Assessing avalanche activity in seismic data with modern machine learning methods.

Monitoring snow avalanche activity is essential for operational avalanche forecasting and the successful implementation of mitigation measures to ensure safety in mountain regions. To facilitate and automate the monitoring process, avalanche detection systems equipped with seismic sensors provide a cost-effective solution. Still, automatically differentiating avalanche signals from other sources in seismic data remains rather challenging. This is mainly due to the complexity of the seismic signals generated by avalanches, the relatively rare occurrence of avalanches and the presence of multiple sources in the continuous recordings.

To discriminate avalanches from other sources in the continuous seismic recordings, we test three random forest classifiers using two feature sets extracted with two autoencoders and a set of 57 statistical features. We extract these features from 10s windows of the seismograms recorded with an array of five seismometers installed in Davos, Switzerland. The statistical feature set includes waveform, spectral and spectrogram attributes. The first autoencoder is composed of convolutional layers and a long short-term memory unit. This neuronal network automatically extracts 64 features from the raw waveform signal. The second autoencoder applies a sequence of fully connected layers to extract the same number of features from the spectrum of the signals. We assess the performance of each classifier and compare the results. To improve the predictive performance of the seismic system, we employ different post-processing, e.g. adaption of classification thresholds and ensembling the predictions from the three classifiers. The final model is tested with the continuous seismic data of the last winter season to potentially be used as an operational, near-real-time detection system.