Combining seismic-based and infrasound-based machine learning models to enhance automatic avalanche detection

Summarise with VIRAT
Authors: Yannik Zeller
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Current avalanche activity data are a crucial source of information for avalanche forecasting. Traditionally, such data are obtained through visual observations, which are limited under poor weather conditions or during the night. To overcome these limitations, seismic and infrasound detection systems can be employed. However, avalanche-related seismic and infrasound signals are highly variable and often obscured by strong background noise, making automatic detection challenging. Previous attempts to implement automatic avalanche detection using traditional machine learning algorithms have not achieved operationally reliable results. In this study, eight existing convolutional autoencoder models, four trained on seismic and four on infrasound data, which achieved strong performance on an event-based dataset but performed poorly on continuous data, were optimized for application to continuous data. The seismic and infrasound data used to train and test the models were acquired at two operational avalanche monitoring test sites in Switzerland: the Vallée de la Sionne test site near Sion and the Dischma test site in Davos. The objective was to reduce false alarms while maintaining high avalanche detection rates. Therefore, two postprocessing strategies which combine the seismic and infrasound model outputs were developed: (1) a Bayesian optimization approach that optimized causal time aggregation (up to 10 seconds), model aggregation, and thresholding; and (2) a metaclassifier approach, in which potential avalanche events identified by a single model were further classified using a random forest model based on statistical and temporal features derived from the output probabilities of all eight models. For both test sites, the metaclassifier outperformed the Bayesian optimization approach, and both postprocessing strategies substantially improved performance relative to all individual baseline models. The best seismic baseline model achieved recalls of 0.50 and 0.90 with corresponding precisions of 0.032 and 0.005 at Vallée de la Sionne and Dischma, respectively. The best infrasound baseline model reached recalls of 0.50 and 0.60 with precisions of 0.003 and 0.002. In contrast, the metaclassifier achieved a recall and precision of 0.75 each at Vallée de la Sionne and a recall of 0.70 with a precision of 0.47 at Dischma. The Bayesian optimization approach yielded recalls of 0.75 and 0.70 with corresponding precisions of 0.14 and 0.03 at Vallée de la Sionne and Dischma, respectively. The corresponding number of false alarms in Vallée de la Sionne and Dischma, respectively, were 60 and 1,734 for the best individual model, 18 and 244 for the Bayesian optimization approach, and only 1 and 8 for the metaclassifier, demonstrating the substantial reduction in false alarms achieved by the metaclassifier. These results demonstrate that combining seismic and infrasound models notably reduces false alarm rates in automatic avalanche detection. This finding suggests that integrating both data types, either through a postprocessing step or directly during model development, could be a key step toward developing an operational avalanche detection system with high detection and low false alarm rates.

Fields of Study & Subjects

Seismology and Earthquake Studies Landslides and related hazards Cryospheric studies and observations

Publication details

Journal / source
Digital Library of the University of Innsbruck (University of Innsbruck)
Publication date
1 January 2026
Language
EN
Document type
Thesis