AI provides debris flow warnings

Measurements at the Illgraben test site

For their study, the scientists selected Illgraben in the canton of Valais. On the steep slopes of the upper catchment area, boulders and sediment often detach and are deposited in the channel. Episodically, these deposits are mobilized as debris and mud avalanches and move for two to three kilometres through the deeply cut Illgraben gorge. They then reach the main Rhone valley and cover another two to three kilometres before reaching the Rhone river. WSL has operated an observatory with measuring stations at Illgraben for more than 20 years in order to study the formation and movement of debris flows and to determine their mass, density and speed.

In the 1960s, the lower Illgraben channel was stabilized and secured through the installation of check dams to keep the debris flows in the channel without endangering the surrounding area. Since various hiking trails run close to or cross the channel, an early warning system has been in operation since 2007.

This system is based on sensors within the channel bed, including geophones, radar and laser devices as well as video cameras. Although the equipment reliably records passing debris flows, it can only be used in the lower section of the valley where the torrent is accessible. This restricts the warning time to just a few minutes.

The result: of the 13 debris flows and smaller flood events that occurred at Illgraben in the summer of 2020, the AI​ algorithm reliably detected every single one, without generating any false alarms. "The algorithm recorded even the first seismic signals of debris flows high up in the catchment," says Walter. At Illgraben, this increased the warning time by at least 20 minutes compared with existing detection systems. "This is a huge improvement," he says.  

Generalist or specialist?

The study provided evidence that debris flows can be detected at an early stage using seismic data and machine learning. Illgraben offers an ideal natural laboratory for this and the method works well there. However, it requires a large set of debris flow signals to train the algorithm. "Such training data are almost never available at other sites," admits the specialist in seismic mass movements.

It is still unclear to what extent the detector trained at Illgraben can also detect debris flows in other catchment areas. The scientists want to further develop the algorithm so that it can function with less or perhaps even with no location-​specific training data.

Collaboration in the early detection of natural hazards

The ambitions of the researchers go even further. The innovative detector is the first milestone in a higher-​level project conducted by WSL and Swisscom Broadcast. The research collaboration, in which Walter's group is also significantly involved, aims to improve the monitoring of mass movements in Alpine regions. To this end, Swisscom Broadcast is developing a platform that brings together data streams from various sources, and analyses them in real time to identify natural hazards at an early stage.

The natural hazards platform is currently fed primarily by Walter's seismic sensors and seismographs from the Swiss Seismological Service. The researchers are working on integrating other relevant data sources in the future – from precipitation data and permafrost measurements to seismic monitoring using fibre optic cables and even numerous Internet of Things sensors. "In order to process such huge amounts of data, you need big data approaches and intelligent algorithms," says Walter. The AI detector for debris flows is the first step in this direction.