Automatic Monitoring of Seismogenic Slope Failure Activity at Brienz (Switzerland) Using Distributed Acoustic Sensing and Semi-Supervised Learning

Abstract

Listen

Translate article

Distributed Acoustic Sensing (DAS) represents a leap in seismic monitoring capabilities. Compared to traditional single-seismometer stations, DAS measures seismic strain at meter to sub-meter intervals along fiber-optic cables thus offering unprecedented temporal and spatial resolution. Leveraging the resolution of DAS enables us to monitor and detect seismogenic processes in the domain of hazardous mass-movements, including catastrophic rock avalanches. Here, we present a semi-supervised neural network algorithm for screening DAS data related to mass movements at the Brienz landslide in Eastern Switzerland, which partially failed on 15 June 2023. A DAS interrogator connected to a 10 km-long dark fiber provided by Swisscom Broadcast AG near the landslide recorded seismic data from 16 May to 30 June 2023, with a sampling frequency of 200 Hz and a channel spacing of 4m. During a test period from June 1 to June 19, 2023, a total of 634 characteristic waveforms potentially related to slope failures, including the 15 June 2023 event, were detected, along with vehicle and other anthropogenic noise sources with characteristic diurnal and weekday/weekend variations. For information extraction, we selected a subset of adjacent DAS channels, which include cable sections that were parallel to the failure event trajectory and thus particularly sensitive to mass movement activity. To facilitate efficient processing, we downsampled the data to 20 Hz, considering that slope failure events predominantly excite seismicity at below 10 Hz. We conceptualize the DAS data as a series of images representing consecutive strain rate data in the two dimensions of time and space. To bring out signal coherence between DAS channels, we transform the waveforms into cross-spectral density matrices (CSDM’s) which serve as the input image for unsupervised feature learning using an autoencoder (AE). Leveraging the features learned from the AE, we focus on activity classification using approximately 1500 samples. As ground truth for the slope failure class, we utilize concurrent Doppler radar data. The radar provides an event magnitude, which scales with failure volume and the number of individual rockfalls. Furthermore, the radar provides a measure of the moving mass’s trajectory length and front speed. The radar detected 516 slope failures during the test period. Our algorithm captures 41.09 % of the slope failures recorded by the Doppler radar. The undetected events mainly have low radar magnitudes suggesting that they are associated with mass movements generating reduced seismic activity. Among the slope failure-type signals detected by DAS, 87.85% are also present in the radar catalogue. Interference from vehicle or human-triggered seismic waves, deteriorating the signal-to-noise ratio significantly, poses a challenge for our algorithm to differentiate between slope failures and those activities. Our study thus provides a benchmark for future natural hazard monitoring and suggests that using existing fiber optic infrastructure has a high potential for early warning purposes.