Mayotte seismic crisis: building knowledge in near real-time by combining land and ocean-bottom seismometers, first results

The brutal onset of seismicity offshore Mayotte island North of the Mozambique Channel, Indian Ocean, that occurred in May 2018 caught the population, authorities, and scientific community off guard. Around 20 potentially felt earthquakes were recorded in the first 5 days, up to magnitude Mw 5.9. The scientific community had little pre-existing knowledge of the seismic activity in the region due to poor seismic network coverage. During 2018 and 2019, the MAYOBS/REVOSIMA seismology group was progressively built between four French research institutions to improve instrumentation and data sets to monitor what we know now as an on-going exceptional sub-marine basaltic eruption. After the addition of 3 medium-band stations on Mayotte island and 1 on Grande Glorieuse island in early 2019, the data recovered from the Ocean Bottom Seismometers were regularly processed by the group to improve the location of the earthquakes detected daily by the land network. We first built a new local 1D velocity model and established specific data processing procedures. The local 1.66 low VP/VS ratio we estimated is compatible with a volcanic island context. We manually picked about 125,000 P and S phases on land and sea bottom stations to locate more than 5,000 events between February 2019 and May 2020. The earthquakes outline two separate seismic clusters offshore that we named Proximal and Distal. The Proximal cluster, located 10km offshore Mayotte eastern coastlines, is 20 to 50 km deep and has a cylindrical shape. The Distal cluster start 5 km to the east of the Proximal cluster and extends below Mayotte’s new volcanic edifice, from 50 km up to 25 km depth. The two clusters appear seismically separated, however our dataset is insufficient to firmly demonstrate this. hour day and night shift. At the end of the pickathon, the relocated events, with additional manual picks and polarities on OBSs data, is merged within the REVOSIMA/MAYOBS database for further use for monitoring and research. The same procedure is applied to on land pickathons, only the work is completed during 2.5 normal work days instead of 24-hour shifts. earthquake locations and 1D velocity model ascertained by using OBSs data provides a solid foundation for future studies. Our anticipation is that our improved earthquake location catalog, coupled with geodetic modelling, petrological studies of rock samples and geochemical analysis of fluids in the water column, will bring understanding of the Mayotte sismo-volcanic crisis and regional tectonics. On-going work using machine learning picking algorithms should provide a much more complete catalog of the seismicity.

), was not considered as a significantly seismically active area (Bertil et al., 1998). The last reported widely felt earthquakes occurred around 30 km west of Mayotte: a moment magnitude (Mw) 5.0 event on September 9 th 2011 (EMS98 intensity V estimated) and a magnitude (M) 5.2 event on December 1 st 1993 (Lambert, 1997) with moderate damages (EMS98 intensity VI estimated).
Together with an unfelt M 5.1 event on March 23 rd 1993, 80 km south-west of Mayotte, these were the only M5+ earthquakes recorded within 100 km of the island since the advent of the global seismological networks in 1964 (Storchak et al., 2017;ISC, 2020). As a consequence, only one real-time seismic station (RA.YTMZ; Résif, 1995) was installed on the island at the onset of the 2018 seismic crisis. This station was deployed by BRGM (Bureau de recherches géologiques et minières, the French geological survey) for the French accelerometric monitoring network (Résif-RAP, Pequegnat et al., 2008).
On May 10 th 2018, the first felt earthquake, quickly followed by many others, surprised inhabitants. More than 130 M4+ earthquakes were recorded in the following months, with the strongest being a Mw 5.9 on May 15 th 2018 (GCMT project, Dziewonski et al., 1981;Ekström et al., 2012, Lemoine et al., 2020a. After about 50 days of very intense seismic activity, this unprecedented seismic sequence continued less intensively. During the summer of 2018, Global Navigation Satellite System (GNSS) data from the locally continuously recording sites began to show rapid surface displacements of the island (Briole, 2018 Lemoine et al., 2020a). Their elastic modeling evidence a large regional deflation centred east of Mayotte's shorelines (Cesca et al., 2020;Lemoine et al., 2020a). On November 11 th 2018, a very low frequency tremor was recorded worldwide (Cesca et al., 2020, Lemoine et al., 2020a, confirming that the seismic crisis was very likely of magmatic origin. This was later confirmed by the discovery of a new submarine volcanic edifice offshore Mayotte during the MAYOBS1 scientific expedition onboard RV Marion Dufresne in May 2019 (Feuillet et al., 2021). This large eruption began either on June 18 th (Cesca et al., 2020) or on July 3 rd 2018 (Lemoine et al., 2020a).
Since the onset of the crisis, collaborations were progressively established between French research institutes to improve the understanding and knowledge of the on-going crisis.
These collaborations enhanced the seismic monitoring of the region, which included installing additional real-time sites onshore, access to real-time data recorded by existing regional stations and offshore deployments. During the first year of the crisis, the monitoring network thus evolved rapidly (Lemoine et al., 2020a). In March 2019, one month after funding, 4 seismic stations were installed onshore (3 on Mayotte island and 1 on Grande Glorieuse island) and 6 Ocean Bottom Seismometers (OBS) were deployed offshore, within a radius of 40 km around the seismically active area. A seismology team was created among the researchers, engineers, and students belonging to the participating French institutions newly acquired data as quickly as possible and to obtain first hand results in almost real-time, to improve the daily monitoring and the knowledge of the volcano-seismic crisis.
In this paper, we review the local seismic network improvements since the beginning of the crisis and our scientific developments. We detail how this collaborative work was orchestrated for maximum efficiency and how it led to an improved local velocity model and a seismic catalog from February 2019 up to May 2020 to better document the Mayotte 2018-ongoing seismo-volcanic crisis.

Seismic network evolution and data processing
From May 2018 to June 2019, the Mayotte local real-time seismic network progressively evolved from 1 to 8 stations. Daily data analysis protocols have also been continuously adapted in several Institutes by our group to take advantage of the increasing number of local stations and to produce better locations for the detected events (see details in section 2.2). In 2019, we also developed a new protocol to efficiently process OBSs data during pickathons, when, at the same place, several analysts dedicate a few days to work together on the newly recovered data.   The OBSs are regularly serviced for maintenance and data recovery (every 3 to 4 months for INSU-IPGP, every month for micrOBSs). As of May 2020, 9 different OBSs deployments have been conducted. When the OBSs are recovered, their data are downloaded and timecorrected for internal clock drift and converted to miniSEED format using L-Cheapo tools (Orcutt and Constable, 1996). These data are then integrated in the main waveform database at the IPGP data center, along with the local and regional land station data.

Daily monitoring
For daily monitoring of the crisis, only the real-time data from land stations can be used.
Several institutes successively participated in the day-to-day seismic data processing, using SeisComP3 (Weber et al., 2007) with slightly different setups. The BRGM office in Mayotte was involved first and maintained a seismic catalog from the beginning of the crisis in May 2018 . The events from  were located from manually picked waveforms using the LocSAT algorithm (Bratt et Bache, 1988) and a slightly modified IASPEI91 velocity model (Kennett & Engdahl, 1991).

OBS integration
After each OBSs data recovery, we manually pick phases on the OBS data and offline land stations to improve the locations in the existing earthquake catalog during dedicated pickathons that continue today. Adding new phase picks significantly improved the location of the earthquakes already detected and first located only by the land network. Since there are already a great number of events to relocate in the existing catalog, we do not search for new events from the OBS data set. Searching for new events in the OBS data set is a work in progress and will be reported in future studies. During each pickathon, the timespan of the OBSs data we need to process is divided across three teams using the same software setup and each team manually locates earthquakes in descending order of magnitude (Saurel et al. 2019, Figure 2). When the OBSs are recovered from an oceanographic research vessel, such as RV Marion Dufresne, large enough to board a team of around 10 analysts, we divided the work in three 4-hour-shifts (i.e., 2 or 3 analysts by shift) we can manually pick earthquakes 24/7. Otherwise, the MAYOBS/REVOSIMA seismology group meet -either virtually or at one of the group institutes -for two days to process the recently-recovered data, produce graphics demonstrating the evolution of the crisis, and interpret the results together.
Whether the phase picking, location, and interpretation are done onboard or on land, we use the same setup/configuration, starting database, and software to analyze the recently recovered data. We assigned uncertainty to each phase, from a common predefined list of values. Time uncertainties assigned to the S-phase were always equal to or larger than the uncertainty assigned to the P-phase. Impulsive P-phase polarity onset is also reported for first motion source mechanisms studies later use. We were able to locate events as small as M 0.8, but despite improvement since July 2019 and the use of more stations in the STA/LTA automatic processing, their detection is far from complete and mainly depends on the land station daily noise level. We estimate that the magnitude of completeness is below M 3.0 (see section 4) whatever the time allocated to the processing during each pickathon and the number of events processed. We typically process around 1,000 earthquakes during pickathons when performed onboard scientific cruises and around 500 earthquakes during pickathons conducted on land. So far, our catalog contains more than 5,000 manually picked earthquakes from February 2019 to May 2020, relocated using combined land and OBS data.
hour day and night shift. At the end of the pickathon, the relocated events, with additional manual picks and polarities on OBSs data, is merged within the REVOSIMA/MAYOBS database for further use for monitoring and research. The same procedure is applied to on land pickathons, only the work is completed during 2.5 normal work days instead of 24-hour shifts.
To enable this collaborative work, we use techniques and configurations developed during the last 10 years for the daily routine processing of earthquakes in IPGP volcanic and seismologic observatories. Waveform data and event databases are held by a SeisComP3 instance (Weber et al., 2007). Each earthquake analyst uses their own laptop and, regardless of their laptop's Operating System, they run a VirtualBox pre-configured Linux machine with the SeisComP3 Origin Locator GUI client (scolv, Weber et al., 2007). We use NLL software and the new local 1D velocity model described in the following section to locate the events.
Magnitudes are computed with the embedded Local Magnitude (M L ) formula in SeisComP3 (Richter, 1958). The horizontal signals are converted to a Wood-Anderson seismometer response (Urhammer & Collins, 1990) before measuring their pick amplitude. We have not yet calibrated this magnitude as very few of the earthquakes since 2019 have been characterized with a moment tensor magnitude by the global monitoring agencies (GCMT project, Dziewonski et al., 1981;Ekström et al., 2012).

Improved 1D local velocity model
A local velocity model is essential to provide precise location of this dense swarm seismicity.
One of the first challenges in improving earthquake locations was then to build a reasonable 1D local velocity model, because only global models were available so far. This was done onboard RV Marion Dufresne during the MAYOBS1 campaign (Feuillet, 2019). We used the data of the first OBSs recovery and of 3 local land stations (see Section 2). We first produced modified Wadati diagrams (Chatelain, 1978) and considered 2 different existing velocity profiles from the area. The first profile, named "Coffin449", is based on a P-wave velocity (V P ) profile derived from a 1980 active-seismic sonobuoy. That experiment was located 100 km south-east of Mayotte (Coffin et al., 1986, instrument 449; Figure 1b)  Hypo71 (Lee and Lahr, 1972) and the "Coffin449" model with a V P /V S ratio of 1.80 extrapolated from Eastern and Central Afar studies (Jacques et al., 1999;Grandin et al., 2011). The modified Wadati diagrams (Chatelain, 1978) indicated a local (OBS and Mayotte land stations) V P /V S ratio of 1.66 and a regional V P /V S ratio of 1.72 (Figure 3a-b).
We then tested different combinations of velocity model parameters (2 velocity profiles and 3 V P /V S ratios) on a more complete dataset of 800 events with OBS phases, using NLL software (Lomax et al., 2014). Contrary to Hypo71, NLL allows the use of depth variations of the V P /V S ratios and different velocity models depending on the station. Its probabilistic approach (Lomax et al., 2014) also makes the reported ellipsoidal errors more meaningful and easier to interpret than estimated horizontal and vertical errors given by Hypo71. We NLL also computes a V P /V S ratio for each event, based on P-and S-arrival times and independent from earthquake location, using a formula described in the HypoEllipse manual (Lahr, 2012). The NLL estimation, using the 800 events, supports the mean V P /V S ratios previously estimated for local and regional stations on the first 100 events.  (Figure 4d). For a reasonable half-space P-wave velocity between 5 and 9 km/s, all events are located between 20 and 60 km depth, with an average RMS lower than 1 s and best results obtained with V P = 6 km/s. Except in the extreme cases, earthquakes were always located deeper than 15 km . Mean and standard variation of the RMS, depth and vertical error distributions for various P-wave velocity speeds and this study's V P /V S ratio of 1.66 using a homogeneous half-space velocity model. All tests were performed using NLL software.

Earthquakes locations
The We did not locate any earthquake with reliable depths shallower than 20km during the period of investigation (Figure 6d). However, because we only relocated events that have already been automatically detected and located using only the land stations network, we might have missed some shallow, local, and low magnitude earthquakes.

Discussion and conclusions
Thanks to the collaborations among several research institutions, we, the MAYOBS/REVOSIMA seismology team, has facilitated the deployment of multiple seismometers, as well as the collection and interpretation of their data since early 2019. The numerous pickathons we have held since then have improved our understanding of the seismicity occurring offshore Mayotte. The seismicity manually-picked and relocated in this study, from February 2019 to May 2020 is in agreement with the analysis of 2018 activity (Cesca et al., 2020;Lemoine et al, 2020a; Table 1), as seismicity clusters identified in this work seems to have been active since summer 2018. The Distal cluster correlates with the first cluster of activity that represents rock fracturing and dyke opening from the center of deflation toward the volcano, leading to the creation of the NVE (Cesca et al, 2020;Lemoine et al, 2020a;Feuillet et al, 2021). However, a noticeable difference with our catalog is the lack of earthquakes in the first 20 km below the volcano. This could be explained by a magma path toward the surface now opened and generating only very small events not detected by the land network while the path fracturing the crust in 2018 implied strong felt earthquakes (Duputel et al, 2019). The Proximal cluster correlates with seismic activity that initiated during July 2018 (Cesca et al, 2020 ;Lemoine et al, 2020a) and became very active at the end of August 2018. This correspond to the beginning of the island subsidence and eastward displacement as recorded in GNSS data, inferred to be due to the drainage of an at least 30 km deep magma chamber (Briole, 2018 ;Lemoine et al, 2020a ;Cesca et al, 2020 ;Feuillet et al, 2021). This proximal cluster appears not connected to the Distal cluster in our dataset and this seems to have been the case since the beginning of the crisis. The Proximal cluster cylindrical shape match existing interpretations of a slowly sagging caldeira piston (Cesca et al, 2020 ;Feuillet et al, 2021), but it could also be an ancient fault zone re-activated by the massive changes in lithosphere constraints due to the eruption. The 2 local and regional velocity models used in this study with their respective V P /V S ratios are consistent with the two main geological setting interpretations of the area. The "Coffin449" model upper low velocity layer and rapid increase up to a Moho at 15 km depth is compatible with an oceanic crust. Its averaged V P /V s ratio can explain an oceanic crust or mixed oceanic and continental crust at regional scale. The "ADofal" model smooth velocity changes and the low local V P /V S ratios can be associated with an heterogenous volcanic island context: hot material and presence of gas or fluid-filled fractured rock. This low V P /V S ratio is also consistent with the H-k stacking from receiver functions performed by Dofal et al. (submitted) and could support their interpretation of a continental crust with underplating similar to the magmatic continental domain of the southeastern coast of Madagascar (Rindraharisaona et al., 2017).
In terms of monitoring, the regular recovery and deployment of OBSs and the subsequent and immediate data analysis have been essential to monitor the Mayotte sismo-volcanic crisis. The wide collaboration between many scientists, engineers, and students, from Onboard the monitoring cruises, information from this rapid processing is used to prioritize operations (bathymetry survey) and focus on the areas of interest and/or crisis. With each pickathon, we progressively increased our knowledge of the area, which in turn also feeds the hazard and risk assessment studies that will help the authorities with the decisionmaking process. For example, the Proximal cluster, while it does not seem directly linked to the magma emission at the new volcano is of particular concern since it is much closer to the island and even slightly expands below Mayotte. Tsunami modeling studies (Lemoine et al., 2020b;Poulain et al., 2020), up to impact mapping, were conducted with tsunamigenic sources derived from this improved knowledge of the seismicity.
The resulting high-quality dataset of manually picked arrivals is now used in several detailed on-going studies (lithosphere structure investigations, seismicity time and space evolution, seismic sources studies, high-resolution locations) that will better constrain the seismicity, active structures, and geological setting of the area. Various non-earthquake signals, such as hydro-acoustic or seismic waves not clearly associated to earthquakes, have been discovered during routine data screening and are currently being investigated. While REVOSIMA has reported more than 30,000 events for the period covered by the catalog after daily manual screening of the continuous land stations data, our refined 5,000 earthquake locations and 1D velocity model ascertained by using OBSs data provides a solid foundation for future studies. Our anticipation is that our improved earthquake location catalog, coupled with geodetic modelling, petrological studies of rock samples and geochemical analysis of fluids in the water column, will bring understanding of the Mayotte sismo-volcanic crisis and regional tectonics. On-going work using machine learning picking algorithms should provide a much more complete catalog of the seismicity.
GMT (Wessel et al., 2019) was used for figure 1 and figure 6. Matlab was used for figure 4 and figure 5.
NonLinLoc software was used for earthquake locations with OBS data. Hypo71 software was used for preliminary earthquake locations of the first MAYOBS1 dataset. VirtualBox software was used on all analyst's computer to run SeisComP3 graphical user interface client.