Remote Single‐Station Seismic Monitoring of the July–October 2022 Earthquake Swarm at Ta‘ū Volcano, American Samoa

The Samoa archipelago consists of volcanic islands formed above the Samoa hotspot in the south Pacific Ocean, as the Pacific plate moves to the northwest (Fig. 1a; McDougall, 2010). Independent Samoa includes the westernmost and oldest islands, Savai‘i and Upolu. American Samoa, an unincorporated U.S. territory, includes the islands of Tutuila (population ∼49,000; Fig. 1b; U.S. Census Bureau, 2021), and the Manu‘a Islands ∼100 km to the east, consisting of Ofu–Olosega (∼280 residents) and Ta‘ū (∼550 residents; Fig. 1b,c; U.S. Census Bureau, 2021).

Despite its intraplate tectonic setting, American Samoa is exposed to natural hazards, including earthquakes, tsunamis, and volcanic eruptions. A deadly 2009 tsunami resulted from $Mw$ 8.1 and 7.8 doublet earthquakes located ∼200 km southwest at the Tonga trench subduction zone (Fig. 1a; Beavan et al., 2010; Lay et al., 2010). Recent volcanism includes: an offshore eruption in Ofu–Olosega in 1866 (Fig. 1b, dashed circle; Stice and McCoy, 1968; Tepp et al., 2019; Deligne et al., 2024), and an eruption at the Vailulu‘u seamount, where the Samoa hotspot is currently located (∼40 km east of Ta‘ū Island), between 2001 and 2005 (Fig. 1b,c; Staudigel et al., 2006). Tutuila, Ofu–Olosega, and Ta‘ū, which are basaltic shield volcanoes, are designated low‐threat volcanoes by the U.S. Geological Survey (USGS) (Ewert et al., 2018); Vailulu‘u seamount has not been assessed.

Starting on 26 July 2022 and continuing through September, Manu‘a Islands residents reported feeling shaking (Table 1; Fig. 2a, green vertical lines), up to several times a day, to the National Oceanic and Atmospheric Administration National Weather Service (NWS) office in Pago Pago, Tutuila. Because of the frequent shaking and residents’ fears of an impending tsunami, the American Samoa Department of Homeland Security activated an emergency response on 8 August 2022. USGS Hawaiian Volcano Observatory (HVO) and NWS Pago Pago began deploying staff and instrumentation to the remote Manu‘a Islands to determine if the earthquakes were volcanic and assess the risks of an eruption, with the team arriving in Ta‘ū on 13 August (Wech et al., 2025).

American Samoa had no local seismic monitoring when the earthquakes started (Tepp et al., 2019; Deligne et al., 2024), leading to a lack of instrumental information about the earthquakes and related volcanic activity. Felt reports (Table 1) were the only indication that earthquakes were happening near the Manu‘a Islands. The Advanced National Seismic System Comprehensive Catalog (ComCat) (USGS Earthquake Hazards Program, 2017) does not contain any earthquakes between 1 July and 12 August 2022 within the region in Figure 1b. Increasing the ComCat search area to the region in Figure 1a only returned magnitude 4–5.8 earthquakes at the Tonga trench (Fig. 3c, red circles), too far to be felt in the Manu‘a Islands. The absence of detectable ComCat events by global seismic monitoring, combined with reports of frequent shaking without a clear mainshock, suggests a nearby swarm of M < 4.5 earthquakes. The nearest available station IU.AFI, ∼250 km away on Upolu, Samoa (Figs. 1a and 3c, black triangle), recorded three‐component continuous seismic data, motivating the characterization of the swarm in a situation when reliable earthquake information was needed promptly, but local observations were not yet available. This study shows how modern automated seismological processing methods, such as deep‐learning phase pickers and template‐matching, when applied to a single remote station, can yield an earthquake catalog tracking a swarm’s evolution (Fig. 2).

Two Raspberry Shake seismometers deployed in the Manu‘a Islands recorded the first local seismic observations starting on 13 August 2022 (Fig. 4); deep‐learning pickers applied to these data helped localize the swarm during the ongoing response. By mid‐August 2022, HVO installed a permanent seismic network in American Samoa, including three‐component broadband seismometers and Raspberry Shakes (Fig. 1b,c, triangles), which remain operational. The permanent network captured the second‐half of the swarm, which ended in October 2022 without an eruption. We extended our single‐station IU.AFI swarm catalog until 15 October 2022, which contains a uniform record of the swarm, both before and after local seismic monitoring was in place (Fig. 2).

Before 13 August 2022 (Fig. 2a–c, pink), the seismic station nearest to American Samoa was IU.AFI (Figs. 1a and 3c, black triangle), ∼250 km from the Manu‘a Islands. IU.AFI, a Global Seismographic Network (Albuquerque Seismological Laboratory [ASL]/USGS, 2014; Wilson et al., 2024) borehole broadband station, has recorded three‐component seismic data since 1993. We selected the BH channels sampled at 40 Hz (Section S1, available in the supplemental material to this article).

Traditional amplitude‐based automatic phase pickers, such as short‐term average to long‐term average (e.g., Allen, 1982), can struggle to detect low signal‐to‐noise ratio events at regional distances. Instead, we applied the EQTransformer deep‐learning model (Mousavi et al., 2020), using parameters in Table S1, to automatically detect events and pick phases in continuous IU.AFI data. We used EQTransformer because it successfully identified uncatalogued small earthquakes in previous studies (e.g., Yoon et al., 2023), and was trained on 60 s time windows from a global data set (Mousavi et al., 2020), which is long enough to record the entire event waveform at ∼250 km distance. First arrivals picked by EQTransformer at IU.AFI, for the swarm events of interest, are Pn and Sn head waves that traveled through the upper mantle below thin oceanic crust (Section S1). To reduce false detections, we kept only events containing both a Pn and Sn pick. For duplicate events, where a pick is within 2 s of another event’s pick, we remove the event with the lower event‐detection probability score.

Felt reports contributed by Manu‘a Islands residents to NWS Pago Pago (Table 1), though missing some earlier reports from late July to early August, provided essential context about the swarm that was used to constrain further analysis. Many felt reports include the time of shaking down to the minute, and descriptions of 2–5 s duration shaking suggested a local swarm of smaller magnitude earthquakes. We visually inspected picks on an EQTransformer‐detected event waveform with an arrival time at IU.AFI within a minute of a felt report time (Fig. 3a). This detection of an earthquake from the swarm had ∼23 s Sn–Pn time. The distribution of Sn–Pn times for early EQTransformer‐detected events at IU.AFI, from 1 July to 12 August 2022 (Fig. 3b, red), also peaks at ∼23 s. Going forward, we kept only 1092 events with Sn–Pn times of 22.5–25 s (Fig. 3b,c, yellow) for our single‐station deep‐learning swarm catalog. We excluded non‐swarm events from the smaller, earlier peak centered at Sn–Pn time of 20 s, which are likely from the Tonga trench (Fig. 3c, red dots). Noisy first‐motion polarization signals prevented us from estimating back azimuths to further distinguish between Tonga trench and swarm earthquakes.

Using 149 EQTransformer‐detected earthquakes between 23 July and 12 August with Sn–Pn times between 23 and 24 s, which includes events that matched felt report times (Fig. 3a), template windows were created using the three‐component IU.AFI station data filtered between 2–10 Hz. Templates were 10 s long and started 2 s prior to each Pn or Sn phase arrival. These templates were then cross‐correlated against the continuous seismic data on IU.AFI between 1 July and 17 October 2022 to detect smaller, unknown earthquakes with similar waveforms. A conservative detection threshold of 12*(CC/MAD) was used, where CC is the station‐normalized cross‐correlation coefficient, and MAD is the daily median absolute deviation of the cross‐correlations. Relative magnitudes for the template‐matching catalog were calculated with singular value decomposition (Rubinstein and Ellsworth, 2010). This template‐matching catalog contained 2095 total swarm events from 1 July to 12 August 2022.

HVO and NWS staff installed the first two local seismometers, both Raspberry Shake RS4D models, in the Manu‘a Islands: AM.R3112 on Ta‘ū on 13 August, then AM.RAA63 on Olosega on 14 August (Fig. 4). Raspberry Shakes are consumer‐grade seismometers, often used for citizen science (Paul et al., 2023), that are lightweight, low cost, and easily deployed on the ground surface. The RS4D model contains a three‐component accelerometer and a vertical‐component (EHZ) short‐period velocity transducer recording continuous seismic data sampled at 100 Hz, available ∼30 min after acquisition on an International Federation of Digital Seismograph Networks compliant data center (Raspberry Shake, 2016).

From 13 to 19 August 2022, EQTransformer detected events and picked phases on EHZ continuous data, independently at AM.RAA63 and AM.R3112. Following the same approach from IU.AFI, we retained only events with both P and S arrivals, removed duplicate events, and identified swarm events where times of EQTransformer detections and felt reports agreed within 5 min (Fig. 4, Table 1). On 14 August 2022, a day after installing the first RS4D, short (1–3 s) S–P times for swarm events recorded at AM.R3112 (Fig. 4c, blue) ruled out Vailulu‘u seamount as the source of shaking (Fig. 4d,e) as Vailulu‘u was too farther away. From the range of RS4D S–P times, we designated swarm events as having S–P times of 1–3 s at AM.R3112 (Fig. 4c, pink), and 2–4 s at AM.RAA63 (Fig. 4c, yellow). We then used these S–P times to estimate the range of epicentral distances from AM.R3112 (Fig. 4d,e, pink rings) and AM.RAA63 (Fig. 4d,e, yellow rings) to swarm events (Section S2), assuming shallower (depth z = 0 km, Fig. 4d), and deeper (depth z = 15 km, Fig. 4e) swarm origins. We constrained the swarm location to the intersection of these two distance ranges: either directly under (Fig. 4e), or ∼10–20 km offshore north or south of (Fig. 4d,e), Ta‘ū Island. The RS4Ds were not properly oriented during setup, precluding a reliable back‐azimuth estimate from the three‐component accelerometers to further localize the swarm. Once the proximity of the swarm to the populated Manu‘a Islands was established, local authorities and the public became more concerned about the potential impacts of volcanic unrest.

HVO staff installed 12 seismic stations across American Samoa (USGS Hawaiian Volcano Observatory [HVO], 1956) starting 13 August through early September 2022 (Fig. 2a–c, purple; Section S3, Tables S2 and S3): four broadband three‐component seismometers (Fig. 1b,c, black triangles) and eight RS4D sensors (Fig. 1b,c, red triangles). In late August, USGS 24/7 duty staff from Hawaii, Alaska, and California Volcano Observatories located the swarm within ∼15 km of Ta‘ū Island, with depths of ∼5–15 km (Fig. 1c–e; Wech et al., 2025), and closely monitored changes in swarm activity until it ended in October 2022 without an eruption. Permanent local seismic monitoring of American Samoa, which started 20 August 2022, is currently operational.

In November 2022, USGS National Earthquake Information Center analysts retrospectively determined event origins for the swarm with the local seismic network (Wech et al., 2025), where hypocenter locations were computed with manual picks (USGS Earthquake Hazards Program, 2017) and the ak135 global velocity model (Kennett et al., 1995). The resulting ComCat swarm catalog has 309 M 2.5+ events between 20 August and 6 October 2022, with the largest $mb$ 4.5 event on 5 September 2022 (Fig. 2b–d, red; USGS Earthquake Hazards Program, 2017). ComCat events captured the second half of the swarm, which declined from peak activity in late August to early September and ended in October (Fig. 2a–c, red).

We extended the single‐station catalogs until 15 October 2022 to cover the entire swarm duration. The resulting deep‐learning catalog had 2091 total events (Fig. 2a–d, blue) with Sn–Pn times between 22.5 and 25 s (Fig. 3b, gray), and 184 deep‐learning catalog events matched a ComCat event (Fig. 2e, green), where the times of ComCat manually picked and EQTransformer‐picked Pn arrivals at IU.AFI agreed within 4 s (Fig. S1). For these 184 events, the deep‐learning catalog local magnitude (equation 1) agreed well with the ComCat‐reported local magnitude, differing by only 0.09 ± 0.14 magnitude units (Fig. S2a). The template‐matching catalog had 6221 total events (Fig. 2a–d, cyan), including 290 events that matched a ComCat event (Fig. S1), where relative magnitudes from the template‐matching catalog differed from ComCat‐reported local magnitudes by −0.07 ± 0.37 magnitude units (Fig. S2b).

ComCat swarm locations were within ∼10–20 km of Ta‘ū Island, with depths ∼10–15 km (Fig. 1c–e; USGS Earthquake Hazards Program, 2017). However, ComCat event locations have large (±10 km) location uncertainties (Fig. 1c–e, gray error bars) due to the linear geometry of the local seismic network, leading to a trade‐off between distance and depth, allowing for a deeper swarm closer to Ta‘ū Island, or a shallower swarm farther offshore (Fig. 1c). The swarm was likely related to volcanic magma movement, based on its location near Ta‘ū Island and observations of deeper volcanic tremor on local broadband seismometers and IU.AFI (Wech et al., 2025).

Deep‐learning and template‐matching, constrained by felt reports (Table 1), automatically created a uniform catalog for the Ta‘ū Island swarm, using one station ∼250 km away, for its entire duration from July to October 2022 (Fig. 2). Template‐matching (Fig. 2a–d, cyan) detected smaller magnitude events than deep‐learning (Fig. 2a–d, blue), but the initial deep‐learning step, which automatically recognized swarm events from their Sn–Pn times, was essential to identify template waveforms at times matching felt reports (Fig. 3, Table 1). Event magnitudes and seismicity rates from this catalog illuminated the first half of the swarm’s temporal evolution (Fig. 2a–c, pink), including its onset and escalation to peak activity—critical information during an ongoing event response, when initial knowledge about the swarm was otherwise limited to felt reports. Swarm activity was at low levels in late July and early August, until a sudden increase in the seismicity rate on 5–6 August 2022, with elevated seismicity over the next week from 7 to 13 August 2022 (Fig. 2a), when the first two local RS4D stations were installed. Magnitude‐of‐completeness (⁠$Mc$⁠) estimates were 3.1 for ComCat, and 2.6 for the deep‐learning and template‐matching catalogs (Section S4). Thus, template‐matching detected more events than deep‐learning but did not improve completeness.

However, the single‐station catalogs have several limitations. First, since IU.AFI was too far away to reliably constrain event locations, it was important to promptly deploy local seismometers in the Manu‘a Islands. Second, seismicity rate estimates (and $Mc$⁠) are sensitive to time‐varying noise levels at IU.AFI (Fig. 2c, blue, cyan), such as diurnal variation in anthropogenic noise (Fig. S3) and lower‐frequency environmental noise from wind and storms (e.g., McNamara et al., 2019); in August 2022, times with lower $Mc$ had higher seismicity rates, and vice versa, for both deep‐learning (Fig. 2a, blue) and template‐matching (Fig. 2a, cyan). Third, EQTransformer is unable to detect emergent or longer‐duration signals indicating volcanic unrest, such as volcanic tremor, which are thus missing from these catalogs. Fourth, these catalogs may be contaminated by a few non‐swarm events, such as Tonga trench events with Sn–Pn times between 22.5 and 25 s, which might account for some earlier larger‐magnitude events in July 2022 (Fig. 2b,c, blue). Finally, the deep‐learning catalog, which contains only 184 of 309 (∼60%) ComCat earthquakes (Fig. 2e, green), is missing many events (Fig. 2e, orange), including the largest $mb$ 4.5 event—where EQTransformer incorrectly picked the Sn phase on IU.AFI only ∼3 s after correctly picking Pn (Fig. S4), and was thus eliminated since the Sn–Pn time was not within 22.5–25 s. This lackluster performance of EQTransformer on IU.AFI data is not surprising, as only 8% of its training data came from stations at regional distances of 110–350 km (Mousavi et al., 2019). Fine‐tuning the EQTransformer model (Mai et al., 2023), applying transfer learning to EQTransformer by including labeled waveforms from IU.AFI (Mai et al., 2023), or using a deep‐learning picker specifically trained for stations at regional distances (e.g., Aguilar Suarez and Beroza, 2024) might improve event detection performance. The template‐matching catalog produced using IU.AFI was more complete, finding 290 of 309 (∼94%) ComCat earthquakes (Fig. S1).

We compare our single‐station catalogs with a 1030‐event hydroacoustic catalog of the American Samoa swarm (Fig. 2a–d, yellow), which was the only other consistent analysis throughout the entire swarm, derived independently using a different type of sensor: T‐phase records of the swarm on the Wake Island hydrophone array, located ∼4500 km away (Wech et al., 2025). Compared to our single‐station deep‐learning catalog, the hydroacoustic catalog has fewer total events and a slightly higher $Mc$ of ∼2.7 (Fig. 2a,d, yellow), but performs more reliably, detecting 264 of 309 (85%) ComCat earthquakes (Wech et al., 2025). Although our single‐station deep‐learning (Fig. 2a, blue) and template‐matching (Fig. 2a, cyan) catalogs indicate peak seismicity rates earlier in August, the hydroacoustic catalog is less sensitive to fluctuating noise levels; the resulting swarm escalation and decline appears more symmetric in time, with the highest seismicity rate in mid‐August (Fig. 2a, yellow). Tracking temporal changes in the largest‐magnitude event per day, which should be less affected by $Mc$ fluctuations, all catalogs indicate increasing magnitude during the first half of August, then decreasing magnitude in late August and early September, though with more variation later in September (Fig. 2b).

This study applied modern automated seismological methods, such as deep‐learning phase pickers and template‐matching, to openly available continuous seismic data at one distant station, rapidly characterizing a volcanic swarm in a sparsely instrumented region. These methods can help situational awareness of unfolding earthquake sequences in other sparsely monitored situations, such as induced seismicity, intraplate settings, and offshore regions. The resulting earthquake catalogs might help improve aftershock forecasts in remote or offshore areas. Setting up data access and processing software to be ready to run on a powerful computer, requiring minimal changes to input data files and parameters, speeds up the availability of useful scientific outputs.

However, local near‐source seismic stations are essential for event location. RS4D sensors were easier and faster to deploy than broadband seismometers in the remote Manu‘a Islands—a key advantage when needing to quickly constrain the swarm location. Deep‐learning pickers, such as EQTransformer, can process RS4D vertical‐component continuous data, available for download ∼30 min after acquisition (Raspberry Shake, 2016). The RS4D short‐period velocity sensor, which is more sensitive to higher‐frequency signals, is well suited to record smaller earthquakes in a swarm at local distances (Anthony et al., 2019; Paul et al., 2023). In the future, we recommend that the RS4D three‐component accelerometer be properly oriented during installation to enable reliable back‐azimuth estimates from particle motion polarization.

During an event response, where updated scientific information might affect emergency response decisions, it is essential to consider all possible data sources, regularly coordinating updates to data and analysis results as the situation evolves. Precise felt report times and locations (Table 1) were an important constraint for the single‐station catalog developed with deep‐learning and template‐matching. The single‐station catalog provides one initial perspective of the swarm before local stations were deployed but has notable limitations such as missing larger (M 3–4.5) earthquakes; it should be integrated within context of, and compared for consistency and discrepancy with, other information about the swarm and volcanic unrest.

Between July and October 2022, a noneruptive volcanic earthquake swarm occurred near Ta‘ū Island, American Samoa, which initially had no local seismic network. Using continuous data from one seismic station ∼250 km away, we created an earthquake catalog for the swarm, with the EQTransformer deep‐learning picker and template‐matching, constrained by felt shaking reports from nearby residents. This single‐station catalog supplied new information about the swarm’s onset and escalation to peak activity before local seismic monitoring commenced in mid‐August; it also captured the decline and end of the swarm in a consistent manner with ComCat, after a field response team installed the local permanent network. Short S–P times of EQTransformer‐detected events, recorded on continuous data from two local Raspberry Shake sensors, quickly narrowed down the swarm location near Ta‘ū Island. This study illustrates how modern seismological processing methods can inform an ongoing event response in a poorly monitored setting.

Earthquakes: Advanced National Seismic System Comprehensive Catalog (ComCat, doi: 10.5066/F7MS3QZH; https://earthquake.usgs.gov/fdsnws/event/1/) (last accessed March 2024). Continuous seismic data: EarthScope Data Center (https://ds.iris.edu/ds/), with data from IU (doi: 10.7914/SN/IU), HV (doi: 10.7914/SN/HV) networks; Raspberry Shake Data Center (https://data.raspberryshake.org/) with data from AM (doi: 10.7914/SN/AM) network. EQTransformer software: https://github.com/smousavi05/EQTransformer (last accessed May 2023). Scripts for single‐station analysis and plotting: https://gitlab.com/cyoon1/AmericanSamoa (last accessed February 2025). ObsPy 1.4.0 (doi: 10.1088/1749-4699/8/1/014003; https://www.obspy.org/, last accessed November 2024) for seismological processing and visualization, Generic Mapping Tools (GMT) 6.4.0 (doi: 10.1029/2019GC008515; Wessel et al., 2019; https://www.generic-mapping-tools.org, last accessed November 2024) for maps. The supplemental material includes Sections S1–S4, Tables S1–S3, and Figures S1–S4.

The authors acknowledge that there are no conflicts of interest recorded.

Waveform data and metadata for this study were accessed through the EarthScope and Raspberry Shake Data Centers. The authors acknowledge contributions of the technical and monitoring staff at U.S. Geological Survey (USGS) Hawaiian Volcano Observatory (HVO) and National Oceanic and Atmospheric Administration (NOAA) National Weather Service (NWS) Pago Pago for seismic data acquisition, and USGS National Earthquake Information Center (NEIC) for the ComCat earthquake catalog used in this study. The authors thank all participants from USGS Volcano Science Center observatories, NOAA NWS, and other partners who contributed to the months‐long scientific response to the volcanic unrest. The authors thank the local community in the Manu‘a Islands who contributed their felt reports, and NWS Pago Pago staff for collecting them in a detailed database (Table 1). The authors thank all developers and maintainers of open‐source seismology software listed in Data and Resources. Will Yeck, two anonymous reviewers, and Editor‐in‐Chief Keith Koper provided helpful article reviews. Gabrielle Tepp provided helpful discussion on an early article draft.

Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.