Toward the Rapid Seismic Assessment of Landslides in Coastal Alaska

In 2015, a landslide in Taan Fiord, Alaska, sent $76Mm3$ of rock into the water below, generating a tsunami that stripped vegetation 190 m above the waterline (Gualtieri and Ekström, 2018; Higman et al., 2018). A similar event in Karat Fjord, Greenland, washed into Nuugaatsiaq, killing four people (Gauthier et al., 2018). As the climate warms, retreating glaciers have exposed oversteepened fjord walls, buttressed for millennia by glacial ice, leading to the hypothesis that glacial retreat is increasing the occurrence of landslides (Coe et al., 2018). This hazard is an emerging concern in much of coastal Alaska—a region dominated by steep, glaciated mountain ranges (Coe et al., 2018; Dai et al., 2020).

These past events highlight the need for monitoring systems that can inform emergency response. However, the mechanics of landslides pose a challenge for real‐time warning. The rapid displacement of massive volumes of rock during landslides has long been known to create distinctive long‐period (<0.1 Hz) seismic waves visible at teleseismic distances (e.g., Kanamori and Given, 1982). The Global Centroid Moment Tensor project has long captured the largest of these through continuous backprojection (Ekström and Stark, 2013). Global approaches, however, are limited by the speed of seismic wave propagation, diminishing their value for real‐time assessment. Chao et al. (2017) present a novel, if complex, approach for landslides in Taiwan with volumes $>1Mm3$ in which cascading modules for detection, identification, and location are able to send alerts within about six minutes.

Techniques using high‐frequency seismic waves are more sensitive to smaller landslides (>1 Hz; e.g., Schneider et al., 2010). Several authors have demonstrated that the shape and amplitude of landslide seismograms can be used to estimate location and volume (e.g., Fuchs et al., 2018; Hibert et al., 2014). These techniques typically require considerable tuning to account for site variability (e.g., Chen et al., 2013; Lee et al., 2019). In geologically active regions, these approaches are challenged by the abundance of competing seismic signals from sources such as local earthquakes, glaciers, and anthropogenic sources. This is a particular issue in coastal fjords where glaciers can create hundreds of prominent seismic events daily (Bartholomaus et al., 2012).

Assessing a landslide in coastal Alaska quickly enough to provide a meaningful tsunami warning is not currently possible. Barry Arm, in Prince William Sound, Alaska, has recently become an example of this hazard in the United States (Dai et al., 2020; Barnhart et al., 2021). Because Barry Glacier has retreated, the newly de‐buttressed fjord wall has slumped, in spurts, more than 200 m over the past few decades. This has prompted concerns that a catastrophic failure could generate a tsunami with several meters of peak wave height reaching nearby communities in just 20 min. The threat has prompted a coordinated interagency response led by the U.S. Geological Survey. Seismic detection of landslides based on the analysis here is one piece of that response.

To maximize warning time and enhance our ability to assess smaller landslides, we use near‐source seismic records for location and volume estimation. We build on approaches demonstrated by the previous work but prioritize the potential for fast warning, algorithmic simplicity, and relevance to coastal Alaska. We assemble a benchmark set of landslides spanning southern Alaska for which proximal seismic records and accurate volume estimates are available (Fig. 1). We use these events to demonstrate an approach to rapid assessment that is portable across different locations and real‐time capable in the presence of sufficient seismic network coverage.

As of early 2023, seismic stations in Barry Arm had helped anecdotally detect three nearby landslides. Barry Arm 1 (⁠$0.6Mm3$⁠) was observed in October 2020 in remote sensing imagery several days after it occurred (Fig. 1a). In August 2021, Barry Arm 3 (⁠$0.5Mm3$⁠) was also observed in imagery 15 km to the north, with a runout onto Barry Glacier. A review of seismic data allowed us to determine when these events occurred. This review also revealed a third significant event in October 2020, Barry Arm 2 (⁠$0.5Mm3$⁠), later corroborated with satellite imagery.

We expand the range of volumes, tectonic settings, and network configurations by adding six other well‐studied landslides (Fig. 1b and Table S1, available in the supplemental material to this article). The 2015 Taan Fiord (Gualtieri and Ekström, 2018; Higman et al., 2018) and the 2016 Lamplugh Glacier (Bessette‐Kirton et al., 2018) landslides each have the largest volume estimates (72 and $64Mm3$⁠, respectively); the 2016 Red Glacier (Toney et al., 2021) and the 2014 Mount La Perouse events have estimates of 16 and $16Mm3$⁠, respectively; and the 2020 Taku River and the 2016 Cowee Creek events have volumes of 0.6 and $0.5Mm3$⁠, respectively. These landslides span complex tectonics with heterogeneous seismic velocities (e.g., Nayak et al., 2020). Barry Arm, Taan Fiord, and Red Glacier have seismic station coverage well distributed in distance and azimuth. Lamplugh Glacier, Mount La Perouse, and Taku River lack observations within 50 km. Lamplugh Glacier, Cowee Creek, and Taku River suffer from significant gaps in azimuthal coverage. We purposefully choose examples with nonideal network configurations, because they serve as realistic edge cases. Details and references for each landslide can be found in Karasozen and West (2023).

Within a few hundred kilometers of the source, the vertical‐component long‐period (<0.05 Hz) waveforms demonstrate remarkable coherence and moveout (Fig. 2a). By contrast, the short‐period waveforms (>1 Hz) are highly variable between stations, do not have clear phase arrivals, attenuate quickly with distance, and share considerable similarities to nearby glacier activity (e.g., Allstadt et al., 2018).

We use waveform coherence to test the ability to back‐project vertical‐component traces using a fixed surface‐wave velocity. We use five‐minute waveforms from stations within 2° (∼222 km) of the source. Seismograms are corrected for instrument response, filtered at 0.01–0.5 Hz, and tapered. We align the traces, assuming velocities of 2.5–6.0 km/s. Finally, we normalize all traces, stack, and divide by the number of traces. The maximum absolute value of this stack is a measure between zero and one of seismogram coherence.

All nine examples demonstrate good coherence when back‐projected to the known source using a velocity of 3.4 km/s (Fig. S1). This agrees well with similar results for Taiwan (Lin et al., 2015). The Cowee Creek event has notably low signal‐to‐noise and lower trace alignment, suggesting significant regional velocity differences. The Red Glacier region includes many low‐coherence stations located on volcanoes, suggesting that it may make sense to exclude these types of stations from monitoring efforts (Figs. S2 and S3). The Taan Fiord landslide shows more azimuthal variation and two significant energy pulses separated by 50 s. Despite these complexities and data limitations, all examples demonstrate the steady moveout of highly coherent vertical‐component energy within just a few degrees of the source (Fig. S3). This fundamental observation is the foundation for the following location and detection discussions.

Techniques that utilize long‐period waveforms locate landslides using backprojection, source inversions, and traditional earthquake location techniques (e.g., Ekström and Stark, 2013; Lin et al., 2015; Chao et al., 2017). We build on these long‐period approaches but with a focus on near‐source observations to reduce detection limits, location errors, and eventual warning time.

We search a 2° × 2° grid of potential sources with a grid spacing of 1 km. Three‐minute waveforms are prepared as in the Long‐Period Waveforms at Close Distances section. We evaluate location quality based on the strength of the coherence coefficient, the absolute location error, and the relative location error. We define the relative error as the distance from the predicted location to where the coherence drops to 90% of its maximum value. We measure this in the east–west and north–south directions and take the larger of the two. This is a relative measure only, but it is attractive because it is sensitive to the presence of multiple local maxima.

We test different methods for measuring waveform coherence. The most straightforward of these is summation or stacking, as described earlier. We also test the amplitude‐normalized version of semblance originally defined by Neidell and Taner (1971). This approach is similar to stacking but includes a square in the amplitude summation that causes the coefficient to decrease more quickly when traces are slightly misaligned. Finally, we test a modified version introduced by Ripepe et al. (2007), in which semblance is defined as the mean of the normalized covariance between each pair of traces. This approach further penalizes slight misalignment. See Text S1 for a full description of coherence measures.

Many semblance studies use envelopes of waveforms instead of the actual waveforms. This is a common adaptation when there is insufficient coherence in the waveforms. We tested the same coherence measures using envelopes and, as expected, found lower resolution in the source locations. We find no reason to reduce the data to envelopes, because the waveforms demonstrate excellent coherence at short distances.

All three methods return broadly consistent locations (Fig. 3 and Fig. S4). The locations for the three Barry Arm landslides are remarkably good (absolute location errors <5 km). This reflects the excellent station coverage. The Red Glacier location is good, but the coherence value is low. This results from the inclusion of low‐signal‐to‐noise volcano stations discussed previously. Lamplugh Glacier, Cowee Creek, and Taku River have reasonable locations but considerable relative error, demonstrating the impact of limited station coverage within a couple of degrees. Mount La Perouse demonstrates nicely how our definition of relative location error accurately captures the impact of local maxima, in this case, because only five poorly distributed stations are being used (Fig. 1b and Figs. S2 and S3). Taan Fjord shows a surprising amount of relative error. This could be because of the azimuthal variation and improved with a more azimuthally balanced station selection. The rest of the analyses we present are based on the amplitude stacking technique, though semblance methods produce comparable results (Fig. S4).

We test location quality as a function of station distance, seismogram window length, and grid spacing. We find that a station inclusion distance of 2° allows both high waveform coherence and minimum location errors (Fig. S5a). These short travel times are also essential in facilitating rapid real‐time performance. A time window of 180 s is sufficient to capture most of the event signals while also allowing for 2° worth of travel time (∼65 s) (Fig. S5b). As expected, grid spacing impacts the resolution of our results, though finer grids significantly increase computational time (Fig. S5c). We describe each test in Text S2.

An initial estimate of landslide size is essential in shaping an emergency response. This is particularly true if the landslide may be tsunamigenic. We demonstrate an approach for approximating landslide volume from seismic amplitudes—a metric readily available during location estimation.

Several authors have reconstructed the mass and movement histories of large landslides (e.g., Allstadt, 2013; Ekström and Stark, 2013). These retrospective studies include careful data selection, parameter testing, and customized modeling efforts—options not amenable to rapid automation (e.g., Norris, 1994; Deparis et al., 2008; Dammeier et al., 2011; Hibert et al., 2014; Lin et al., 2015; Manconi et al., 2016; Fuchs et al., 2018; Le Roy et al., 2019). It is challenging to reconcile these studies, because they derive from quite different analyses of seismograms and require differing assumptions about size, distance, and event source mechanism. However, the highly consistent waveforms exemplified in Figure 2 demonstrate the ability to use these same signals for volume estimation. We focus on volume, because this is the primary control on landslide hazard and related tsunami potential. This is aided by the wide availability of landslide volume estimates in the literature.

To develop an amplitude measure, we examine the maximum absolute seismic amplitude on the same set of stations used for location estimation. We adjust for geometric spreading by multiplying waveform amplitudes by the square root of the distance. As expected, we observe a trend of larger volume landslides producing larger seismic amplitudes (Fig. S6). The seismic energy is influenced by volume, height, runout, and duration. However, volume is sufficiently dominant to allow a meaningful regression against seismic amplitude (e.g., Fuchs et al., 2018).

Volume estimates for all three Barry Arm landslides are quite good, with minor deviations (⁠$∼0.1Mm3$⁠) from the remote sensing volume estimates, thanks to their excellent station distribution and relatively simpler source characteristics (Fig. 4). Lamplugh Glacier, Mount La Perouse, and Cowee Creek landslides have equally good volume estimates within the acceptable error range, despite their varying station distribution (Table S1). Taan Fjord and Red Glacier fall outside the volume estimation error range, likely due to their complicated, long‐duration sources, unbalanced station distribution, and low signal‐to‐noise ratio, resulting in an underestimation of their volumes. Although Taku River’s fit is less precise, it still falls within an acceptable, albeit wide, error range. Volume estimates for landslides often span a wide error range (e.g., Collins et al., 2022). The approach we suggest in this study falls within a comparable error range (e.g., the error range in Table S1) and can be updated in the future as more data becomes available. Even this level of imprecision, however, is sufficient to enable a prompt assessment of the scale of the hazard and estimate its potential consequences.

To test the operational potential of our approach, we set up a real‐time implementation using a grid centered on Barry Arm (Fig. 3). Once per minute, we compute the grid of coherence values for data in the past 180 s. We search for coherence that exceeds an empirically defined threshold of 0.5. If this criterion is met, we declare an alert and send the results to the amplitude and volume routines. The location, volume estimate, and diagnostic information are then pushed to cell phone and email notifications with supporting figures. Although the threshold approach in this test system is simplistic, we find it to be effective.

We began running this real‐time test version while finalizing this article. During a few‐month period in the summer of 2023, we identified seven landslides in near real time that were later confirmed by overflight or satellite imagery (Fig. S7 and Table S2). Continued refinement is needed to reduce false triggers from large earthquakes by testing a range of waveform similarity measures, exploring different signal processing criteria, and assessing various detection thresholds.

We also ran the continuous scanning version on time periods surrounding our benchmark events. The panels in Figure 5 show the 24 hr period surrounding the three Barry Arm events. All the three landslides demonstrate peak coherence values that separate clearly from the rest of the day. Regional and teleseismic surface waves manifest as periods of elevated coherence, though well below the landslide values.

Our results indicate that it should be possible to evaluate landslides within about three minutes of initiation using existing seismic data streams and that a rapid volume estimate can be obtained directly from seismic amplitudes. We find success with landslide volumes as low as $∼0.5Mm3$⁠.

The landslide volume relationship we develop spans a critical range in tsunami warning. The low end of this scale, $∼0.5Mm3$⁠, is well below the volumes estimated to generate regionally damaging tsunamis (Barnhart et al., 2021), whereas the higher end of this range is well demonstrated to be tsunamigenic. An order of magnitude volume estimate that can be derived rapidly without human refinement is sufficient to inform whether or not a tsunami advisory should be issued. These estimates can be augmented after the fact with additional force or moment estimates, as available.

The combination of real‐time and historical testing demonstrates clearly that it is possible to detect and assess the location and size of a landslide within ∼3 min, which is well under the 5 min goal for National Oceanic and Atmospheric Administration (NOAA) tsunami warning protocols. Most landslides we test suffer from poor network coverage, noisy waveforms, and/or clipped data. These challenges are typical in real‐time monitoring systems. There are considerable opportunities to develop more sophisticated detection criteria, harden codes against latent data, and improve the rejection of false alarms. Even so, the real‐time implementation is a motivating proof of concept. With growing evidence for increased hazard from coastal landslides and a significant interagency effort currently underway to track these events, the toolset described here is achievable, and compatible with existing real‐time earthquake monitoring and early warning systems.

Seismic data are from the AK, AT, AV, and CN networks (https://www.fdsn.org/networks, last accessed August 2023). Landslide information includes data from the Incorporated Research Institutions for Seismology (IRIS) Exotic Events Catalog (https://ds.iris.edu/ds/products/esec, last accessed September 2023; Collins et al., 2022). Codes are available through Zenodo (doi: 10.5281/zenodo.7510980; version 1.0.0). The supplemental material includes Text S1 and S2 detailing waveform coherence measures and additional parameter testing information for landslide location estimation; Figures S1 to S7, depicting various aspects such as moveout velocity, station maps, waveform alignments, landslide locations, and parameter test results, including real‐time landslide detections in summer 2023; and Tables S1 and S2 summarizing event details of landslides analyzed and detected in near real‐time during summer 2023.

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

Support for this work was provided by U.S. Geological Survey (USGS) Agreement Number G00014775 and the Office of Alaska State Seismologist. Data from the Alaska Geophysical Network (doi: 10.7914/SN/AK) is made possible with support from the USGS Advanced National Seismic System (ANSS) program cooperative Agreement Number G22AC00001. The authors are grateful to two anonymous reviewers for thorough and thoughtful reviews of this article.