Abstract. Snow avalanches generate seismic signals as many other mass movements. Detection of avalanches by seismic monitoring is highly relevant to assess avalanche danger. In contrast to other seismic events, signals generated by avalanches do not have a characteristic first arrival nor is it possible to detect different wave phases. In addition, the moving source character of avalanches increases the intricacy of the signals. Although it is possible to visually detect seismic signals produced by avalanches, reliable automatic detection methods for all types of avalanches do not exist yet. We therefore evaluate whether hidden Markov models (HMMs) are suitable for the automatic detection of avalanches in continuous seismic data. We analyzed data recorded during the winter season 2010 by a seismic array deployed in an avalanche starting zone above Davos, Switzerland. We re-evaluated a reference catalogue containing 385 events by grouping the events in seven probability classes. Since most of the data consist of noise, we first applied a simple amplitude threshold to reduce the amount of data. As first classification results were unsatisfying, we analyzed the temporal behavior of the seismic signals for the whole data set and found that there is a high variability in the seismic signals. We therefore applied further post-processing steps to reduce the number of false alarms by defining a minimal duration for the detected event, implementing a voting-based approach and analyzing the coherence of the detected events. We obtained the best classification results for events detected by at least five sensors and with a minimal duration of 12 s. These processing steps allowed identifying two periods of high avalanche activity, suggesting that HMMs are suitable for the automatic detection of avalanches in seismic data. However, our results also showed that more sensitive sensors and more appropriate sensor locations are needed to improve the signal-to-noise ratio of the signals and therefore the classification.

No abstract

Abstract. We used continuous data from a seismic monitoring system to automatically determine the avalanche activity at a remote field site above Davos, Switzerland. The approach is based on combining a machine learning algorithm with array processing techniques to provide an operational method capable of near real-time classification. First, we used a recently developed method based on hidden Markov models (HMMs) to automatically identify events in continuous seismic data using only a single training event. For the 2016–2017 winter period, this resulted in 117 events. Second, to eliminate falsely classified events such as airplanes and local earthquakes, we implemented an additional HMM-based classifier at a second array 14 km away. By cross-checking the results of both arrays, we reduced the number of classifications by about 50 %. In a third and final step we used multiple signal classification (MUSIC), an array processing technique, to determine the direction of the source. As snow avalanches recorded at our arrays typically generate signals with small changes in source direction, events with large changes were dismissed. From the 117 initially detected events during the 4-month period, our classification workflow removed 96 events. The majority of the remaining 21 events were on 9 and 10 March 2017, in line with visual avalanche observations in the Davos region. Our results suggest that the classification workflow presented could be used to identify major avalanche periods and highlight the importance of array processing techniques for the automatic classification of avalanches in seismic data.

Accurate avalanche occurrence data are of crucial importance for avalanche forecasting, since recent avalanching provides direct evidence on snowpack instability. We therefore explore how avalanche activity data obtained through seismic monitoring can be used for avalanche forecasting. By visually inspecting data from a seismic sensor deployed in an avalanche starting zone, we obtained three avalanche catalogues for two entire winters and one period of 10 days with intense wet-snow avalanche activity. Avalanche activity was clustered in time for all catalogues, and diurnal periodicity was clearly present during spring. In winter, when dry-snow avalanches predominantly release, rather weak long-term correlations on the order of several days were found between past and future avalanche activity. We investigated the performance of a simple model to predict future avalanches based on past avalanche activity. Model performance was better in spring than in winter, especially for very short time scales of up to three hours, and for time scales around

Abstract.We use a seismic monitoring system to automatically determine the avalanche activity at a remote field site near Davos, Switzerland. By using a recently developed approach based on hidden Markov models (HMMs), a machine learning algorithm, we were able to automatically identify avalanches in continuous seismic data by providing as little as one single training event. Furthermore, we implemented an operational method to provide near real-time classification results. For the 2016-2017 5 winter period 117 events were automatically identified. False classified events such as airplanes and local earthquakes were filtered using a new approach containing two additional classification steps. In a first step, we implemented a second HMM based classifier at a second array 14 km away to automatically identify airplanes and earthquakes. By cross-checking the results of both arrays we reduced the amount of false classifications by about 50 %. In a second step, we used multiple signal classifications (MUSIC), an array processing technique to determine the direction of the source. Although avalanche events 10 have a moving source character only small changes of the source direction are common for snow avalanches whereas false classifications had large changes in the source direction and were therefore dismissed. From the 117 initially detected events during the 4 month period we were able to identify 90 false classifications based on these two additional steps. The obtained avalanche activity based on the remaining 27 avalanche events was in line with visual observations performed in the area of Davos.