Severity Classification of a Seismic Event based on the Magnitude-Distance Ratio Using Only One Seismological Station

Abstract

Seismic event characterization is often accomplished using algorithms based only on information received at seismological stations located closest to the particular event, while ignoring historical data received at those stations. These historical data are stored and unseen at this stage. This characterization process can delay the emergency response, costing valuable time in the mitigation of the adverse effects on the affected population. Seismological stations have recorded data during many events that have been characterized by classical methods, and these data can be used as previous "knowledge" to train such stations to recognize patterns. This knowledge can be used to make faster characterizations using only one three-component broadband station by applying bio-inspired algorithms or recently developed stochastic methods, such as kernel methods. We trained a Support Vector Machine (SVM) algorithm with seismograph data recorded by INGEOMINAS's National Seismological Network at a three-component station located near Bogota, Colombia. As input model descriptors, we used the following: (1) the integral of the Fourier transform/power spectrum for each component, divided into 7 windows of 2 seconds and beginning at the P onset time, and (2) the ratio between the calculated logarithm of magnitude (Mb) and epicentral distance. We used 986 events with magnitudes greater than 3 recorded from late 2003 to 2008. The algorithm classifies events with magnitude-distance ratios (a measure of the severity of possible damage caused by an earthquake) greater than a background value. This value can be used to estimate the magnitude based on a known epicentral distance, which is calculated from the difference between P and S onset times. This rapid (< 20 seconds) magnitude estimate can be used for rapid response strategies. The results obtained in this work confirm that many hypocentral parameters and a rapid location of a seismic event can be obtained using a few seconds of signal registered at a single station. A cascade scheme of SVMs or other appropriate algorithms can be used to completely classify an event in a very short time with acceptable accuracy using data from only one station.

Keywords

• Machine learning
• seismology
• single station
• magnitude
• distance.