated with a bending of the Vésubie river channel, a ∼2.5 km long downstream reach of the Vésubie river that aligns with the estimated azimuth, or the confluence of the Venanson stream with the Vésubie river which lies in the estimated direction ( Figure B2B). This provides evidence that the commonly made assumption that the recorded seismic signals are associated with the river segments located closest to the station (e.g.,  Both seismic power and peak frequency are site-dependent seismic parameters, i.e. they depend on seismic quality factor, the velocity of Rayleigh waves, and the source-station distance (Aki and Richards, 2002). However, according to a modified Tsai et al. (2012) model for hazardous flow monitoring from Lai et al. (2018), the seismic power is strongly sensitive to particle sediment size and flow speed, while the peak frequency mostly depends on the distance from the seismic source to 160 the receiver. Also, previous observations reported no significant shift in peak frequency with varying runoff (Schmandt et al., 2013;Burtin et al., 2016). Therefore, the observed drop in the peak frequency (down to 4 Hz) that temporally correlates with the occurrence of the second seismic power maximum at the SPIF station (Figure 2A, E) can potentially be generated by an additional, more distant river segment or by a slope failure. Indeed, the flash flood impacted the adjacent hill slopes through undercutting and destabilization of the riverbanks, leading to bank, road, and bridge collapses, and landslides distributed along 165 the river network ( Figure B1). Another possible explanation could be a tributary that becomes a dominant seismic source at this moment. However, the results of the rainfall-runoff modeling for large tributaries (Boréon and Madone de Fenestre river) do not confirm this hypothesis. Also, the backazimuth analysis does not show value changes during the second seismic power maximum. This can be due to (1) changes in the seismic source location that lie in the same general azimuthal direction, (2) the difference in time scale between backazimuth estimates made over 30 minutes versus peak frequency calculations made 170 over 200 s windows, or, perhaps most likely, (3) the low degree of polarization of the surface waves due to spatial spread of the source or to wave scattering.
Since river flow turbulence is expected to preferentially generate ground motion at low frequencies compared to bedload transport (e.g., Burtin et al., 2011;Schmandt et al., 2013;Gimbert et al., 2014), the relationships between seismic power at low versus high frequencies can tell us whether our observations may be sensitive to bedload transport (Bakker et al., 2020).

175
As the flood develops we observe a change in scaling between low-and high-frequency seismic power, materialized by a transition from a 0.8 to a 1.3 scaling exponent as high frequency seismic power becomes higher than -158 dB ( Figure 2G). We interpret this observation as an indication that high frequency seismic power above the -158 dB threshold is mostly bedload induced. This is consistent with expectation of enhanced bedload transport from this stage onwards due to increased bed shear stress and/or the activation of additional sediment supply sources from river bed destabilization or bank erosion (Cook et al.,180 2018). Interestingly, after peak seismic energy has been reached, high frequency seismic power drops drastically compared to low frequency seismic power (with a scaling exponent of about 2), consistent with an abrupt decrease in sediment transport.
At night, the low-versus high-frequency power scaling relation comes back to that observed during the early rising phase, consistent with higher frequencies over this timeframe getting back to being mostly sensitive to water flow. We also note that after the flood, the low-frequency seismic power is higher compared to before the flood (∼10 dB difference, see also the 185 spectrogram of SPIF station in Figure 1), which could be due to flood-induced changes in river bed geometry and/or flow conditions (e.g. river roughness, Roth et al. (2017)) that may preferentially affect low frequency power.
About six hours passed between the beginning of storm Alex and the first flash-flood peak flow. The two seismic power maxima visible at the near-river stations (SPIF and BELV; the first maximum is marked in pink, and the second one in orange in Figures 2A, B and G) occurred in what we identified as the bedload transport phase in Figure 2G. Under the hypothesis 190 that the two peaks associated with seismic power maxima represent the same moving source, we estimate their propagation velocity at 5.8 (+/-1.2) m/s and 4.8 (+/-1.5) m/s, respectively. The details of the propagation velocity calculation and the error propagation calculation are given in appendix B. These peaks overlap in time with the first maximum of runoff simulations (Figure 2A-B). Such elevated and short-live peaks could be generated by flood waves. Similar peaks in seismic power generated by flood waves were observed during glacial lake outburst floods in the Himalayas by Cook et al. (2018) and Maurer et al. 195 (2020). These peaks may share similar characteristics to sediment pulses reproduced experimentally by Piantini et al. (2021) in a torrential river setting as investigated here. Such sediment pulses can be generated by sudden destabilization of debris deposits produced by mass wasting and accumulated at the based of slopes and cliffs.
The absence of the two main maxima on the TURF station can be related to a lack of sensitivity of this station to the bedload transport due to its large distance from the river (∼6 km). Farther distance means stronger geometrical attenuation at higher 200 frequencies versus lower frequencies, and thus lower sensitivity to bedload compared to water flow (Gimbert et al., 2014).
Moreover, due to the location of the TURF station further to the east, this station can be also influenced by the flood on the Roya river that is located ∼10 km away from the station. The timing of the main seismic power maximum at the TURF station and the third seismic power maximum of the BELV station are well correlated with the runoff simulations and can be related to the maximum runoff. From maximum 1 to 3 there is a shift from short-lived peaks to a much more spread distribution of 205 power through time. That could be potentially related to different dynamics of the first two maxima (associated with two fast propagating flood waves causing a sudden rise in seismic power) and a progressive increase in the seismic power associated with a progressive increase in the runoff. Finally, the differences between the observed seismic power and the runoff simulations indicate that the simple runoff simulation cannot fully explain the flash-flood dynamics.

Earthquake swarm in the Tinée valley 210
The spatial coincidence between the maximum rainfall of storm Alex in the Tinée valley and the seismic sequence a few hours later ( Figure 1A) raises the question of whether the earthquakes were triggered by the heavy rainfall. Three different hypotheses can be proposed for the triggering of seismicity by meteorological forcing. The first hypothesis is a pore pressure increase at depth caused by fluid migration from the surface through hydraulically connected fractures. In this case, the time lag between rainfall at the surface and earthquakes at depth is dependent on the hydraulic diffusivity along with the fractures (Saar and 215 Manga, 2003;Kraft et al., 2006). The second hypothesis is an elastic stress perturbation in the crust induced by hydrological loadings, such as groundwater level increase after rainfall (Rigo et al., 2008). The third hypothesis is a pore pressure increase in deep fluid-saturated poroelastic rocks in response to overlying hydrological loading (Miller, 2008;D'Agostino et al., 2018).
The time lag between the onset of the rain (October 02, 06:00 UTC) and the onset of the first earthquake swarm (October 4, 00:52 UTC, southern swarm) is ∆t=43 h. Taking a seismicity depth of z=5,000m ± 2,000 m below the surface and using a time-220 distance dependent equation for a propagating pore pressure front, z = √ 4πD∆t (Shapiro et al., 1997), we find a hydraulic diffusivity ranging from D=4.6 to 25.2 m 2 /s. This diffusivity range is unrealistically large to indicate earthquakes triggered by fluid migration. Indeed, with such a mechanism, earthquake activity following exceptional rainfall episodes or snowmelt is characterized by a delay of several days to several months and a lower hydraulic diffusivity ranging from D=0.01 to 5 m 2 /s unlikely for the first earthquake swarm.
The geology of the Tinée valley consists of limestone formations topped by a sandstone layer (Grès d'Annot). These rocks can store large volumes of water, which might support the hypothesis of seismicity triggered by groundwater weight. Rigo et al. (2008) describes earthquake triggering in a karstic (made up of limestone) region at depths smaller than 10 km, 43 h after the onset of heavy rainfall. The authors interpret this earthquake activity as the response of the crust to an elastic stress 230 increase caused by a vertical loading because of the groundwater level rise. On the other hand, Miller (2008) shows that a sharp increase of the hydraulic loading in karst can also produce an instantaneous increase of the pore pressure in the underlying fluid-saturated crust, able to trigger earthquakes.
The central and northern swarms occur with a larger time delay than the first swarm, 6 and 22 days respectively (5.6 mm and 36.0 mm of cumulative rain after storm Alex), which may be more compatible with the surface to depth fluid migration 235 mechanism. Yet, as these swarms are at the same depth as the southern swarm, this would imply a rather large spatial variation of the hydraulic diffusivity from D=1.4-7.5 m 2 /s for the middle swarm to D=0.4-2 m 2 /s for the northern swarm. Alternatively, the successive activation of the swarms might highlight the horizontal propagation of a pore pressure front coming from the southern cluster area and propagating northward along a permeable pathway such as a fault parallel to the valley. The migration velocity of the earthquake swarms from the south to the north is around 20-30 m/h, which is comparable to the values given 240 by Chen et al. (2012) for fluid-driven earthquakes migration. The pore pressure front could be possibly initiated by a pore pressure increase in the hypocentral area of the southern swarm in response to the overlying hydrological loading (hypothesis two). Therefore, we conclude that the short-term between rainfall and the southern swarm (∆t=43h) is likely compatible with hypotheses two or three.

245
Our results show that seismometers can constrain interaction between the different Earth's systems, time-and space-dependent processes during the flood and rainfall-runoff relationship at the catchment scale. That is particularly important in the absence of traditional hydrological measurements, as the study presented here. Observations from permanent seismological stations in the Maritime Alps provide the timing and velocity propagation of the flood waves. They reveal bedload and turbulence dominated phases of the flood that occurred on the Vésubie River. Our observations also suggest that 114 earthquakes between 250 local magnitude ML 0.5 and 2.5 were triggered by the hydrological loading and/or by the resulting in-situ pore pressure increase in the Tinée valley. Heavy rainfall occurs regularly in autumn in the Mediterranean region, and its intensity is increasing due to climate change (Tramblay and Somot, 2018;Ribes et al., 2019). In the future, the installation of a seismic array dedicated to flood could help further detect and constrain flood dynamics and triggered earthquakes (e.g., Meng and Ben-Zion, 2017;Eibl et al., 2020;Chmiel et al., 2021). Finally, the results from this study pave the way to further analysis based on the presented  Figures 1D and 2A-C), but no overlap is used in Figures 2E and 2G. The windowing function window is applied to each segment and the PSD is calculated by Welch's average periodogram method (Solomon, 1991a). Then for the SPIF station, we follow previous work on debris flows (Lai et al., 2018) and we investigate signal's peak frequency in individual 200 s time windows between 2-50 Hz. We also analyse seismic power recorded on the SPIF station in two different frequencies bands: 2-10 Hz and 10-45 Hz. For that, we estimate PSD using again 275 Welch's method with time segments of 2 s and no overlap, and then we calculate a median over 30 s time windows.

A2 Azimuth analysis
We perform a frequency-dependent polarization analysis to determine the dominant backazimuth assuming that the seismic signature of the flood is dominated by surface waves on the SPIF station (Goodling et al., 2018). The horizontal azimuth and degree of polarization are determined based on the dominant eigenvector of the spectral covariance matrix of the 3 measured 280 components (N, E and Z), following the approach of Park et al. (2005) and its recent application by Goodling et al. (2018). We determine these variables for 30-minute intervals using 9 subwindows with 50% overlap. The dominant azimuth per frequency (θ) is obtained and given for a range 0-180°as there is a 180°ambiguity in this value.

A3 Rainfall-runoff models
Runoff is firstly estimated using the Soil Conservation Service-Curve Number (SCS-CN) production function method. The

285
SCS-CN function allows to estimate the runoff from a rainfall event depending on the catchment saturation conditions. A simplified unit hydrograph routing function is then used to produce temporal runoff series. This analysis aims at estimating, for each studied catchment, the distances between each Digital Elevation Model (DEM) grid cell and the considered outlet and use the distance to root the runoff at the studied catchment outlets. A distinction is made between the distance travelled on the slopes and the distance travelled in the river (i.e., within the hydrographic network): the flow velocity on the slopes (fixed here 290 at 0.2 m/s) is assumed to be slower than that in the river (fixed here at 5 m/s). These distances are used to calculate, for each grid cell "x" belonging to a studied watershed, the transfer time τ [in (s)] between this grid cell x and the considered outlet: These transfer times are used to calculate the simulated flow, at time step t, at each studied outlet (denoted Q and expressed in m 3 /s) by the following expression (no initial base flow is considered in this study): where: A: catchment area upstream of the grid cell "x" (km 2 ), q: runoff estimated at timestep "t" and at the grid cell "x" 300 (m/s). The runoff is simulated for three locations along the Vésubie river which are the closest to the seismic stations ( Figure   B2).

A4 Earthquake swarm detection
Previous studies have shown that template matching (e.g., Gibbons and Ringdal, 2006) has a higher detection sensitivity than threshold-based methods such as the STA/LTA used in the Seiscomp3 system. We use template matching to detect low-305 magnitude earthquakes that belong to the earthquake swarms. Template matching is performed on the broadband station MVIF (10 km to the south of the swarm). We verified that this station was little affected by the seismic noise generated by the increased river flow during and after the storm. We use the following approach. Data are bandpass filtered in the 5-30 Hz frequency band. We use as templates the 23 earthquakes detected by SeisComP3. The templates are constructed using a 5-s window that includes P and S waves. Next, each template is cross-correlated with daily continuous seismic data. We use only 310 vertical components of the seismograms and we automatically scan the seismic data between September 27 and December 10, 2020. A new earthquake is detected if the cross-correlation coefficient exceeds a threshold of 0.6. This value allows the detection of earthquake waveforms that might slightly differ from the templates (if, for example, the origin location is not the same) while minimizing the number of false detections. Finally, the magnitude of detected earthquakes is estimated from the ratio between its maximum amplitude and the maximum amplitude of the best-correlated template (local magnitude ML). An 315 example of templates and detected events by template matching are presented in Figure B9-B11.

A5 Peak propagation velocity and uncertainty calculation
The peak arrival times are manually picked by taking the beginning of the maximum above fixed seismic power (PSD) thresholds ( Figure B3,B8). Also, we verify the time delay between the two PSDs using cross-correlation ( Figure B8). We find two maxima of 0.3 and 0.15 at time lag values of 19 and 28 min, respectively. We calculate the peak propagation velocity as a 320 ratio between the distance (d) of the two nearest river coordinates to the SPIF and BELV station (8,012 m) to manually pick the propagation time of the peaks (t). To calculate the distance, we use the nearest river coordinates to the stations, and we integrate the distance following the Vésubie river coordinates (8,012 m). Then, we use error propagation to estimate the uncertainty of the estimated velocity propagation. For that, we use the variance formula assuming that the distance and time measurements are independent: where d is the distance between the two nearest river coordinates to the SPIF and BELV station (   Time (s) Counts Figure B10. Same as Figure B9, but for a template located in the middle swarm.