Search

With seismic catalogues becoming progressively larger, extracting information becomes challenging and calls upon using sophisticated statistical analysis. Data are typically clustered by machine learning algorithms to find patterns or identify regions of interest that require further exploration. Here, we investigate two density-based clustering algorithms, DBSCAN and OPTICS, for their capability to analyse the spatial distribution of seismicity and their effectiveness in discovering highly active seismic volumes of arbitrary shapes in large data sets. In particular, we study the influence of varying input parameters on the cluster solutions. By exploring the parameter space, we identify a crossover region with optimal solutions in between two phases with opposite behaviours (i.e. only clustered and only unclustered data points). Using a synthetic case with various geometric structures, we find that solutions in the crossover region consistently have the largest clusters and best represent the individual structures. For identifying strong anisotropic structures, we illustrate the usefulness of data rescaling. Applying the clustering algorithms to seismic catalogues of recent earthquake sequences (2016 Central Italy and 2016 Kumamoto) confirms that cluster solutions in the crossover region are the best candidates to identify 3-D features of tectonic structures that were activated in a seismic sequence. Finally, we propose a list of recipes that generalizes our analyses to obtain such solutions for other seismic sequences. [ABSTRACT FROM AUTHOR]

A rockfall (RF) is a ubiquitous geohazard that is difficult to monitor or predict and poses a significant risk for people and transportation in several hilly and mountainous environments. The seismic signal generated by RF carries abundant physical and mechanical information. Thus, signals can be used by researchers to reconstruct the event location, onset time, volume and trajectory, and develop an efficient early warning system. Therefore, the precise automatic detection and classification of RF events are important objectives for scientists, especially in seismic monitoring arrays. An algorithm called DESTRO (DEtection and STorage of ROckfalls) aimed at combining seismic event automatic detection and classification was implemented ad hoc within the MATLAB environment. In event detection, the STA/LTA (short-time-average through long-time-average) method combined with other parameters, such as the minimum duration of an RF and the minimum interval time between two continuous seismic events is used. Furthermore, nine significant features based on the frequency, amplitude, seismic waveform, duration and multiple station attributes are newly proposed to classify seismic events in a RF environment. In particular, a three-step classification method is proposed for the discrimination of five different source types: RFs, earthquakes (EQs), tremors, multispike events (MSs) and subordinate MS events. Each component (vertical, east–west and north–south) at each station within the monitoring network is analysed, and a three-step classification is performed. At a given time, the event series detected from each component are integrated and reclassified component by component and station by station into a final event-type series as an output result. By this algorithm, a case study of the seven-month-long seismic monitoring of a former quarry in Central Italy was investigated by means of four triaxial velocimeters with continuous acquisition at a sampling rate of 200 Hz. During this monitoring period, a human-induced RF simulation was performed, releasing 95 blocks (in which 90 blocks validated) of different sizes from the benches of the quarry. Consequently, 64.9 per cent of EQs within 100 km were confirmed in a one-month monitoring period, 88 blocks in the RF simulation were classified correctly as RF events and 2 blocks were classified as MSs given their small energy. Finally, an ad hoc section of the algorithm was designed specifically for RF classification combined with EQ recognition. The algorithm could be applied in slope seismic monitoring to monitor the dynamic states of rock masses, as well as in slope instability forecasting and risk evaluation in EQ-prone areas. [ABSTRACT FROM AUTHOR]