← Back to Sessions

Can Machine Learning improve Debris Flow Warning?

Session: Critical Zone, Environmental and Cryospheric Seismology

Type: Oral

Date: 4/22/2021

Presentation Time: 05:15 PM Pacific

Description: 

Debris flows are complex mixtures of water, fragmented rock, and sediments moving rapidly down steep torrents. They are mobilized by intense precipitation and their volumes often exceed tens of thousands of cubic meters. Debris flows have a high destructive potential which is amplified at the flow front, where large boulders concentrate. The significant hazard to human life and infrastructure in alpine regions calls for reliable warning systems to reduce risk in vulnerable areas.

The basic assumption for most detection systems is that the debris flows move at moderate velocities (below 10 m/s) and cover several kilometer distances before they hit infrastructure. For that reason, rapid detection is indispensable to maximize warning times. Seismic monitoring has received particular attention in this regard as debris flows induce ground unrest over kilometer distances. However, automatic identification of debris flow signals in continuous seismic records remains a challenge.

Here, we introduce a machine-learning approach to detect debris flows and hazardous debris floods based on their seismic signature. For the Illgraben torrent, Switzerland, seismic records from an 8-station network allow detecting debris flows in the upper catchment area, which is practically inaccessible. A machine learning model based on the random forest algorithm recognizes early stages of debris flow formation with an accuracy exceeding 90%. Trained with 20 events from previous years, our detection algorithm detected all 13 hazardous torrential events with no false alarms during a three-months, realtime test phase in 2020. The independently confirmed detections include mostly debris flows, but also smaller floods with high water content. Our approach provides up to an additional hour of warning time to the earliest possible in-torrent detection. The proposed seismic machine-learning detector increases warning times using simpler instrumentation compared to existing operational systems, which is a major step towards the next generation of debris-flow warning systems.

Presenting Author: Małgorzata Chmiel

Student Presenter: No

Authors