Detecting and discriminating rockfall signals from anthropogenic noise using Distributed Acoustic Sensing along a Norwegian railway

Rørstadbotnen, Robin André; Hunnestad, Kasper; Truoung, Khanh; Landrø, Martin; Eidsvik, Jo; Pettersen, Aadne Ingvald Austigar; Brannfjell, Rune

doi:10.5194/egusphere-2026-4460

Preprints

https://doi.org/10.5194/egusphere-2026-4460

Preprints

21 Aug 2026

| 21 Aug 2026

Status: this preprint is open for discussion and under review for Natural Hazards and Earth System Sciences (NHESS).

Detecting and discriminating rockfall signals from anthropogenic noise using Distributed Acoustic Sensing along a Norwegian railway

Robin André Rørstadbotnen, Kasper Hunnestad, Khanh Truoung, Martin Landrø, Jo Eidsvik, Aadne Ingvald Austigar Pettersen, and Rune Brannfjell

Abstract. In mountainous regions, railway lines frequently traverse terrain with elevated rockfall and avalanche risk. In recent years, several such events have struck railways, resulting in casualties and significant infrastructure damage. Continuous monitoring along railway corridors is therefore essential for early warning and preventive measures. Given that fibre-optic telecommunications cables are already installed along many railways, Distributed Acoustic Sensing (DAS) represents an operationally attractive solution for continuous, large-scale rockfall monitoring. Avalanches and rockfalls release seismic energy as they propagate, which can be recorded by seismic sensors installed along the railway. Here, we use a fibre-optic cable running parallel to a railway in Norway to acquire DAS data and characterise the seismic signatures of rockfall events. The DAS system also recorded a range of anthropogenic signals, including those from road vehicles and trains, and the frequency distribution are found to be distinctly different from rockfalls. Rockfall signals are characterised by higher dominant frequencies and apparent propagation velocities than anthropogenic noise sources. Furthermore, moving vehicles generate persistent low-frequency energy that is absent in rockfall signals, providing an additional basis for discrimination. In addition to natural rockfall events, we carried out controlled rockfall experiments in which rocks of various sizes were dropped along the railway. The smallest rockfall detected had a mass of approximately 50 kg and was observable over approximately 400 m of fibre, whereas the largest boulder could be observed over several kilometres. Combined, these characteristics indicate that rockfalls can be effectively discriminated from anthropogenic signals, forming a promising foundation for the development of automated detection algorithms using cost-efficient DAS system.

Received: 24 Jul 2026 – Discussion started: 21 Aug 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Robin André Rørstadbotnen, Kasper Hunnestad, Khanh Truoung, Martin Landrø, Jo Eidsvik, Aadne Ingvald Austigar Pettersen, and Rune Brannfjell

Status: open (until 02 Oct 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2026-4460', Anonymous Referee #1, 01 Sep 2026 reply
The manuscript presents an interesting application of Distributed Acoustic Sensing (DAS) to rockfall-induced seismic events and anthropogenic events. The study applied processing and analysing approaches on both experimental and observed rockfall events. In general, I think this study has the potential for publication and is relevant to NHESS. However, several important aspects require further clarification, some key information needs to be added to the manuscript, and some interpretations require further discussion and explanation.
My comments are listed below:
Given the long distance of fibre (~98km), it would be useful to have additional information on the DAS acquisition. For example, I was wondering about the optical setup, and whether the optical amplification or repeaters were used. It would be also very useful to clarify if the signal quality varies with distance from the interrogator. For example, did the authors check the signal-to-noise ratio or sensitivity along different sections of the fibre, especially near the far end of the 98 km fibre? If this variability exists, could the detectability of rockfalls and other events be affected?

Lines 83-84: it is unclear to me how the tap tests were used to identify the sections with higher avalanche risk. What observable metrics were derived from the tap test and used for this assessment?

How were the geological effects from the near surface materials distinguished from the effects from fibre coupling and fibre orientation, etc.? How confident can we be that the main cause of the spatial variation in recorded amplitude is near-surface geological factors?

The subsection “data processing” is too short. A lot of methodological details are missing from this paragraph. Was the low-pass filter applied before the downsampling (for avoiding aliasing)? Why were the 5-70 Hz and 0.15-2 Hz frequency bands selected for the events of avalanches and traffic? Are these frequency bands based on spectral characteristics observed in the data or previous studies? What is the Butterworth filter order and did you apply the zero-phase filtering? Please better specify the preprocessing procedure and organize the section.

The hyperbolic-fitting workflow would need a more complete description. Why is the 5-50 Hz band used here but not the 5-70 Hz used in the general processing? It would be helpful to indicate how these processing parameters were selected.

It is not always clear which processing steps and criteria are applied to which types of events (e.g., controlled cases, natural rockfall, avalanches, or other detected signals). Some of this important information appears at different points throughout the text, which makes the overall processing workflow difficult to follow. I therefore suggest reorganizing the part of “Methods and signal classification”.

The reduced surface-waves amplitudes observed along the tunnel sections are attributed to the absence of surficial deposits beneath the fibre (around Line 140). However, DAS amplitudes can also be strongly influenced by fibre coupling, depth, orientation, and other installation conditions. Could the authors discuss these alternative factors and provide more explanation?

It is unclear for me how the predicted curves in Figure 7b and 7d were calculated and how were they compared with the observations quantitatively. For example, how was the amplitude scaling factor (A0) determined for the prediction curves? How was the best-fitting function identified (e.g., least-squares fitting)?

Lines 232-235: It may require more justification for the link between the unclear seismic phases in the f-k plot of the nature rockfall to the near-surface geology. In particular, natural rockfalls may represent more complex and spatially distributed sources than the controlled experiment through multiple impacts, rolling, etc. In addition, spatial variations in DAS coupling or sensitivity can also influence the f-k representation. Could the authors provide more evidence or explanation to support near-surface geology as the dominant cause?

The apparent velocities reported in the manuscript appear to be plausible, however, the procedure used to estimate them (and also their associated uncertainties) requires further clarification. In the manuscript, it is mentioned that the hyperbolas were manually overlaid on the data to determine the apparent velocities (~Line 263). However, how was the best-fitting velocity selected? What uncertainty is associated with these choices?

It is also mentioned in the manuscript that the difference between the fitted velocity and the manually estimated velocity for the controlled single rock-drop experiment (Table 1) reveals a limitation of the fitting procedure, despite the very low MAE. Could uncertainty estimates, pick-consistency criteria, or other quality metrics for the inferred velocities be provided to help strengthen the analysis?

Line 88: it may help if the author can provide the actual length of the Robel maintenance vehicle.

Figure 2: was the normalization or scaling applied when plotting? Please clarify the quantity plotted here.

Figure 10: it is difficult to interpret the figure in the current form and combination. It is difficult for me to understand why the events in panels (a) and (d) are presented differently from those in panels (b), (c), (e) and (f). Please mention the reason for using different representations, or simply presenting them in a more consistent way. Furthermore, additional information is needed to better understand the figure. Please clarify whether the quantity has been normalized, what is “c” in panels (b,c,e,f) and “d” in panels (a,d), what is represented by the colorbar at the right of (c) and (f), and what is “Value” 1 to 0.

A more detailed caption would improve the readability of the figure.

Several figure captions do not provide sufficient information to interpret the figures independently. I recommend that the authors revise the captions throughout the manuscript.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-4460-RC1
RC2:
'Comment on egusphere-2026-4460', Anonymous Referee #2, 21 Sep 2026 reply
Recommendation: Reject
The manuscript does not demonstrate its central claim: reliable discrimination of rockfall signals from anthropogenic noise. It presents selected recordings from two natural events and two controlled experiments, together with selected traffic examples, but provides neither a validated discrimination procedure nor sufficient experimental information to reproduce the analysis. More seriously, its principal quantitative comparison mixes vehicle translation speeds with seismic-wave propagation speeds. The processing also depends on the event class already being known.
The topic is relevant to NHESS. However, the manuscript does not establish a transferable result for hazard monitoring or provide the acquisition documentation and methodological validation needed by the DAS community. The concerns below affect the scientific argument itself. Addressing them would require substantial reanalysis and validation, rather than clarification and editing alone. I therefore recommend rejection.

The central velocity comparison measures different physical quantities (Section 3.2; lines 259–274; Table 1).

The automatic fits return 24–25 m/s for cars and trucks, corresponding to vehicle movement, and 1240–2642 m/s for rockfalls, interpreted as seismic propagation velocities. The manuscript explicitly acknowledges this distinction, yet subsequently presents the approximately two-order-of-magnitude separation as evidence supporting discrimination. This is not a comparison of seismic-wave velocities between source classes.
Using the manually assigned traffic-wave velocity of 430 m/s instead gives ratios of approximately 2.9–6.1 relative to the automatic rockfall estimates. Those comparisons still require consistent phase identification and control of site effects. Source speed and wave speed might both be useful features, but they must be defined, extracted, and evaluated separately.
The controlled single drop also yields 1240 m/s automatically against 2200 m/s by manual overlay, despite a normalized MAE of 0.01. The manual value is not independent ground truth, but this large disagreement demonstrates that a small fitting residual does not establish a correct wavefront or velocity. The conclusion that rockfall velocities exceed 2 km/s then contradicts the automatic result in Table 1.

The fitting procedure uses the event label before the claimed discrimination and lacks uncertainty analysis (lines 105–130).

The fractional-change operation is applied specifically to rockfalls, and their fits receive onset-time and source-offset constraints derived from prior knowledge. Vehicle signals are treated differently. How would an unknown event be processed without first knowing its class? A workflow conditioned on known labels cannot establish classification performance unless a procedure for selecting among those alternatives is defined and independently tested.
The assertion that independent channel picking makes the method insensitive to coupling and geology is also unjustified. These factors affect waveform shape and signal-to-noise ratio, not merely an overall amplitude scale. Huber loss cannot correct systematic selection of later arrivals or mixed phases.
Furthermore, Equation (1) assumes an appropriate source–receiver geometry and an effective constant velocity. Its applicability to curved fibre and heterogeneous terrain must be assessed. Report fitted positions and offsets, parameter uncertainties, and errors against the known controlled-source locations. Normalizing MAE by the analysis-window duration does not supply this information and makes scores from different windows difficult to compare.

Reliable discrimination is asserted without being tested, while the most relevant counterexamples are deferred (lines 58–64 and 275–309).

The manuscript gives no explicit decision rule, independent test set, false-alarm rate over representative continuous data, or assessment of missed events. A long deployment does not compensate for a small, selectively presented event sample. State how much data were examined, how events were selected, and how their labels were independently established. Photographs of deposited rocks do not by themselves establish the timing and identity of the selected DAS signals.
Most importantly, lines 294–300 acknowledge unexplained impulsive signals with rockfall-like hyperbolic moveout, but leave their characterization to future work. These are precisely the negative examples needed to evaluate the central claim. Comparing rockfalls mainly with prolonged vehicle passages avoids the difficult discrimination problem.
Nor does the presence of multiple impacts establish a unique rockfall signature. Repeated anthropogenic impacts must be considered. Restricting detection to mapped steep slopes cannot replace signal validation: anthropogenic events also occur near slopes, and spatial exclusion can hide missed hazards. Include these confounding signals in the analysis and quantify the resulting errors.

Acquisition, source characterization, and processing are insufficiently documented (Sections 2.1, 2.2, and 3.1).

Naming OptoDAS is not a reproducible acquisition description. Provide the manufacturer and configuration, channel spacing, relevant instrument response, and an event-by-event record of gauge length and sampling rate. Explain when and why 8 or 16 m gauges and 400, 800, or 1000 Hz sampling were used. Section 3.1 describes only resampling from 800 to 200 Hz, leaving the other acquisition settings unresolved.
The fibre installation is equally important: cable construction, burial or duct/tray placement, rail and road offsets, tunnel mounting, slack loops, and coupling conditions are inadequately described. These omissions directly affect the amplitude, velocity, and geological interpretations.
“Sledgehammer shots” sounds more like refreshments on the maintenance journey than a reproducible source description. Specify whether these were manual hammer blows or mechanically generated impacts, and document the equipment, impact surface, source position, stop spacing, and number of strikes. A stopping vehicle is not an impulsive tap; explain its separate role in channel mapping. The quoted ±50 m and ±25 m uncertainties also require a coordinate-survey method and an actual error assessment.
Finally, provide the phase-to-strain conversion, integration treatment, anti-alias filtering, filter order and phase response, tapering, and f–k normalization. These details are necessary to evaluate the reported results.

The spectral comparisons do not separate source characteristics from measurement and processing effects (Figures 5–11).

Changing gauge length changes the spatial response and can affect slow and fast apparent waves differently at the same frequency. This is directly relevant when traffic and rockfall spectra are compared. Without assigning acquisition settings to each event and accounting for their response, the observed contrast cannot simply be attributed to source type.
Figure 11 also uses different colour scales and visibly different low-frequency resolutions. Report the time windows, spatial apertures, spectral normalization, decibel reference, and effective resolution. A short impact window cannot be compared with a long vehicle passage without addressing those differences. Explain whether spatial-median subtraction preserves the low-frequency, low-wavenumber physical signals used in the interpretation.
The manuscript further conflates absence of a moving-source ridge with absence of low-frequency energy. These are different claims; low-frequency power is visible in the rockfall panels. The reported velocity ranges also overlap: vehicle-associated energy extends to 2.2 km/s, while natural rockfall energy extends down to 0.5 km/s. Quantify distributions and overlap rather than infer distinct classes from selected images.

Geological interpretation and hazard screening are unsupported, and the NGU citation is inadequate (lines 83–95, 133–150, and 316–320).

The cited NGU map link does not identify the actual evidence used for the manuscript’s claims. It provides neither a reproducible account of the hazard delineation nor sufficient support for local deposit thickness and fibre–ground contact. Identify the exact layers or map extracts, scale, version, relevant units, and surveyed locations. Show how these data support each inference. For hazard screening, provide explicit slope/runout criteria and an independent inventory. For cable coupling, provide installation evidence. A generic map link cannot substitute for either.
The promised tap-test characterization of sensitivity versus geology is not presented. Qualitative tunnel/open-section amplitude differences do not fulfil that objective. A tunnel cable is not necessarily directly coupled to bedrock, and lower recorded amplitude does not demonstrate the absence of surface waves or trapped energy. Installation, orientation, coupling, and geology remain confounded.
The conclusion nevertheless attributes a clear geological correlation to the tap-test. That claim requires results which the manuscript does not show.

The physical interpretation of directional sensitivity and wave types is not established (lines 247–258).

The statement that incidence near 90° produces maximum axial-strain sensitivity requires an explicit angle definition and derivation. If measured relative to the fibre axis, this is inconsistent with the usual longitudinal plane-wave response, which varies as cos²θ. Rayleigh-wave sensitivity additionally depends on polarization and propagation direction. Source offset alone does not establish the claimed amplitude relationship.
Likewise, apparent velocities of approximately 1.2–2.6 km/s do not, by themselves, identify body waves. Short duration and small rock size do not demonstrate negligible surface-wave generation. Support the phase assignments with independent observations, an appropriate velocity model, or forward modelling.
The claim that rockfalls do not move along the fibre is also an unjustified generalization. Impact positions evolve downslope, with a geometry-dependent projection onto fibre chainage. Conversely, passing vehicles can generate localized impulses at fixed features. These possibilities must be tested before stationary versus moving signatures are treated as reliable class distinctions.

Detectability and warning claims exceed the demonstrated evidence (lines 285–309 and conclusions).

The smallest observed rock is not a minimum detectable mass. Its recorded footprint depends on impact energy, substrate, offset, coupling, processing, and background noise. Define a reproducible detection criterion before comparing the reported extents and durations. Kilometres of responding fibre are also not equivalent to kilometres of stand-off detection range.
Train-related saturation or masking is a central operational limitation. Establish whether the observed problem is actual instrument saturation, phase-retrieval failure, or signal masking, and quantify its spatial extent, duration, and frequency. The monitoring system’s blind intervals occur when trains are present; this cannot simply be deferred while effective railway monitoring is advertised.
Finally, distinguish detection after material reaches the track from detection before impact or prediction of failure. No demonstrated latency, location accuracy, or actionable warning time supports the present claims. The data and code needed to assess these results should be available during review, not promised only after acceptance.
The manuscript establishes that selected rockfall events were recorded on existing railway fibre. It does not establish reliable discrimination, validated localization, or an operational warning capability. Because the principal quantitative argument is inconsistent and the decisive validation is missing, I recommend rejection.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-4460-RC2

Robin André Rørstadbotnen, Kasper Hunnestad, Khanh Truoung, Martin Landrø, Jo Eidsvik, Aadne Ingvald Austigar Pettersen, and Rune Brannfjell

Viewed

Total article views: 341 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
172	96	73	341	81	74

HTML: 172
PDF: 96
XML: 73
Total: 341
BibTeX: 81
EndNote: 74

Views and downloads (calculated since 21 Aug 2026)

Month	HTML	PDF	XML	Total
Aug 2026	44	17	10	71
Sep 2026	126	74	61	261
Oct 2026	2	5	2	9

Cumulative views and downloads (calculated since 21 Aug 2026)

Month	HTML	PDF	XML	Total
Aug 2026	44	17	10	71
Sep 2026	126	74	61	261
Oct 2026	2	5	2	9

Viewed (geographical distribution)

Total article views: 331 (including HTML, PDF, and XML) Thereof 331 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 02 Oct 2026

Short summary

Railway lines in mountainous regions face serious risks from rockfalls, which can derail trains and damage infrastructure. We used fibre cables along a Norwegian railway to detect rockfalls using Distributed Acoustic Sensing, a technology turning cables into thousands of sensors. Comparing rockfall signals with those from vehicles and trains revealed differences in frequency and wave speed. Rocks as small as 50 kilograms were detected, laying the groundwork for future automated warning systems.


Total:	0
HTML:	0
PDF:	0
XML:	0