Item: Automatic classification of continuous seismic data for avalanche monitoring purposes
-
-
Title: Automatic classification of continuous seismic data for avalanche monitoring purposes
Proceedings: International Snow Science Workshop Proceedings 2018, Innsbruck, Austria
Authors:
- Matthias Heck [ WSL Institute for Snow and Avalanche Research SLF, Davos, Switzerland ]
- Conny Hammer [ Swiss Seismological Service SED, ETH Zurich, Switzerland ]
- Manuel Hobiger [ Swiss Seismological Service SED, ETH Zurich, Switzerland ]
- Alec van Herwijnen [ WSL Institute for Snow and Avalanche Research SLF, Davos, Switzerland ]
- Jürg Schweizer [ WSL Institute for Snow and Avalanche Research SLF, Davos, Switzerland ]
- Donat Fäh [ Swiss Seismological Service SED, ETH Zurich, Switzerland ]
Date: 2018-10-07
Abstract: Seismic monitoring systems are well suited to detect avalanches independent of weather conditions. Nevertheless, seismic monitoring systems are not yet used operationally, as developing algorithms to automatically detect avalanches is far from trivial. Thus far, attempts to automatically identify avalanches in seismic data have focused on using machine learning algorithms with varying degrees of complexity, requiring extensive training data sets and generally resulting in rather high false alarm rates. Recently, a promising new approach was introduced using so-called hidden Markov models (HMMs), a statistical pattern recognition tool commonly used for speech recognition. With this method, the abundance of background noise data is exploited and only one training event is required. We adapted this method to automatically detect avalanches in data recorded by a small aperture seismic array deployed above Davos, Switzerland. While preliminary results were very encouraging, the number of false alarms remained rather high. To eliminate false detections, primarily produced by regional earthquakes or distant airplanes, we introduced a two-step approach to reduce the number of false alarms. First, using HMMs trained at a second array at a distance of 14 km, we compared the automatically detected events at both sites. Any co-detected events were removed. Second, for the remaining events, we used multiple signal classification (MUSIC), an array processing technique, to determine the back-azimuth and the apparent velocity of the incoming wave-fields to obtain information on the direction of the source of the events. In contrast to avalanches, falsely classified events had much larger changes in back-azimuth and could thus be dismissed. We applied this method on data recorded from January to April 2017 and automatically obtained an avalanche activity pattern in line with visual observations performed by the avalanche warning service in the area of Davos. Overall, our new classification approach shows that seismic monitoring systems can be used to automatically provide timely information of large avalanches occurring within a distance of 2-3 km.
Object ID: ISSW2018_O07.10.pdf
Language of Article: English
Presenter(s):
Keywords: Snow avalanche, avalanche forecasting, seismic monitoring, automatic detection, event localization
Page Number(s): 631-635
-