Automatically Detecting Avalanche Events in Passive Seismic Data

• DOI: 10.1109/icmla.2012.12
• Publication Date: 2013-01-17T15:36:00Z
• Authors: Marc J. Rubin; Tracy Camp; Alec Van Herwijnen; Jurg Schweizer
• Keywords: Geophone

During the 2010-2011 winter season, we deployed seven geophones on a mountain outside of Davos, Switzerland and collected over 100 days seismic data containing 385 possible avalanche events (33 confirmed slab avalanches). In this article, describe our efforts to develop pattern recognition workflow automatically detect snow from passive data. Our initial consisted frequency domain feature extraction, cluster-based stratified subsampling, runs training testing 12 different classification algorithms. When tested entire season single sensor, all twelve machine learning algorithms resulted in mean accuracies above 84%, with classifiers reaching 90%. We then experimented voting based paradigm that combined information sensors. This method increased overall accuracy precision, but performed quite poorly terms classifier recall. We, therefore, decided pursue other signal preprocessing methodologies. focused improving performance sensor detection, employed spectral flux event selection identify significant instantaneous increases energy. With threshold 90% relative increase, correctly selected 32 33 avalanches reduced problem space by nearly 98%. trained set only events, decision stump achieved 93% accuracy, 89.5% recall, improved precision 2.8% 13.2%.