Artificial intelligence for avalanche warning

At the WSL Institute for Snow and Avalanche Research SLF, man and machine now work hand in hand. Both expert reports are consolidated in the avalanche bulletin, which the institute publishes daily on behalf of the federal government. First, three avalanche forecasters independently produce a conventional, expert-based forecast of the avalanche danger for the following day. They use current information on the development and forecast of the weather, data from automatic measuring stations, but also from observers in the mountain areas and feedback from mountain guides. They assign these findings to one of five hazard levels by region. Only then do they consult the computer's automatically generated forecast, which is based on machine learning (ML). He was taught to make such forecasts by SLF researchers who work closely with the Swiss Data Science Center (SDSC). Their research has been immediately taken up by the operational warning team.

"A forecast on my phone about the snow slope I'm cruising down? Now that would be a fascinating vision of the future!" - JÜRG SCHWEIZER, SLF MANAGER

During the first winter, an SLF forecaster had gained initial experience with the numerical forecast model. Since the winter of 2021/2022, the automatically generated avalanche danger forecast has been available to the entire warning team. Using measurement and model data, the ML method is used to generate a human-independent prediction. Data from automatic measuring stations, the SNOWPACK numerical snowpack model and COSMO, the weather forecast model of MeteoSwiss, are combined to forecast the danger level for a dry avalanche situation for the next 24 hours.

Automatic forecasts

In the meantime, SLF researchers have further developed the ML-based computer model and expanded the range of models. In the winter of 2022 / 2023, this will make it possible for the first time to generate automatic forecasts for wet snow avalanches and the stability of the snowpack. "Thanks to the cooperation with SDSC, after only 18 months from the start of the project, the model chain was semi-operational and we were able to test the automated forecasts. The results were promising," emphasizes project manager Jürg Schweizer. The model forecasts are subsequently incorporated into the man-made forecasts. "We hope that the digitally and automatically generated forecasts will improve the consistency of our warnings. They are a valuable independent second opinion," states SLF head Schweizer.

"In very large avalanches, we take drone images to capture the outline of the downslope." - JÜRG SCHWEIZER, SLF MANAGER

The higher the avalanche danger, the larger the avalanches to be expected. In addition, they point to a critical situation. But information on current avalanches is hardly available in real time. Detection systems based on seismic or infrasound close this gap. To filter out the avalanche signals, ML comes into play. In the meantime, initial tests are also underway at the SLF to obtain information on the spatial distribution or size of avalanches using satellite images or images from drones. These remote sensing methods have the great advantage that they record avalanche activity over a wide area, ideally for the entire Swiss Alps. Remote sensing data is collected in avalanche warning and is also used to validate and improve models, such as for avalanche dynamic models in hazard zoning. "For very large avalanches, we take drone images to capture the outline of the downslope, calculate avalanche volume and mass balance."

Maps thanks to Deep Learning

Artificial intelligence (AI) is also found in other areas of WSL research. In cooperation with the ETH Zurich and financed by the Federal Offices for the Environment and Roads, WSL developed a system that automatically detects the narrow-leaved ragwort and the tree of gods along freeways. Both species are on the "black list" of invasive neophytes in Switzerland. A custom-developed, Deep Learning (DL)-based, pinpoint method was used for the mapping, which involved filming from a moving car. This approach provided well-reproducible, high-spatial-resolution distribution maps that are also being produced for other species.

Modern digital possibilities in environmental research make it possible to obtain a greater abundance and breadth of data in an automated way and to evaluate them scientifically in a more targeted manner. WSL's use of these not only supports its reputation as a leader in modern environmental research, but is also due to the interdisciplinary complexity of the issues surrounding climate change.