Automated discrimination of seismo-acoustic avalanche signals

The unpredictable nature and destructive power of snow avalanches demand reliable, real-time detection systems of the events in mountain regions. Remote detection systems based on seismic and infrasound sensors have been increasingly used to monitor avalanches at a rather low economic cost. The seismo-acoustic wave field generated by avalanches enables the detection of natural avalanches in a large area, independently of the weather and visibility conditions. One approach for the automatization of avalanche detection is the discrimination of seismic and infrasound signals in the continuous recordings by applying machine learning classification methods. In this study, we evaluated the automatic classification of avalanche signals recorded by a seismo-acoustic detection system installed in Davos (Switzerland) since the winter season 2020-2021. We tested three feature extraction methods to classify the signals based on a Random Forest algorithm. The first RF classifier was trained with a set of features extracted from the individual components of the array. This set of features included waveform properties, spectral features and spectrogram attributes. The second classifier used input features extracted from the amplitude, backazimut and apparent slowness time series of the array-processing outputs. In addition, we tested an autoencoder feature extraction method based on a convolutional neural network with long short-term memory. This automated set of input features was used to train another RF classifier using the same labels. We compared the predictive performance of the three classifiers. Our final goal is to develop an effective classification algorithm combining the different methods to automatically detect snow avalanches in near-real time.