Self-supervised learning for automated avalanche detection from seismic and infrasound data

Snow avalanches are among the deadliest natural hazards in mountainous regions. Yet avalanche activity is often still documented manually, and accurate avalanche release times are mostly missing. Automated monitoring systems equipped with seismic and infrasound sensors, combined with detection algorithms, could help record avalanche occurrences and provide accurate data on release time, size, and type. This comprehensive data on avalanche activity is indispensable for improving and validating avalanche forecasts and for implementing mitigation measures. At the Vallée de la Sionne (VDLS) test site in Switzerland, a combination of avalanche monitoring systems has been deployed for over two decades, including radars, cameras, seismic and infrasound stations. Additionally, avalanche researchers have manually documented and verified most avalanche events over the past 14 winter seasons to compile a unique avalanche catalogue.
To facilitate and automate avalanche detection, we aimed to implement two deep learning-based methods that scan continuous seismic and infrasound data separately in (near) real-time to detect and classify avalanche signals. Therefore, we leveraged the large volume of continuous data collected every winter at VDLS by adopting concepts from recent, powerful language models. Specifically, we pre-trained transformer networks in a self-supervised manner (i.e. without using expert labels) on a wide variety of signals mined from continuous seismic and infrasound data streams. The models receive fixed-length waveforms as input, partition them into sequential patches and compute patch-wise spectrograms. By training the models to reconstruct a portion of randomly masked patches, they learn to extract meaningful representations from the data, achieving silhouette scores of up to 0.6. This indicates good separability between avalanche and non-avalanche signals. Thus, these representations can later be used to automatically detect avalanches by fine-tuning a classifier on top. Moreover, combining predictions from the seismic and infrasound models has the potential to further improve (near) real-time avalanche detection.