3:30 PM - 4:30 PM
[S10-P-01] The technology for automatic probabilistic prediction of earthquakes
We provide the technology of automatic probabilistic prediction of earthquakes. All types of data on seismic process are transformed into spatio-temporal grid-based fields. The predictive field is constructed from several initial fields retrospectively with help of machine learning methods. We construct the predictive field so that only the grid nodes with the values near to maximum or minimum of the seismological field may be precursors of earthquakes. Then the forecasting can be performed using the threshold detection.
Further we will consider only positive anomalies in seismological fields to simplify the explanation of the decision making algorithm. If the value of predictive field in some grid point f(x,y,t) exceeds the upper threshold, the alarm is claimed in the cylinder volume with the center of base (x,y,t+dt), radius of base equals R and the central line [(x,y,t+dt),( x,y,t+dt+T)]. Here dt is the time step. The forecasting of an earthquake is successful if the epicenter occurs in the domain consisting from the union of alarm cylinders. So the forecasting of an event with epicenter coordinates (x,y,t+du), du<dt, is successful if and only if in the cylinder volume with the center of base (x,y,t-T), radius of base R and the central line [(x,y,t-T),( x,y,t-dt)], there is at least one point with the forecasting field value more than threshold. This property is of importance for machine learning procedure. We use it for selecting the trainings sample set.
We present the results on simulation of the technology on GIS GeoTime 3 (www.geo.iitp.ru/GT3). In considered case studies, we used earthquake catalogues loaded from the sites of ISC, NEIC and Kamchatka Branch of the Geophysical Service RAS as initial data.
The research is supported by RFBR projects 16-07-0326 and 17-07-00494.
Further we will consider only positive anomalies in seismological fields to simplify the explanation of the decision making algorithm. If the value of predictive field in some grid point f(x,y,t) exceeds the upper threshold, the alarm is claimed in the cylinder volume with the center of base (x,y,t+dt), radius of base equals R and the central line [(x,y,t+dt),( x,y,t+dt+T)]. Here dt is the time step. The forecasting of an earthquake is successful if the epicenter occurs in the domain consisting from the union of alarm cylinders. So the forecasting of an event with epicenter coordinates (x,y,t+du), du<dt, is successful if and only if in the cylinder volume with the center of base (x,y,t-T), radius of base R and the central line [(x,y,t-T),( x,y,t-dt)], there is at least one point with the forecasting field value more than threshold. This property is of importance for machine learning procedure. We use it for selecting the trainings sample set.
We present the results on simulation of the technology on GIS GeoTime 3 (www.geo.iitp.ru/GT3). In considered case studies, we used earthquake catalogues loaded from the sites of ISC, NEIC and Kamchatka Branch of the Geophysical Service RAS as initial data.
The research is supported by RFBR projects 16-07-0326 and 17-07-00494.