Analysis of Seismic Activity Using Self-Organizing Map: Implications for Earthquake Prediction

The temporal and spatial distribution of seismic activity is analyzed using a self-organizing map (SOM) to find the relation between large, hazardous and other, smaller earthquakes. The seismic activity is represented by the number of shallow earthquakes with magnitude greater than 1.0 that are counted in lists of hypocenters every 3 months for suitably divided spatial segments of the Tohoku and Kanto areas in northern and central Japan, respectively. The input vector of the SOM, consisting of the spatial distribution of seismic activity, is assigned to each time interval and given a label that specifies the time relative to the occurrence of large earthquakes. Unsupervised learning of the SOM produces a two-dimensional map that can separate the input vectors into active and inactive seismic periods. Since an active period begins with a precursory stage preceding large earthquakes, the possible occurrence of a large earthquake can be predicted by identifying the current state of the seismic activity as either an active or inactive period in the SOM.