Image Recognition Techniques for Earthquake Early Warning

When monitoring on his/her PC a map of seismic stations, whose colors scale with the real-time transmitted ground motions amplitudes observed in a dense seismic network, an experienced person will fairly easily recognize when and where an earthquake occurs. Using the maximum amplitudes at stations at close epicentral distances, he/she might even be able to roughly estimate the size of the event. From the number and distribution of stations turning 'red', the person might also be able to recognize the rupturing fault in a large earthquake (M>>7.0), and to estimate the rupture dimensions while the rupture is still developing. Following this concept, we are adopting techniques for automatic image recognition to provide earthquake early warning. We rapidly correlate a set of templates with real-time ground motion observations in a seismic network. If a 'suspicious' pattern of ground motion amplitudes is detected, the algorithm starts estimating the location of the earthquake and its magnitude. For large earthquakes the algorithm estimates finite source dimensions and the direction of rupture propagation. These predictions are continuously up-dated using the current 'image' of ground motion observations. A priori information, such as on the orientation of mayor faults, helps enhancing estimates in less dense networks. The approach will be demonstrated for multiple simulated and real events in California.