Towards the automatic detection of seismic and infrasound signals generated by snow avalanches
Examensarbete för masterexamen
Avalanches are the cause of many fatal accidents every winter in the European Alps. Established procedures to forecast the regional avalanche danger level use data on the weather, the snow cover and recent avalanche activity. Avalanche activity data is today obtained by visual observations. As avalanches generally occur during snowstorms when visibility is poor, these data are often incomplete in space and time. To solve this issue, arrays of infrasound or seismic sensors can be deployed to monitor avalanches at distances up to 4 kilometers, independent of visibility or time of day. Today, only infrasound systems are used operationally to automatically localize and detect avalanches. The main limitation of these systems is that only larger dry-snow avalanches can be detected, as it is generally assumed that wet-snow or smaller avalanches do not generate enough energetic infrasound signals. Since seismic signals are generated by a different process than infrasound signals, seismic systems can be used to detect both dry- and wet-snow avalanches. However, methods to automatically detect avalanches in near real-time have not been established yet. In this project, seismic and infrasound signals generated by avalanches from two winter seasons were investigated at a field site above Davos, Switzerland. With an array processing method, specifically a sliding window three dimensional beamforming algorithm, it was possible to localize the source of avalanches and track changes over time. The method worked for localizing avalanches from both infrasound and seismic data. It was found that avalanches generate seismic activity for a longer duration and that seismic sensors also record infrasound. Using a frequency-wavenumber-analysis to apply a velocity filter to separate seismic and infrasound wave fields did not further improve the accuracy of the localization method. To automatically detect avalanches, 18 parameters were derived based on the beamforming method and pattern recognition procedures. For the testing season (2016-2017), 55% of all visually confirmed avalanches were detected from the seismic data. The learning season (2017-2018) detected 100% of all visually confirmed avalanches from both seismic and infrasound data with a low ratio of detected unassigned events. These results clearly suggest that automatic detection of avalanches can also reliably be used on seismic data. Since the processing method is computationally efficient, it could be implemented for near realtime avalanche detection using seismic monitoring systems.
Infrasound; Seismic; signal processing; pattern recognition; automatic avalanche event detection; event localization; FK filter