29.06.23 - Deep Snow: Improving Forecasts by supporting Avalanche Experts with data driven methods

Avalanche forecasting in Switzerland is still mainly carried out by highly qualified and trained experts. They have to examine a large amount of relevant data to carry out this task once or twice a day. This heavy task is therefore prone to bias and potential error. Yet its accuracy is crucial not only for winter activities, but also for land-use planning and the management of areas at risk from natural hazards.

This is why the Swiss Data Science Center and Eidg. Forschungsanstalt WSLhave launched the "DeepSnow: Improving snow avalanche forecasting by data-driven automated predictions" project. Their goal is to support human experts in determining the danger level.
 
This task is very difficult to translate into a machine learning model, because the danger level is defined as an ordinal scale that does not depend directly on the physical properties of the snowpack. The experience and interpretation of forecasters are important in assigning a value. This leads to variability in the past data needed to form the model. In addition, there is a significant imbalance in the number of events in each danger level.

Two different models were trained, the first for dry-snow avalanches and the second for wet-snow avalanches. The accuracy of the models for dry-snow avalanches was around 75%. This corresponds to the estimated rate of agreement between the forecasts and the nowcasts of well-trained observers. For wet-snow avalanches, the Spearman correlation coefficient between model forecasts and avalanche activity over a test dataset was 0.71.

Both models went through operational pre-testing during one winter season and showed promising results as decision support tools. Further research is done to improve further the model and prepare to a possible deployment.

#avalanche #forecasting #snow