DEAPSnow | AI for avalanche forecasting

“DEAPSnow”, a project by the Swiss Data Science Center (SDSC) and the WSL Institute for Snow and Avalanche Research SLF, developed an Artificial Intelligence (AI) to support avalanche forecasters in creating the avalanche bulletin. This product provides essential information about the prevailing snow and avalanche conditions in the Swiss Alps and the Jura.

Snow avalanches rank among the deadliest natural hazards in the mountainous regions of Switzerland. In the winter season 2022/23, 230 people were affected by avalanches, resulting in 23 fatalities, which is consistent with the yearly average. Coupled with the increasing popularity of mountain sports, this underscores the vital importance of accurate avalanche forecasts with a high temporal and spatial resolution. As of today, avalanche forecasting is still an expert-based and time-consuming decision-making process. It involves the evaluation and interpretation of a variety of data describing weather and snow conditions from which expected avalanche conditions are inferred. Over the past years, it has been shown that AI models are capable of efficiently processing large amounts of data, and hence established itself as an indispensable tool in various fields.

The development of AI models for avalanche forecasting proved challenging, as avalanche formation is a complex interaction of terrain features, meteorological conditions, snowpack characteristics, and triggering mechanisms. Furthermore, avalanche danger can not be measured in the field. The project pursued two main approaches: one for the direct prediction of danger levels and another for the prediction of the probability of avalanche activity. These approaches utilized meteorological data, simulated measurements, avalanche catalogs, and avalanche forecasts from the past twenty years. 

‍

It is always fascinating to push what AI can do for Environmental and Earth Sciences, and in particular for natural hazards. Solutions are often far from trivial, and require specific adjustments of standard methods – Dr. Michele Volpi, Lead Data Scientist at the SDSC

‍

In the past three winter seasons, the developed AI models were used operationally by the SLF in Davos to compare with the traditional human-made forecasts. While these models have not yet attained human-level predictive accuracy, they do deliver consistent outcomes when exposed to identical inputs. Furthermore, they exhibit distinct error patterns compared to humans, making them an apt choice for evaluating human predictions from a different perspective, even if the question of how to best integrate AI into the forecasting process remains open. Nonetheless, the project represents a significant advancement in avalanche forecasting and lays the foundation for more precise and timely predictions that can enhance safety in mountainous regions.

‍

The AI model may not yet quite reach the level of human predictions, but it delivers consistent results with the same input and makes different errors than humans. This makes it well suitable for questioning human predictions from a different perspective. – Dr. Frank Techel, Avalanche forecaster at the SLF

‍

Further reading / Bibliography:

• Pérez-Guillén, C., Techel, F., Hendrick, M., Volpi, M., van Herwijnen, A., Olevski, T., & Schweizer, J. (2022). Data-driven automated predictions of the avalanche danger level for dry-snow conditions in Switzerland. Natural Hazards and Earth System Sciences, 22(6), 2031-2056. https://doi.org/10.5194/nhess-22-2031-2022
• Hendrick, M., Techel, F., Volpi, M., Olevski, T., Pérez-Guillén, C., van Herwijnen, A., & Schweizer, J. (2023). Automated prediction of wet-snow avalanche activity in the Swiss Alps. Journal of Glaciology (14 pp.). https://doi.org/10.1017/jog.2023.24