Landslide hazard cascades can trigger earthquakes

Mass movements such as landslides, avalanches, and debris flows are major hazards and key geomorphic agents in mountainous regions. They can be triggered by rainfall, snowmelt, and human activities, and in seismically active regions large earthquakes can induce widespread mass wasting, form landslide-dammed lakes (LDLs), and lead to catastrophic outburst floods after dam breach. Historical examples (e.g., the 1933 M7.5 Diexi earthquake) illustrate how earthquake-triggered landslides and subsequent LDL outbursts amplify disaster impacts. Prior research has largely focused on how earthquakes trigger mass movements, quantify their spatial distribution, and assess post-earthquake geomorphic evolution. However, processes that alter crustal stress via surface loading and/or fluid diffusion—such as earth tides, dam impoundment, hydraulic fracturing, wastewater disposal, and lake filling—can also trigger or induce earthquakes when stress increases on nearby active faults exceed critical thresholds. Because hazardous mass movements and their cascades involve sediment redistribution and changes in water storage and surface loads, they may similarly change stress states and potentially trigger earthquakes. Here, using the 2018 Baige landslide hazard cascades on the Tibetan plateau as a case study, the authors investigate whether LDLs can trigger earthquakes by quantifying the spatiotemporal evolution of seismicity and stress changes on surrounding faults, and by modeling Coulomb stress using a global LDL database.

The paper situates its research within several lines of prior work: (1) Extensive studies document that earthquakes trigger mass movements and hazard cascades, with analyses of landslide failure mechanisms, spatial distributions, and geomorphic impacts in mountainous regions. (2) Numerous studies demonstrate that surface loading and fluid diffusion processes—earth tides, water storage behind dams, hydraulic fracturing, wastewater disposal, and lake filling—can induce or trigger earthquakes when fault stresses surpass thresholds. (3) Reservoir-triggered seismicity is well documented globally, with cases including large events; the mechanisms involve direct gravitational loading and pore pressure diffusion affecting surrounding faults. (4) Mass movements and their cascades cause sediment redistribution and significant, transient changes in water storage and loading, implying that analogous stress changes might trigger seismicity. The paper leverages these literatures to hypothesize that LDLs can trigger earthquakes and that such bidirectional interactions may be more widespread than recognized.

Earthquake catalog building: Continuous seismic data from 13 nearby permanent stations (Jan 2014–Jan 2023) were processed using an AI-based workflow. PhaseNet was used to pick P- and S-wave arrivals, which were associated into events using PyOcto with thresholds of at least two P picks, one S pick, and a total of five P/S picks, yielding 6692 earthquakes within 30°–32.5°N, 97°–100°E. Around the target area (30.7°–31.7°N, 98.1°–99.2°E), 859 events were manually inspected to remove 112 non-tectonic/unreasonable events. Remaining 747 events were located with Hypoinverse, with uncertainties estimated. Magnitude estimation: Local magnitudes were estimated from S-wave amplitudes using a Sichuan-specific national standard magnitude scale. After instrument response removal and Wood–Anderson simulation, maximum horizontal amplitudes were measured in windows starting 0.5 s before P and of length twice S–P. Magnitudes were computed for 669 events with ≥2 stations having S-wave SNR > 3 (range ML 0.1–4.2). Template matching: To improve completeness, GPU-M&L template matching used the 669 events as templates (6 s windows starting 1 s before P and S; 2–8 Hz filtering; local 1-D velocity model). Detections were relocated by grid search (1.2×1.2 km area, 0.06 km spacing) when ≥3 stations recorded both template and continuous waveforms; otherwise, template locations were used. Detection thresholds: if common component number ≥9, CC ≥0.3 and 12×MAD; else CC ≥0.5 and 15×MAD. GPU-M&L yielded 4294 detections including templates. Magnitudes of detections were scaled from template magnitudes by amplitude ratios. Due to sparse pre-May 2018 coverage, 22 routine-catalog earthquakes (Mar 2014–May 2018) were manually added. Declustering and seismicity-rate analysis: Magnitude of completeness within 10 km of LDLs was estimated as Mc = 1.1. Using Reasenberg declustering with parameters: look-ahead 0.5–3 days, confidence 0.9, effective min magnitude 1.1, increase in lower cutoff during clusters 0, crack radii multiplier 5, five clusters (13 events) were identified over ~5 years, leaving 54 events. Statistical significance of rate changes was tested with Habermann’s Z statistic and Marsan & Nalbant’s P statistic. For the week 10–16 Nov 2018, after declustering N=9, prior 27 weeks M=8 (rate ~0.3/week): Z=2.90 (significant at 99.62%). For P, using prior weekly rate 0.3: P=0.9982, implying 0.18% chance of random occurrence (>99.82% significance). Additional tests for low background rates: a conservative Poisson probability using u=3 events/week (largest outside the sequence) gave P=0.0027 (99.73% significance); an empirical windowed count (>1700 weekly windows with 1-day slide) found the >99% threshold at 3 events/week, whereas 9 were observed in the trigger week. Robustness to alternative declustering parameters was also verified (significance >~90%). Coulomb stress modeling: Coulomb failure stress change (ΔCFS) from LDLs was modeled as the combined effects of direct gravitational loading and pore pressure diffusion using GeoTaos. The LDL surface load was computed by convolving Green’s functions for vertical point forces on a homogeneous elastic half-space with distributed surface forces derived from spatiotemporal water level evolution. The LDL surface was discretized into 200×200 m cells; cell elevations from ASTER GDEM V3 (30 m) yielded water depths and loads. Gravitational loading from displaced landslide sediments was neglected (local redistribution). Receiver faults were assumed planar at 4.5 km depth (~0.5 km asl), left-lateral strike-slip with strike ~N50°W, dip ~75°NE, consistent with four regional M>3.5 focal mechanisms (16–35 km from LDLs) and first-motion polarities of the two largest local events (ML 2.6 and 1.9). Friction coefficient was 0.5. Hydraulic diffusivity was taken as 0.3 m²/s (within documented range 0.01–5 m²/s), and sensitivity to diffusivity, fault type, geometry, and depth was explored. Uncertainty analyses: Station coverage was sparse before May 2018. Location uncertainties average ~4 km horizontal and ~8 km vertical; during the second LDL, ~50% and ~95% of hypocenters were shallower than ~1.2 km and ~4.5 km, respectively; median depth of ML≥1.1 events was ~1.5 km (average uncertainty ~6 km). Modeled ΔCFS exceeded ~0.01 MPa down to ~7 km and was larger at shallower depths. Sensitivity tests showed fault dip primarily affects ΔCFS distribution and amplitude but not its sign; friction coefficient had minimal effect on distribution and amplitude; hydraulic diffusivity controls timing and spatial extent of pore-pressure-driven ΔCFS but ΔCFS exceeded 0.01 MPa for diffusivities 0.02–2 m²/s.

• Two landslides (Oct 10 and Nov 3, 2018) on the Tibetan plateau formed LDLs and subsequent outburst floods. The second LDL (peak water level Nov 13, 2018) coincided with a sharp, localized increase in seismicity within 10 km of the LDLs.
• In the week 10–16 Nov 2018, 61 earthquakes (maximum ML ~2.6) occurred within 10 km of the LDLs. With Mc=1.1, 16 events with ML≥1.1 occurred that week; after declustering, 9 events remained. Background rates were ~0.3 events/week (previous 27 weeks) and ~0.2 events/week (subsequent ~224 weeks). Only five other events (all ML<2.0) occurred within 60 km during that week, and no ML≥3.5 earthquakes occurred within 50 km in the prior 3 years, indicating the sequence was not triggered by larger surrounding events.
• Statistical tests consistently show the seismicity-rate increase is highly significant: Z=2.90 (99.62%), P=0.9982 (>99.82% significance), Poisson probability 0.0027 (99.73% significance), and empirical window analysis (>99% threshold is 3 events/week vs observed 9 after declustering).
• Modeled ΔCFS increased with LDL filling and decreased after breach. First LDL peak ΔCFS was ~0.007 MPa (<0.01 MPa threshold commonly associated with triggering), consistent with no significant rate increase. Second LDL peak ΔCFS reached ~0.024 MPa near the peak water level, coincident with maximum seismicity.
• About 90% of earthquakes occurred in regions of positive ΔCFS from combined direct loading and pore pressure diffusion; pore pressure sustained elevated ΔCFS for days after the dam breach, explaining continued activity.
• Generalization: A typical LDL with 30 m dam height (global median) would produce ΔCFS > 0.01 MPa at a representative fault point after ~11 days if not breached, via combined loading and pore-pressure diffusion. Sensitivity tests show fault type/geometry mainly affect spatial distribution, not amplitude, and stress changes are larger at shallower depths.
• Global context: Of ~300 LDLs with documented heights, ~73% exceed the minimum height of reservoirs known to induce earthquakes and ~10% exceed the median such height. Of 145 LDLs with documented duration, ~64% lasted >11 days (no correlation with height), indicating many LDLs worldwide could trigger earthquakes where critically stressed faults exist.

The observed seismicity tightly tracks the evolution of LDL water levels and modeled ΔCFS, supporting LDL-triggered earthquakes. Direct gravitational loading can reduce ΔCFS beneath the LDL but increases it in surrounding regions, while pore-pressure diffusion generally increases ΔCFS and dominates near the LDL and after breach. During 10–13 Nov, the relative contribution of gravitational loading to ΔCFS decreased from ~0.45 to ~0.41; after the 13 Nov breach and peak seismicity, it dropped sharply to ~0.19, indicating pore pressure became the dominant mechanism. The relative contribution is sensitive to hydraulic diffusivity, which ranges widely in nature; higher diffusivity accelerates and broadens pore-pressure effects. Because the pore-pressure source is an elongated river/lake rather than a point, seismicity shows no clear migration pattern. The low background seismicity near Baige enabled detection of a sharp, localized increase, but in more active regions LDLs could potentially trigger more and larger events (as seen for reservoir-triggered seismicity). Examples like the 2008 Wenchuan earthquake, which produced hundreds of LDLs including large and long-lived ones, suggest potential feedback loops between earthquakes and landslide hazards. The findings extend the known earthquake–surface hazard interactions to bidirectional triggering, implying that hazard cascades can both be initiated by earthquakes and, in turn, modulate seismicity through stress and pore-pressure changes. This has significant implications for hazard assessment and management in mountainous regions.

The study presents evidence that landslide-dammed lakes can trigger earthquakes through combined direct gravitational loading and pore-pressure diffusion. For the 2018 Baige case, seismicity increased significantly as the second LDL approached peak water level, with ~90% of events occurring in regions of positive modeled ΔCFS that peaked at ~0.024 MPa. The first, smaller LDL produced insufficient stress change (~0.007 MPa) to elevate seismicity. Modeling indicates that many LDLs globally have the potential to trigger earthquakes, given their typical heights and durations and the presence of critically stressed faults. The results establish that earthquake–surface hazard interactions can be bidirectional. Future work should integrate LDL-triggered seismicity into multi-hazard risk assessments, improve real-time monitoring of LDL water levels and local seismicity, refine coupled poroelastic models (including potential coupling between loading and diffusion), and assess similar triggering potential from other surface water bodies such as expanding glacial lakes under climate change.

• Seismic network coverage was limited before May 2018, potentially missing smaller events; however, the main sequence occurred when coverage improved.
• Earthquake location uncertainties are relatively large (~4 km horizontal, ~8 km vertical); focal mechanisms for small events are not fully constrained.
• Receiver fault geometry (especially dip) is inferred from nearby larger events; smaller-event geometries may differ. Sensitivity tests indicate dip mainly affects ΔCFS distribution, not its sign.
• Hydraulic diffusivity is poorly constrained and spatially variable; conclusions are robust over a plausible range (0.02–2 m²/s), but timing and spatial extent of ΔCFS depend on diffusivity.
• The modeling neglects coupling between pore-pressure diffusion and gravitational loading and ignores gravitational loading from redistributed landslide sediments (assumed negligible at broader scales).
• The region’s low background seismicity aids detection; in more active areas, LDL-triggered changes might be masked by ongoing earthquake–earthquake triggering.