Volcano-seismic event classification using wavelet scattering transforms

To learn more about the physical processes related to volcanic activity, more and more data from extensive networks of seismic stations is being collected and analyzed. Conventionally, this data is identified and classified manually – a labor-intensive and time-consuming process. Here, we propose a classification method based on the clustering of wavelet scattering transforms of the volcanic events, which are embedded into a lower dimensional space, using t-distributed stochastic neighbor embedding (t-SNE). Wavelet scattering is chosen because of its advantageous properties, such as the invariance of the representation, the high information content, and its stability. For clustering, the spectral clustering method is used. By embedding the data to a pre-trained t-SNE scaffolding a supervised classification method is also possible. For classification, a simple k-nearest neighbor-classifier is used. The method is tested on events from the Llaima volcano in Chile, under supervised and also unsupervised conditions. These lead to promising results with a classification accuracy of 97% in the unsupervised and 99% in the supervised case, respectively.