Earthquake prediction by data mining and visualization.

New techniques based on cluster analysis of the multi–resolution structure of earthquake patterns is developed and applied to observed synthetic seismic catalogs. The synthetic data were generated by numerical simulations for various cases. At the highest resolution, analysis of the local cluster structure in the data space of seismic events for the two types of catalogs by using an agglomerative clustering algorithm is carried out. Seismic event, quantized in space and time, generate the multi – dimensional feature space of the earthquake parameters. Using a non - hierarchical clustering algorithm and multi – dimensional scaling, the multitudinous earthquakes by real time 3D visualization and inspection of multivariate clusters is explored. The resolutions characteristic of the earthquake parameters, all of the ongoing seismic activity before and after largest events accumulate to a global structure consisting of a few separate clusters in the feature space. By combining the clustering results from low and high resolution spaces, we can recognize precursory events more precisely and decode vital information that cannot be discerned at a single level of resolution

The understanding of earthquake dynamics and development of forecasting algorithms require a knowledge and skill in both measurement and analysis that cover various types of data, such as seismic, electromagnetic, gravitational, geodetic, geochemical, etc. The Gutenberg - Richter power law distribution of earthquake sizes implies that the largest events are surrounded (in space and time) by a large number of small events. The multi - dimensional and multi - resolutional structure of this global cluster depend strongly on geological and geophysical conditions. Past seismic activities are closely associated events (e.g., volcano eruptions) and time sequence of the earthquakes forming isolated events, patches, swarms etc.  Investigations on earthquake predictions are based on the assumption that all of the regional factors can be filtered out and general Information about the earth quake precursory patterns can be extracted. This extraction process is usually performed by using classical statistical or pattern recognition methodology. Feature extraction involves a pre selection process of various statistical properties of data and generation of a set of the seismic parameters, which correspond to linearly independent coordinates in the feature space. The seismic parameters in the form of time series can be analyzed by using various pattern recognition techniques ranging from fuzzy sets theory and expert systems, multi – dimensional wavelets to neural networks. The prediction of the earthquakes is a very difficult and challenging task; we cannot operate on only one level of resolution. The coarse graining of the original data can destroy the local dependences between the events and the isolated earthquakes by neglecting their spatial localization.  In this manner, the subtle correlations between the earthquakes and preceding patches of events can be dissolved in the background of uncorrelated and noisy data.

We can extract local spatio - temporal clusters of low magnitude events and identify correlations between the clusters and the earthquakes. These clusters could reflect clearly the short term trends in seismic activities followed by isolated large events. However, local clustering of seismic events is not able to extract an overall picture concerning the precursory patterns. Data mining techniques, include not only various clustering algorithms but also feature extraction and visualization techniques. Multi – dimensional scaling procedures for visualization of multi - dimensional events in 3D space is used. This visual analysis helps greatly in detecting the subtle structures, which escape the classical clustering techniques.

Earth and Planetary Sci. Letters, August, 2003 - Earthquakes over Space, Time and Feature Space  by Cluster Analysis, Data-Mining and Multi dimensional Visualization.