When a lahar is created from a lake breakout or outburst flood event, the transition from flood or streamflow torrent depends on the erosivity of the channel and the supply of sediment being entrained within the flow (e.g., Scott, 1988; Doyle et al., 2011). An event may start as a highly turbulent low-sediment flow, then transform into a hyperconcentrated flow, and may even eventually “bulk up” to exhibit characteristics of a debris flow with the possibility of plug-like (limited internal motion and collisions) or laminar behavior (Scott, 1988, Pierson et al., 1990). At Mt. Ruapehu, the propagational differences of lahars down channel have been observed and characterized in the past (e.g., Cronin et al., 1996, 1999, 2000; Manville et al., 2000; Procter et al., 2010a; Lube et al., 2012). From these studies, models of how lahars bulk up and transition throughout the run-out distance have been postulated. For the lahars in the Whangaehu channel, Cronin et al. (1999) created three four-phase conceptual models based on source distances of 23.5, 42, and >55 km. The first two models are for lahar regimes, whereas the third model describes a lahar almost at its peak run-out distance. In each model, the first phase consists of a super charged streamflow pulse that flows ahead of the head of the flow and is considered the front of the lahar. This phenomenon has also been noted for debris flows interacting with streamflow (Arattano and Moia, 1999). Furthermore, discharge is maximum at the transition between phase 1 and phase 2 (Cronin et al., 1999) and is described as the head of the flow. Phase 2 is described as a mixing zone between streamflow and increasing sediment content, where the peak sediment concentration usually occurs at the end of phase 2 or at the beginning of phase 3 (e.g., Pierson and Scott, 1985). Cronin et al. (1999) defined phase 3 as the lahar body, which has the least amount of the original streamflow contained within. Phase 3 is also characterized by coarse sediment suspensions and is the most likely location for debris flow rheology. Finally, phase 4 is the tail of the lahar where debulking and dilution occurs transforming the lahar back into a hyperconcentrated, mixed, or streamflow.

Mt. Ruapehu (2797 ma.s.l.) is the largest stratovolcano in the central North Island of Aotearoa / New Zealand (Fig. 1), which sits at the southwestern end of the Taupō Volcanic Zone (TVZ). The volcano has a volume of 110 km3, which is composed of several overlapping cone-building formations and surrounding ring plain volcaniclastics (Carrivick and Manville, 2009; Pardo et al., 2012). On top of the volcano, above the currently active vent sits a 1×107 m3 acidic crater lake (Procter et al., 2010a). The Whangaehu channel is the preferred outlet for Crater Lake water and lahars in recent history (Procter et al., 2012, 2021). The Whangaehu channel is on the eastern flank of Mt. Ruapehu where it runs down across the volcanic ring plane and eventually heads southwest for ∼200 km reaching the Tasman Sea (Fig. 1).

Prior to the events that took place in the morning local time on 18 March 2007, a heavy rainstorm occurred accumulating about 256 mm of water over the 10 h prior to the dam breach that led to the outburst flood (Massey et al., 2010). The intense rain caused the Crater Lake level to rise an extra 6.4 m, which caused seepage and extra water to overtop the natural lava formation ledge and enter the Whangaehu channel (Carrivick and Manville, 2009). At ∼ 23:18 UTC, the tephra dam collapsed causing 1.3×106 m3 of water to flow out of the lake and into the Whangaehu channel (Procter et al., 2010a). The dam was eroded and undercut in multiple stages resulting in a series of retrogressing landslides along with the main lahar.

Since the lahar was caused by lake-breakout dynamics and thus contained an abundance of water, the event was classified as a hyperconcentrated streamflow (Procter et al., 2010b). At ∼8.0 km from source, the lahar velocity was recorded at ∼9.5 ms-1 and had an estimated 6 m of downcutting, showing the capability of the lahar to deposit and erode massive amounts of material (Procter et al., 2010a, b). Furthermore, the 18 March 2007 lahar was one of the most thoroughly monitored lahars ever (Manville and Cronin 2007). In total there were 21 monitoring locations (only three of which had three-component seismometers) set up to measure various lahar properties (e.g., flow monitor, camera, stage height, flow sampling, pore pressure, seismic) along the channel (Keys and Green, 2008; Lube et al., 2012), with the lahar taking over 16 h to eventually travel out to the Aotearoa / New Zealand coast, ∼200 km from the original crater lake source.

Here, we delve into the properties of the 18 March 2007 lake-breakout hyperconcentrated streamflow that bulked up to a volume of ∼4.4×106 m3 (Procter et al., 2010a) over the course of 83 km along the Whangaehu channel, originating from Mt. Ruapehu, Aotearoa / New Zealand. The combination of seismic analysis (frequency and directionality) with supplementary measurements (e.g., video, sediment concentration) shows how a lahar transforms over time and distance and how using these seismic techniques can help monitor the ever-changing dynamics and properties of a flow event. Furthermore, we examine previous models of the evolution of a lahar and compare the model with the seismic data available.

In order to examine the PSFs for all three components at each site along the channel, we use the frequency recorded at the maximum amplitude of the frequency spectra for each 10 s running time window (i.e., non-overlapping windowed FFT). The PSF for RTMT (7.4 km from source) shows similar patterns between all three components (Fig. 3). The 5 min prior to the arrival of the front of the lahar is characterized by scattered PSFs between 20–40 Hz for the cross-channel (Fig. 3a) and parallel (Fig. 3b) directions, while in the vertical direction (Fig. 3c) the PSF is ∼30 Hz. When the front (streamflow pulse/bow wave) of the lahar arrives at the station, the PSF in all three components decreases to ∼ 5–10 Hz for about 1 min before increasing again to higher frequencies. After the head (peak seismic amplitude) of the lahar passes the station, the PSF in the cross-channel and parallel directions remain between 30 and 40 Hz for the rest of the recording window. In the vertical component, the PSF is scattered between 20 and 40 Hz for ∼15 min after the arrival of the head of the lahar and then becomes narrower, similar to both the cross-channel and parallel components with PSFs between 30 and 40 Hz.

Further down the channel at station TRAN (28 km from source), the PSFs for all three components show a similar overall pattern (Fig. 4). The pre-lahar PSF distribution in all three components is between 20 and 32 Hz. Like RTMT higher up the channel, the PSF for the front of the lahar at TRAN drops down to around 10 Hz, and when the lahar head arrives (∼10 min, Fig. 4) the PSF increases to ∼30 Hz for parallel (Fig. 4b) and cross-channel (Fig. 4a) directions and between 20 and 30 Hz in the vertical component (Fig. 4c). This decrease to lower frequencies before the head of the lahar at TRAN lasts for about 5 min. After the head of the lahar passes the recording station, the PSF content decreases for ∼ 15 min to 10–20 Hz for the parallel and cross-channel components and between 10 and 25 Hz for the vertical components. The PSF after the 30 min mark in Fig. 4 displays a bimodal pattern with frequencies between 10 and 35 Hz, with PSF time windows concentrating most at ∼30 Hz.

At the COLL recording station (83 km from source), the PSF distribution shows differing patterns for all three components (Fig. 5). The PSF in the cross-channel direction (Fig. 5a) depicts a bimodal pattern throughout with a strong lower concentration of time windows at ∼18 Hz and a higher PSF at ∼25 Hz. For the parallel component (Fig. 5b), the pre-lahar signal has a wide PSF range between 12 and 30 Hz. When the lahar arrives, the PSF becomes concentrated at ∼22 Hz for ∼8 min before transforming into a bimodal pattern similar to that of the cross-channel PSF, with frequencies between 20 and 30 Hz. In the vertical component (Fig. 5c), the pre-lahar PSF is scattered between 22 and 30 Hz, then as the front of the lahar passes the station the PSF stabilizes around 28 Hz for about 12 min. When the lahar head arrives, the PSF again transitions to more of a scattered pattern during the highest energy stage of the lahar (Fig. 5, 25–40 min).

When recording mass flows with three-component sensors, the directionality may be examined due to the sensor being able to record signals in the two horizontal directions. The directionality ratio allows for the determination of which horizontal component has stronger energy over the course of the recording window. This is possible because, in channel side deployments for mass flow monitoring systems, the sensor is either installed so that the north component is aligned to be parallel and the east component aligned as perpendicular to the flow, or the components are rotated during the data processing stage to align with the channel orientation. Furthermore, with the channel side installations, attenuational factors can mostly be ignored due to the close proximity to the channel and energy output of the flow event. The directionality ratio (DR) can be defined as the cross-channel amplitude divided by the flow-parallel amplitude. A DR > 1 indicates that the cross-channel amplitude is larger than that of the flow parallel, and vice versa for a DR <1. Directionality ratios have been used in the past to show rheology changes within flows, where the DR increases when streamflow transitions into a lahar (Walsh et al., 2020), and have been hypothesized to be an indicator for flow properties such as sediment concentration, wetted perimeter, and/or the number of particle collisions within a lahar (Doyle et al., 2010).

The directionality ratios estimated from 10 s non-overlapping running time windows of the rms amplitudes at each seismic station for the 18 March 2007 lake-breakout lahar are shown in Fig. 6. The DR for RTMT (Fig. 6a) displays a DR ≤1 (0.8–1.0) pre-lahar, then decreases (0.7–0.8) as the lahar arrives at the recording station (Fig. 6, dashed line), then as soon as the lahar head arrives, the DR increases to above DR =1 for ∼2 min. After the peak lahar flood pulse passes RTMT, the DR then proceeds to decrease below a DR =1 for the rest of the recording window. Similar to RTMT, the DR for TRAN starts out with a DR <1 (0.7–0.8) and as the lahar front passes, the DR similarly decreases to 0.6–0.7 before increasing to a DR >1 for ∼5 min when the lahar is at peak energy output starting at about the 10 min mark (Fig. 6d, red line). After the passing of the peak energy, the DR for TRAN decreases below 1 again for the remainder of the recording window. Further down the channel at COLL (Fig. 6c), the DR before the lahar arrives has a wide range of values between 0.8 and 1.2. When the front of the lahar passes (Fig. 6, dashed line), the DR stabilizes between 0.8 and 1, before increasing slightly when the peak energy of the lahar passes the monitoring site at about the 25 min mark.

In order to obtain an understanding of if PSFs are able to properly describe the lahar dynamics (i.e., the weight of the spectral amplitude at the PSF), frequency constraints must be analyzed. To complete this, normalized spectrograms along with spectral centroidal frequency (SCF) and spectral spreads are computed (e.g., Rubin et al., 2012; Saló et al., 2018). The normalized spectrograms are estimated by normalizing (using the maximum) the spectral amplitude for each 10 s time window of the lahar individually. By normalizing each time window, ranges of dominant frequencies can be visualized. SCFs are used because they represent the weighted average of the spectra and yield the location (i.e., frequency) of the center of the spectral mass. The SCF of each time window is estimated similar to that of Saló et al. (2018), in which 1SCF=∑f1f2f⋅A(f)∑f1f2A(f), where f is the frequency and A(f) is the spectral amplitude associated with each frequency bin. The spectral spread measures the width of the spectral energy around the SCF (i.e., standard deviation), thus yielding information about the quality of the PSFs (e.g., Rubin et al., 2012; Giannakopoulos and Pikrakis, 2014; Saló et al., 2018). Spectral spread can be estimated by 2SS=∑f1f2(f-SCF)2⋅A(f)∑f1f2A(f). The computed normalized spectrograms along with SCFs and spectral spreads for each of the three monitoring stations are shown in Fig. 7. For simplicity and comparison, only the flow-parallel data are shown. The normalized spectrograms for every station and component can be seen in Figs. S1–S3 in the Supplement, as well as the values in Table S1.

The normalized spectrogram for RTMT (Fig. 7a) yields very similar results to that of the PSF (Fig. 3b), where most of the higher spectral amplitudes are at the same frequencies as those of the PSF. Notably, the low ∼10 Hz signal immediately before the arrival of the head of the lahar is not only seen in the dominant normalized spectra, but also through the decrease in SCF. Additionally, the PSFs at these time windows are contained within the spectral spread (Fig. 7a, black lines). For TRAN, the normalized spectrogram (Fig. 7b) is, again, very similar to the PSF in Fig. 4b. The SCF mirrors the pattern of the PSF with higher frequencies for the streamflow, a decrease for the front of the lahar, increase for the head of the lahar, decrease after the passing of the head, and finally a slight increase later in the lahar body. The normalized spectra yield this same pattern, with the late lahar body displaying the only time frame with increased spectral amplitude distributed throughout the spectral spread (Fig. 7b, after 30 min). This most likely explains the bimodal distribution of PSFs for TRAN in Fig. 4 after the ∼30 min mark. Continuing, the normalized spectrogram for COLL (Fig. 7c) also shows similarities to that of the PSFs in Fig. 5b. The PSFs for COLL range between ∼ 20 and 30 Hz with a slight bimodal pattern. This same pattern can be seen where the higher spectral amplitudes are located (Fig. 7c). Furthermore, the SCF for COLL splits the PSF range and stays at ∼25 Hz during the bimodal phase of the PSF. Overall, with the analysis of the normalized spectrograms, SCFs, and spectral spreads, we confirm that the use of PSFs to describe mass flow dynamics is concise for the 18 March 2007 lake-breakout lahar.

The main goal of this research is to contribute to defining better monitoring criteria for dangerous mass flow events. The data described above are part of a larger collection of monitoring data collected over the entire length of the Whangaehu channel, consisting of 21 monitoring sites and years of preparation (e.g., Manville and Cronin, 2007; Keys and Green, 2008). Due to this, the ability to accurately estimate the properties of the lahar at various stages along its path is possible. When it comes to flow events of any size, the ability to understand how the dynamics change with distance along the channel is important for warning and future hazard mitigation. We show here that a lake-breakout event can start out as an outburst flood, bulk up into a hyperconcentrated flow, then eventually elongate and entrap enough sediment to transform into a plug-like slurry flow. Each of these flow types yields differing PSF ranges and patterns due to the relationship between the channel geometry, sediment concentration, turbulence, and bedload transport. While the lahar at different stations along the channel may have differing PSF content, we also show that the lahar elongates, and a predictable model (e.g., Cronin et al., 1999) can be used with and shown in the seismic data. Being able to apply such a model may yield some relevance of universality in terms of warning systems at different distances away from the mass flow source, whereas, shown above, the flow phases at each monitoring station can be seen, but at differing lengths and times in the seismic signal (e.g., Fig. 6). To better visualize this concept, conceptual models based off of the Cronin et al. (1999) models are created for each of the three seismic stations for the 18 March 2007 lahar (Fig. 9). In the conceptual models for the 2007 lahar, the aforementioned elongation of the frontal pulse or bow wave (phase 1) and head of the lahar (phase 2) is shown, along with the differences and similarities between the properties of the lahar at the three seismic monitoring sites.

Another implication for future warning is the implementation of three-component sensors and the use of DRs for channels that have streamflow. Walsh et al. (2020) showed that for lahars flowing in La Lumbre channel at Volcán de Colima the DR for streamflow is <1 and then increases when the head of the lahar arrives. This same feature can be seen at each of the three monitoring sites for the 18 March 2007 event (Fig. 6), indicating that differing flow types will still show this DR pattern within the same flow and at other channels. To further show this, there were three natural non-lake-breakout eruption-based lahars that occurred in the Whangaehu channel in September 2007 (for more details on the lahars see Cole et al., 2009; Kilgour et al., 2010) and recorded on the seismometer at RTMT. The DR for the September events starts with streamflow with a DR <1, and when the first lahar arrives the DR increases to >1 and as the lahar fully passes, the DR decreases to < 1 again (Fig. 10a). As the second lahar arrives at RTMT (Fig. 10, second dashed line), the DR increases to >1 again. After the second lahar passes the DR deceases once again back below DR <1. Finally, as the third lahar arrives (Fig. 10, third dashed line) the DR yet again increases above 1 for the entirety of the event.

For many mass flows and especially those that flow into channels with preexisting streamflow, the peak seismic amplitude does not always coincide with the arrival of the mass flow, and thus may not be the most reliable for event detection or warning (e.g., Arattano and Moia, 1999; Cole et al., 2009). These observations may be due to a frontal surge, the lag in sediment concentration, or differences in peak amplitude with peak discharge. Phase 1 (frontal streamflow surge) of the model proposed by Cronin et al. (1999) was based on a hyperconcentrated flow interacting with streamflow, but has also been shown for debris flows as well (e.g., Arattano and Moia, 1999). Arattano and Moia (1999) showed at Moscardo Torrent, Italy, through a hydrograph that there was a precursory surge ahead of the debris flow that was not seen in the seismic record. Similarly, at Ruapehu, for the 18 March 2007 lahar, at each of the three stations there is little evidence or rise in the seismic amplitude that would indicate that there was a precursory surge or phase 1 (Figs. 3–5, bottom panel), which could be problematic for detection methods that use amplitude thresholds or short-time-average vs. long-time-average (STA/LTA) algorithms. Conversely, the surge ahead of the lahar can be seen in both the PSF analysis (drop to low frequencies) and in the DR (decrease in DR) right before the peak seismic amplitude arrives. This shows that when monitoring for future events not only the amplitude should be used but other analysis (e.g., PSF, DR) as well, otherwise there could be a delay in the detection of an event.

Using all three components of the seismometer can be very beneficial in lahar monitoring. The above-mentioned DR analysis can only be completed with horizontal recording, and analyzing PSF in each component can yield critical information about the flow properties and dynamics. Examining the seismic amplitude differences can generate significant discoveries; for example, when the vertical component is stronger than the horizontal components, bedload may dominate over turbulence noise (Burtin et al., 2010). Greater flow-parallel signals may indicate higher water transport noises (Barriere et al., 2015), while higher cross-channel signals could be caused by increased interflow particle collisions and flow-channel wall interactions (Doyle et al., 2010). While using the differences in each component can be useful, there are also some concerns. Channel geometry and bed conditions can alter the seismic signal (e.g., Coviello et al., 2019; Marchetti et al., 2019). Additionally, the flow-parallel direction can be influenced by the lahar that has already passed, the lahar at the station, and the lahar arriving. Furthermore, the tilt of the seismometer may play a large role in determining which component is stronger (e.g., Anthony et al., 2018). In the case of the 18 March 2007 lahar, a large pulse of water passed the monitoring stations which may explain why the parallel component is stronger than the other two components at RTMT (Fig. 3d) and TRAN (Fig. 4d). At COLL, the lahar had elongated, lost energy, and thus showed more decreased flow-parallel energy compared to the previous two stations (Fig. 5d). In the cross-channel direction, if a flow overtops the channel, the amplitude would presumably be dampened. This may be the case at TRAN where both the flow-parallel and vertical directions are more energetic than the cross-channel amplitude after the passing of the head and breaking out of the channel occurred (Figs. 2d and 4d). Another concern when using the horizontal components of a seismometer is the effects shallow layers may have on the site response of the sensor. This is especially true when a sensor is installed on soft or loose sediment (e.g., soil, fluvial/alluvial deposits). To test for potential effects by shallow layer fundamental frequencies, H/V analysis of ambient noise (streamflow dominant) was conducted (see Supplement). For RTMT, the H/V results depict a broad-frequency peak between 5 and 15 Hz with a local maximum at ∼8 Hz (Fig. S4a). Comparing the H/V frequency with the PSF of RTMT (Fig. 3), the only overlap is when the front of the lahar passes the station where the PSF decreases for ∼1 min before the head of the lahar arrives. The H/V analysis for TRAN has a multi-broad-peak shape, with frequency peaks at ∼14 and ∼ 28–35 Hz (Fig. S4b). While these frequencies are similar to PSF values for TRAN (Fig. 4), the H/V analysis has no distinguishable fundamental frequency and contains large errors, and no frequency peak has a H/V amplification >2. In order for a H/V frequency peak to be considered ideal, generally the amplification must be greater than 2 and the standard deviation lower than a factor of 2 (SESAME, 2004). The H/V amplification for COLL displays a broad-frequency peak between 13 and 18 Hz, with a local maximum at ∼18 Hz (Fig. S4c). Comparing the PSFs at COLL (Fig. 5), only the cross-channel direction has significant PSF values in the same frequency range (∼18 Hz band). With all three stations not yielding distinct H/V fundamental frequencies, we surmise that the PSF content for the 18 March 2007 lake-breakout lahar is most likely dominated by the large flow passing by the seismic sensor rather than large site amplification effects from a shallow layer. While this may be the case, there is still the possibility that some of the PSF values could be due to local effects and should not be considered in the lahar analysis, e.g., the low PSFs at RTMT between 15 and 20 min (Fig. 3a and b), at TRAN contributing to some of the “jumping” in PSF content (Fig. 4a–c), or in the mostly dominant 15–20 Hz PSF in the cross-channel direction at COLL (Fig. 5a). Conversely, SCF values at each station do not reside in the broad H/V frequency range at any station (Fig. 7), which may further support the hypothesis that almost all of the recorded frequencies are indeed produced by the lahar. With the use of horizontal components becoming common in mass flow monitoring, future three-component analyses of mass flows should consider estimating H / V ratios or use other site response methods (e.g., spectral ratio analysis) in order to identify whether near-surface structures may affect the recorded flow data. Overall, all these concerns can and should be tested to estimate potential error in three-component methods. Nevertheless, using all three components of the seismometer can enhance the productivity of warning systems and, if possible, should be used instead of single-component sensors.

Finally, implementation of these new results into new or existing mass flow warning systems must be discussed. In an ideal setup, to remove any doubt about the recorded signal, machine learning techniques should be used to separate the mass flow noise from other non-flow noises (e.g., environmental, human induced, earthquakes). For instance, recently Wenner et al. (2021) used a supervised random forest algorithm to classify differing sources in a debris flow setting. Once the mass flow source has been classified, integrating automated DR and PSF analysis would be quick and straightforward. Implementation of these techniques would be similar to other seismic analysis or detection methods, such as a STA/LTA (e.g., Coviello et al., 2019) or a number of frequency detection algorithms (e.g., Rubin et al., 2012) where real-time analysis of set time windows is used to determine if there has been a change in the seismicity along the channel. The system could be programmed to identify changing features in the flow automatically by analyzing the content of each window, as well as comparing previous time windows. The analysis of continual data could then be fed into machine learning algorithms (e.g., Rubin et al., 2012; Wenner et al., 2021) to increase not only the confidence of detection, but also the characterization of flow behavior. One of the main discoveries of this research was the evolution of seismic signals produced by the lahar as the flow moved further from its source. The changes in the seismic signal along with the flow characteristics may be able to help hazard and forecast modeling through the use of numerical models (e.g., Mead et al., 2021). Modern flow hazard assessment is based on numerical models that use potential energy equations of a large non-changing mass sliding down slope with limited inputs for how the flow may evolve over time as starting inputs. This can lead to errors in risk mitigation and hazard assessments. The findings shown above on how the 18 March 2007 lahar evolved over 83 km at Mt. Ruapehu will help to improve mass flow modeling in the future by enabling modelers to add constraints or more inputs on how a mass flow might evolve, leading to improved forecasts and hazard assessment.