Artificial intelligence and avalanche warning
Human experts will continue to produce the avalanche bulletin in future, but computer assessments based on weather data and other measurements have a valuable role to play as a “second opinion”.
The avalanche warning service has been using computer-generated avalanche danger assessments produced by artificial intelligence on the basis of measured and weather data in its operations since last winter (2021/22). The way in which the model predicts avalanche danger levels is described in the magazine DIAGONAL 1/22 and, most recently, in an academic paper published in the the journal Natural Hazards and Earth System Science.
Every day, the three SLF avalanche forecasters on duty are responsible for generating the avalanche bulletin. In the mornings they gather information: How has the weather changed? What do the outlooks say? What are the observers reporting, and what feedback is being received from mountain guides and backcountry tourers? Each of the three forecasters independently assesses the regional danger levels and the terrain types that are particularly affected – and thus generates a forecast for the following day. In the daily 3 pm briefing they consolidate their assessments (for more detail).
Starting last winter, they then consult the computer-generated appraisal based on artificial intelligence. In many cases it matches the forecasters’ own assessments, but sometimes there are variants. As forecaster Frank Techel comments, “The computer analyses the data in a different way than we do. That’s why it occasionally arrives at a slightly different conclusion.” Techel and his colleagues respond to the computer’s “variant” opinion by critically re-examining and, if necessary, revising their consolidated assessment. “Yes, that does happen,” he says, “the computer forecast is very useful especially for drawing exact boundaries between regions in which different danger levels prevail.” In view of their findings, the forecasters will continue to use the model.
The computer model adopted last winter is suitable only for predicting dry-snow avalanches, but the SLF’s researchers have been hard at work. Based on the same machine-learning methods, models for wet-snow avalanches and snowpack stability are now available as well. These will be subjected to operational trials next winter.
