Snow avalanches are a major natural hazard in mountainous areas in winter. Since most recreationists and many public authorities rely on avalanche bulletins for their decision making, the ability to appropriately predict the avalanche danger accurately is of paramount importance. Avalanche forecasters greatly rely on avalanche activity data to determine the avalanche hazard since recent avalanche occurrences provide unambiguous information on snow cover instability. Avalanche activity is usually estimated based on visual observations, which are incomplete and impossible at night or when visibility is limited. This often leads to uncertainties in the number and exact timing of avalanches.
Listening to the snow
Automatically detecting avalanches in near real-time and in any visibility is of great interest to avalanche forecasting. Since the mid 1970s, it has been known that seismic sensors are very well suited for the remote detection of snow avalanches. Seismic sensors record ground vibrations and are widely used to detect earthquakes. However, many things also create ground vibrations, such as explosions, airplanes, helicopters, and ski lifts. It is therefore very important to distinguish avalanches signals from other signals. To do so, we installed cameras to compare pictures from our field site with the seismic data. In this way we can 'learn' what typical characteristics seismic signals produces by avalanches have (Figure 2).
During the last decades, seismic instrumentation has greatly improved and now allows us to deploy very sensitive seismic sensors in alpine environments to continuously record ground vibrations. To monitor avalanche activity, we have therefore installed seismic sensors in or near avalanche start zones and we can detect ground vibrations produced by avalanches in a range of about 2 km2 (Figure 1). The range of the system strongly depends on the size and type of avalanches. Large avalanches can be detected when they are quite far away, whereas small avalanches cannot be detected, unless they pass directly above the sensors. To improve the range, we needed to reduce the signal to noise ratio. Therefore we decided to anchor the geophones to rocks (Fig. 4).
Automatic detection of avalanches
Since we have learned many of the characteristics of seismic signals produced by avalanches from the data we have collected over the last 4 years, we now intend to design an automatic detection system for avalanches. To this aim we will use a machine learning algorithm to build up a computer model which can decide if a signal was produced by an avalanche or not. First tests showed that such a model was able to correctly identify 16 avalanches in data collected during a period of intense wet-snow avalanche activity. Furthermore, the model did not misclassify any events, meaning that it did not identify events which were not avalanches (Figure 6). However, improvements are still required, since there were still many misclassifications when using the same model during a different period. Once the automatic avalanche classification model is reliable, our next goal will be to implement it in the remote field stations to obtain real-time avalanche activity data.
Localisation automatique des avalanches
Avec le système automatique de détection des avalanches décrit plus haut, il est possible de déterminer exactement quand une avalanche s’est déclenchée. Par contre, on ne sait pas encore localiser correctement les phénomènes. Pour obtenir là aussi de meilleures données, le SLF teste une autre disposition : les scientifiques ont groupé sept géophones en cercle, et ont utilisé des techniques de traitement du signal déjà appliquées par les sismologues. Mais comme les signaux des avalanches ne sont pas exactement similaires à ceux des séismes, ils doivent encore améliorer cette méthode pour les avalanches.