Automatic detection of alpine rockslides in continuous seismic data using hidden Markov models

Abstract

Data from continuously recording permanent seismic networks can contain information about rockslide occurrence and timing complementary to eyewitness observations and thus aid in construction of robust event catalogs. However, detecting infrequent rockslide signals within large volumes of continuous seismic waveform data remains challenging and often requires demanding manual intervention. We adapted an automatic classification method using hidden Markov models to detect rockslide signals in seismic data from two stations in central Switzerland. We first processed 21 known rockslides, with event volumes spanning 3 orders of magnitude and station event distances varying by 1 order of magnitude, which resulted in 13 and 19 successfully classified events at the two stations. Retraining the models to incorporate seismic noise from the day of the event improved the respective results to 16 and 19 successful classifications. The missed events generally had low signal-to-noise ratio and small to medium volumes. We then processed nearly 14 years of continuous seismic data from the same two stations to detect previously unknown events. After postprocessing, we classified 30 new events as rockslides, of which we could verify three through independent observation. In particular, the largest new event, with estimated volume of 500,000 m 3 , was not generally known within the Swiss landslide community, highlighting the importance of regional seismic data analysis even in densely populated mountainous regions. Our method can be easily implemented as part of existing earthquake monitoring systems, and with an average event detection rate of about two per month, manual verification would not significantly increase operational workload.

Automatic classification of endogenous landslide seismicity using the Random Forest supervised classifier

“…In these studies, the seismic events are classified manually after detection and rely on the personal experience of the human operator which can be subjective and time-consuming. Automatic classification methods have been developed for detecting the sources in volcanic areas [Langer et al, 2006;Curilem et al, 2009] to differentiate earthquakes and blasts [Fäh and Koch, 2002;Laasri et al, 2015] or for characterizing large rockslides [Dammeier et al, 2016]. For multiclass problems, many classifiers were used such as hidden Markov models (HMMs), artificial neural networks, and support vector machines (SVMs), mainly on a reduced number of seismic attributes [Curilem et al, 2009;Hibert et al, 2014;Ruano et al, 2014].…”

“…Recently, some studies [Beyreuther and Wassermann, 2008;Ruano et al, 2014;Quang et al, 2015] focused on the classification of continuous seismic records discriminating the background noise from the signal of interest. HMM was modified to detect one type of signal from few to one single example which is interesting for the detection of rare seismic sources [Hammer et al, 2012[Hammer et al, , 2013Dammeier et al, 2016]. However, the use of a unique seismic signal as reference lacks to capture the influence of the travel path effects on the waveform and the frequency content of the recorded signal [Hammer et al, 2013].…”

Real‐time detection, location, and characterization of rockslides using broadband regional seismic networks

“…The analysis of rockslide-induced seismicity may provide significant information about the source of mass movement volumes in the order of 10 7 m 3 , as well as about their dynamics [Ekström and Stark, 2013]. Despite a growing number of studies exploring the use of broadband seismic recordings to trace the evolution of landslide events in time and space, strategies for automatic detection and characterization within an operational and real-time framework are still challenging [Yamada et al, 2013;Hibert et al, 2014;Iverson et al, 2015;Dammeier et al, 2016]. Despite a growing number of studies exploring the use of broadband seismic recordings to trace the evolution of landslide events in time and space, strategies for automatic detection and characterization within an operational and real-time framework are still challenging [Yamada et al, 2013;Hibert et al, 2014;Iverson et al, 2015;Dammeier et al, 2016].…”

Single-block rockfall dynamics inferred from seismic signal analysis

“…Hence, this allows a seismic detection of the events that do not generate long-period seismic waves (e.g. Deparis et al, 2008;Helmstetter and Garambois, 2010;Dammeier et al, 2011Dammeier et al, , 2016Hibert et al, 2011Hibert et al, , 2014bClouard et al, 2013;Chen et al, 2013;Burtin et al, 2013;Tripolitsiotis et al, 2015;Zimmer and Sitar, 2015). The limitation of this approach is that high-frequency seismic waves are more prone to be influenced by propagation effects (attenuation, dispersion, scattering) and, more importantly, that the source of the high-frequency seismic waves associated with gravitational instabilities is not yet well understood.…”