Unsupervised Feature Selection for Pattern Discovery in Seismic Wavefields

This study presents an unsupervised feature selection approach for the discovery of significant patterns in seismic wavefields. We iteratively reduce the number of features generated from seismic time series by first considering significance of individual features. Significance testing is done by assessing the randomness of the time series with the Wald-Wolfowitz runs test and by comparing observed and theoretical variability of features. In a second step the in-between feature dependencies are assessed based on correlation hunting in feature subsets using Self-Organizing Maps (SOMs). We show the improved discriminative power of our procedure compared to manually selected feature subsets by cross-validation applied to synthetic seismic wavefield data. Furthermore, we apply the method to real-world data with the aim to define suitable features for earthquake detection and seismic phase classification in seismic recordings.