Inferring geo-spatial neutral similarity from earthquake data using mixture and state clustering models

Traditionally, earthquake events are identified by prescribed and well formed geographical region boundaries. However, fixed regional schemes are subject to overlook seismic patterns typified by cross boundary relations that deem essential to seismological research. Rather, we investigate a statistically driven system that clusters earthquake bound places by similarity in seismic feature space, and is impartial to geo-spatial proximity constraints. To facilitate our study, we acquired hundreds of thousands recordings of earthquake episodes that span an extended time period of forty years, and split them into groups singled out by their corresponding geographical places. From each collection of place affiliated event data, we have extracted objective seismic features expressed in both a compact term frequency of scales format, and as a discrete signal representation that captures magnitude samples in regular time intervals. The distribution and temporal typed feature vectors are further applied towards our mixture model and Markov chain frameworks, respectively, to conduct clustering of shake affected locations. We performed extensive cluster analysis and classification experiments, and report robust results that support the intuition of geo-spatial neutral similarity.