Development of a deep learning-based seismic waveform classifier to automatically detect snow avalanches.

Avalanche forecasters mostly rely on human observations of avalanche activity, but these reports are typically incomplete during periods of poor visibility, delayed, and not automated. Automated detection systems equipped with seismic sensors can improve monitoring efficiency, providing accurate avalanche data to support avalanche forecasting, regardless of visibility and weather conditions. Additionally, such systems could be implemented as early warning tools to enhance safety measures in mountain regions. While seismic detection systems have been widely used for avalanche monitoring, there is currently no automated method to reliably identify signals originating from avalanches. To address this, we developed a deep neural network to automatically detect avalanches in real-time continuous seismic recordings. This model was trained using seismic data collected over 13 winter seasons at the avalanche test site of Vallée de la Sionne, in Switzerland. Avalanches of varying sizes and paths are monitored using four seismometers strategically placed within and outside the avalanche path. The system simultaneously acquires continuous seismic data and event alarms. The alarms are based on amplitude thresholds recorded by the two seismometers near frequent release zones. While these alarms provide preliminary insights into avalanche activity, they require manual verification to filter out false alarms caused by events such as earthquakes or other unknown sources.

 

To overcome this limitation, we trained an end-to-end deep learning-based seismic waveform classifier on normalized, 40-second signal snippets extracted from the event database. The model architecture includes a convolutional encoder, a convolutional feature extractor with attention mechanisms, and a fully connected classification head. The network reliably distinguishes between avalanches and non-avalanche signals, achieving an accuracy of 0.97 on held-out events from the 2022/23 and 2023/24 winter seasons. The model was also deployed during the latest winter season to classify signals in near real-time, providing daily avalanche detection rates and demonstrating its feasibility as an automatic detection system. Finally, we plan to investigate the transferability of the classifier to seismic data collected by a distributed acoustic sensing (DAS) system installed on the avalanche test site, exploring its potential for broader applications.