Volume 51, Issue 19 e2024GL110672
Research Letter
Open Access

Automatic Monitoring of Rock-Slope Failures Using Distributed Acoustic Sensing and Semi-Supervised Learning

Jiahui Kang

Corresponding Author

Jiahui Kang

Swiss Federal Institute for Forest, Snow and Landscape Research, Zürich, Switzerland

Faculty of Geosciences and Environment, University of Lausanne, Lausanne, Switzerland

Correspondence to:

J. Kang,

[email protected]

Contribution: Conceptualization, Methodology, Validation, Formal analysis, Data curation

Search for more papers by this author
Fabian Walter

Fabian Walter

Swiss Federal Institute for Forest, Snow and Landscape Research, Zürich, Switzerland

Contribution: Conceptualization, Methodology, Validation, Formal analysis, Data curation, Supervision, Funding acquisition

Search for more papers by this author
Patrick Paitz

Patrick Paitz

Swiss Federal Institute for Forest, Snow and Landscape Research, Zürich, Switzerland

Contribution: Conceptualization, Methodology, Validation, Formal analysis, Data curation

Search for more papers by this author
Johannes Aichele

Johannes Aichele

Department of Earth Sciences, ETH Zürich, Zürich, Switzerland

Contribution: Methodology, Validation, Formal analysis, Data curation

Search for more papers by this author
Pascal Edme

Pascal Edme

Department of Earth Sciences, ETH Zürich, Zürich, Switzerland

Contribution: Methodology, Validation, Formal analysis, Data curation

Search for more papers by this author
Lorenz Meier

Lorenz Meier

Geopraevent AG, Zürich, Switzerland

Contribution: Methodology, Validation, Formal analysis, Data curation

Search for more papers by this author
Andreas Fichtner

Andreas Fichtner

Department of Earth Sciences, ETH Zürich, Zürich, Switzerland

Contribution: Methodology, Validation, Formal analysis, Data curation

Search for more papers by this author
First published: 28 September 2024

Abstract

Effective use of the wealth of information provided by Distributed Acoustic Sensing (DAS) for mass movement monitoring remains a challenge. We propose a semi-supervised neural network tailored to screen DAS data related to a series of rock collapses leading to a major failure of approximately 1.2 million m 3 ${\mathrm{m}}^{3}$ on 15 June 2023 in Brienz, Eastern Switzerland. Besides DAS, the dataset from 16 May to 30 June 2023 includes Doppler radar data for partially ground-truth labeling. The proposed algorithm is capable of distinguishing between rock-slope failures and background noise, including road and train traffic, with a detection precision of over 95 % $95\%$ . It identifies hundreds of precursory failures and shows sustained detection hours before and during the major collapse. Event size and signal-to-noise ratio (SNR) are the key performance dependencies. As a critical part of our algorithm operates unsupervised, we suggest that it is suitable for general monitoring of natural hazards.

Key Points

  • A semi-supervised neural network is developed for rock-slope failure monitoring with Distributed Acoustic Sensing at Brienz, Switzerland

  • Our model achieves over 95% precision for rock slope failures detected by a Doppler radar system over 45 days

  • The sustained detection of slope failures before the major collapse highlights the potential of our approach for early warning

Plain Language Summary

Steep mountains and hills produce dangerous rockfalls and similar phenomena such as landslides and debris flows. A major collapse is typically preceded by a series of rockfalls over days or months. It is therefore crucial to reliably detect these events and recognize the warning signs of an impending major collapse. When rocks bounce on the ground they release seismic waves, which generate vibrations that propagate long distances. Such vibrations stretch and compress fiber optic cables within the ground enough so they can be measured with a novel technique called Distributed Acoustic Sensing (DAS). Here we show how to identify such DAS signals using machine learning algorithms to detect precursory rockfall activity and a major collapse on a slope in Switzerland. We compare our detections with radar measurements, which are highly reliable but also come at a greater cost for installation. Since we can apply DAS to unused fiber within the ground, our approach may pave the way for the next generation of natural hazard warning.

1 Introduction

Detecting rockfalls, landslides, debris flows and other mass movements in alpine regions is essential for the protection of human lives, livestock and infrastructure. Seismic monitoring has emerged as a valuable tool for detecting events and characterizing their dynamics (Allstadt et al., 2018). Seismic techniques can distinguish between free-fall rockfalls and granular flows and deduce their loss of potential energy (Hibert et al., 2011). Signal metrics are related to fundamental mass movement properties (Dammeier et al., 2011) and low-frequency signals can be inverted to reconstruct trajectories (Ekström & Stark, 2013). Moreover, accelerated seismic event activity before major rock failure likely elucidates precursory slip episodes (Poli, 2017). These findings highlight the importance of seismic records for monitoring unstable slopes (Del Gaudio & Wasowski, 2011; Deparis et al., 2008; Larose et al., 2015; Mainsant et al., 2012; Provost et al., 2018), and for the study of major collapses (Allstadt, 2013; Hibert et al., 2017; M. Walter et al., 2012; F. Walter et al., 2020).

To boost seismic monitoring capabilities with denser sensor coverage, Distributed Acoustic Sensing (DAS) offers new perspectives. With this technique, fiber optic cables are turned into arrays of strain or strain rate sensors (Dou et al., 2017). A DAS interrogation unit injects light pulses into optical fibers and interprets the backscattered Rayleigh light to detect changes in the local axial strain along the fiber (Parker et al., 2014). DAS not only provides exceptional spatiotemporal resolution but also offers cost-effectiveness by using “dark fibers”, initially deployed for communication but currently unused (Lindsey & Martin, 2021). The DAS amplitude and phase responses remain stable over 17 octaves from 1 / 3000 $1/3000$ to 60 Hz (Lindsey et al., 2020; Paitz et al., 2020). Edme et al. (2023) show that a “dark fiber” can be used for near real-time avalanche detection and location along a 10 km mountain pass road. At the same time, processing DAS records in real-time or for post-event analysis remains challenging, especially since mass movement seismograms are often attenuated and emergent, thus evading traditional threshold-based detection (Provost et al., 2017).

Machine learning models have demonstrated success in detecting complicated seismic signals. Hammer et al. (2012) pioneered the use of Hidden Markov Models for volcanic monitoring using seismicity patterns with minimal training data. Chmiel et al. (2021) demonstrate the accuracy of supervised algorithms for debris flow detection via their seismic signatures. For DAS signals and to alleviate the dependence on training data, Paitz et al. (2023) propose an automatic detection algorithm using an unsupervised Bayesian Gaussian Mixture Model. A central machine learning task is to identify mass movement signals in the presence of background noise. For DAS records, Martin et al. (2018) and van den Ende et al. (2023) have proposed the suppression of coherent and incoherent noise, respectively. Nevertheless, effective processing of DAS data in noisy environments remains a challenge, especially along roads where mass movement signals interfere with seismic waves of anthropogenic origin, such as traffic.

In this study, we present a 45 day DAS monitoring of an unstable slope near the village of Brienz in Switzerland during summer 2023. The investigation period includes numerous rather small (tens to hundreds of cubic meters) slope failures and a major rock collapse adding up to a total volume of approximately 1.2 million m 3 ${\mathrm{m}}^{3}$ . Using semi-supervised machine learning to exploit the seismic signal coherence on DAS data, we compile an event catalog that compares favorably with a co-located Doppler radar. The proposed DAS-based algorithm achieves high detection precision and offers new perspectives for mass movement monitoring, especially in regions with pre-existing fiber-optic infrastructure.

2 Study Site and Data

Located above the village of Brienz in Switzerland's canton of Grisons, our study site comprises a dynamic landslide complex that is divided into lower (Landslide “Dorf”) and upper (Landslide “Berg”) sections (Figure 1a). Landslide Berg above the village encompasses a high steep slope extending about 3 6 o $3{6}^{\mathrm{o}}$ (Gojcic et al., 2021; Häusler et al., 2021) from an elevation of 1 800 ${1}^{\prime }800$ m to 1 200 ${1}^{\prime }200$ m a.s.l (Figi et al., 2022). It consists of talus-like sediment deposits, fragmented material and rock outcrops, which up to June 2023 included the “Insel” rock mass. Within the landslide Berg, terrain displacements increased from 0.6 m/year in October 2011 to 3.8 m/year in November 2020. Landslide Dorf hosts the village of Brienz and consists of the terrain between the foot of landslide Berg and the Albula River, about 800 m to the south. It is less inclined with a slope of about 1 0 o $1{0}^{\mathrm{o}}$ and also exhibits terrain displacement with velocities increasing from 0.2 m/year before November 2011 to 1.2 m/year by November 2020 (Häusler et al., 2021).

Details are in the caption following the image

Site and instrumentation overview. (a) Digital Color Orthophotomosaic of Brienz (Federal Office of Topography swisstopo, 2024). The DAS interrogator is positioned near the Albula River to the southwest of the village and the cable runs adjacent to the cantonal traffic road and a train line (Ch201 and Ch712). The DAS records capture both slope failures and anthropogenic signals such as roads and railway traffic. (b) After-collapse image (Keystone/Michael, 2023) (c) Fourier spectrogram and frequency decomposition of the seismic signal at Ch510 around the major collapse with the colormap representing the Power Spectral Density (PSD). (d) Radar data in range-time domain obtained by summing velocity bins for the same period, with the y-axis representing the distance to the Doppler radar (range) and the magnitude denoting the power of the summed velocity spectrum. For explanation of Doppler magnitude, see the main text. (e) Cartoon of combined recording of Doppler radar and DAS cable.

Since 2011, an extensive monitoring system has been installed including a Doppler radar to detect rapid mass movements, which triggers automatic road closures upon detection (Geopraevent, 2022; Kenner et al., 2022). In May 2023, additional acceleration within Landslide Berg ( > 100 ${ >} 100$ mm per day) was accompanied by increased activity in failure events and prompted the evacuation of the village on 12 May 2023 (Petley, 2023). Slope failures, characterized as structural failures resulting in rapid mass movements such as rockfalls or granular flows escalated in June with the displacement of the Insel part reaching up to 80 m/day, and led to the collapse of the Insel rock formation in the night between 15 and 16 June. This event left deposits on a village access road, but the absence of sunlight inhibited documentation of event numbers and progression of sub-events as well as their volumes (Albula/Alvra, G, 2023).

2.1 DAS Records of the Brienz Slope Failure 2023

From 16 May to 30 June 2023, we connected a Febus A1 interrogator to the end of a dark fiber linking the villages of Tiefencastel and Filisur. The 10 km-long fiber runs parallel to the Rhaetian Railway's UNESCO World Heritage line and Cantonal road along the Albula River (Figure 1a). The interrogator measures strain rate, averaged over a 8 m gauge length (Daley et al., 2016), at a sampling frequency of 200 Hz, and a channel spacing of 4 m.

We select the DAS channels closest to Landslide Berg from Ch201 to Ch712 (Figure 1a). This section is mainly oriented perpendicular to slope failure trajectories on Landslide Berg. Nevertheless, it also includes a short trajectory-parallel part providing higher sensitivity to slope failure for Rayleigh waves polarized in the radial direction (Kennett et al., 2024). DAS data are downsampled to 20 Hz for processing efficiency. DAS seismograms feature typical slope failure characteristics like emergent onsets, signal content above 1 Hz and no discernible seismic phases (Allstadt et al., 2018). One failure event at 21:38, 15 June 2023 [UTC] during the night of the Insel collapse (henceforth referred to as “major collapse”) exhibits particularly concentrated energy between 2 and 8 Hz (Figure 1c).

2.2 Doppler Radar Data

Doppler radars identify objects moving faster than approximately one meter per second (Meier et al., 2017), measuring the Doppler shift of transmitted and reflected signals to determine the velocity and range of moving targets (Carlà et al., 2024). These radars capture slope failures such as those encountered on Landslide Berg. The Doppler radar detections therefore serve as a partial ground truth and are provided by the company Geopraevent AG. Attached to the wall of the former school in Brienz (Figure 1a), the Doppler radar continuously monitors the rockfall area from a distance of approximately 1 km.

For data storage considerations, the raw radar data are only retained for approximately 10 min around a detection defined by the exceedance of a predefined threshold (Carlà et al., 2024). We process the range-Doppler matrix by integrating the velocity spectrum for each 3 s Doppler radar image and 3.75 m range bin. The resulting Doppler magnitude provides a measure of event size (Figure 1d). Higher magnitudes indicate a broader velocity spectrum or elevated amplitudes in selected bins, or both. In addition, the detection start and end times, along with three metrics describing event characteristics are retained. The latter are size category, length, and front speed. Size category is a key metric in analyzing the DAS detections below. It is empirically derived from comparisons with snow avalanches: events ranging from 10 to 100 m 3 ${\mathrm{m}}^{3}$ are categorized as less than 1, while events in the range of several tens of thousands to hundreds of thousands of cubic meters fall within size categories four to five (Geopraevent, unpublished information). During the monitoring period, the Doppler radar documented 727 slope failures. Whereas the documented events represent confirmed rock-slope failures, the radar system categorizes short-duration events lasting less than 5 s as noise and filters out some weak failures that do not exceed the predefined threshold. Consequently, the radar ground truth should be regarded as a partially labeled dataset rather than an absolute physical ground truth.

3 Methods

We treat seismic event detection as an image classification task. In view of real-time monitoring applications, we classify successive time windows rather than event-based records. Recognizing the pivotal role of data representation in the effectiveness of machine learning algorithms (Bengio et al., 2013), we design our detection algorithm as two interconnected tasks: unsupervised representation learning for feature extraction and dimensionality reduction, and supervised classification using a minimal number of labeled samples. We use Convolutional Neural Networks (CNNs) that are recognized as key tools for extracting image features (Alzubaidi et al., 2021). Specifically, we adopt autoencoder-based representation learning algorithms to map high-dimensional observations to a lower-dimensional representation space (Tschannen et al., 2018).

3.1 Cross-Spectral Density Matrices (CSDMs) and Labeling

3.1.1 CSDMs

To implement the seismic event detection as an image classification task, we segment the DAS data into a sequence of images, each with a resolution of 512 × 512 $512\times 512$ pixels. These images represent strain rate time series in both temporal and spatial dimensions, covering 25.6 s and 512 channels. To mitigate potential boundary effects, we introduce an overlap window of 25 % $25\%$ , resulting in 188,400 samples from 12:50:00, 16 May, to 04:36:51, 30 June 2023 [UTC]. The processed data are used to calculate the phase of the Cross-Spectral Density Matrices (CSDMs), denoted by ϕ CSDM ${\phi }_{\mathit{CSDM}}$ . Mathematically, the calculation of ϕ CSDM ${\phi }_{\mathit{CSDM}}$ between DAS channels x and y over frequency ω $\omega $ is expressed as (Rabiner et al., 1978):
ϕ CSDM ( ω ) = < X ( ω ) , Y ( ω ) > angle ${\phi }_{\mathit{CSDM}}(\omega )=< X(\omega ),Y{(\omega ) > }_{\mathit{angle}}$ (1)
where X ( ω ) = F ( x ( t ) ) $X(\omega )=F(x(t))$ and Y ( ω ) = F ( y ( t ) ) $Y(\omega )=F(y(t))$ represent the Fourier transforms from channels x and y, respectively, and < > angle ${< } {\cdot > }_{\mathit{angle}}$ denotes the complex phase of the outer product operation. Accordingly, ϕ CSDM ${\phi }_{\mathit{CSDM}}$ is an N × ${\times} $ N matrix, where N is the number of observation locations, in our case 512 channels.

We chose CSDMs after experiments with 2D waveforms, spectrograms, and f-k analysis failed to extract sufficient representative features for robust classification. They either contain too much noise and detail or lack discernible features. Compared to other data representations, ϕ CSDM ${\phi }_{\mathit{CSDM}}$ has the advantage of amplifying the coherence between channels (Camacho et al., 2009). It therefore contains information on seismic waves traveling along the DAS cable with specific apparent velocities. This allows the distinction between different types of coherent signals and non-coherent signals.

Seismic waves induced by slope failure experience scattering as they propagate through the shallow, heterogeneous subsurface with topographical variations (Wu & Aki, 1988). Much of this scattering occurs outside the terrain covered by the DAS cable, which as a result is struck by incident seismic waves from different directions. This induces complex coherence patterns along the entire cable section. In contrast, the surface waves emitted by quasi-point sources of road and train traffic only experience the local ground structure with fewer scatterers and propagate with reduced interruption toward the cable. Incoherent random noise appears with little coherence between channels. The ϕ CSDM ${\phi }_{\mathit{CSDM}}$ values are averaged over three empirically chosen frequency bands: 2 4 $2-4$ Hz, 4 6 $4-6$ Hz, and 6 8 $6-8$ Hz.

3.1.2 Labeling

Following the calculation of ϕ CSDM ${\phi }_{\mathit{CSDM}}$ , we labeled a data subset. From the radar catalog, a subset of 385 samples were selected manually by prioritizing those DAS records with minimal noise and the most distinctive coherence features. In addition, 1,106 DAS signals unrelated to slope failure were chosen. Approximately 1,500 slots were labeled into the following four DAS signal classes (Figures 2a–2d).
  1. Vehicle noise C 0 $C0$ : Cars or trains which generate surface waves propagating with little interruption toward the cable.

  2. Random noise C 1 $C1$ : Random seismic noise with minimal or no inter-channel coherence.

  3. Slope Failure C 2 $C2$ : Rock-slope failure generating complex coherent signals between channels.

  4. Narrow-band noise C 3 $C3$ : Additional noise signals with predominant energy around 4 H z $4Hz$ . This may be related to instrumental, installation or structural site effects.

Details are in the caption following the image

Labeling and neural network architecture: (a)  C 0 $C0$ , (b)  C 1 $C1$ , (c)  C 2 $C2$ , (d)  C 3 $C3$ . Each image includes raw records (top row), and phase of CSDMs (second to fourth row). The different textures in the CSDM images distinguish coherent signals ( C 0 $C0$ and C 2 $C2$ ) from those with little or no inter-channel coherence ( C 1 $C1$ and C 3 $C3$ ). All values of the raw data and ϕ CSDM ${\phi }_{\mathit{CSDM}}$ are normalized. Note the linear, diagonal coherence pattern especially in the 2 4 $2-4$ Hz range of the vehicle class (C0) indicating propagation of surface waves. (e) The architecture of the proposed semi-supervised algorithm. Top: Feature extraction using the VQ-VAE network, repeated for 7 selected sub-images per slot. Bottom: Classification process trained with a minimal labeled dataset, outputting the class type and corresponding probability for each time window.

3.2 Semi-Supervised Classification

Figure 2e depicts the architecture of the proposed algorithm. To enhance the efficiency of representation learning, we start by optimizing feature extraction. Each image is divided into 64 × 64 $64\ \times \ 64$ segments and selecting seven diagonally adjacent grids out of 64, which allows the algorithm to focus on critical features and reduce feature dimensionality (see Text S1 in Supporting Information S1 for implementation details).

We applied VQ-VAE (van den Oord et al., 2017) which has demonstrated its effectiveness in handling complex image representations. Its architecture is shown in Figure 2e. Consider the i t h ${i}^{th}$ input phase image x i ${x}_{i}$ . The encoder network transforms x i ${x}_{i}$ into a posterior distribution q z | x i $q\left(z\vert {x}_{i}\right)$ which encapsulates the continuous latent variable z e x i ${z}_{e}\left({x}_{i}\right)$ . The distribution is assumed to follow a normal distribution: q z | x i = N μ en , σ en I $q\left(z\vert {x}_{i}\right)=\mathcal{N}\left({\mu }_{\text{en}},{\sigma }_{\text{en}}I\right)$ where μ en ${\mu }_{\text{en}}$ and σ en ${\sigma }_{\text{en}}$ denote the mean and standard deviation parameters obtained from the encoder and I $I$ is the identity matrix. The continuous latent representation z e x i ${z}_{e}\left({x}_{i}\right)$ is then quantized to discrete codes using a codebook E $\mathcal{E}$ . A codebook is a set of vectors e j ${e}_{j}$ that act as the discrete representation: E = e 1 , e 2 , e k $\mathcal{E}={e}_{1},{e}_{2},\text{\ldots }{e}_{k}$ . This quantization process involves mapping z e x n ${z}_{e}\left({x}_{n}\right)$ to the nearest neighbor in the discrete codebook, yielding the discretized representation z q x i = e k ${z}_{q}\left({x}_{i}\right)={e}_{k}$ , where k = arg min j z e x i e j 2 $k=\mathrm{arg}{\min }_{j}\Vert {z}_{e}\left({x}_{i}\right)-{e}_{j}{\Vert }_{2}$ . The discretized representation is then used for classification.

In preparation for classification within a high-dimensional feature space, data augmentation was employed to avoid overfitting (Shorten & Khoshgoftaar, 2019). We use Extreme Gradient Boosting (XGBoost) (Chen & Guestrin, 2016) for classification. Both representation learning and classification training processes utilize data from 1 June to 14 June. The classification model achieves a precision of 95.6 % $95.6\%$ when tested on the labeled dataset on and after 15 June ( 25 % $25\%$ of the labeled dataset). Finally, we apply the trained classifier to the entire dataset's feature space, yielding the probability for each class and determining the final class type. Further details are described in Text S1 in Supporting Information S1.

4 Classification Results

As expected, the daytime period between 05:00 and 19:00 [UTC] (07:00 to 21:00 local time) is characterized by the predominance of vehicle noise (Figure 3a). Our algorithm does not differentiate between cars and trains, as both act as quasi-point sources moving at nearly constant velocity along the cable while generating seismic waves with the same coherence patterns. In contrast, during night times, the random noise class is dominant.

Details are in the caption following the image

Detection results: (a) Overview of DAS detections, with dates on y-axis and time of day on x-axis. Each detection represents a time window of 25.6 s. Event classes are color-coded (legend). (b) Detection results for slope failure incorporating DAS and radar data by binary values (1 for detection, 0 for non-detection), with cumulative detection numbers on a secondary y-axis. The blue boxes represent a highlighted detection period observed between 20:00 and 23:59 [UTC] on 15 June, shown in an enlarged view above. (c) Heatmap illustrating grouped detection rate, with x-axis representing RMS noise level quantile (step size 0.1) and y-axis indicating size category (step size 0.1). As the noise level increases, the detection rate decreases. (d) Scatter plots of detected time slots, with y-axis representing time and signal duration (DUR), the peak value of the strain rate envelope (ePSR) and the strain rate envelope area (EA) on x-axis. The major collapse is denoted by a red dot. (e) Probabilistic power spectral density (PPSD) of radar records. The major collapse, distinguished by its unique distribution and extremely low occurrence probability, is encircled by the cyan lines.

Within this noise background, our algorithm identifies 993 time windows with slope failure class C 2 $C2$ probabilities above 0.7. Due to our temporal resolution limited by the 25.6 s window length, several DAS detections may be associated with a single radar detection. To facilitate comparison between DAS and radar detections, we segment radar detections into corresponding DAS segments based on their start and end times. Cumulative numbers are calculated with each detection representing a duration of 25.6 s. The detection results from 12 June to 18 June 2023 reveal a similar trend in detection for both DAS and radar data (Figure 3b, see Figure S2 in Supporting Information S1 for the entire period). There is a notable increase in detection numbers toward 15 June 2023, particularly evident in DAS data. During this period, sustained detection (detection break less than 5 min) is observed from 20:43 to 22:21 [UTC]. After this period, slope failure detection rate decreases substantially from 127 to only 15 detections per hour.

For quantitative analysis of the results, we introduce a criterion for overlapping instead of window-wise comparison. This adjustment is made because while the trend in detection numbers remains consistent across both techniques, the documented warning times for radar and DAS may not align precisely. All DAS detections on 15 June are considered hits due to continuous activities throughout the day, of which the radar only documents a subset. Apart from 15 June, a hit is defined when DAS time slots fall within the radar detection time window, with a 1 min tolerance. This results in 896 ( P 1 ) $(P1)$ hits out of 993 ( S 1 ) $(S1)$ where P 1 $P1$ denotes DAS detections matching radar detections and S 1 $S1$ is total DAS detections. Among the remaining 97 detections, 54 events are confirmed manually in the DAS records as slope failure missed by the Doppler radar. We define the detection precision as P = P 1 S 1 $P=\frac{P1}{S1}$ which reaches 95.7 % $95.7\%$ including events not captured by radar. In contrast, the number of total radar detections ( S 2 $S2$ , 727) and the number of radar detections matching DAS detections ( P 2 $P2$ , 295) give a DAS detection rate R = P 2 S 2 = 40.6 % $R=\frac{P2}{S2}=40.6\%$ (see Table S2 in Supporting Information S1).

During the monitoring period, DAS also captures 10 local and regional earthquakes (distance to epicenter between 4.52 and 218.67 km) of which five are classified as slope failure and the other five as vehicle noise. Each earthquake is divided into several 25.6 s time windows and only those containing a substantial separation between P- and S-wave arrivals are classified as slope failure. This indicates increased scattering effects experienced by P-waves (Biondi et al., 2017). We furthermore reviewed the signals of two teleseismic earthquakes (M7.7 and M7.1 earthquakes southeast of the Loyalty Islands on 19 May 2023 and 20 May 2023). These events show high-frequency energy concentration within the 2–4 Hz band and are classified as vehicle noise. Examples of classified earthquakes are shown in Figure S3 in Supporting Information S1.

Whereas the detection metrics are good performance indicators, there is no absolute ground truth. Instead, radar data are an approximation of the ground truth. The precision at 95.7 % $95.7\%$ indicates the algorithm's proficiency to correctly identify slope failure mainly by avoiding false positives.

4.1 Event Size and Noise Level

Given a detection rate of 40.6 % $40.6\%$ , we examine the events our algorithm missed. Approximately 74 % $74\%$ of the missed events are classified as vehicle noise ( C 0 ) $(C0)$ , while 17 % $17\%$ are labeled as random noise ( C 1 ) $(C1)$ . We next compare the DAS detection rate of different event size categories indicated by radar detection and noise levels by averaging seismic power spectral density (Peterson, 1993) within 2–4 Hz over 30 min intervals.

The detection rate increases with event size category (Figure 3c). Events with size categories over 3.3 exhibit a detection rate above 90 % $90\%$ across all noise levels. We attribute this correlation to the effect of seismic background noise. This background noise is more often exceeded by large failures as the power spectral density of seismic signals grows with the mass of individual blocks and the total impact rate (Dammeier et al., 2011; Hibert et al., 2011; Tsai et al., 2012). Figure S4b in Supporting Information S1 illustrates a typical weak failure event classified as C 1 $C1$ .

Inbal et al. (2018) demonstrate that transient signals generated by vehicles act as moving seismic sources, that may obscure tectonic signals. In our study, this phenomenon complicates the differentiation between vehicle-induced and slope failure-induced signals, as it disrupts the coherence pattern across channels as indicated in Figure S4a in Supporting Information S1. Traffic noise generally contributes to seismic power, thereby elevating the noise level (Riahi & Gerstoft, 2015). In our case, reduced noise levels are associated with an increased detection rate (Figure 3c). This explains the 74 % $74\%$ of “missed” detections classified as C 0 $C0$ .

Having predominantly included strong events in the labeled dataset, we trained another classification model using an additional dataset comprising approximately 40 weaker slope failure events (size category < 3.3 ${< } 3.3$ ) to explore potential improvements in detection rate. The results show only a marginal 1.9 % p t $1.9\%pt$ . increase in detection rate, with a 10 % p t $10\%pt$ . decrease in precision (see Figure S5 in Supporting Information S1 and Table S3 in Supporting Information S1). This suggests that the primary factors influencing the detection rate are event size and traffic noise, rather than the method itself. Additionally, we revised the frequency bands to 1 3 $1-3$ Hz, 3 5 $3-5$ Hz, and 5 7 $5-7$ Hz, isolating narrow-band ( C 3 ) $(C3)$ noise signals. The results, shown in Table S4 in Supporting Information S1, yielded a precision of 95.9 % $95.9\%$ and a detection rate of 40.4 % $40.4\%$ , indicating minimal impact from C 3 $C3$ noise and stable detection performance across different bandwidth configurations. In order to favor generalizability independent of local noise signals, we retain the original frequency bands.

4.2 Major Collapse Characteristics

The major collapse distinguishes itself by the strongest low-frequency signals between 0.01 and 0.03 Hz across multiple channels of the trajectory-parallel section of the cable (Figure 1c). In the context of mass movements, such low-frequency signals are typically associated with the forces generated by the acceleration and deceleration of a flowing/liquefied granular mass (Hibert et al., 2014). The exact nature of such signals should be the subject of future investigations.

We investigate the seismic signal duration (DUR), the peak value of the strain rate envelope (ePSR) and the strain rate envelope area (EA) (Dammeier et al., 2011) (detailed in Text S2 in Supporting Information S1). The major collapse is the highest-ranking event across all three metrics and stands out for ePSR and EA (Figure 3d). We explain this by an exceptionally large failing rock volume compared to the other events. Since the seismic power spectral density scales with a power of grain size (Zhang et al., 2021), the major collapse likely involves particularly large blocks. The elevated Doppler magnitude spectrum extends from the top of the slope to the road, also suggesting an exceptionally large displacement volume as well as a long trajectory (Figure 3e).

5 Discussion and Conclusions

Machine learning enhances rock-slope failure detection by capitalizing on the spatiotemporal information within DAS records. Our algorithm delivers a measure of slope failure activity with a clear increase of precursory events before the major collapse. Its primary limitation is the low detection rate for smaller events whereas larger rockfalls are detected with higher reliability. Given the widespread existence of accelerated rockfall activity before catastrophic collapses (Intrieri et al., 2019; Rosser et al., 2007; Royán et al., 2015), our approach is an attractive solution for monitoring or even early warning within hours.

At the heart of our approach lies the classification of CSDMs for contiguous DAS records which elucidate signal correlation between channels. The classification algorithm reacts to signal coherence between channels with the possibility of phase shifts generated by finite wave velocities. Consequently, the algorithm distinguishes the two key classes of vehicular traffic and slope failures based on how they emit seismic waves and thus generate different coherence patterns along the cable.

Geological settings and site effects modify seismic waves, which is the reason why conventional seismic detection algorithms based on single-station records cannot be easily transferred from one region of interest to another. In contrast, we argue that point-like vehicle sources and more distant slope failures generate seismic waves, which can be distinguished qualitatively by their coherence along the interrogated cable. Our learned features of cross-spectral density matrices should therefore also be useable at other sites, especially in urban environments, where dark fibers typically follow multiple orientations along streets and offer varying sensitivities to rock failure signals. The only condition is that spatial scaling between train track/road-cable distance, mass movement-cable distance, and cable length is similar. Some modifications may be necessary as, for example, additional noise classes may require the definition of additional classes. On the other hand, there exist more signal features, like amplitude information, which we discarded in our case, but which could be leveraged for fine-tuning the algorithm when applied to a new site.

Acknowledgments

We are thankful to Davide Piciucco for his help in fieldwork. We also thank Małgorzata Chmiel and Léonard Seydoux for the insightful discussion. Special thanks to Nayan Gurung for the beautiful cartoon design. We also appreciate the thorough and constructive review comments provided by Thomas Poulet. Swisscom Broadcast AG generously provided the fiber-optic cable, with Anton Poschung as the contact person. Our project has received funding from Horizon Europe MSCA Doctoral Network EnvSeis under grant agreement No. 101073148.

    Data Availability Statement

    The DAS dataset (including both raw data and cross-spectral density matrices), radar detection files, extracted features, labeled dataset, two trained models (feature extraction model and XGBoost classification model), along with scripts to reproduce the entire training and classification processes, and a notebook to replicate the result analysis, are provided at https://www.doi.org/10.16904/envidat.541. The training process was performed in Python (v3.11.4), using PyTorch (Paszke et al., 2019), pytorch-lightning (Falcon & The PyTorch Lightning team, 2019), Numpy (Harris et al., 2020), xgboost (Chen & Guestrin, 2016) and jupyter (Kluyver et al., 2016).