Coherent Source Subsampling of Seismic Noise for Distributed Acoustic Sensing in the Swiss Alps

Distributed Acoustic Sensing (DAS) provides dense measurements of seismic noise along fiber-optic cables and offers new opportunities for subsurface characterization. In environments where controlled sources are unavailable, conventional noise interferometry workflows for DAS construct virtual shot gathers via cross-correlation and average them over long time windows to obtain coherent surface waves for dispersion analysis and subsequent shear-wave velocity (V_s) inversion. In noise-based interferometric imaging, the distribution of noise sources controls the quality of the retrieved interstation response. In practice, seismic sources are highly anisotropic and intermittent, and so simply averaging all available time windows produces interferometric responses that are difficult to interpret and lead to unstable dispersion curves and biased V_s estimates. We present a data-driven coherent source subsampling (CSS) framework that automatically identifies and selects the time windows of seismic noise that contribute constructively to the physically interpretable interstation response.

We demonstrate the method using DAS data acquired along 30 km of pre-existing telecommunication fiber deployed by the Swiss Federal Railways (SBB) in a major alpine valley floor, recorded with a Sintela interrogator at 3 m channel spacing with 6 m gauge length. Our objective is to recover stable Rayleigh-wave dispersion curves and a shallow V_s structure in the upper 50 m. The fiber runs along the railway track in surface cable ducts, providing a realistic test bed with complex ambient noise, including car traffic, factories, quarry blasts, in addition to the train-generated signals. Subsampling strategies based on prior knowledge of the sources, such as train schedules or velocity-based filtering, can partly mitigate this problem. However, these strategies are tedious, strongly location-dependent along the fiber, and do not guarantee that the retained windows contribute coherently to the interstation response of the segment under investigation.

Here, we use a symmetric variational autoencoder (SymVAE) to perform coherent source subsampling. Trained on virtual shot gathers from multiple time windows, the SymVAE groups windows according to the similarity of their correlation wavefields and enables the selection of those windows that consistently exhibit symmetric surface-wave contributions on both the causal and acausal sides. Averaging only these subsampled windows yields interstation responses that are substantially denoised and symmetric. We interpret these cleaner and symmetric cross-correlations as being associated with the stationary-phase contributions for the fiber segment under investigation. The same framework also identifies fiber segments that lack coherent, dispersive Rayleigh waves, indicating where robust subsurface imaging is not feasible.

Applying CSS to the SBB DAS data produces stable Rayleigh-wave dispersion curves along the cable, which we invert for two-dimensional V_s profiles. Although demonstrated here on railway-generated noise, the proposed CSS framework can be extended to any uncontrolled settings, such as road-traffic-dominated areas, where source variability and non-uniformity may be even more severe.

• ¹Centre for Earth Sciences, Indian Institute of Science, Bangalore, India
• ²Department of Earth and Planetary Sciences, ETH Zurich, 8092 Zurich, Switzerland
• ³ SBB CFF FFS