Seismic Monitoring and Geomorphic Impacts of the Catastrophic 2018 Baige Landslide Hazard Cascades in the Tibetan Plateau

Surface hazards can form hazard cascades which expand their reach. However, our understanding of their complex dynamics and ability to mitigate their impacts remain limited. In 2018, two landslides dammed the Jinsha River in the Tibetan plateau and formed landslide‐dammed lakes. Subsequent dam breaches prompted the evacuation of >120,000 people. An early warning system for floods using a regional seismic network has been proposed on the basis of catastrophic floods having been detected ∼100 km away, with seismic energy proportional to discharge. Surprisingly, we find that this catastrophic outburst flood was undetectable beyond a few kilometers, with peak seismic energy preceding peak discharge. We propose that river channel stability also controls seismic energy generation and should be considered for accurate monitoring of catastrophic floods. In contrast, we find that the various processes during dam breach can be well‐characterized seismically further away and provide warning ∼60 min before discharge exceeds monsoon flood levels. We also show that numerical modeling of dam breaches which typically lacks in situ measurements can benefit from incorporating seismic data as constraints. Finally, we show that this event drastically increased sediment fluxes ∼670 km downstream for years and may significantly reduce the capacity of hydropower plants. Our results reveal ways to improve early warning of catastrophic outburst floods and the need to consider surface hazards' long‐term impact when managing infrastructure in mountainous regions.

and population growth and infrastructure development contribute to increasing exposure (Ozturk et al., 2022), our ability to manage these surface hazards is critical for sustainable development in mountainous regions.
Surface hazard management would benefit from developing early warning systems and understanding the complex physical mechanisms and long-term impact of these hazards.Typically, the monitoring of these surface hazards relies on satellite remote sensing (Irons et al., 2012;Tralli et al., 2005;Wasowski & Bovenga, 2014) and in situ sensors (Arattano & Marchi, 2008;McCoy et al., 2010).While satellite remote sensing can track these events at meter-scale resolution over most of our planet, typical satellite revisit times of the order of days preclude real-time monitoring and early warning.In comparison, in situ sensors such as force plates and flow stage sensors used to detect and monitor floods (Habersack et al., 2017;McArdell et al., 2007) can provide high temporal resolution data but have much smaller spatial coverage and can be challenging to install and maintain (Le Guern et al., 2020;Vericat et al., 2006).In addition, while numerical models can be used to simulate certain surface hazards (Khrapov et al., 2013), they require sufficient understanding of the underlying physical mechanisms and well-constrained physical parameters as inputs (Anees et al., 2016;Yavari-Ramshe & Ataie-Ashtiani, 2016).
In recent years, seismometers have been shown to be capable of tracking various surface hazards in real-time at remote distances (Chmiel et al., 2022;Dietze et al., 2022;Ekström & Stark, 2013;Zhang et al., 2021), and are a promising alternative tool to quantify the dynamics and evolution of these processes.Early warning systems for catastrophic floods using regional seismic networks have also been proposed based on these events having been detected ∼100 km away (Cook et al., 2021;Maurer et al., 2020).However, while models explaining how certain surface processes radiate seismic energy have been proposed, our understanding of extreme events remains limited and incomplete.For example, while existing models propose that floods' seismic signals are produced by the impacts of moving bedload particles (Tsai et al., 2012) and the energy transmitted to the river bed and banks by turbulent flow (Gimbert et al., 2014), whether seismic energy always increases with discharge as these models predict and whether there are other important physical processes/factors that also affect the generation of catastrophic floods' seismic signals (Maurer et al., 2020) are still open questions.In addition, no model has yet been proposed to explain seismic signals generated during dam breaches.This makes it difficult to ascertain the general effectiveness of currently proposed early warning systems (Cook et al., 2021;Maurer et al., 2020).
Since surface hazards can mobilize coarse sediments as well as erode and reorganize river channels (Hovius et al., 1997;Schaller et al., 2001), they are also well-known to impact geomorphic evolution (Cook et al., 2018;Schaller et al., 2001;Tramblay et al., 2020).However, previous studies have typically focused on the short-term impact on geomorphic systems by quantifying terrain changes immediately before and after these events (Cook et al., 2018;Shugar et al., 2021).Therefore, the lack of longer-term in situ observations (Garcia-Castellanos & O'Connor, 2018;Shugar et al., 2021) means that our understanding of the lasting impact of such hazards on geomorphic systems, which might also affect downstream hydropower plants and reservoirs (Hewawasam, 2010;Zhang et al., 2015), remains limited.
On 10 October and 3 November 2018, two successive landslides in the Tibetan plateau (Figure 1a; Zhang, He, et al., 2019;Zhang, Xiao, et al., 2019;Zhong et al., 2020;Fan, Yang, et al., 2020;Gao et al., 2021;Chen et al., 2021) blocked the Jinsha River, which is the upper course of the Yangtze River, and created two LDLs.The subsequent dam breaches resulted in massive outburst floods that caused the immediate evacuation of >120,000 people (NDRCC, 2018b(NDRCC, , 2018a) ) and severe damage/destruction of 27,000 houses, 8 bridges, 4 hydropower plants, and 33,000 ha of farmland up to 670 km downstream (Chen et al., 2021;Fan, Yang, et al., 2020;Gao et al., 2021;Zhang, Xiao, et al., 2019;Zhong et al., 2020).The second landslide lake outburst flood (LLOF), with peak discharge of ∼31,000 m 3 /s and volume of 5.8 × 10 8 m 3 (Chen et al., 2021;Fan, Yang, et al., 2020;Gao et al., 2021;Zhong et al., 2020), is the largest recorded LLOF of the 21st century to date and provides a unique opportunity to understand such hazard cascades.Here, we integrate seismic observations with hydrological data, videos/photographs, satellite observations, and numerical modeling to study the detailed seismic signatures, evolution, and impact of the hazard cascades.Our results reveal ways to improve seismic monitoring and early warning of catastrophic floods and the potential long-term impact of surface hazards on river systems and downstream hydropower plants.

10.1029/2023JF007363
3 of 15 Fan, Yang, et al., 2020;Gao et al., 2021;Zhong et al., 2020).This ∼61 m high landslide dam created a large LDL that inundated terrain up to ∼45 km upstream (Figure 1b).The LDL naturally breached the west side of the landslide dam at 17:15 CST on 12 October.With the dam breach, the discharge rapidly increased to a maximum of 10,000 m 3 /s at 06:00 CST on 13 October before gradually decreasing back to the same level as the flow upstream feeding the LDL at 22:00 CST on 13 October.
Three weeks later at 17:40 CST on 3 November, a second landslide in Baige transported an additional 8.7 × 10 6 m 3 of materials into the Jinsha River.This smaller landslide deposited atop the residual dam from the first landslide and created a higher dam that was at least 96 m high that inundated terrain up to ∼70 km upstream (Figure 1b; Chen et al., 2021;Fan, Yang, et al., 2020;Gao et al., 2021;Zhong et al., 2020).To mitigate the potential damage from an unexpected outburst flood, the Yangtze River Flood Control and Drought Relief Headquarters dug a ∼15 m deep drainage trench on the west side of the dam (Zhong et al., 2020).Water started flowing into the drainage trench at 04:45 CST on 12 November, which initiated the erosion of the top and downstream face and subsequent sediment collapse at the breach side slope of the landslide dam (Figure S2 in Supporting Information S1).The dam breach was well-recorded by a dense network of seismometers operated by the China Earthquake Administration (Figure 1a).

Seismic Data Processing
All stations used in this study (Figure 1a) are operated by the China Earthquake Administration and equipped with a 3-component broadband seismometer with a sampling rate of 100 Hz.Due to the rapid attenuation of high-frequency seismic energy, >5 Hz seismic signals generated by dam breach and flood were not visible above the background noise level at these stations.Hence, we focus on 0.5-5 Hz seismic signals for our study.After removing the instrumental response, using Welch's method, we calculate the power spectral density (PSD) of seismic signals with time window of 200 s.We then estimated the signal's peak frequencies for the 200-s moving time windows, excluding time windows that included seismic signals from local earthquakes.

Landslide Force Vector Inversion
Acceleration/deceleration of the two landslides, accompanying landslide initiation, movement, and deposition, generates long-period seismic signals that attenuate slowly and are less sensitive to heterogeneities; hence, they can be recorded by seismic stations several hundred kilometers away (Ekström & Stark, 2013;Zhang, He, et al., 2019).The measured seismic signals contain information about the source and its propagation path.In the case where the landslide path is much shorter than the landslide-to-station distance, the single-point force well approximates the force exerted by the landslide on the hillslope during run-out (Zhang, He, et al., 2019).We treat the Green's function as the effect of the seismic wave path through the Earth and calculate it using the 1-D ak135 Earth velocity and an elastic attenuation model based on the matrix propagation method (Wang, 1999).Next, according to the damped least squares approach, we use the long-period (from 20 to 80 s for the first landslide and from 20 to 60 s for the second landslide) seismic signals recorded by 12 seismic stations within 200 km to invert for the force vector exerted by the landslide on the slope (Figure 2).Information about these seismic stations is shown in Table S1 in Supporting Information S1.

Dam Breach Modeling
We model the breach process of the second landslide dam due to overtopping using the DLBreach software (Figure S1 in Supporting Information S1).This landslide dam breach is simplified as a two-way breach of a homogeneous non-cohesive dam by overtopping (Wu, 2013(Wu, , 2016)).The simulation includes hydrodynamics, sediment erosion, and breach enlargement processes.For the hydrodynamics, the Weir equation is used to calculate the breach flow during the intensive breaching or erosion stage (Singh & Snorrason, 1984; Wang & Bowles, 2006), while the Keulegan equation is used to estimate the breach flow during the general breach or inlet evolution stage (Wu, 2013(Wu, , 2016)).For the sediment erosion, a non-equilibrium total-load transport model is used for the erosion of the second landslide dam (Armanini & Di Silvio, 1988;Sánchez & Wu, 2011).For breach development, the simulation was run to calculate the volumetric bed change using time steps of 0.2 s.The volume change can then be converted to bed change and width change along the breach cross-section, which involves the sediment collapse/mass failure of breach side slopes (Figure S1 in Supporting Information S1).Since there are already sufficient field surveys, laboratory experiments, and numerical simulations on this dam breach (e.g., Chen et al., 2020;Zhong et al., 2020), the key input parameters used to simulate the dam breach are well documented and are listed in Table S2 in Supporting Information S1.For the less well-constrained input parameters such as porosity and adaption coefficient, while changes in these parameters affect the temporal evolution of discharge and sediment transport, the peak sediment transport always preceded peak in measured discharge by hours (Figure S2 in Supporting Information S1).Therefore, our main conclusion is robust and our preferred model is the one where the simulated and measured discharges are aligned (Figure 3d).

Measured Discharge and Seismic Signals of LLOF
Although seismic stations GYXJ and BTA are located close to hydrological stations that can provide measured discharge data (Figure 1a), the nearest points of the river channel to stations GYXJ and BTA are still ∼1.5 and 7.8 km upstream of the Yebatan hydropower plant and Batang hydropower plant, respectively.We thus calculate the arrival time difference of the flood passing through the hydropower plants and the nearest point of the river channel to the seismic stations using the average flow propagation speed.
Based on the hydrological measurements, we know that it took 120 and 110 min (Figure 1c) for the peak flows of the two floods to travel ∼54 km from the breach to the Yebatan hydropower plant.Hence, the average flow propagation speeds of the two flood peaks passing through the Yebatan plant are 7.5 and 8.2 m/s, respectively.Therefore, the arrival time differences between the two floods passing by seismic station GYXJ and the Yebatan hydropower plant are 200 and 180 s, respectively.In comparison, it took 70 and 105 min (Figure 1c) for the peak flows of the two floods to travel ∼25 km from the Lawa to the Batang hydropower plant.Hence, the average flow propagation speeds of the two flood peaks passing through the Batang station are 6.0 and 4.0 m/s, respectively.Therefore, the arrival time differences between the two floods passing by seismic station BTA and the Batang hydropower plant are 1,300 and 1,950 s.The estimated arrival time differences are then used to align the timings of the recorded discharges and seismic signals.Finally, we assume the flow discharges recorded by the Yebatan and Batang hydropower plants as the flow discharges when the outburst floods passed by stations GYXJ and BTA, respectively.

Sediment Flux Data and Water Surface Width Changes
Measured sediment flux data at the Batang and Shigu hydrological stations were extracted from the Chinese Hydrological Data Yearbook published by the Ministry of Water Resources, China.Note that, like most rivers in the world, the bedload flux in the Jinsha River has not been systematically monitored.Therefore, we only focus on the suspended sediment flux.
The water surface width changes were estimated from Sentinel-2 RGB images which have 10 m resolution.In order to reduce the effect from changes in discharge and vegetation, we selected images from the same month (December) each year from 2015 to 2021, which had similar discharges (Table S3 in Supporting Information S1), to estimate the river water surface widths.We confirm that data from other winter months show similar trends.We focus on the region from 30 km upstream to 100 km downstream of the landslide dam, which is covered by two tiles.The date of the selected images and the associated discharges are listed in Table S3 in Supporting Information S1.For each image, we manually delineate the margins of the river water surface and then smooth the data with a 1-km moving window.Finally, we take the river water surface width in December 2017 as the reference to calculate annual water surface width changes.

Landslide Dynamics
The shape of the force-time functions of the two landslides are similar, suggesting that the two landslides have similar travel processes (Figure 2).The failure of the material at about 3,700 m elevation initiated at 22:05:46 October 10 (Beijing time, UTC + 8; all subsequent time are given in Beijing time) for the first event and at 17:21:30 November 3 for the second event, and moved downward rapidly, generating an upward force.As the material on the hillslope is eroded and entrained, the volume of the two landslides gradually increased to approximately 24 million and 8.7 million m 3 (Chen et al., 2021;Fan, Yang, et al., 2020;Gao et al., 2021) before entering the Jinsha river at 22:06:21 October 10 ( Zhang, He, et al., 2019) and 17:21:53 November 3 respectively, generating the downward subvertical forces with a northeastward component, consistent with the boundary of the impact area.As the first landslide rushed into the river, a landslide dam with a length of ∼1,200 m, width of 143 m, and minimum height of ∼51 m, formed (Chen et al., 2021;Gao et al., 2021).Although the volume of the second landslide is smaller, this event blocked the natural drainage channel of the first dam and formed a larger dam with minimum height of ∼96 m (Chen et al., 2021;Gao et al., 2021;Zhong et al., 2020).

Evolution of Dam Breaches
While some dam breaches have been recorded seismically (Feng, 2009;Yang et al., 2022), how the evolution of various dam breach processes can be inferred from their seismic signals remains underexplored.The first dam breach event generated 0.5-5 Hz seismic energy that was clearly recorded by station JSZJ at ∼15 km away starting 00:45 CST on 13 Oct and lasting for ∼6 hr before the station failed due to unknown reasons (Figure S3 in Supporting Information S1).The seismic signal's peak frequencies are mainly at 0.8 and 2.1 Hz.Similarly, the second dam breach event generated seismic energy that was clearly recorded starting 14:03 CST on 13 Nov and lasting for ∼6.5 hr (Figure 3).However, the peak seismic frequencies varied between 0.8, 1.5, and 2.1 Hz at different time periods.
We interpret that the three peak frequencies were generated by (a) the collapse of breach side slope which is similar to landslides that can generate low-frequency (<1 Hz) seismic signals (Ekström & Stark, 2013), (b) turbulent flow which tends to radiate high-frequency (>1 Hz) seismic signals (Gimbert et al., 2014), and (c) bedload transport which tends to radiate even higher-frequency seismic energy than turbulent flow (Gimbert et al., 2014;Tsai et al., 2012).The 1.5 Hz turbulent flow seismic energy is apparent at certain stages during the second dam breach because unlike the first dam which only contained easily erodible sediments, the second dam contained huge boulders that are difficult to transport (Chen et al., 2021;Fan, Yang, et al., 2020) and can radiate stronger turbulent seismic energy through interaction with water flow.
The variability of peak frequency content divides the second dam breach process into three distinct stages (Figure 3).The first stage, which lasted ∼1.5 hr, contains peak frequencies of 0.8 and 1.5 Hz, which suggests that both slope collapse and turbulent flow are the dominant processes emitting seismic energy.Subsequently, as the discharge increased, the seismic signal became dominated primarily by the 0.8 Hz peak frequency over the next 2 hr.This is consistent with field observations (Zhong et al., 2020) of large numbers of materials on the breach side walls collapsing due to increasing trench erosion which resulted in rapid trench enlargement and maximum seismic energy during this time period (Figure 3).Eventually, the 0.8 Hz seismic energy dropped out, which suggests that the trench enlargement process slowed down and eventually stopped.At this point, the discharge reached a peak of ∼31,000 m 3 /s (Figure 3c) which far exceeds the 10,000-year return period discharge of ∼11,500 m 3 /s at the site (Zhang, Xiao, et al., 2019) and is more than 10 times larger than previously reported peak discharges of lake outburst floods (LOFs) that were recorded by regional seismic networks, such as the ∼2,500 and 2,400 m 3 /s for the 1994 Lugge Tsho (Maurer et al., 2020) and the 2016 Gongbatongshacuo LOFs (Cook et al., 2018), respectively.As the discharge reached its maximum at 18:00 CST on 13 October, there was a second smaller peak in seismic energy (Figure 3c) with the seismic signal dominated by peak frequencies of 1.5 and 2.1 Hz, which reflect turbulent flow and bedload transport.This is consistent with existing fluvial seismic models (Gimbert et al., 2014;Tsai et al., 2012) which showed that both the bedload sediment transport and the interaction between the turbulent flow and bed/bank generate seismic signals and they correlate with discharge (Bakker et al., 2020).
Our numerical modeling of the dam breach shows that the simulated sediment flux's temporal evolution correlates with the PSD of the recorded 2-5 Hz seismic signal (Figure 3d).This supports our interpretation that the highest peak frequency (2.1 Hz) seismic signals are dominated by bedload sediment transport.This also suggests that seismic recordings can potentially improve numerical modeling of dam breaches by constraining sediment flux's temporal evolution, since monitoring of sediment flux is a significant challenge even for well-documented dam breaches due to their catastrophic nature (Wahl, 2004).In addition, due to the unpredictability of dam breaches, monitoring the temporal evolution of drainage trenches and discharge during dam breaches is also a challenge.Since seismic data contain information on multiple physical processes of dam breaches, such as the drainage trench's evolution and discharge, the numerical modeling of dam breaches can benefit from incorporating seismic data as constraints.Finally, we find that the dam breach seismic signal at station JSZJ is clearly observed at 14:03 which is ∼4 hr before the time of peak discharge and ∼60 min before the resulting LLOF's measured discharge exceeds the typical monsoon flood discharge level (Figure 3).Furthermore, the dam breach seismic signals are clearly detected outside the anthropogenic noise frequency band by stations up to ∼66 km away (Figure S4 in Supporting Information S1).

Evolution of Outburst Floods
As the first LLOF traveled down to the Jinsha River, its peak flow decreased from 7,850 m 3 /s at Batang hydrological station at 15:20 CST on 13 November to 5,880 m 3 /s at Benzilan hydrological station at 10:00 CST on 14 October to 5,220 m 3 /s at Shigu hydrological station at 01:00 CST on 15 October (Figure 1c).Finally, the peak flow decreased to 4,850 m 3 /s at 10:00 CST on 15 October when it entered the Liyuan hydropower plant.As the second LLOF traveled down to the Jinsha River, its peak flow decreased to 20,900 m 3 /s at 1:00 CST on 14 Nov, 15,700 m 3 /s at 13:15 CST on 14 Nov, and 7,170 m 3 /s at 8:40 CST on 15 Nov, according to discharge measurements at the Batang, Benzilan, and Shigu hydrological stations ∼190, 380, and 560 km downstream of the landslide dam, respectively (Figure 1c).Finally, the peak flow decreased to 7,200 m 3 /s at 12:30 on 15 November when entering the Liyuan hydropower plant ∼670 km downstream of the landslide dam.
As the second LLOF traveled downstream and passed through four hydropower plants, its peak discharge along most of this flow path (Figure 1c) was significantly larger than the catastrophic 2021 Chamoli, India (Shugar et al., 2021-14,200 m 3 /s) flood that was detected by regional seismic stations ∼100 km away.However, seismic signals of this LLOF were only visible at two stations within 10 km (Figure 4) and not visible at stations beyond 10 km from the river (Figure S5 in Supporting Information S1).Similarly, the first LLOF was not visible at any station (Figure S6 in Supporting Information S1).
The 1-2.5 Hz average power during the catastrophic Chamoli flood was −148.4 dB at station BNG, which was located ∼100 km from the river (Cook et al., 2021).This is 3.7 dB higher than the average power during our second LLOF recorded by stations within 10 km of the river, although our LLOF's discharge when passing by stations GYXJ, BAT, CLZJ, and HNXJ (Figure 1c) was >1.5 times than that of the Chamoli flood (Shugar et al., 2021).Furthermore, we find that the background noise levels at stations GYXJ and BTA (<−162.4dB) were lower than that in the Chamoli region (>−159.1 ± 3.2 dB; Cook et al., 2021).Therefore, the low seismic detectability of our LLOFs compared with the previous study (Cook et al., 2021) is not due to differences in the background noise level and instrument sensitivity.
Furthermore, previous studies assumed that the source of peak seismic energy tracks the point of peak discharge since existing fluid-seismic models (Gimbert et al., 2014;Tsai et al., 2012) predict a power-law relation between discharge and seismic energy.However, our direct in situ discharge measurements show that the peak seismic energy recorded at station BTA preceded the LLOF's peak discharge by ∼1.4 hr (Figure 4b).
Using satellite-based observations (Figure S7 in Supporting Information S1), we also find that after the hazard cascades, the water surface widths of the Jinsha River from 30 km upstream to 100 km downstream of the landslide dams are larger than before the hazard cascades under similar discharge conditions (Figure 6a).Furthermore, while the average water surface width of the 100 km section downstream from the landslide dams remains greater than before the hazard cascades, under similar discharge conditions, this downstream average water surface width has been decreasing continuously from 2018 to 2021 (Figure 6b).

Seismic Monitoring and Early Warning of Hazard Cascades
Compared with existing physical models (Gimbert et al., 2014;Tsai et al., 2012) which predict that seismic energy increases with discharge, we find that the second LLOF's peak seismic energy significantly preceded peak discharge (Figures 3 and 4).Although these existing models can explain the seismic energy peak at 18:00 CST on 13 Oct corresponding to the maximum discharge, PSD calculated from the existing models cannot explain our observation of the seismic energy peak that occurred at 16:30 CST on 13 Oct which precedes peak discharge.Furthermore, the second LLOF is relatively quiet compared to a previously recorded LOF with smaller peak discharge (Cook et al., 2021;Maurer et al., 2020).Note that since there are many uncertainties in calculating the propagation media response of high-frequency (>1 Hz) seismic signals for regions where a high-resolution seismic velocity model is not available, and the PSD simulations are in this frequency domain (Gimbert et al., 2014;Tsai et al., 2012), we do not directly calculate the seismic energy radiated by the LLOFs based on these existing models.Nevertheless, it is clear that there exist other physical process(es) that can control the seismic energy radiation during catastrophic flood events.We suggest that channel erodibility and bank slope stability are important controls of seismic energy radiation from LOFs.Before the peak flood discharge, the peak in seismic energy might reflect the channel being significantly eroded which also triggered widespread collapse of the bank slope.Similar channel erosion and bank collapses also occur during dam breaches, and our numerical simulation of the dam breach shows that the peaks of breach expansion and sediment flux preceded peak discharge (Figure 2d).Subsequently, the weaker seismic energy radiated by the flood may be due to the more stable banks of the Jinsha River because the unstable bank slopes had already failed.The channel stability of the Jinsha River near the landslides has been confirmed by many geological surveys (Fan, Lin, et al., 2020;Liu et al., 2017;Zhao et al., 2018), hence ∼10 hydropower stations are being planned/constructed within 20 km upstream and 370 km downstream of the landslides (Wang et al., 2015;Wu et al., 2020).Our conceptual model can explain why during the 1994 Bhutan glacial LOF, only certain river segments radiated detectable seismic energy (Maurer et al., 2020) since they may reflect differences in channel erodibility and/or riverbank stability.This implies that while catastrophic floods along erodible channels with unstable riverbanks might be readily detectable by regional seismic networks, floods along stable channels might be relatively quiet and hence require a denser seismic network to monitor.
While an early warning system for floods using a regional seismic network has been proposed based on the idea that catastrophic floods have been detected ∼100 km away (Cook et al., 2021;Maurer et al., 2020), our observations show that the largest recorded outburst flood in the 21st century to date was undetectable beyond a few kilometers, probably due to variations of river channel stability.Therefore, seismic monitoring and early warning of catastrophic floods need to account for spatiotemporal changes in river channel erodibility, bank stability, and sediment flux-discharge relationship for an accurate estimate of the location, timing, and size of flood events.In comparison, we show that the dam breach processes (slope collapse, turbulent flow, and bedload transport) can be characterized seismically further away and potentially provide warning ∼60 min before discharge exceeds typical monsoon flood levels (Figure 3d).The seismic signals are also more easily distinguishable from other natural and inferred anthropogenic noise sources due to differences in frequency content (Figure 3 and Figure S4 in Supporting Information S1).Thus, seismic monitoring of dam breach can provide earlier and clearer warnings of catastrophic outburst flood events compared to currently proposed early warning systems (Cook et al., 2021;Maurer et al., 2020).These seismic observations can also provide important constraints for numerical modeling of dam breaches.

Long-Term Impact of Hazard Cascades on River Morphology and Hydrology
Our observations demonstrate that hazard cascades can have a long-term impact on river morphology and hydrology that lasts at least several years.The two Baige landslides transported 3.4 × 10 7 m 3 of sediments into the Jinsha River.Following the landslide dam breaches, the catastrophic outburst floods not only entrained sediments from the landslide dams but may have also eroded and scraped the river channel, bank, and hillslope (Yang et al., 2021), which resulted in a significant increase in the downstream sediment flux in 2018.As these sediments are transported downstream, some are deposited along the way before being successively re-transferred downstream during the next monsoon seasons, which explains why the recorded sediment fluxes after the hazard cascades are higher than the expected sediment flux level given annual runoff in the following years (Figure 5).On the other hand, the greater water surface width of the 30 km section upstream from the landslide dams (Figure 6a) indicates that the residual dam after the dam breach still blocks the channel, which reduces the flow velocity in this section and thus intercepts some of the incoming sediments (Korup & Montgomery, 2008).A new river knickpoint may thus form here.Therefore, the sediment flux at sections downstream from the landslide dams depends on the relative amount of sediments intercepted by the residual dam and the re-transported sediments that were deposited along the channel during the hazard cascades.If the amount of intercepted sediments is similar to the re-transported sediments, the sediment flux downstream would be similar to the expected sediment flux level given the runoff, as observed at Batang in 2019 (Figure 5a).However, while the amount of intercepted sediments is the same for any site downstream from the landslide dams, more sediments deposited along the river channel after the hazard cascades can be re-transported for sites located farther downstream from the landslide dams.This could explain why the recorded sediment flux at Shigu in 2019 is higher than the expected sediment flux level given the runoff (Figure 5c).Furthermore, when the annual runoff at the Batang and Shigu hydrological stations in 2020 exceeds that in 2018 and 2019, sediments that were still deposited along the channel due to the hazard cascades were remobilized and thus the recorded sediment fluxes at these two stations in 2020 exceeded the expected sediment flux level given the runoff (Figure 5).
As the sediments deposited along the channel by the hazard cascades were gradually transported further downstream, the amount of sediments that still remained which can be re-transported will decrease over time.This can explain why the sediment fluxes at the Batang and Shigu hydrological stations in 2021 were within the expected sediment flux levels given the annual runoff, though this may also be attributed to the lower annual runoff in 2021 compared to previous years.This is also consistent with our observation that the average water surface width of the 100 km section of Jinsha River downstream from the landslide dams has been decreasing over time, given similar discharge, due to re-transportation of sediments downstream from the landslide dams (Figure 6b).As the downstream sediments deposited along the channel after the hazard cascades are completely transported in the next few years, the interception effect of the residual dam may make the downstream sediment flux lower than the expected sediment flux level given annual runoff, which may benefit the downstream hydropower plants temporarily.However, eventually the residual dam will likely breach completely, at which point the downstream infrastructures will be affected by a significant increase in sediment flux within a short period of time.This increase in sediment flux and its associated organic carbon may change the water quality and affect hundreds of thousands of people and the river ecosystem downstream from the landslide dam.
Assuming a sediment bulk density of 1,300 kg/m 3 (Yang et al., 2014), the increased sediment flux at the Shigu station from 2018 to 2021 due to these hazard cascades is equivalent to a reduction of the storage capacity of the Liyuan hydropower plant ∼670 km downstream of the landslide dam (Figure 1a) by ∼3.8 × 10 7 ± 1.5 × 10 7 m 3 which is ∼22% ± 9% of the designed regulation storage capacity (1.73 × 10 8 m 3 ; Xie et al., 2012).This effect might also be borne by other hydropower plants downstream from the hazard cascades (Figure 1), including three of the four largest existing/planned hydropower plants in the world (Three Gorges, Baihetan, and Xiluodu hydropower plants; Hu et al., 2009).Furthermore, continued hillslope deformation around the Baige landslide scars suggests that more landslides might occur soon (Chen et al., 2021;Fan, Yang, et al., 2020;Zhang, Xiao, et al., 2019) which would further increase the sediment flux.Therefore, catastrophic outburst floods cannot only directly damage/destroy hydropower plants, but the resulting increased sediment fluxes in the years after the hazard cascades can also significantly reduce the designed storage capacity of hydropower plants and reservoirs downstream from the landslide dams, which threatens the region's irrigation capacity and energy security.

Conclusion
The hazard cascades initiated by the 2018 Baige landslides in the Tibetan plateau provided a unique opportunity to evaluate our current understanding of seismic monitoring of surface hazards as well as these events' complex dynamics and impact on river systems.Surprisingly, we find that the largest recorded outburst flood in the 21st century to date was seismically undetectable beyond a few kilometers.In addition, contrary to current understanding, the peak seismic energy preceded peak discharge by hours.Therefore, we suggest that accurate seismic monitoring of catastrophic floods requires accounting for spatiotemporal changes in the sediment flux-discharge relationship.In addition, seismic monitoring of dam breaches may provide some of the earliest warnings of potential outburst floods as well as important constraints on dam breach evolution.Finally, our observations demonstrate that such hazard cascades can drastically change the geomorphology and hydrology of river systems up to at least 670 km downstream for at least a few years.Accordingly, the increased sediment fluxes in the years following the hazard cascades may have significantly reduced the storage capacity of a hydropower plant 670 km downstream.Therefore, our results highlight the need to consider the long-term effects of surface hazards on river systems during hydropower plant and land use planning.the manuscript.We thank L. L.

Figure 1 .
Figure 1.Landslide-dammed lakes and outburst floods.(a) Study area with the locations of hydropower plants, hydrological stations, and seismic stations.The names of the seismic stations used in our analyses are marked.(b) Inundation areas of the landslide-dammed lakes.(c) Timings and amplitudes of peak discharges of the first (red box) and second (blue box) outburst floods recorded at different hydrological stations and hydropower plants.Digital elevation model source: Advanced Spaceborne Thermal Emission and Reflection Radiometer.

Figure 2 .
Figure 2. Landslide force vector during runout.(a) and (b) are the first and the second landslides' forces inverted from seismic signals, respectively.

Figure 3 .
Figure 3. Seismic signals and discharge during the second landslide dam breach.(a) and (b) 0.5-5 Hz seismic signals recorded at station JSZJ and its spectrogram.(c) Peak frequencies calculated for 200 s moving time windows.(d) Measured and simulated discharge at the breach, simulated sediment transport, and power spectral density (PSD) of 2-5 Hz seismic signals.The PSD is smoothed with a 200 s moving time window.Gray bar indicates the range of annual maximum discharges (1,620-5,610 m 3 /s) recorded at the Batang hydrological station (Figure 1a) from 1963 to 2021, excluding 2018 and years when data are not available.Timings of some key events are marked.

Figure 4 .
Figure 4. Discharge versus the seismic signals generated by the second outburst flood.(a) and (b) 0.5-5 Hz seismic signals during the second outburst flood recorded at stations GYXJ and BTA, their spectrograms, and measured discharge of the flood.The distances from the stations to the river channel are listed.White dashed boxes mark the inferred anthropogenic noise.

Figure 5 .
Figure 5. Measured sediment flux at hydrological stations.(a) and (c) Measured annual runoff versus sediment flux at the Batang and Shigu hydrological stations ∼190 and 560 km downstream from the landslides.Black lines indicate the linear fit using measurements from 1963 to 2017 in Batang and from 1960 to 2017 in Shigu (excluding 1969-1970 and 1988-2006 when data are unavailable).Gray bars indicate the standard deviation of the measured sediment fluxes of ±5.3 × 10 9 kg in Batang and ±6.6 × 10 9 kg in Shigu, respectively.(b) and (d) Measured sediment flux difference, compared with the linear fit relationship between annual runoff and sediment flux, at the Batang and Shigu hydrological stations.Gray dashed lines indicate the timing of the Baige landslides.

Figure 6 .
Figure 6.Width changes of the Jinsha River.(a) 2015 to 2021 water surface width changes relative to 2017, from 30 km upstream to 100 km downstream of the landslide dams.Water surface widths are estimated from Sentinel-2 images and have an uncertainty of ±10 m.(b) 2015 to 2021 measured discharge at the Batang (orange triangle) and Gangtuo (orange square) hydrological stations, and average water surface width changes relative to 2017 of the 100 km section downstream from the landslide dams (blue circle).
Zhu for valuable discussions and the Changjiang Water Resources Commission for providing the gauge data.This work was supported by the Croucher Tak Wah Mak Innovation Award, CUHK Direct Grant for Research (Grant 4053594), CUHK Research Fellowship Scheme (Grant 4200720), CUHK Postdoctoral Fellowship Scheme (Grant 3135043), the National Key Research and Development program of China (Grant 2022YFF0800604), and the Major Program of the National Natural Science Foundation of China (Grant 42090051).The salary of M. Chmiel was paid by the WSL program Climate Change Impacts on Alpine Mass Movements (CCAMM) and Observatoire de la Côte d'Azur Henri Poincaré Postdoctoral fellowship.